arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar ,...

49
WHO 2016 subtyping and automated segmentation of glioma using multi-task deep learning Sebastian R. van der Voort 1,, Fatih Incekara 2,3,, Maarten MJ Wijnenga 4 , Georgios Kapsas 2 , Renske Gahrmann 2 , Joost W Schouten 3 , Rishi Nandoe Tewarie 5 , Geert J Lycklama 6 , Philip C De Witt Hamer 7 , Roelant S Eijgelaar 7 , Pim J French 4 , Hendrikus J Dubbink 8 , Arnaud JPE Vincent 3 , Wiro J Niessen 1,9 , Martin J van den Bent 4 , Marion Smits 2,†† , and Stefan Klein 1,††,* 1 Biomedical Imaging Group Rotterdam, Department of Radiology and Nuclear Medicine, Erasmus MC University Medical Centre Rotterdam, Rotterdam, the Netherlands 2 Department of Radiology and Nuclear Medicine, Erasmus MC University Medical Centre Rotterdam, Rotterdam, the Netherlands 3 Department of Neurosurgery, Brain Tumor Center, Erasmus MC University Medical Centre Rotterdam, Rotterdam, the Netherlands 4 Department of Neurology, Brain Tumor Center, Erasmus MC Cancer Institute, Rotterdam, the Netherlands 5 Department of Neurosurgery, Haaglanden Medical Center, the Hague, the Netherlands 6 Department of Radiology, Haaglanden Medical Center, the Hague, the Netherlands 8 Department of Pathology, Brain Tumor Center at Erasmus MC Cancer Institute, Rotterdam, the Netherlands 9 Imaging Physics, Faculty of Applied Sciences, Delft University of Technology, Delft, the Netherlands 7 Department of Neurosurgery, Cancer Center Amsterdam, Brain Tumor Center, Amsterdam UMC, Amsterdam, Netherlands These authors contributed equally †† These authors contributed equally * Corresponding author; [email protected] 1 arXiv:2010.04425v1 [eess.IV] 9 Oct 2020

Transcript of arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar ,...

Page 1: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

WHO 2016 subtyping and automated

segmentation of glioma using multi-task

deep learning

Sebastian R van der Voort1dagger Fatih Incekara23dagger Maarten MJWijnenga4 Georgios Kapsas2 Renske Gahrmann2 Joost W

Schouten3 Rishi Nandoe Tewarie5 Geert J Lycklama6 Philip CDe Witt Hamer7 Roelant S Eijgelaar7 Pim J French4 HendrikusJ Dubbink8 Arnaud JPE Vincent3 Wiro J Niessen19 Martin J

van den Bent4 Marion Smits2daggerdagger and Stefan Klein1daggerdagger

1Biomedical Imaging Group Rotterdam Department of Radiologyand Nuclear Medicine Erasmus MC University Medical Centre

Rotterdam Rotterdam the Netherlands2Department of Radiology and Nuclear Medicine Erasmus MC

University Medical Centre Rotterdam Rotterdam the Netherlands3Department of Neurosurgery Brain Tumor Center Erasmus MCUniversity Medical Centre Rotterdam Rotterdam the Netherlands

4Department of Neurology Brain Tumor Center Erasmus MCCancer Institute Rotterdam the Netherlands

5Department of Neurosurgery Haaglanden Medical Center theHague the Netherlands

6Department of Radiology Haaglanden Medical Center theHague the Netherlands

8Department of Pathology Brain Tumor Center at Erasmus MCCancer Institute Rotterdam the Netherlands

9Imaging Physics Faculty of Applied Sciences Delft University ofTechnology Delft the Netherlands

7Department of Neurosurgery Cancer Center Amsterdam BrainTumor Center Amsterdam UMC Amsterdam Netherlands

daggerThese authors contributed equallydaggerdaggerThese authors contributed equally

Corresponding author skleinerasmusmcnl

1

arX

iv2

010

0442

5v1

[ee

ssI

V]

9 O

ct 2

020

Abstract

Accurate characterization of glioma is crucial for clinical decision mak-ing A delineation of the tumor is also desirable in the initial decisionstages but is a time-consuming task Leveraging the latest GPU capabil-ities we developed a single multi-task convolutional neural network thatuses the full 3D structural pre-operative MRI scans to can predict theIDH mutation status the 1p19q co-deletion status and the grade of a tu-mor while simultaneously segmenting the tumor We trained our methodusing the largest most diverse patient cohort to date containing 1508glioma patients from 16 institutes We tested our method on an indepen-dent dataset of 240 patients from 13 different institutes and achieved anIDH-AUC of 090 1p19q-AUC of 085 grade-AUC of 081 and a meanwhole tumor DICE score of 084 Thus our method non-invasively pre-dicts multiple clinically relevant parameters and generalizes well to thebroader clinical population

1 Introduction

Glioma is the most common primary brain tumor and is one of the deadliestforms of cancer [1] Differences in survival and treatment response of glioma areattributed to their genetic and histological features specifically the isocitratedehydrogenase (IDH) mutation status the 1p19q co-deletion status and thetumor grade [2 3] Therefore in 2016 the World Health Organization (WHO)updated its brain tumor classification categorizing glioma based on these ge-netic and histological features [4] In current clinical practice these features aredetermined from tumor tissue While this is not an issue in patients in whomthe tumor can be resected this is problematic when resection can not safelybe performed In these instances surgical biopsy is performed with the solepurpose of obtaining tissue for diagnosis which although relatively safe is notwithout risk [5 6] Therefore there has been an increasing interest in comple-mentary non-invasive alternatives that can provide the genetic and histologicalinformation used in the WHO 2016 categorization [7 8]

Magnetic resonance imaging (MRI) has been proposed as a possible candi-date because of its non-invasive nature and its current place in routine clinicalcare [9] Research has shown that certain MRI features such as the tumor het-erogeneity correlate with the genetic and histological features of glioma [10 11]This notion has popularized in addition to already popular applications suchas tumor segmentation the use of machine learning methods for the predictionof genetic and histological features known as radiomics [12 13 14] Althougha plethora of such methods now exist they have found little translation to theclinic [12]

An often discussed challenge for the adoption of machine learning methodsin clinical practice is the lack of standardization resulting in heterogeneity ofpatient populations imaging protocols and scan quality [15 16] Since machinelearning methods are prone to overfitting this heterogeneity questions the va-lidity of such methods in a broader patient population [16] Furthermore it has

2

been noted that most current research concerns narrow task-specific methodsthat lack the context between different related tasks which might restrict theperformance of these methods [17]

An important technical limitation when using deep learning methods is thelimited GPU memory which restricts the size of models that can be trained[18] This is a problem especially for clinical data which is often 3D requiringeven more memory than the commonly used 2D networks This further limitsthe size of these models resulting in shallower models and the use of patches ofa scan instead of using the full 3D scan as an input which limits the amount ofcontext these methods can extract from the scans

Here we present a new method that addresses the above problems Ourmethod consists of a single multi-task convolutional neural network (CNN)that can predict the IDH mutation status the 1p19q co-deletion status andthe grade (grade IIIIIIV) of a tumor while also simultaneously segmenting thetumor see Figure 1 To the best of our knowledge this is the first method thatprovides all of this information at the same time allowing clinical experts to de-rive the WHO category from the individually predicted genetic and histologicalfeatures while also allowing them to consider or disregard specific predictionsas they deem fit Exploiting the capabilities of the latest GPUs optimizing ourimplementation to reduce the memory footprint and using distributed multi-GPU training we were able to train a model that uses the full 3D scan as aninput We trained our method using the largest most diverse patient cohortto date with 1508 patients included from 16 different institutes To ensurethe broad applicability of our method we used minimal inclusion criteria onlyrequiring the four most commonly used MRI sequences pre- and post-contrastT1-weighted (T1w) T2-weighted (T2w) and T2-weighted fluid attenuated in-version recovery (T2w-FLAIR) [19 20] No constraints were placed on thepatientsrsquo clinical characteristics such as the tumor grade or the radiologicalcharacteristics of scans such as the scan quality In this way our method couldcapture the heterogeneity that is naturally present in clinical data We testedour method on an independent dataset of 240 patients from 13 different insti-tutes to evaluate the true generalizability of our method Our results show thatwe can predict multiple clinical features of glioma from MRI scans in a diversepatient population

3

Convolutionalneural network

IDH status

Wildtype Mutated

1p19q status

Intact Co-deleted

Grade

II III IV

WHO 2016categorization

MRI scansPreprocessed

scansSegmentation

Figure 1 Overview of our method Pre- and post-contrast T1w T2w and T2w-FLAIR scans are used as an input The scans are registered to an atlas biasfield corrected skull stripped and normalized before being passed through ourconvolutional neural network One branch of the network segments the tumorwhile at the same time the features are combined to predict the IDH status1p19q status and grade of the tumor

4

2 Results

21 Patient characteristics

We included a total of 1748 patients in our study 1508 as a train set and240 as an independent test set The patients in the train set originated fromnine different data collections and 16 different institutes and the test set wascollected from two different data collections and 13 different institutes Table 1provides a full overview of the patient characteristics in the train and test setand Figure 2 shows the inclusion flowchart and the distribution of the patientsover the different data collections in the train set and test set

Table 1 Patient characteristics for the train set and test set

Train set Test setN N

Patients 1508 240IDH status

Mutated 226 150 88 367Wildtype 440 292 129 537Unknown 842 558 23 96

1p19q co-deletion statusCo-deleted 103 68 26 108Intact 337 224 207 863Unknown 1068 708 7 29

GradeII 230 153 47 196III 114 76 59 246IV 830 550 132 550Unknown 334 221 2 08

WHO 2016 categorizationOligodendroglioma 96 64 26 108Astrocytoma IDH wildtype 31 21 22 92Astrocytoma IDH mutated 98 64 57 237GBM IDH wildtype 331 219 106 442GBM IDH mutated 16 11 5 21Unknown 936 621 24 100

SegmentationManual 716 475 240 100Automatic 792 525 0 0

IDH isocitrate dehydrogenase WHO World Health Organization GBMglioblastoma

5

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 2: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

Abstract

Accurate characterization of glioma is crucial for clinical decision mak-ing A delineation of the tumor is also desirable in the initial decisionstages but is a time-consuming task Leveraging the latest GPU capabil-ities we developed a single multi-task convolutional neural network thatuses the full 3D structural pre-operative MRI scans to can predict theIDH mutation status the 1p19q co-deletion status and the grade of a tu-mor while simultaneously segmenting the tumor We trained our methodusing the largest most diverse patient cohort to date containing 1508glioma patients from 16 institutes We tested our method on an indepen-dent dataset of 240 patients from 13 different institutes and achieved anIDH-AUC of 090 1p19q-AUC of 085 grade-AUC of 081 and a meanwhole tumor DICE score of 084 Thus our method non-invasively pre-dicts multiple clinically relevant parameters and generalizes well to thebroader clinical population

1 Introduction

Glioma is the most common primary brain tumor and is one of the deadliestforms of cancer [1] Differences in survival and treatment response of glioma areattributed to their genetic and histological features specifically the isocitratedehydrogenase (IDH) mutation status the 1p19q co-deletion status and thetumor grade [2 3] Therefore in 2016 the World Health Organization (WHO)updated its brain tumor classification categorizing glioma based on these ge-netic and histological features [4] In current clinical practice these features aredetermined from tumor tissue While this is not an issue in patients in whomthe tumor can be resected this is problematic when resection can not safelybe performed In these instances surgical biopsy is performed with the solepurpose of obtaining tissue for diagnosis which although relatively safe is notwithout risk [5 6] Therefore there has been an increasing interest in comple-mentary non-invasive alternatives that can provide the genetic and histologicalinformation used in the WHO 2016 categorization [7 8]

Magnetic resonance imaging (MRI) has been proposed as a possible candi-date because of its non-invasive nature and its current place in routine clinicalcare [9] Research has shown that certain MRI features such as the tumor het-erogeneity correlate with the genetic and histological features of glioma [10 11]This notion has popularized in addition to already popular applications suchas tumor segmentation the use of machine learning methods for the predictionof genetic and histological features known as radiomics [12 13 14] Althougha plethora of such methods now exist they have found little translation to theclinic [12]

An often discussed challenge for the adoption of machine learning methodsin clinical practice is the lack of standardization resulting in heterogeneity ofpatient populations imaging protocols and scan quality [15 16] Since machinelearning methods are prone to overfitting this heterogeneity questions the va-lidity of such methods in a broader patient population [16] Furthermore it has

2

been noted that most current research concerns narrow task-specific methodsthat lack the context between different related tasks which might restrict theperformance of these methods [17]

An important technical limitation when using deep learning methods is thelimited GPU memory which restricts the size of models that can be trained[18] This is a problem especially for clinical data which is often 3D requiringeven more memory than the commonly used 2D networks This further limitsthe size of these models resulting in shallower models and the use of patches ofa scan instead of using the full 3D scan as an input which limits the amount ofcontext these methods can extract from the scans

Here we present a new method that addresses the above problems Ourmethod consists of a single multi-task convolutional neural network (CNN)that can predict the IDH mutation status the 1p19q co-deletion status andthe grade (grade IIIIIIV) of a tumor while also simultaneously segmenting thetumor see Figure 1 To the best of our knowledge this is the first method thatprovides all of this information at the same time allowing clinical experts to de-rive the WHO category from the individually predicted genetic and histologicalfeatures while also allowing them to consider or disregard specific predictionsas they deem fit Exploiting the capabilities of the latest GPUs optimizing ourimplementation to reduce the memory footprint and using distributed multi-GPU training we were able to train a model that uses the full 3D scan as aninput We trained our method using the largest most diverse patient cohortto date with 1508 patients included from 16 different institutes To ensurethe broad applicability of our method we used minimal inclusion criteria onlyrequiring the four most commonly used MRI sequences pre- and post-contrastT1-weighted (T1w) T2-weighted (T2w) and T2-weighted fluid attenuated in-version recovery (T2w-FLAIR) [19 20] No constraints were placed on thepatientsrsquo clinical characteristics such as the tumor grade or the radiologicalcharacteristics of scans such as the scan quality In this way our method couldcapture the heterogeneity that is naturally present in clinical data We testedour method on an independent dataset of 240 patients from 13 different insti-tutes to evaluate the true generalizability of our method Our results show thatwe can predict multiple clinical features of glioma from MRI scans in a diversepatient population

3

Convolutionalneural network

IDH status

Wildtype Mutated

1p19q status

Intact Co-deleted

Grade

II III IV

WHO 2016categorization

MRI scansPreprocessed

scansSegmentation

Figure 1 Overview of our method Pre- and post-contrast T1w T2w and T2w-FLAIR scans are used as an input The scans are registered to an atlas biasfield corrected skull stripped and normalized before being passed through ourconvolutional neural network One branch of the network segments the tumorwhile at the same time the features are combined to predict the IDH status1p19q status and grade of the tumor

4

2 Results

21 Patient characteristics

We included a total of 1748 patients in our study 1508 as a train set and240 as an independent test set The patients in the train set originated fromnine different data collections and 16 different institutes and the test set wascollected from two different data collections and 13 different institutes Table 1provides a full overview of the patient characteristics in the train and test setand Figure 2 shows the inclusion flowchart and the distribution of the patientsover the different data collections in the train set and test set

Table 1 Patient characteristics for the train set and test set

Train set Test setN N

Patients 1508 240IDH status

Mutated 226 150 88 367Wildtype 440 292 129 537Unknown 842 558 23 96

1p19q co-deletion statusCo-deleted 103 68 26 108Intact 337 224 207 863Unknown 1068 708 7 29

GradeII 230 153 47 196III 114 76 59 246IV 830 550 132 550Unknown 334 221 2 08

WHO 2016 categorizationOligodendroglioma 96 64 26 108Astrocytoma IDH wildtype 31 21 22 92Astrocytoma IDH mutated 98 64 57 237GBM IDH wildtype 331 219 106 442GBM IDH mutated 16 11 5 21Unknown 936 621 24 100

SegmentationManual 716 475 240 100Automatic 792 525 0 0

IDH isocitrate dehydrogenase WHO World Health Organization GBMglioblastoma

5

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 3: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

been noted that most current research concerns narrow task-specific methodsthat lack the context between different related tasks which might restrict theperformance of these methods [17]

An important technical limitation when using deep learning methods is thelimited GPU memory which restricts the size of models that can be trained[18] This is a problem especially for clinical data which is often 3D requiringeven more memory than the commonly used 2D networks This further limitsthe size of these models resulting in shallower models and the use of patches ofa scan instead of using the full 3D scan as an input which limits the amount ofcontext these methods can extract from the scans

Here we present a new method that addresses the above problems Ourmethod consists of a single multi-task convolutional neural network (CNN)that can predict the IDH mutation status the 1p19q co-deletion status andthe grade (grade IIIIIIV) of a tumor while also simultaneously segmenting thetumor see Figure 1 To the best of our knowledge this is the first method thatprovides all of this information at the same time allowing clinical experts to de-rive the WHO category from the individually predicted genetic and histologicalfeatures while also allowing them to consider or disregard specific predictionsas they deem fit Exploiting the capabilities of the latest GPUs optimizing ourimplementation to reduce the memory footprint and using distributed multi-GPU training we were able to train a model that uses the full 3D scan as aninput We trained our method using the largest most diverse patient cohortto date with 1508 patients included from 16 different institutes To ensurethe broad applicability of our method we used minimal inclusion criteria onlyrequiring the four most commonly used MRI sequences pre- and post-contrastT1-weighted (T1w) T2-weighted (T2w) and T2-weighted fluid attenuated in-version recovery (T2w-FLAIR) [19 20] No constraints were placed on thepatientsrsquo clinical characteristics such as the tumor grade or the radiologicalcharacteristics of scans such as the scan quality In this way our method couldcapture the heterogeneity that is naturally present in clinical data We testedour method on an independent dataset of 240 patients from 13 different insti-tutes to evaluate the true generalizability of our method Our results show thatwe can predict multiple clinical features of glioma from MRI scans in a diversepatient population

3

Convolutionalneural network

IDH status

Wildtype Mutated

1p19q status

Intact Co-deleted

Grade

II III IV

WHO 2016categorization

MRI scansPreprocessed

scansSegmentation

Figure 1 Overview of our method Pre- and post-contrast T1w T2w and T2w-FLAIR scans are used as an input The scans are registered to an atlas biasfield corrected skull stripped and normalized before being passed through ourconvolutional neural network One branch of the network segments the tumorwhile at the same time the features are combined to predict the IDH status1p19q status and grade of the tumor

4

2 Results

21 Patient characteristics

We included a total of 1748 patients in our study 1508 as a train set and240 as an independent test set The patients in the train set originated fromnine different data collections and 16 different institutes and the test set wascollected from two different data collections and 13 different institutes Table 1provides a full overview of the patient characteristics in the train and test setand Figure 2 shows the inclusion flowchart and the distribution of the patientsover the different data collections in the train set and test set

Table 1 Patient characteristics for the train set and test set

Train set Test setN N

Patients 1508 240IDH status

Mutated 226 150 88 367Wildtype 440 292 129 537Unknown 842 558 23 96

1p19q co-deletion statusCo-deleted 103 68 26 108Intact 337 224 207 863Unknown 1068 708 7 29

GradeII 230 153 47 196III 114 76 59 246IV 830 550 132 550Unknown 334 221 2 08

WHO 2016 categorizationOligodendroglioma 96 64 26 108Astrocytoma IDH wildtype 31 21 22 92Astrocytoma IDH mutated 98 64 57 237GBM IDH wildtype 331 219 106 442GBM IDH mutated 16 11 5 21Unknown 936 621 24 100

SegmentationManual 716 475 240 100Automatic 792 525 0 0

IDH isocitrate dehydrogenase WHO World Health Organization GBMglioblastoma

5

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 4: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

Convolutionalneural network

IDH status

Wildtype Mutated

1p19q status

Intact Co-deleted

Grade

II III IV

WHO 2016categorization

MRI scansPreprocessed

scansSegmentation

Figure 1 Overview of our method Pre- and post-contrast T1w T2w and T2w-FLAIR scans are used as an input The scans are registered to an atlas biasfield corrected skull stripped and normalized before being passed through ourconvolutional neural network One branch of the network segments the tumorwhile at the same time the features are combined to predict the IDH status1p19q status and grade of the tumor

4

2 Results

21 Patient characteristics

We included a total of 1748 patients in our study 1508 as a train set and240 as an independent test set The patients in the train set originated fromnine different data collections and 16 different institutes and the test set wascollected from two different data collections and 13 different institutes Table 1provides a full overview of the patient characteristics in the train and test setand Figure 2 shows the inclusion flowchart and the distribution of the patientsover the different data collections in the train set and test set

Table 1 Patient characteristics for the train set and test set

Train set Test setN N

Patients 1508 240IDH status

Mutated 226 150 88 367Wildtype 440 292 129 537Unknown 842 558 23 96

1p19q co-deletion statusCo-deleted 103 68 26 108Intact 337 224 207 863Unknown 1068 708 7 29

GradeII 230 153 47 196III 114 76 59 246IV 830 550 132 550Unknown 334 221 2 08

WHO 2016 categorizationOligodendroglioma 96 64 26 108Astrocytoma IDH wildtype 31 21 22 92Astrocytoma IDH mutated 98 64 57 237GBM IDH wildtype 331 219 106 442GBM IDH mutated 16 11 5 21Unknown 936 621 24 100

SegmentationManual 716 475 240 100Automatic 792 525 0 0

IDH isocitrate dehydrogenase WHO World Health Organization GBMglioblastoma

5

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 5: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

2 Results

21 Patient characteristics

We included a total of 1748 patients in our study 1508 as a train set and240 as an independent test set The patients in the train set originated fromnine different data collections and 16 different institutes and the test set wascollected from two different data collections and 13 different institutes Table 1provides a full overview of the patient characteristics in the train and test setand Figure 2 shows the inclusion flowchart and the distribution of the patientsover the different data collections in the train set and test set

Table 1 Patient characteristics for the train set and test set

Train set Test setN N

Patients 1508 240IDH status

Mutated 226 150 88 367Wildtype 440 292 129 537Unknown 842 558 23 96

1p19q co-deletion statusCo-deleted 103 68 26 108Intact 337 224 207 863Unknown 1068 708 7 29

GradeII 230 153 47 196III 114 76 59 246IV 830 550 132 550Unknown 334 221 2 08

WHO 2016 categorizationOligodendroglioma 96 64 26 108Astrocytoma IDH wildtype 31 21 22 92Astrocytoma IDH mutated 98 64 57 237GBM IDH wildtype 331 219 106 442GBM IDH mutated 16 11 5 21Unknown 936 621 24 100

SegmentationManual 716 475 240 100Automatic 792 525 0 0

IDH isocitrate dehydrogenase WHO World Health Organization GBMglioblastoma

5

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 6: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

Patient screening

Train set2181 Glioma patients

1241 Erasmus MC491 Haaglanden Medical Center168 BraTS130 REMBRANDT66 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set461 Glioma patients

199 TCGA-LGG262 TCGA-GBM

Patient inclusion

Train set1508 Patients in train set

816 Erasmus MC279 Haaglanden Medical Center168 BraTS109 REMBRANDT51 CPTAC-GBM39 Ivy GAP20 Amsterdam UMC20 Brain-Tumor-Progression6 University Medical Center Utrecht

Test set240 Patients in test set

107 TCGA-LGG133 TCGA-GBM

Patient exclusion

Train set673 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

425 Erasmus MC212 Haaglanden Medical Center

0 BraTS21 REMBRANDT15 CPTAC-GBM0 Ivy GAP0 Amsterdam UMC0 Brain-Tumor-Progression0 University Medical Center Utrecht

Test set221 No pre-operative

pre- or post-contrast T1wT2w or T2w-FLAIR

92 TCGA-LGG129 TCGA-GBM

Figure 2 Inclusion flowchart of the train set and test set

6

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 7: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

22 Algorithm performance

We used 15 of the train set as a validation set and selected the model pa-rameters that achieved the best performance on this validation set where themodel achieved an area under receiver operating characteristic curve (AUC) of088 for the IDH mutation status prediction an AUC of 076 for the 1p19qco-deletion prediction an AUC of 075 for the grade prediction and a meansegmentation DICE score of 081 The selected model parameters are shown inAppendix F We then trained a model using these parameters and the full trainset and evaluated it on the independent test set

For the genetic and histological feature predictions we achieved an AUC of090 for the IDH mutation status prediction an AUC of 085 for the 1p19qco-deletion prediction and an AUC of 081 for the grade prediction in thetest set The full results are shown in Table 2 with the corresponding receiveroperating characteristic (ROC)-curves in Figure 3 Table 2 also shows the resultsin (clinically relevant) subgroups of patients This shows that we achieved anIDH-AUC of 081 in low grade glioma (LGG) (grade IIIII) an IDH-AUC of064 in high grade glioma (HGG) (grade IV) and a 1p19q-AUC of 073 in LGGWhen only predicting LGG vs HGG instead of predicting the individual gradeswe achieved an AUC of 091 In Appendix A we provide confusion matrices forthe IDH 1p19q and grade predictions as well as a confusion matrix for thefinal WHO 2016 subtype which shows that only one patient was predicted asa non-existing WHO 2016 subtype In Appendix C we provide the individualpredictions and ground truth labels for all patients in the test set to allow forthe calculation of additional metrics

For the automatic segmentation we achieved a mean DICE score of 084 amean Hausdorff distance of 189 mm and a mean volumetric similarity coeffi-cient of 090 Figure 4 shows boxplots of the DICE scores Hausdorff distancesand volumetric similarity coefficients for the different patients in the test set InAppendix B we show five patients that were randomly selected from both theTCGA-LGG and TCGA-GBM data collections to demonstrate the automaticsegmentations made by our method

23 Model interpretability

To provide insight into the behavior of our model we created saliency mapswhich show which parts of the scans contributed the most to the predictionThese saliency maps are shown in Figure 5 for two example patients from thetest set It can be seen that for the LGG the network focused on a bright rim inthe T2w-FLAIR scan whereas for the HGG it focused on the enhancement in thepost-contrast T1w scan To aid further interpretation we provide visualizationsof selected filter outputs in the network in Appendix D which also show thatthe network focuses on the tumor and these filters seem to recognize specificimaging features such as the contrast enhancement and T2w-FLAIR brightness

7

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 8: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin

Table 2 Evaluation results of the final model on the test set

Patientgroup

Task AUC Accuracy Sensitivity Specificity

All IDH 090 084 072 0931p19q 085 089 039 095Grade (IIIIIIV) 081 071 NA NAGrade II 091 086 075 089Grade III 069 075 017 094Grade IV 091 082 095 066LGG vs HGG 091 084 072 093

LGG IDH 081 074 073 0771p19q 073 076 039 089

HGG IDH 064 094 040 096

Abbreviations AUC area under receiver operating characteristic curve IDHisocitrate dehydrogenase LGG low grade glioma HGG high grade glioma

Figure 3 Receiver operating characteristic (ROC)-curves of the genetic andhistological features evaluated on the test set The crosses indicate the locationof the decision threshold for the reported accuracy sensitivity and specificity

8

Figure 4 DICE scores Hausdorff distances and volumetric similarity coeffi-cients for all patients in the test set The DICE score is a measure of theoverlap between the ground truth and predicted segmentation (where 1 indi-cates perfect overlap) The Hausdorff distance is a measure of the agreementbetween the boundaries of the ground truth and predicted segmentation (loweris better) The volumetric similarity coefficient is a measure of the agreementin volume (where 1 indicates perfect agreement)

9

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Saliency maps of a low grade glioma patient (TCGA-DU-6400) This is an IDH mu-tated 1p19q co-deleted grade II tumor The network focuses on a rim of brightnessin the T2w-FLAIR scan

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) Saliency maps of a high grade glioma patient (TCGA-06-0238) This is an IDHwildtype grade IV tumor The network focuses on enhancing spots around the necrosison the post-contrast T1w scan

Figure 5 Saliency maps of two patients from the test set showing areas thatare relevant for the prediction

10

24 Model robustness

By not excluding scans from our train set based on radiological characteristicswe were able to make our model robust to low scan quality as can be seen in anexample from the test set in Figure 6 Even though this example scan containedimaging artifacts our method was able to properly segment the tumor (DICEscore of 087) and correctly predict the tumor as an IDH wildtype grade IVtumor

Figure 6 Example of a T2w-FLAIR scan containing imaging artifacts and theautomatic segmentation (red overlay) made by our method It was correctlypredicted as an IDH wildtype grade IV glioma This is patient TCGA-06-5408from the TCGA-GBM collection

Finally we considered two examples of scans that were incorrectly predictedby our method see Figure 7 These two examples were chosen because ournetwork assigned high prediction scores to the wrong classes for these casesFigure 7a shows an example of a grade II IDH mutated 1p19q co-deletedglioma that was predicted as grade IV IDH wildtype by our method Ourmethodrsquos prediction was most likely caused by the hyperintensities in the post-contrast T1w scan being interpreted as contrast enhancement Since these hy-perintensities are also present in the pre-contrast T1w scan they are most likelycalcifications and the radiological appearance of this tumor is indicative of anoligodendroglioma Figure 7b shows an example of a grade IV IDH wildtypeglioma that was predicted as a grade III IDH mutated glioma by our method

11

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) TCGA-DU-6410 from the TCGA-LGG collection The ground truth histopatho-logical analysis indicated this glioma was grade II IDH mutated 1p19q co-deletedbut our method predicted it as a grade IV IDH wildtype

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(b) TCGA-76-7664 from the TCGA-HGG collection Histopathologically this gliomawas grade IV IDH wildtype but our method predicted it as grade III IDH mutated

Figure 7 Examples of scans that were incorrectly predicted by our method

12

3 Discussion

We have developed a method that can predict the IDH mutation status 1p19qco-deletion status and grade of glioma while simultaneously providing the tu-mor segmentation based on pre-operative MRI scans For the genetic andhistological feature predictions we achieved an AUC of 090 for the IDH mu-tation status prediction an AUC of 085 for the 1p19q co-deletion predictionand an AUC of 081 for the grade prediction in the test set

In an independent test set which contained data from 13 different instituteswe demonstrated that our method predicts these features with good overall per-formance we achieved an AUC of 090 for the IDH mutation status predictionan AUC of 085 for the 1p19q co-deletion prediction and an AUC of 081 forthe grade prediction and a mean whole tumor DICE score of 084 This per-formance on unseen data that was only used during the final evaluation of thealgorithm and that was purposefully not used to guide any decisions regardingthe method design shows the true generalizability of our method Using thelatest GPU capabilities we were able to train a large model which uses thefull 3D scan as input Furthermore by using the largest most diverse patientcohort to date we were able to make our method robust to the heterogeneitythat is naturally present in clinical imaging data such that it generalizes forbroad application in clinical practice

By using a multi-task network our method could learn the context betweendifferent features For example IDH wildtype and 1p19q co-deletion are mu-tually exclusive [21] If two separate methods had been used one to predict theIDH status and one to predict the 1p19q co-deletion status an IDH wildtypeglioma might be predicted to be 1p19q co-deleted which does not stroke withthe clinical reality Since our method learns both of these genetic features si-multaneously it correctly learned not to predict 1p19q co-deletion in tumorsthat were IDH wildtype there was only one patient in which our algorithm pre-dicted a tumor to be both IDH wildtype and 1p19q co-deleted Furthermoreby predicting the genetic and histological features individually instead of onlypredicting the WHO 2016 category it is possible to adopt updated guidelinessuch as cIMPACT-NOW future-proofing our method [22]

Some previous studies also used multi-task networks to predict the geneticand histological features of glioma [23 24 25] Tang et al [23] used a multi-tasknetwork that predicts multiple genetic features as well as the overall survivalof glioblastoma Since their method only works for glioblastoma patients thetumor grade must be known in advance complicating the use of their methodin the pre-operative setting when tumor grade is not yet known Furthermoretheir method requires a tumor segmentation prior to application of their methodwhich is a time-consuming expert task In a study by Xue et al [24] a multi-task network was used with a structure similar to the one proposed in thispaper to segment the tumor and predict the grade (LGG or HGG) and IDHmutation status However they do not predict the 1p19q co-deletion statusneeded for the WHO 2016 categorization Lastly Decuyper et al [25] useda multi-task network that predicts the IDH mutation and 1p19q co-deletion

13

status and the tumor grade (LGG or HGG) Their method requires a tumorsegmentation as input which they obtain from a U-Net that is applied earlierin their pipeline thus their method requires two networks instead of the singlenetwork we use in our method These differences aside the most importantlimitation of each of these studies is the lack of an independent test set forevaluating their results It is now considered essential that an independent testset is used to prevent an overly optimistic estimate of a methodrsquos performance[15 26 27 28] Thus our study improves on this previous work by providinga single network that combines the different tasks being trained on a moreextensive and diverse dataset not requiring a tumor segmentation as an in-put providing all information needed for the WHO 2016 categorization andcrucially by being evaluated in an independent test set

An important genetic feature that is not predicted by our method is theO6-methylguanine-methyltransferase (MGMT) methylation status Althoughthe MGMT methylation status is not part of the WHO 2016 categorizationit is part of clinical management guidelines and is an important prognosticmarker in glioblastoma [4] In the initial stages of this study we attemptedto predict the MGMT methylation status however the performance of thisprediction was poor Furthermore the methylation cutoff level which is usedto determine whether a tumor is MGMT methylated shows a wide varietybetween institutes leading to inconsistent results [29] We therefore opted not toinclude the MGMT prediction at all rather than to provide a poor prediction ofan unsharply defined parameter Although some methods attempted to predictthe MGMT status with varying degrees of success there is still an ongoingdiscussion on the validity of MR imaging features of the MGMT status [23 3031 32 33]

Our method shows good overall performance but there are noticeable per-formance differences between tumor categories For example when our methodpredicts a tumor as an IDH wildtype glioblastoma it is correct almost all of thetime On the other hand it has some difficulty differentiating IDH mutated1p19q co-deleted low-grade glioma from other low-grade glioma The sensitiv-ity for the prediction of grade III glioma was low which might be caused bythe lack of a central pathology review Because of this there were differences inmolecular testing and histological analysis and it is known that distinguishingbetween grade II and grade III has a poor observer reliability [34] Althoughour method can be relevant for certain subgroups our methodrsquos performancestill needs to be improved to ensure relevancy for the full patient population

In future work we aim to increase the performance of our method by includ-ing perfusion-weighted imaging (PWI) and diffusion-weighted imaging (DWI)since there has been an increasing amount of evidence that these physiologi-cal imaging modalities contain additional information that correlates with thetumorrsquos genetic status and aggressiveness [35 36] They were not included inthis study since PWI and to a lesser extent DWI are not as ingrained in theclinical imaging routine as the structural scans used in this work [19 20] Thusincluding these modalities would limit our methodrsquos clinical applicability andsubstantially reduce the number of patients in the train and test set However

14

PWI and DWI are increasingly becoming more commonplace which will allowincluding these in future research and which might improve performance

In conclusion we have developed a non-invasive method that can predict theIDH mutation status 1p19q co-deletion status and grade of glioma while atthe same time segmenting the tumor based on pre-operative MRI scans withhigh overall performance Although the performance of our method might needto be improved before it will find widespread clinical acceptance we believe thatthis research is an important step forward in the field of radiomics Predict-ing multiple clinical features simultaneously steps away from the conventionalsingle-task methods and is more in line with the clinical practice where multipleclinical features are considered simultaneously and may even be related Fur-thermore by not limiting the patient population used to develop our method toa selection based on clinical or radiological characteristics we alleviate the needfor a priori (expert) knowledge which may not always be available Althoughsteps still have to be taken before radiomics will find its way into the clinicespecially in terms of performance our work provides a crucial step forward byresolving some of the hurdles of clinical implementation now and paving theway for a full transition in the future

4 Methods

41 Patient population

The train set was collected from four in-house datasets and five publicly avail-able datasets In-house datasets were collected from four different institutesErasmus MC (EMC) Haaglanden Medical Center (HMC) Amsterdam UMC(AUMC) [37] and University Medical Center Utrecht (UMCU) Four of the fivepublic datasets were collected from The Cancer Imaging Archive (TCIA) [38]the Repository of Molecular Brain Neoplasia Data (REMBRANDT) collection[39] the Clinical Proteomic Tumor Analysis Consortium Glioblastoma Multi-forme (CPTAC-GBM) collection [40] the Ivy Glioblastoma Atlas Project (IvyGAP) collection [41 42] and the Brain-Tumor-Progression collection [43] Thefifth dataset was the 2019 Brain Tumor Segmentation challenge (BraTS) chal-lenge dataset [44 45 46] from which we excluded the patients that were alsoavailable in the TCGA-LGG and TCGA-GBM collections [47 48]

For the internal datasets from the EMC and the HMC manual segmentationswere available which were made by four different clinical experts For patientswhere segmentations from more than one observer were available we randomlypicked one of the segmentations to use in the train set The segmentationsfrom the AUMC data were made by a single observer of the study by Visseret al [37] From the public datasets only the BraTS dataset and the Brain-Tumor-Progression dataset provided manual segmentations Segmentations ofthe BraTS dataset as provided in the 2019 training and validation set wereused For the Brain-Tumor-Progression dataset the segmentations as providedin the TCIA data collection were used

15

Patients were included if pre-operative pre- and post-contrast T1w T2wand T2w-FLAIR scans were available no further inclusion criteria were setFor example patients were not excluded based on the radiological characteristicsof the scan such as low imaging quality or imaging artifacts or the gliomarsquosclinical characteristics such as the grade If multiple scans of the same contrasttype were available in a single scan session (eg multiple T2w scans) the scanupon which the segmentation was made was selected If no segmentation wasavailable or the segmentation was not made based on that scan contrast thescan with the highest axial resolution was used where a 3D acquisition waspreferred over a 2D acquisition

For the in-house data genetic and histological data were available for theEMC HMC and UMCU dataset which were obtained from analysis of tumortissue after biopsy or resection Genetic and histological data of the publicdatasets were also available for the REMBRANDT CPTAC-GBM and IvyGAP collections Data for the REMBRANDT and CPTAC-GBM collectionswas collected from the clinical data available at the TCIA [40 39] For theIvy GAP collection the genetic and histological data were obtained from theSwedish Institute at httpsivygapswedishorghome

As a test set we used the TCGA-LGG and TCGA-GBM collections from theTCIA [47 48] Genetic and histological labels were obtained from the clinicaldata available at the TCIA Segmentations were used as available from theTCIA based on the 2018 BraTS challenge [45 49 50] The inclusion criteriafor the patients included in the BraTS challenge were the same as our inclusioncriteria the presence of a pre-operative pre- and post-contrast T1w T2w andT2w-FLAIR scan Thus patients from the TCGA-LGG and TCGA-GBM wereincluded if a segmentation from the BraTS challenge was available Howeverfor three patients we found that although they did have manual segmentationsthey did not meet our inclusion requirements TCGA-08-0509 and TCGA-08-0510 from TCGA-GBM because they did not have a pre-contrast T1w scan andTCGA-FG-7634 from TCGA-LGG because there was no post-contrast T1wscan

42 Automatic segmentation in the train set

To present our method with a large diversity in scans we wanted to include asmany patients in the train set as possible from the different datasets Thereforewe performed automatic segmentation in patients that did not have manualsegmentations To this end we used an initial version of our network (presentedin Section 44) without the additional layers that were needed for the predictionof the genetic and histological features This network was initially trained usingall patients in the train set for whom a manual segmentation was availableand this trained network was then applied to all patients for which a manualsegmentation was not available The resulting automatic segmentations wereinspected and if their quality was acceptable they were added to the trainset The network was then trained again using this increased dataset and wasapplied to scans that did not yet have a segmentation of acceptable quality

16

This process was repeated until an acceptable segmentation was available forall patients which constituted our final complete train set

43 Pre-processing

For all datasets except for the BraTS dataset for which the scans were alreadyprovided in NIfTI format the scans were converted from DICOM format toNIfTI format using dcm2niix version v1020190410 [51] We then registeredall scans to the MNI152 T1w and T2w atlases version ICBM 2009a whichhad a resolution of 1x1x1 mm3 and a size of 197x233x189 voxels [52 53] Thescans were affinely registered using Elastix 50 [54 55] The pre- and post-contrast T1w scans were registered to the T1w atlas the T2w and T2w-FLAIRscans were registered to the T2w atlas When a manual segmentation wasavailable for patients from the in-house datasets the registration parametersthat resulted from registering the scan used during the segmentation were usedto transform the segmentation to the atlas In the case of the public datasetswe used the registration parameters of the T2w-FLAIR scans to transform thesegmentations

After the registration all scans were N4 bias field corrected using SimpleITKversion 124 [56] A brain mask was made for the atlas using HD-BET both forthe T1w atlas and the T2w atlas [57] This brain mask was used to skull stripall registered scans and crop them to a bounding box around the brain maskreducing the amount of background present in the scans resulting in a scan sizeof 152x182x145 voxels Subsequently the scans were normalized such that foreach scan the average image intensity was 0 and the standard deviation of theimage intensity was 1 within the brain mask Finally the background outsidethe brain mask was set to the minimum intensity value within the brain mask

Since the segmentation could sometimes be rugged at the edges after regis-tration especially when the segmentations were initially made on low-resolutionscans we smoothed the segmentation using a 3x3x3 median filter (this was onlydone in the train set) For segmentations that contained more than one la-bel eg when the tumor necrosis and enhancement were separately segmentedall labels were collapsed into a single label to obtain a single segmentation ofthe whole tumor The genetic and histological labels and the segmentations ofeach patient were one-hot encoded The four scans ground truth labels andsegmentation of each patient were then used as the input to the network

44 Model

We based the architecture of our model on the U-Net architecture with someadaptations made to allow for a full 3D input and the auxiliary tasks [58] Ournetwork architecture which we have named PrognosAIs Structure-Net or PS-Net for short can be seen in Figure 8

To use the full 3D scan as an input to the network we replaced the firstpooling layer that is usually present in the U-Net with a strided convolutionwith a kernel size of 9x9x9 and a stride of 3x3x3 In the upsampling branch of

17

32

32 64

128

256

512 256

7x8x7256 128

128 64

64 32

32 2

Segmentation

145x182x152

49x61x51

25x31x26

13x16x13

1472

512 2IDH

512 2

1p19q

512 3Grade

Batch normalization Concatenation Convolution amp ReLU3x3x3

Convolution amp Softmax1x1x1

(De)convolution amp ReLU9x9x9

stride 3x3x3

Dense amp ReLU Dense amp Softmax Dropout

Max pooling2x2x2

Up-convolution amp ReLU2x2x2

Global maxpooling

Figure 8 Overview of the PrognosAIs Structure-Net (PS-Net) architecture usedfor our model The numbers below the different layers indicate the number offilters dense units or features at that layer We have also indicated the featuremap size at the different depths of the network

the network the last up-convolution is replaced by a deconvolution with thesame kernel size and stride

At each depth of the network we have added global max-pooling layersdirectly after the dropout layer to obtain imaging features that can be used topredict the genetic and histological features We chose global pooling layers asthey do not introduce any additional parameters that need to be trained thuskeeping the memory required by our model manageable The features from thedifferent depths of the network were concatenated and fed into three differentdense layers one for each of the genetic and histological outputs

l2 kernel regularization was used in all convolutional layers except for thelast convolutional layer used for the output of the segmentation In total thismodel contained 27042473 trainable an 2944 non-trainable parameters

18

45 Model training

Training of the model was done on eight NVidia RTX2080Tirsquos with 11GB ofmemory using TensorFlow 220 [59] To be able to use the full 3D scan asinput to the network without running into memory issues we had to optimizethe memory efficiency of the model training Most importantly we used mixed-precision training which means that most of the variables of the model (such asthe weights) were stored in float16 which requires half the memory of float32which is typically used to store these variables [60] Only the last softmaxactivation layers of each output were stored as float32 We also stored ourpre-processed scans as float16 to further reduce memory usage

However even with these settings we could not use a batch size larger than1 It is known that a larger batch size is preferable as it increases the stabilityof the gradient updates and allows for a better estimation of the normalizationparameters in batch normalization layers [61] Therefore we distributed thetraining over the eight GPUs using the NCCL AllReduce algorithm whichcombines the gradients calculated on each GPU before calculating the updateto the model parameters [62] We also used synchronized batch normalizationlayers which synchronize the updates of their parameters over the distributedmodels In this way our model had a virtual batch size of eight for the gradientupdates and the batch normalization layers parameters

To provide more samples to the algorithm and prevent potential overtrain-ing we applied four types of data augmentation during training croppingrotation brightness shifts and contrast shifts Each augmentation was appliedwith a certain augmentation probability which determined the probability ofthat augmentation type being applied to a specific sample When an image wascropped a random number of voxels between 0 and 20 was cropped from eachdimension and filled with zeros For the random rotation an angle betweenminus30 and 30 degrees was selected from a uniform distribution for each dimen-sion The brightness shift was applied with a delta uniformly drawn between0 and 02 and the contrast shift factor was randomly drawn between 085 and115 We also introduced an augmentation factor which determines how ofteneach sample was parsed as an input sample during a single epoch where eachtime it could be augmented differently

For the IDH 1p19q and grade output we used a masked categorical cross-entropy loss and for the segmentation we used a DICE loss see Appendix E fordetails We used AdamW as an optimizer which has shown improved general-ization performance over Adam by introducing the weight decay parameter as aseparate parameter from the learning rate [63] The learning rate was automat-ically reduced by a factor of 025 if the loss did not improve during the last fiveepochs with a minimum learning rate of 1 middot 10minus11 The model could train for amaximum of 150 epochs and training was stopped early if the average loss overthe last five epochs did not improve Once the model was finished training theweights from the epoch with the lowest loss were restored

19

46 Hyperparameter tuning

Hyperparameters involved in the training of the model needed to be tuned toachieve the best performance We tuned a total of six hyper parameters the l2-norm the dropout rate the augmentation factor the augmentation probabilitythe optimizerrsquos initial learning rate and the optimizerrsquos weight decay A fulloverview of the trained parameters and the values tested for the different settingsis presented in Appendix F

To tune these hyperparameters we split the train set into a hyperparametertraining set (851282 patients of the full train data) and a hyperparametervalidation set (15226 patients of the full train data) Models were trained fordifferent hyperparameter settings via an exhaustive search using the hyperpa-rameter train set and then evaluated on the hyperparameter validation set Nodata augmentation was applied to the hyperparameter validation to ensure thatresults between trained models were comparable The hyperparameters that ledto the lowest overall loss in the hyperparameter validation set were chosen asthe optimal hyperparameters We trained the final model using these optimalhyperparameters and the full train set

47 Post-processing

The predictions of the network were post-processed to obtain the final predictedlabels and segmentations for the samples Since softmax activations were usedfor the genetic and histological outputs a prediction between 0 and 1 was out-putted for each class where the individual predictions summed to 1 The finalpredicted label was then considered as the class with the highest prediction scoreFor the prediction of LGG (grade IIIII) vs HGG (grade IV) the predictionscores of grade II and grade III were combined to obtain the prediction scorefor LGG the prediction score of grade IV was used as the prediction score forHGG If a segmentation contained multiple unconnected components we onlyretained the largest component to obtain a single whole tumor segmentation

48 Model evaluation

The performance of the final trained model was evaluated on the independenttest set comparing the predicted labels with the ground truth labels For thegenetic and histological features we evaluated the AUC the accuracy the sen-sitivity and the specificity using scikit-learn version 0231 for details see Ap-pendix G [64] We evaluated these metrics on the full test set and in subcate-gories relevant to the WHO 2016 guidelines We evaluated the IDH performanceseparately in the LGG (grade IIIII) and HGG (grade IV) subgroups the 1p19qperformance in LGG and we also evaluated the performance of distinguishingbetween LGG and HGG instead of predicting the individual grades

To evaluate the performance of the segmentation we calculated the DICEscores Hausdorff distances and volumetric similarity coefficient comparing theautomatic segmentation of our method and the manual ground truth segmen-

20

tations for all patients in the test set These metrics were calculated usingthe EvaluateSegmentation toolbox version 20170425 [65] for details see Ap-pendix G

To prevent an overly optimistic estimation of our modelrsquos predictive valuewe only evaluated our model on the test set once all hyperparameters werechosen and the final model was trained In this way the performance in thetest set did not influence decisions made during the development of the modelpreventing possible overfitting by fine-tuning to the test set

To gain insight into the model we made saliency maps that show whichparts of the scan contribute the most to the prediction of the CNN [66] Saliencymaps were made using tf-keras-vis 052 changing the activation function of alloutput layers from softmax to linear activations using SmoothGrad to reducethe noisiness of the saliency maps [66]

Another way to gain insight into the networkrsquos behavior is to visualize thefilter outputs of the convolutional layers as they can give some idea as to whatoperations the network applies to the scans We visualized the filter outputs ofthe last convolutional layers in the downsample and upsample path at the firstdepth (at an image size of 49x61x51) of our network These filter outputs werevisualized by passing a sample through the network and showing the convolu-tional layersrsquo outputs replacing the ReLU activation with linear activations

49 Data availability

An overview of the patients included from the public datasets used in the train-ing and testing of the algorithm and their ground truth label is available inAppendix H The data from the public datasets are available in TCIA un-der DOIs 107937K9TCIA2015588OZUZB 107937k9tcia20183rje41q1107937K9TCIA2016XLwaN6nL and 107937K9TCIA201815quzvnb Datafrom the BraTS are available at httpbraintumorsegmentationorg Datafrom the in-house datasets are not publicly available due to participant privacyand consent

410 Code availability

The code used in this paper is available on GitHub under an Apache 2 license athttpsgithubcomSvdvoortPrognosAIs_glioma This code includes thefull pipeline from registration of the patients to the final post-processing of thepredictions The trained model is also available on GitHub along with code toapply it to new patients

21

Appendices

A Confusion matrices

Tables 3 4 and 5 show the confusion matrices for the IDH 1p19q and gradepredictions and Table 6 shows the confusion matrix for the WHO 2016 subtypes

Table 4 shows that the algorithm mainly has difficulty recognizing 1p19qco-deleted tumors which are mostly predicted as 1p19q intact Table 5 showsthat most of the incorrectly predicted grade III tumors are predicted as gradeIV tumors

Table 6 shows that our algorithm often incorrectly predicts IDH-wildtypeastrocytoma as IDH-wildtype glioblastoma The latest cIMPACT-NOW guide-lines propose a new categorization in which IDH-wildtype astrocytoma thatshow either TERT promoter methylation or EFGR gene amplification or chro-mosome 7 gainchromsome 10 loss are classified as IDH-wildtype glioblastoma[22] This new categorization is proposed since the survival of patients withthose IDH-wildtype astrocytoma is similar to the survival of patients with IDH-wildtype glioblastoma [22] From the 13 IDH-wildtype astrocytoma that werewrongly predicted as IDH-wildtype glioblastoma 12 would actually be cate-gorized as IDH-wildtype glioblastoma under this new categorization Thusalthough our method wrongly predicted the WHO 2016 subtype it might ac-tually have picked up on imaging features related to the aggressiveness of thetumor which might lead to a better categorization

Table 3 Confusion matrix of the IDH predictions

Predicted

Wildtype Mutated

Actu

al

Wildtype 120 9

Mutated 25 63

Table 4 Confusion matrix of the 1p19q predictions

Predicted

Intact Co-deleted

Actu

al

Intact 197 10

Co-deleted 16 10

22

Table 5 Confusion matrix of the grade predictions

Predicted

Grade II Grade III Grade IV

Actu

al Grade II 35 6 6

Grade III 19 10 30

Grade IV 2 5 125

Table 6 Confusion matrix of the WHO 2016 predictions The rsquootherrsquo categoryindicates patients that were predicted as a non-existing WHO 2016 subtype forexample IDH wildtype 1p19q co-deleted tumors Only one patient (TCGA-HT-A5RC) was predicted as a non-existing category It was predicted as anIDH wildtype 1p19q co-deleted grade IV tumor

Predicted

Oligodendrogliom

a

IDH-m

utated

astrocytoma

IDH-w

ildtype

astrocytoma

IDH-m

utated

glioblastoma

IDH-w

ildtype

glioblastoma

Other

Actu

al

Oligodendroglioma 10 8 1 0 7 0

IDH-mutatedastrocytoma 6 34 4 3 10 0

IDH-wildtypeastrocytoma 1 2 3 2 13 1

IDH-mutatedglioblastoma 0 1 0 0 3 0

IDH-wildtypeglioblastoma 0 3 3 1 96 0

Oligodendroglioma are IDH-mutated 1p19q co-deleted grade IIIII giomaIDH-mutated astrocytoma are IDH-mutated 1p19q intact grade IIIII gliomaIDH-wildtype astrocytoma are IDH-wildtype 1p19q intact grade IIIII gliomaIDH-mutated glioblastoma are IDH-mutated grade IV gliomaIDH-wildtype glioblastoma are IDH-wildtype grade IV glioma

23

B Segmentation examples

To demonstrate the automatic segmentations made by our method we randomlyselected five patients from both the TCGA-LGG and the TCGA-GBM datasetThe scans and segmentations of the five patients from the TCGA-LGG datasetand the TCGA-GBM dataset are shown in Figures 9 and 10 respectively TheDICE score Hausdorff distance and volumetric similarity coefficient for thesepatients are given in Table 7 The method seems to mostly focus on the hyper-intensities of the T2w-FLAIR scan Despite the registrations issues that can beseen for the T2w scan in Figure 10d the tumor was still properly segmenteddemonstrating the robustness of our method

Patient DICE HD (mm) VSC

TCGA-LGG

TCGA-DU-7301 089 103 095TCGA-FG-5964 080 58 082TCGA-FG-A713 073 78 088TCGA-HT-7475 087 149 090TCGA-HT-8106 088 112 099

TCGA-GBM

TCGA-02-0037 082 226 099TCGA-08-0353 091 130 098TCGA-12-1094 090 73 093TCGA-14-3477 090 165 099TCGA-19-5951 073 197 073

Table 7 The DICE score Hausdorff distance (HD) and volumetric similaritycoefficient (VSC) for the randomly selected patients from the TCGA-LGG andTCGA-GBM data collections

24

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-DU-7301 from the TCGA-LGG data collection

(b) Patient TCGA-FG-5964 from the TCGA-LGG data collection

(c) Patient TCGA-FG-A713 from the TCGA-LGG data collection

(d) Patient TCGA-HT-7475 from the TCGA-LGG data collection

(e) Patient TCGA-HT-8106 from the TCGA-LGG data collection

Figure 9 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-LGG data collection

25

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Patient TCGA-02-0037 from the TCGA-GBM data collection

(b) Patient TCGA-08-0353 from the TCGA-GBM data collection

(c) Patient TCGA-12-1094 from the TCGA-GBM data collection

(d) Patient TCGA-14-3477 from the TCGA-GBM data collection

(e) Patient TCGA-19-5951 from the TCGA-GBM data collection

Figure 10 Examples of scans and automatic segmentations of five patients thatwere randomly selected from the TCGA-GBM data collection

26

C Prediction results in the test set

27

D Filter output visualizations

Figures 11 and 12 show the output of the convolution filters for the same LGGpatient as shown in Figure 5a and Figures 13 and 14 show the output of theconvolution filters for the same HGG patient as shown in Figure 5b Figures 11and 13 show the outputs of the last convolution layer in the downsample pathat the feature size of 49x61x51 (the fourth convolutional layer in the network)Figures 12 and 14 show the outputs of the last convolution layer in the upsamplepath at the feature size of 49x61x51 (the nineteenth convolutional layer in thenetwork)

Comparing Figure 11 to Figure 12 and Figure 13 to Figure 14 we can seethat the convolutional layers in the upsample path do not keep a lot of detailfor the healthy part of the brain as this region seems blurred However withinthe tumor different regions can still be distinguished The different parts ofthe tumor from the scans can also be seen such as the contrast-enhancing partand the high signal intensity on the T2w-FLAIR For the grade IV glioma inFigure 14 some filters such as filter 26 also seem to focus on the necrotic partof the tumor

28

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 11 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

29

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 12 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-DU-6400 This is an IDH mutated 1p19q co-deleted grade II glioma

30

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 13 Filter output visualizations of the last convolutional layer in thedownsample path of the network at feature map size 49x61x51 for patientTCGA-06-0238 This is an IDH wildtype grade IV glioma

31

Pre-contrast T1w Post-contrast T1w T2w T2w-FLAIR

(a) Scans used to derive the convolutional layer filter output visualizations

Filter 1Filter 1 Filter 2Filter 2 Filter 3Filter 3 Filter 4Filter 4 Filter 5Filter 5 Filter 6Filter 6 Filter 7Filter 7 Filter 8Filter 8

Filter 9Filter 9 Filter 10Filter 10 Filter 11Filter 11 Filter 12Filter 12 Filter 13Filter 13 Filter 14Filter 14 Filter 15Filter 15 Filter 16Filter 16

Filter 17Filter 17 Filter 18Filter 18 Filter 19Filter 19 Filter 20Filter 20 Filter 21Filter 21 Filter 22Filter 22 Filter 23Filter 23 Filter 24Filter 24

Filter 25Filter 25 Filter 26Filter 26 Filter 27Filter 27 Filter 28Filter 28 Filter 29Filter 29 Filter 30Filter 30 Filter 31Filter 31 Filter 32Filter 32

Filter 33Filter 33 Filter 34Filter 34 Filter 35Filter 35 Filter 36Filter 36 Filter 37Filter 37 Filter 38Filter 38 Filter 39Filter 39 Filter 40Filter 40

Filter 41Filter 41 Filter 42Filter 42 Filter 43Filter 43 Filter 44Filter 44 Filter 45Filter 45 Filter 46Filter 46 Filter 47Filter 47 Filter 48Filter 48

Filter 49Filter 49 Filter 50Filter 50 Filter 51Filter 51 Filter 52Filter 52 Filter 53Filter 53 Filter 54Filter 54 Filter 55Filter 55 Filter 56Filter 56

Filter 57Filter 57 Filter 58Filter 58 Filter 59Filter 59 Filter 60Filter 60 Filter 61Filter 61 Filter 62Filter 62 Filter 63Filter 63 Filter 64Filter 64

(b) Filter output visualizations

Figure 14 Filter output visualizations of the last convolutional layer in theupsample path of the network at feature map size 49x61x51 for patient TCGA-06-0238 This is an IDH wildtype grade IV glioma

32

E Training losses

During the training of the network we used masked categorical cross-entropy lossfor the IDH 1p19q and grade outputs The normal categorical cross-entropyloss is defined as

LCEbatch = minus 1

Nbatch

sumj

sumiisinC

yij log (yij) (1)

where LCEbatch is the total cross-entropy loss over a batch yij is the ground truth

label of sample j for class i yij is the prediction score for sample j for class iC is the set of classes and Nbatch is the number of samples in the batch Here itis assumed that the ground truth labels are one-hot-encoded thus yij is either0 or 1 for each class In our case the ground truth is not known for all sampleswhich can be incorporated in Equation (1) by setting yij to 0 for all classesfor a sample for which the ground truth is not known That sample wouldthen not contribute to the overall loss and would not contribute to the gradientupdate However this can skew the total loss over a batch since the loss is stillaveraged over the total number of samples in a batch regardless of whether theground truth is known resulting in a lower loss for batches that contained moresamples with unknown ground truth Therefore we used a masked categoricalcross-entropy loss

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

yij log (yij) (2)

where

microbatchj =

Nbatchsumij yij

sumi

yij (3)

is the batch weight for sample j In this way the total batch loss is only averagedover the samples that actually have a ground truth

Since there was an imbalance between the number of ground truth samplesfor each class we used class weights to compensate for this imbalance Thusthe loss becomes

LCEbatch = minus 1

Nbatch

sumj

microbatchj

sumiisinC

microclassi yij log (yij) (4)

where

microclassi =

N

Ni |C|(5)

is the class weight for class i N is the total number of samples with knownground truth Ni is the number of samples of class i and |C| is the number ofclasses By determining the class weight in this way we ensured that

microclassi Ni =

N

|C|= constant (6)

33

Thus each class would have the same contribution to the overall loss Theseclass weights were (individually) determined for the IDH output the 1p19qoutput and the grade output

For the segmentation output we used the DICE loss

LDICEbatch =

sumj

1minus 2 middotsumvoxels

k yjk middot yjksumvoxelsk yjk + yjk

(7)

where yjk is the ground truth label in voxel k of sample j and yjk is theprediction score outputted for voxel k of sample j

The total loss that was optimized for the model was a weighted sum of thefour individual losses

Ltotal =summ

micromLm (8)

with

microm =1

Xm (9)

where Lm is the loss for output m microm is the loss weight for loss m (either theIDH 1p19q grade or segmentation loss) and Xm is the number of sampleswith known ground truth for output m In this way we could counteract theeffect of certain outputs having more known labels than other outputs

34

F Parameter tuning

Table 8 Hyperparameters that were tuned and the values that were testedValues in bold show the selected values used in the final model

Tuning parameter Tested values

Dropout rate 015 02 025 030 035 040l2-norm 00001 000001 0000001Learning rate 001 0001 00001 000001 00000001Weight decay 0001 00001 000001Augmentation factor 1 2 3Augmentation probability 025 030 035 040 045

35

G Evaluation metrics

We calculated the AUC accuracy sensitivity and specificity metrics for thegenetic and histological features for the definitions of these metrics see [67]

For the IDH and 1p19q co-deletion outputs the IDH mutated and the1p19q co-deleted samples were regarded as the positive class respectively Sincethe grade was a multi-class problem no single positive class could be determinedFor the prediction of the individual grades that grade was seen as the positiveclass and all other grades as the negative class (eg in the case of the gradeIII prediction grade III was regarded as the positive class and grade II and IVwere regarded as the negative class) For the LGG vs HGG prediction LGGwas considered as the positive class and HGG as the negative class For theevaluation of these metrics for the genetic and histological features only thesubjects with known ground truth were taken into account

The overall AUC for the grade was a multi-class AUC determined in a one-vs-one approach comparing each class against the others in this way thismetric was insensitive to class imbalance [68] A multi-class accuracy was usedto determine the overall accuracy for the grade predictions [67]

To evaluate the performance of the automated segmentation we evaluatedthe DICE score the Hausdorff distance and the volumetric similarity coefficientThe DICE score is a measure of overlap between two segmentations where avalue of 1 indicates perfect overlap and the Hausdorff distance is a measureof the closeness of the borders of the segmentations The volumetric similaritycoefficient is a measure of the agreement between the volumes of two segmen-tations without taking account the actual location of the tumor where a valueof 1 indicates perfect agreement See [65] for the definitions of these metrics

36

H Ground truth labels of patients included frompublic datasets

Acknowledgments

Sebastian van der Voort and Fatih Incekara acknowledge funding by the DutchCancer Society (KWF project number EMCR 2015-7859)

Data used in this publication were generated by the National Cancer Insti-tute Clinical Proteomic Tumor Analysis Consortium (CPTAC)

The results published here are in whole or part based upon data generatedby the TCGA Research Network httpcancergenomenihgov

Author contributions

SRvdV FI WJN MS and SK contrived the study and designed theexperiments FI MMJW JWS RNT GJL PCDWH RSEAJPEV MJvdB and MS included patients in the different studiesSRvdV FI MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV MJvdB and MS collected the dataSRvdV carried out the experiments SRvdV FI MS and SK inter-preted the results SRvdV FI MJvdB MS and SK created the initialdraft of the paper MMJW GK RG JWS RNT GJL PCDWHRSE PJF HJD AJPEV WJN and MJvdB revised the paper

References

[1] OFFICE FOR NATIONAL STATISTICS CANCER SURVIVAL IN ENG-LAND Adult Stage at Diagnosis and Childhood-Patients Followed Up to2018 DANDY BOOKSELLERS Limited 2019

[2] Hendrikus J Dubbink Peggy N Atmodimedjo Johan M Kros Pim JFrench Marc Sanson Ahmed Idbaih Pieter Wesseling Roelien EntingWim Spliet Cees Tijssen Winand N M Dinjens Thierry Gorlia andMartin J van den Bent Molecular classification of anaplastic oligoden-droglioma using next-generation sequencing a report of the prospectiverandomized EORTC brain tumor group 26951 phase III trial Neuro-Oncology 18(3)388ndash400 March 2016 ISSN 1522-8517 URL https

doiorg101093neuoncnov182

[3] Jeanette E Eckel-Passow Daniel H Lachance Annette M Molinaro Kyle MWalsh Paul A Decker Hugues Sicotte Melike Pekmezci Terri Rice Matt LKosel Ivan V Smirnov Gobinda Sarkar Alissa A Caron Thomas M

37

Kollmeyer Corinne E Praska Anisha R Chada Chandralekha Halder He-len M Hansen Lucie S McCoy Paige M Bracci Roxanne Marshall ShichunZheng Gerald F Reis Alexander R Pico Brian P OrsquoNeill Jan C BucknerCaterina Giannini Jason T Huse Arie Perry Tarik Tihan Mitchell SBerger Susan M Chang Michael D Prados Joseph Wiemels John KWiencke Margaret R Wrensch and Robert B Jenkins Glioma groupsbased on 1p19q IDH and TERT promoter mutations in tumors NewEngland Journal of Medicine 372(26)2499ndash2508 June 2015 ISSN 0028-4793 URL httpsdoiorg101056NEJMoa1407279

[4] David N Louis Arie Perry Guido Reifenberger Andreas von Deimling Do-minique Figarella-Branger Webster K Cavenee Hiroko Ohgaki Otmar DWiestler Paul Kleihues and David W Ellison The 2016 world health orga-nization classification of tumors of the central nervous system a summaryActa Neuropathologica 131(6)803ndash820 June 2016 ISSN 0001-6322 URLhttpsdoiorg101007s00401-016-1545-1

[5] Ching-Chang Chen Peng-Wei Hsu Tai-Wei Erich Wu Shih-Tseng LeeChen-Nen Chang Kuo-chen Wei Chih-Cheng Chuang Chieh-Tsai WuTai-Ngar Lui Yung-Hsin Hsu Tzu-Kang Lin Sai-Cheung Lee and Yin-Cheng Huang Stereotactic brain biopsy Single center retrospective anal-ysis of complications Clinical Neurology and Neurosurgery 111(10)835ndash839 December 2009 ISSN 0303-8467 URL httpsdoiorg101016

jclineuro200908013

[6] Robert J Jackson Gregory N Fuller Dima Abi-Said Frederick F LangZiya L Gokaslan Wei Ming Shi David M Wildrick and Raymond SawayaLimitations of stereotactic biopsy in the initial management of gliomasNeuro-Oncology 3(3)193ndash200 July 2001 ISSN 1522-8517 URL https

doiorg101093neuonc33193

[7] M Zhou J Scott B Chaudhury L Hall D Goldgof KW Yeom M IvY Ou J Kalpathy-Cramer S Napel R Gillies O Gevaert and R GatenbyRadiomics in brain tumor Image assessment quantitative feature de-scriptors and machine-learning approaches American Journal of Neu-roradiology 39(2)208ndash216 February 2018 ISSN 0195-6108 URL https

doiorg103174ajnrA5391

[8] Wenya Linda Bi Ahmed Hosny Matthew B Schabath Maryellen LGiger Nicolai J Birkbak Alireza Mehrtash Tavis Allison Omar ArnaoutChristopher Abbosh Ian F Dunn Raymond H Mak Rulla M TamimiClare M Tempany Charles Swanton Udo Hoffmann Lawrence H SchwartzRobert J Gillies Raymond Y Huang and Hugo J W L Aerts Artificialintelligence in cancer imaging Clinical challenges and applications CAA Cancer Journal for Clinicians 69(2)caac21552 February 2019 ISSN0007-9235 URL httpsdoiorg103322caac21552

38

[9] Ahmad Chaddad Michael Jonathan Kucharczyk Paul Daniel SihamSabri Bertrand J Jean-Claude Tamim Niazi and Bassam AbdulkarimRadiomics in glioblastoma Current status and challenges facing clinicalimplementation Frontiers in Oncology 9374ndash374 May 2019 ISSN 2234-943X URL httpsdoiorg103389fonc201900374

[10] Marion Smits Imaging of oligodendroglioma The British Journal of Ra-diology 89(1060)20150857 April 2016 ISSN 0007-1285 URL https

doiorg101259bjr20150857

[11] Rachel L Delfanti David E Piccioni Jason Handwerker Naeim BahramiAnithaPriya Krishnan Roshan Karunamuni Jona A Hattangadi-GluthTyler M Seibert Ashwin Srikant Karra A Jones Vivian S Snyder An-ders M Dale Nathan S White Carrie R McDonald and Nikdokht FaridImaging correlates for the 2016 update on WHO classification of gradeIIIII gliomas implications for IDH 1p19q and ATRX status Journal ofNeuro-Oncology 135(3)601ndash609 December 2017 ISSN 0167-594X URLhttpsdoiorg101007s11060-017-2613-7

[12] Sonal Gore Tanay Chougule Jayant Jagtap Jitender Saini and MadhuraIngalhalikar A review of radiomics and deep predictive modeling in gliomacharacterization Academic Radiology July 2020 ISSN 1076-6332 URLhttpsdoiorg101016jacra202006016

[13] Hugo J W L Aerts Emmanuel Rios Velazquez Ralph T H LeijenaarChintan Parmar Patrick Grossmann Sara Carvalho Johan BussinkRene Monshouwer Benjamin Haibe-Kains Derek Rietveld Frank HoebersMichelle M Rietbergen C Rene Leemans Andre Dekker John Quacken-bush Robert J Gillies and Philippe Lambin Decoding tumour pheno-type by noninvasive imaging using a quantitative radiomics approach Na-ture Communications 5(1)4006 September 2014 ISSN 2041-1723 URLhttpsdoiorg101038ncomms5006

[14] Sarah Chihati and Djamel Gaceb A review of recent progress in deeplearning-based methods for MRI brain tumor segmentation In 202011th International Conference on Information and Communication Sys-tems (ICICS) pages 149ndash154 Institute of Electrical and Electronics Engi-neers (IEEE) April 2020 ISBN 9781728162270 URL httpsdoiorg

101109icics494692020239550

[15] Robert J Gillies Paul E Kinahan and Hedvig Hricak Radiomics Imagesare more than pictures they are data Radiology 278(2)563ndash577 Febru-ary 2016 ISSN 0033-8419 URL httpsdoiorg101148radiol

2015151169

[16] James H Thrall Xiang Li Quanzheng Li Cinthia Cruz Synho Do KeithDreyer and James Brink Artificial intelligence and machine learning in ra-diology Opportunities challenges pitfalls and criteria for success Journal

39

of the American College of Radiology 15(3 Part B)504ndash508 March 2018ISSN 1546-1440 URL httpsdoiorg101016jjacr201712026

[17] Ahmed Hosny Chintan Parmar John Quackenbush Lawrence H Schwartzand Hugo J W L Aerts Artificial intelligence in radiology Nature ReviewsCancer 18(8)500ndash510 August 2018 ISSN 1474-175X URL https

doiorg101038s41568-018-0016-5

[18] Okan Kopuklu Neslihan Kose Ahmet Gunduz and Gerhard Rigoll Re-source efficient 3D convolutional neural networks In 2019 IEEECVFInternational Conference on Computer Vision Workshop (ICCVW) Insti-tute of Electrical and Electronics Engineers (IEEE) October 2019 ISBN9781728150239 URL httpsdoiorg101109iccvw201900240

[19] S C Thust S Heiland A Falini H R Jager A D Waldman P C SundgrenC Godi V K Katsaros A Ramos N Bargallo M W Vernooij T YousryM Bendszus and M Smits Glioma imaging in europe A survey of 220centres and recommendations for best clinical practice European Radiology28(8)3306ndash3317 August 2018 ISSN 0938-7994 URL httpsdoiorg

101007s00330-018-5314-5

[20] Christian F Freyschlag Sandro M Krieg Johannes Kerschbaumer DanielPinggera Marie-Therese Forster Dominik Cordier Marco Rossi GabrieleMiceli Alexandre Roux Andres Reyes Silvio Sarubbo Anja Smits JoannaSierpowska Pierre A Robe Geert-Jan Rutten Thomas Santarius TomaszMatys Marc Zanello Fabien Almairac Lydiane Mondot Asgeir S JakolaMaria Zetterling Adria Rofes Gord von Campe Remy Guillevin DanieleBagatto Vincent Lubrano Marion Rapp John Goodden Philip C De WittHamer Johan Pallud Lorenzo Bello Claudius Thome Hugues Duffauand Emmanuel Mandonnet Imaging practice in low-grade gliomas amongeuropean specialized centers and proposal for a minimum core of imagingJournal of Neuro-Oncology 139(3)699ndash711 September 2018 ISSN 0167-594X URL httpsdoiorg101007s11060-018-2916-3

[21] M Labussiere A Idbaih X-W Wang Y Marie B Boisselier C FaletS Paris J Laffaire C Carpentier E Criniere F Ducray S El HallaniK Mokhtari K Hoang-Xuan J-Y Delattre and M Sanson All the 1p19qcodeleted gliomas are mutated on IDH1 or IDH2 Neurology 74(23)1886ndash1890 June 2010 ISSN 0028-3878 URL httpsdoiorg101212WNL

0b013e3181e1cf3a

[22] David N Louis Pieter Wesseling Kenneth Aldape Daniel J Brat DavidCapper Ian A Cree Charles Eberhart Dominique Figarella-BrangerMaryam Fouladi Gregory N Fuller Caterina Giannini Christine Haber-ler Cynthia Hawkins Takashi Komori Johan M Kros HK Ng Brent AOrr Sung-Hye Park Werner Paulus Arie Perry Torsten Pietsch GuidoReifenberger Marc Rosenblum Brian Rous Felix Sahm Chitra SarkarDavid A Solomon Uri Tabori Martin J Bent Andreas Deimling Michael

40

Weller Valerie A White and David W Ellison cIMPACT-NOW up-date 6 new entity and diagnostic principle recommendations of thecIMPACT-Utrecht meeting on future CNS tumor classification and grad-ing Brain Pathology 30(4)844ndash856 July 2020 ISSN 1015-6305 URLhttpsdoiorg101111bpa12832

[23] Zhenyu Tang Yuyun Xu Zhicheng Jiao Junfeng Lu Lei Jin Abudumi-jiti Aibaidula Jinsong Wu Qian Wang Han Zhang and Dinggang ShenPre-operative overall survival time prediction for glioblastoma patients us-ing deep learning on both imaging phenotype and genotype In Ding-gang Shen Tianming Liu Terry M Peters Lawrence H Staib CarolineEssert Sean Zhou Pew-Thian Yap and Ali Khan editors Lecture Notesin Computer Science pages 415ndash422 Springer Science and Business Me-dia LLC 2019 ISBN 9783030322380 URL httpsdoiorg101007

978-3-030-32239-7_46

[24] Zhiyuan Xue Bowen Xin Dingqian Wang and Xiuying Wang Radiomics-enhanced multi-task neural network for non-invasive glioma subtypingand segmentation In Hassan Mohy-ud Din and Saima Rathore editorsRadiomics and Radiogenomics in Neuro-oncology pages 81ndash90 SpringerScience and Business Media LLC 2020 ISBN 9783030401238 URLhttpsdoiorg101007978-3-030-40124-5_9

[25] Milan Decuyper Stijn Bonte Karel Deblaere and Roel Van Holen Au-tomated MRI based pipeline for glioma segmentation and prediction ofgrade IDH mutation and 1p19q co-deletion 2020 Preprint at https

arxivorgabs200511965

[26] Stefania Rizzo Francesca Botta Sara Raimondi Daniela Origgi Cris-tiana Fanciullo Alessio Giuseppe Morganti and Massimo Bellomi Ra-diomics the facts and the challenges of image analysis European Ra-diology Experimental 2(1)36 December 2018 ISSN 2509-9280 URLhttpsdoiorg101186s41747-018-0068-z

[27] Philipp Lohmann Norbert Galldiks Martin Kocher Alexander HeinzelChristian P Filss Carina Stegmayr Felix M Mottaghy Gereon R FinkN Jon Shah and Karl-Josef Langen Radiomics in neuro-oncology Basicsworkflow and applications Methods June 2020 ISSN 1046-2023 URLhttpsdoiorg101016jymeth202006003

[28] Stephen S F Yip and Hugo J W L Aerts Applications and limitationsof radiomics Physics in Medicine and Biology 61(13)R150ndashR166 July2016 ISSN 0031-9155 URL httpsdoiorg1010880031-915561

13r150

[29] Annika Malmstrom Ma lgorzata Lysiak Bjarne Winther Kristensen Eliz-abeth Hovey Roger Henriksson and Peter Soderkvist Do we really knowwho has an MGMT methylated glioma Results of an international survey

41

regarding use of MGMT analyses for glioma Neuro-Oncology Practice 7(1)68ndash76 09 2019 ISSN 2054-2577 URL httpsdoiorg101093

nopnpz039

[30] Takahiro Sasaki Manabu Kinoshita Koji Fujita Junya Fukai NobuhideHayashi Yuji Uematsu Yoshiko Okita Masahiro Nonaka Shusuke Mo-riuchi Takehiro Uda Naohiro Tsuyuguchi Hideyuki Arita Kanji MoriKenichi Ishibashi Koji Takano Namiko Nishida Tomoko Shofuda EmaYoshioka Daisuke Kanematsu Yoshinori Kodama Masayuki ManoNaoyuki Nakao and Yonehiro Kanemura Radiomics and MGMT pro-moter methylation for prognostication of newly diagnosed glioblastomaScientific Reports 9(1)14435 December 2019 ISSN 2045-2322 URLhttpsdoiorg101038s41598-019-50849-y

[31] A Gupta A Prager RJ Young W Shi AMP Omuro and JJ GraberDiffusion-weighted MR imaging and MGMT methylation status in glioblas-toma A reappraisal of the role of preoperative quantitative ADC measure-ments American Journal of Neuroradiology 34(1)E10ndashE11 January 2013ISSN 0195-6108 URL httpsdoiorg103174ajnrA3467

[32] JA Carrillo A Lai PL Nghiemphu HJ Kim HS Phillips S KharbandaP Moftakhar S Lalaezari W Yong BM Ellingson TF Cloughesy andWB Pope Relationship between tumor enhancement edema IDH1 muta-tional status MGMT promoter methylation and survival in glioblastomaAmerican Journal of Neuroradiology 33(7)1349ndash1355 August 2012 ISSN0195-6108 URL httpsdoiorg103174ajnrA2950

[33] Vilde Elisabeth Mikkelsen Hong Yan Dai Anne Line Stensjoslashen Erik Mag-nus Berntsen Oslashyvind Salvesen Ole Solheim and Sverre Helge TorpMGMT promoter methylation status is not related to histological or radio-logical features in IDH wild-type glioblastomas Journal of Neuropathologyamp Experimental Neurology 79(8)855ndash862 August 2020 ISSN 0022-3069URL httpsdoiorg101093jnennlaa060

[34] Martin J van den Bent Interobserver variation of the histopathologicaldiagnosis in clinical trials on glioma a clinicianrsquos perspective Acta Neu-ropathologica 120(3)297ndash304 September 2010 ISSN 1432-0533 URLhttpsdoiorg101007s00401-010-0725-7

[35] Ji Eun Park Ho Sung Kim Youngheun Jo Roh-Eul Yoo Seung Hong ChoiSoo Jung Nam and Jeong Hoon Kim Radiomics prognostication modelin glioblastoma using diffusion- and perfusion-weighted MRI ScientificReports 10(1)4250 December 2020 ISSN 2045-2322 URL httpsdoi

org101038s41598-020-61178-w

[36] Minjae Kim So Yeong Jung Ji Eun Park Yeongheun Jo Seo YoungPark Soo Jung Nam Jeong Hoon Kim and Ho Sung Kim Diffusion-and perfusion-weighted MRI radiomics model may predict isocitrate dehy-drogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade

42

glioma European Radiology 30(4)2142ndash2151 April 2020 ISSN 0938-7994URL httpsdoiorg101007s00330-019-06548-3

[37] M Visser DMJ Muller RJM van Duijn M Smits N Verburg EJ HendriksRJA Nabuurs JCJ Bot RS Eijgelaar M Witte MB van Herk F BarkhofPC de WittHamer and JC de Munck Inter-rater agreement in glioma seg-mentations on longitudinal MRI NeuroImage Clinical 22101727 2019ISSN 2213-1582 URL httpsdoiorg101016jnicl2019101727

[38] Kenneth Clark Bruce Vendt Kirk Smith John Freymann Justin KirbyPaul Koppel Stephen Moore Stanley Phillips David Maffitt MichaelPringle Lawrence Tarbox and Fred Prior The cancer imaging archive(TCIA) Maintaining and operating a public information repository Jour-nal of Digital Imaging 26(6)1045ndash1057 December 2013 ISSN 0897-1889URL httpsdoiorg101007s10278-013-9622-7

[39] Lisa Scarpace Adam E Flanders Rajan Jain Tom Mikkelsen and David WAndrews Data from REMBRANDT The Cancer Imaging Archive 2015URL httpsdoiorg107937K9TCIA2015588OZUZB

[40] National Cancer Institute Clinical Proteomic Tumor Analysis Consortium(CPTAC) Radiology data from the clinical proteomic tumor analysisconsortium glioblastoma multiforme CPTAC-GBM collection The CancerImaging Archive 2018 URL httpsdoiorg107937k9tcia2018

3rje41q1

[41] Nameeta Shah Xu Feng Michael Lankerovich Ralph B Puchalski andBart Keogh Data from Ivy GAP The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016XLwaN6nL

[42] Ralph B Puchalski Nameeta Shah Jeremy Miller Rachel Dalley Steve RNomura Jae-Guen Yoon Kimberly A Smith Michael Lankerovich DarrenBertagnolli Kris Bickley Andrew F Boe Krissy Brouner Stephanie But-ler Shiella Caldejon Mike Chapin Suvro Datta Nick Dee Tsega DestaTim Dolbeare Nadezhda Dotson Amanda Ebbert David Feng Xu FengMichael Fisher Garrett Gee Jeff Goldy Lindsey Gourley Benjamin WGregor Guangyu Gu Nika Hejazinia John Hohmann Parvinder HothiRobert Howard Kevin Joines Ali Kriedberg Leonard Kuan Chris LauFelix Lee Hwahyung Lee Tracy Lemon Fuhui Long Naveed Mastan ErikaMott Chantal Murthy Kiet Ngo Eric Olson Melissa Reding Zack RileyDavid Rosen David Sandman Nadiya Shapovalova Clifford R Slaughter-beck Andrew Sodt Graham Stockdale Aaron Szafer Wayne WakemanPaul E Wohnoutka Steven J White Don Marsh Robert C RostomilyLydia Ng Chinh Dang Allan Jones Bart Keogh Haley R GittlemanJill S Barnholtz-Sloan Patrick J Cimino Megha S Uppin C Dirk KeeneFarrokh R Farrokhi Justin D Lathia Michael E Berens Antonio IavaroneAmy Bernard Ed Lein John W Phillips Steven W Rostad Charles Cobbs

43

Michael J Hawrylycz and Greg D Foltz An anatomic transcriptional at-las of human glioblastoma Science 360(6389)660ndash663 May 2018 ISSN0036-8075 URL httpsdoiorg101126scienceaaf2666

[43] Kathleen Schmainda and Melissa Prah Data from Brain-Tumor-Progression The Cancer Imaging Archive 2018 URL httpsdoiorg

107937K9TCIA201815quzvnb

[44] B H Menze A Jakab S Bauer J Kalpathy-Cramer K FarahaniJ Kirby Y Burren N Porz J Slotboom R Wiest L Lanczi E GerstnerM Weber T Arbel B B Avants N Ayache P Buendia D L CollinsN Cordier J J Corso A Criminisi T Das H Delingette C Demi-ralp C R Durst M Dojat S Doyle J Festa F Forbes E GeremiaB Glocker P Golland X Guo A Hamamci K M IftekharuddinR Jena N M John E Konukoglu D Lashkari J A Mariz R MeierS Pereira D Precup S J Price T R Raviv S M S Reza M RyanD Sarikaya L Schwartz H Shin J Shotton C A Silva N SousaN K Subbanna G Szekely T J Taylor O M Thomas N J Tusti-son G Unal F Vasseur M Wintermark D H Ye L Zhao B ZhaoD Zikic M Prastawa M Reyes and K Van Leemput The mul-timodal brain tumor image segmentation benchmark (BRATS) IEEETransactions on Medical Imaging 34(10)1993ndash2024 2015 URL https

doiorg101109TMI20142377694

[45] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin S Kirby John B Freymann Keyvan Farahani and ChristosDavatzikos Advancing the cancer genome atlas glioma MRI collectionswith expert segmentation labels and radiomic features Scientific Data 4(1)170117 December 2017 ISSN 2052-4463 URL httpsdoiorg10

1038sdata2017117

[46] Spyridon Bakas Mauricio Reyes Andras Jakab Stefan Bauer MarkusRempfler Alessandro Crimi Russell Takeshi Shinohara Christoph BergerSung Min Ha Martin Rozycki Marcel Prastawa Esther Alberts JanaLipkova John Freymann Justin Kirby Michel Bilello Hassan Fathallah-Shaykh Roland Wiest Jan Kirschke Benedikt Wiestler Rivka ColenAikaterini Kotrotsou Pamela Lamontagne Daniel Marcus MikhailMilchenko Arash Nazeri Marc-Andre Weber Abhishek Mahajan UjjwalBaid Elizabeth Gerstner Dongjin Kwon Gagan Acharya Manu Agar-wal Mahbubul Alam Alberto Albiol Antonio Albiol Francisco J AlbiolVarghese Alex Nigel Allinson Pedro H A Amorim Abhijit AmrutkarGanesh Anand Simon Andermatt Tal Arbel Pablo Arbelaez Aaron Av-ery Muneeza Azmat Pranjal B W Bai Subhashis Banerjee Bill BarthThomas Batchelder Kayhan Batmanghelich Enzo Battistella AndrewBeers Mikhail Belyaev Martin Bendszus Eze Benson Jose Bernal Ha-landur Nagaraja Bharath George Biros Sotirios Bisdas James BrownMariano Cabezas Shilei Cao Jorge M Cardoso Eric N Carver Adria

44

Casamitjana Laura Silvana Castillo Marcel Cata Philippe Cattin Al-bert Cerigues Vinicius S Chagas Siddhartha Chandra Yi-Ju ChangShiyu Chang Ken Chang Joseph Chazalon Shengcong Chen Wei ChenJefferson W Chen Zhaolin Chen Kun Cheng Ahana Roy ChoudhuryRoger Chylla Albert Clerigues Steven Colleman Ramiro German Ro-driguez Colmeiro Marc Combalia Anthony Costa Xiaomeng Cui Zhen-zhen Dai Lutao Dai Laura Alexandra Daza Eric Deutsch ChangxingDing Chao Dong Shidu Dong Wojciech Dudzik Zach Eaton-Rosen GaryEgan Guilherme Escudero Theo Estienne Richard Everson JonathanFabrizio Yong Fan Longwei Fang Xue Feng Enzo Ferrante Lucas Fi-don Martin Fischer Andrew P French Naomi Fridman Huan Fu DavidFuentes Yaozong Gao Evan Gates David Gering Amir Gholami WilliGierke Ben Glocker Mingming Gong Sandra Gonzalez-Villa T Gros-ges Yuanfang Guan Sheng Guo Sudeep Gupta Woo-Sup Han Il SongHan Konstantin Harmuth Huiguang He Aura Hernandez-Sabate EvelynHerrmann Naveen Himthani Winston Hsu Cheyu Hsu Xiaojun Hu Xi-aobin Hu Yan Hu Yifan Hu Rui Hua Teng-Yi Huang Weilin HuangSabine Van Huffel Quan Huo Vivek HV Khan M Iftekharuddin FabianIsensee Mobarakol Islam Aaron S Jackson Sachin R Jambawalikar An-drew Jesson Weijian Jian Peter Jin V Jeya Maria Jose Alain JungoB Kainz Konstantinos Kamnitsas Po-Yu Kao Ayush Karnawat ThomasKellermeier Adel Kermi Kurt Keutzer Mohamed Tarek Khadir Mahen-dra Khened Philipp Kickingereder Geena Kim Nik King Haley KnappUrspeter Knecht Lisa Kohli Deren Kong Xiangmao Kong Simon Kop-pers Avinash Kori Ganapathy Krishnamurthi Egor Krivov Piyush Ku-mar Kaisar Kushibar Dmitrii Lachinov Tryphon Lambrou Joon LeeChengen Lee Yuehchou Lee M Lee Szidonia Lefkovits Laszlo LefkovitsJames Levitt Tengfei Li Hongwei Li Wenqi Li Hongyang Li XiaochuanLi Yuexiang Li Heng Li Zhenye Li Xiaoyu Li Zeju Li XiaoGang LiWenqi Li Zheng-Shen Lin Fengming Lin Pietro Lio Chang Liu Bo-qiang Liu Xiang Liu Mingyuan Liu Ju Liu Luyan Liu Xavier LladoMarc Moreno Lopez Pablo Ribalta Lorenzo Zhentai Lu Lin Luo ZhigangLuo Jun Ma Kai Ma Thomas Mackie Anant Madabushi Issam Mah-moudi Klaus H Maier-Hein Pradipta Maji CP Mammen Andreas MangB S Manjunath Michal Marcinkiewicz S McDonagh Stephen McKennaRichard McKinley Miriam Mehl Sachin Mehta Raghav Mehta RaphaelMeier Christoph Meinel Dorit Merhof Craig Meyer Robert Miller Sush-mita Mitra Aliasgar Moiyadi David Molina-Garcia Miguel A B Mon-teiro Grzegorz Mrukwa Andriy Myronenko Jakub Nalepa Thuyen NgoDong Nie Holly Ning Chen Niu Nicholas K Nuechterlein Eric Oer-mann Arlindo Oliveira Diego D C Oliveira Arnau Oliver AlexanderF I Osman Yu-Nian Ou Sebastien Ourselin Nikos Paragios Moo SungPark Brad Paschke J Gregory Pauloski Kamlesh Pawar Nick PawlowskiLinmin Pei Suting Peng Silvio M Pereira Julian Perez-Beteta Vic-tor M Perez-Garcia Simon Pezold Bao Pham Ashish Phophalia GemmaPiella G N Pillai Marie Piraud Maxim Pisov Anmol Popli Michael P

45

Pound Reza Pourreza Prateek Prasanna Vesna Prkovska Tony P Prid-more Santi Puch Elodie Puybareau Buyue Qian Xu Qiao MartinRajchl Swapnil Rane Michael Rebsamen Hongliang Ren Xuhua RenKarthik Revanuru Mina Rezaei Oliver Rippel Luis Carlos Rivera Char-lotte Robert Bruce Rosen Daniel Rueckert Mohammed Safwan MostafaSalem Joaquim Salvi Irina Sanchez Irina Sanchez Heitor M Santos Em-mett Sartor Dawid Schellingerhout Klaudius Scheufele Matthew R ScottArtur A Scussel Sara Sedlar Juan Pablo Serrano-Rubio N Jon ShahNameetha Shah Mazhar Shaikh B Uma Shankar Zeina Shboul HaipengShen Dinggang Shen Linlin Shen Haocheng Shen Varun Shenoy FengShi Hyung Eun Shin Hai Shu Diana Sima M Sinclair Orjan SmedbyJames M Snyder Mohammadreza Soltaninejad Guidong Song MehulSoni Jean Stawiaski Shashank Subramanian Li Sun Roger Sun JiaweiSun Kay Sun Yu Sun Guoxia Sun Shuang Sun Yannick R Suter Las-zlo Szilagyi Sanjay Talbar Dacheng Tao Dacheng Tao Zhongzhao TengSiddhesh Thakur Meenakshi H Thakur Sameer Tharakan Pallavi Ti-wari Guillaume Tochon Tuan Tran Yuhsiang M Tsai Kuan-Lun TsengTran Anh Tuan Vadim Turlapov Nicholas Tustison Maria VakalopoulouSergi Valverde Rami Vanguri Evgeny Vasiliev Jonathan Ventura LuisVera Tom Vercauteren C A Verrastro Lasitha Vidyaratne Veronica Vi-laplana Ajeet Vivekanandan Guotai Wang Qian Wang Chiatse J WangWeichung Wang Duo Wang Ruixuan Wang Yuanyuan Wang ChunliangWang Guotai Wang Ning Wen Xin Wen Leon Weninger Wolfgang WickShaocheng Wu Qiang Wu Yihong Wu Yong Xia Yanwu Xu XiaowenXu Peiyuan Xu Tsai-Ling Yang Xiaoping Yang Hao-Yu Yang JunlinYang Haojin Yang Guang Yang Hongdou Yao Xujiong Ye ChangchangYin Brett Young-Moxon Jinhua Yu Xiangyu Yue Songtao Zhang AngelaZhang Kun Zhang Xuejie Zhang Lichi Zhang Xiaoyue Zhang YazhuoZhang Lei Zhang Jianguo Zhang Xiang Zhang Tianhao Zhang SichengZhao Yu Zhao Xiaomei Zhao Liang Zhao Yefeng Zheng Liming ZhongChenhong Zhou Xiaobing Zhou Fan Zhou Hongtu Zhu Jin Zhu YingZhuge Weiwei Zong Jayashree Kalpathy-Cramer Keyvan Farahani Chris-tos Davatzikos Koen van Leemput and Bjoern Menze Identifying the bestmachine learning algorithms for brain tumor segmentation progression as-sessment and overall survival prediction in the BRATS challenge 2018Preprint at httpsarxivorgabs181102629

[47] Nancy Pedano Adam E Flanders Lisa Scarpace Tom Mikkelsen Jen-nifer M Eschbacher Beth Hermes Victor Sisneros Jill Barnholtz-Sloanand Quinn Ostrom Radiology data from the cancer genome atlas lowgrade glioma [TCGA-LGG] collection The Cancer Imaging Archive 2016URL httpsdoiorg107937K9TCIA2016L4LTD3TK

[48] Lisa Scarpace Tom Mikkelsen Soonmee Cha Sujaya Rao SangeetaTekchandani David Gutman Joel H Saltz Bradley J Erickson NancyPedano Adam E Flanders Jill Barnholtz-Sloan Quinn Ostrom DanielBarboriak and Laura J Pierce Radiology data from the cancer genome

46

atlas glioblastoma multiforme [TCGA-GBM] collection The Cancer Imag-ing Archive 2016 URL httpsdoiorg107937K9TCIA2016

RNYFUYE9

[49] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello MartinRozycki Justin Kirby John Freymann Keyvan Farahani and ChristosDavatzikos Segmentation labels and radiomic features for the pre-operativescans of the TCGA-LGG collection [data set] The Cancer Imaging Archive2017 URL httpsdoiorg107937K9TCIA2017GJQ7R0EF

[50] Spyridon Bakas Hamed Akbari Aristeidis Sotiras Michel Bilello Mar-tin Rozycki Justin Kirby John Freymann Keyvan Farahani and Chris-tos Davatzikos Segmentation labels and radiomic features for the pre-operative scans of the TCGA-GBM collection [data set] The CancerImaging Archive 2017 URL httpsdoiorg107937K9TCIA2017

KLXWJJ1Q

[51] Xiangrui Li Paul S Morgan John Ashburner Jolinda Smith and Christo-pher Rorden The first step for neuroimaging data analysis DICOM toNIfTI conversion Journal of Neuroscience Methods 26447ndash56 May 2016ISSN 0165-0270 URL httpsdoiorg101016jjneumeth201603

001

[52] Vladimir Fonov Alan C Evans Kelly Botteron C Robert Almli Robert CMcKinstry and D Louis Collins Unbiased average age-appropriate at-lases for pediatric studies NeuroImage 54(1)313ndash327 January 2011ISSN 1053-8119 URL httpsdoiorg101016jneuroimage2010

07033

[53] VS Fonov AC Evans RC McKinstry CR Almli and DL Collins Unbiasednonlinear average age-appropriate brain templates from birth to adulthoodNeuroImage 47S102 July 2009 ISSN 1053-8119 URL httpsdoi

org101016S1053-8119(09)70884-5 Organization for Human BrainMapping 2009 Annual Meeting

[54] S Klein M Staring K Murphy MA Viergever and J Pluim elastix Atoolbox for intensity-based medical image registration IEEE Transactionson Medical Imaging 29(1)196ndash205 January 2010 ISSN 0278-0062 URLhttpsdoiorg101109TMI20092035616

[55] Denis Shamonin Fast parallel image registration on CPU and GPU fordiagnostic classification of alzheimerrsquos disease Frontiers in Neuroinfor-matics 750 2013 ISSN 1662-5196 URL httpsdoiorg103389

fninf201300050

[56] Bradley C Lowekamp David T Chen Luis Ibanez and Daniel Blezek Thedesign of SimpleITK Frontiers in Neuroinformatics 745 2013 ISSN1662-5196 URL httpsdoiorg103389fninf201300045

47

[57] Fabian Isensee Marianne Schell Irada Pflueger Gianluca Brugnara DavidBonekamp Ulf Neuberger Antje Wick Heinz-Peter Schlemmer SabineHeiland Wolfgang Wick Martin Bendszus Klaus H Maier-Hein andPhilipp Kickingereder Automated brain extraction of multisequence MRIusing artificial neural networks Human Brain Mapping 40(17)4952ndash4964December 2019 ISSN 1065-9471 URL httpsdoiorg101002hbm

24750

[58] Olaf Ronneberger Invited talk U-net convolutional networks for biomed-ical image segmentation In geb Fritzsche K Maier-Hein geb Lehmann TDeserno Heinz Handels and Thomas Tolxdorff editors Bildverarbeitungfur die Medizin 2017 Informatik aktuell pages 3ndash3 Springer Scienceand Business Media LLC 2017 ISBN 9783662543443 URL https

doiorg101007978-3-662-54345-0_3

[59] Martın Abadi Paul Barham Jianmin Chen Zhifeng Chen Andy DavisJeffrey Dean Matthieu Devin Sanjay Ghemawat Geoffrey Irving MichaelIsard Manjunath Kudlur Josh Levenberg Rajat Monga Sherry MooreDerek G Murray Benoit Steiner Paul Tucker Vijay Vasudevan PeteWarden Martin Wicke Yuan Yu and Xiaoqiang Zheng Tensor-Flow A system for large-scale machine learning In 12th USENIXSymposium on Operating Systems Design and Implementation (OSDI16) pages 265ndash283 USENIX Association November 2016 ISBN978-1-931971-33-1 URL httpswwwusenixorgconferenceosdi16

technical-sessionspresentationabadi

[60] Dipankar Das Naveen Mellempudi Dheevatsa Mudigere Dhiraj KalamkarSasikanth Avancha Kunal Banerjee Srinivas Sridharan KarthikVaidyanathan Bharat Kaul Evangelos Georganas Alexander HeineckePradeep Dubey Jesus Corbal Nikita Shustrov Roma Dubtsov EvaristFomenko and Vadim Pirogov Mixed precision training of convolutionalneural networks using integer operations In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=H135uzZ0-

[61] Samuel L Smith Pieter-Jan Kindermans and Quoc V Le Donrsquot decaythe learning rate increase the batch size In International Conference onLearning Representations 2018 URL httpsopenreviewnetforum

id=B1Yy1BxCZ

[62] Cliff Woolley NCCL accelarated multi-GPU collective communicationsURL httpsimagesnvidiacomeventssc15pdfsNCCL-Woolley

pdf Accessed on 2020-09-30

[63] Ilya Loshchilov and Frank Hutter Decoupled weight decay regularization2017 Preprint at httpsarxivorgabs171105101

48

[64] F Pedregosa G Varoquaux A Gramfort V Michel B Thirion O GriselM Blondel P Prettenhofer R Weiss V Dubourg J Vanderplas A Pas-sos D Cournapeau M Brucher M Perrot and E Duchesnay Scikit-learn Machine learning in python Journal of Machine Learning Re-search 122825ndash2830 2011 URL httpswwwjmlrorgpapersv12

pedregosa11ahtml

[65] Abdel Aziz Taha and Allan Hanbury Metrics for evaluating 3d medicalimage segmentation analysis selection and tool BMC Medical Imaging15(1)29 August 2015 ISSN 1471-2342 URL httpsdoiorg101186

s12880-015-0068-x

[66] Daniel Smilkov Nikhil Thorat Been Kim Fernanda Viegas and MartinWattenberg SmoothGrad removing noise by adding noise 2017 Preprintat httpsarxivorgabs170603825

[67] Alaa Tharwat Classification assessment methods Applied Computing andInformatics 2018 ISSN 2210-8327 URL httpsdoiorg101016j

aci201808003

[68] David J Hand and Robert J Till A simple generalisation of the areaunder the ROC curve for multiple class classification problems MachineLearning 45(2)171ndash186 November 2001 ISSN 1573-0565 URL https

doiorg101023A1010920819831

49

  • 1 Introduction
  • 2 Results
    • 21 Patient characteristics
    • 22 Algorithm performance
    • 23 Model interpretability
    • 24 Model robustness
      • 3 Discussion
      • 4 Methods
        • 41 Patient population
        • 42 Automatic segmentation in the train set
        • 43 Pre-processing
        • 44 Model
        • 45 Model training
        • 46 Hyperparameter tuning
        • 47 Post-processing
        • 48 Model evaluation
        • 49 Data availability
        • 410 Code availability
          • A Confusion matrices
          • B Segmentation examples
          • C Prediction results in the test set
          • D Filter output visualizations
          • E Training losses
          • F Parameter tuning
          • G Evaluation metrics
          • H Ground truth labels of patients included from public datasets
Data Collection Patient IDH_mutated 1p19q_codeleted Grade
BTumorP PGBM-001 -1 -1 -1
BTumorP PGBM-002 -1 -1 -1
BTumorP PGBM-003 -1 -1 -1
BTumorP PGBM-004 -1 -1 -1
BTumorP PGBM-005 -1 -1 -1
BTumorP PGBM-006 -1 -1 -1
BTumorP PGBM-007 -1 -1 -1
BTumorP PGBM-008 -1 -1 -1
BTumorP PGBM-009 -1 -1 -1
BTumorP PGBM-010 -1 -1 -1
BTumorP PGBM-011 -1 -1 -1
BTumorP PGBM-012 -1 -1 -1
BTumorP PGBM-013 -1 -1 -1
BTumorP PGBM-014 -1 -1 -1
BTumorP PGBM-015 -1 -1 -1
BTumorP PGBM-016 -1 -1 -1
BTumorP PGBM-017 -1 -1 -1
BTumorP PGBM-018 -1 -1 -1
BTumorP PGBM-019 -1 -1 -1
BTumorP PGBM-020 -1 -1 -1
BraTS 2013_0 -1 -1 -1
BraTS 2013_10 -1 -1 -1
BraTS 2013_11 -1 -1 -1
BraTS 2013_12 -1 -1 -1
BraTS 2013_13 -1 -1 -1
BraTS 2013_14 -1 -1 -1
BraTS 2013_15 -1 -1 -1
BraTS 2013_16 -1 -1 -1
BraTS 2013_17 -1 -1 -1
BraTS 2013_18 -1 -1 -1
BraTS 2013_19 -1 -1 -1
BraTS 2013_1 -1 -1 -1
BraTS 2013_20 -1 -1 -1
BraTS 2013_21 -1 -1 -1
BraTS 2013_22 -1 -1 -1
BraTS 2013_23 -1 -1 -1
BraTS 2013_24 -1 -1 -1
BraTS 2013_25 -1 -1 -1
BraTS 2013_26 -1 -1 -1
BraTS 2013_27 -1 -1 -1
BraTS 2013_28 -1 -1 -1
BraTS 2013_29 -1 -1 -1
BraTS 2013_2 -1 -1 -1
BraTS 2013_3 -1 -1 -1
BraTS 2013_4 -1 -1 -1
BraTS 2013_5 -1 -1 -1
BraTS 2013_6 -1 -1 -1
BraTS 2013_7 -1 -1 -1
BraTS 2013_8 -1 -1 -1
BraTS 2013_9 -1 -1 -1
BraTS CBICA_AAB -1 -1 -1
BraTS CBICA_AAG -1 -1 -1
BraTS CBICA_AAL -1 -1 -1
BraTS CBICA_AAP -1 -1 -1
BraTS CBICA_ABB -1 -1 -1
BraTS CBICA_ABE -1 -1 -1
BraTS CBICA_ABM -1 -1 -1
BraTS CBICA_ABN -1 -1 -1
BraTS CBICA_ABO -1 -1 -1
BraTS CBICA_ABY -1 -1 -1
BraTS CBICA_ALN -1 -1 -1
BraTS CBICA_ALU -1 -1 -1
BraTS CBICA_ALX -1 -1 -1
BraTS CBICA_AME -1 -1 -1
BraTS CBICA_AMH -1 -1 -1
BraTS CBICA_ANG -1 -1 -1
BraTS CBICA_ANI -1 -1 -1
BraTS CBICA_ANP -1 -1 -1
BraTS CBICA_ANV -1 -1 -1
BraTS CBICA_ANZ -1 -1 -1
BraTS CBICA_AOC -1 -1 -1
BraTS CBICA_AOD -1 -1 -1
BraTS CBICA_AOH -1 -1 -1
BraTS CBICA_AOO -1 -1 -1
BraTS CBICA_AOP -1 -1 -1
BraTS CBICA_AOS -1 -1 -1
BraTS CBICA_AOZ -1 -1 -1
BraTS CBICA_APK -1 -1 -1
BraTS CBICA_APR -1 -1 -1
BraTS CBICA_APY -1 -1 -1
BraTS CBICA_APZ -1 -1 -1
BraTS CBICA_AQA -1 -1 -1
BraTS CBICA_AQD -1 -1 -1
BraTS CBICA_AQG -1 -1 -1
BraTS CBICA_AQJ -1 -1 -1
BraTS CBICA_AQN -1 -1 -1
BraTS CBICA_AQO -1 -1 -1
BraTS CBICA_AQP -1 -1 -1
BraTS CBICA_AQQ -1 -1 -1
BraTS CBICA_AQR -1 -1 -1
BraTS CBICA_AQT -1 -1 -1
BraTS CBICA_AQU -1 -1 -1
BraTS CBICA_AQV -1 -1 -1
BraTS CBICA_AQY -1 -1 -1
BraTS CBICA_AQZ -1 -1 -1
BraTS CBICA_ARF -1 -1 -1
BraTS CBICA_ARW -1 -1 -1
BraTS CBICA_ARZ -1 -1 -1
BraTS CBICA_ASA -1 -1 -1
BraTS CBICA_ASE -1 -1 -1
BraTS CBICA_ASF -1 -1 -1
BraTS CBICA_ASG -1 -1 -1
BraTS CBICA_ASH -1 -1 -1
BraTS CBICA_ASK -1 -1 -1
BraTS CBICA_ASN -1 -1 -1
BraTS CBICA_ASO -1 -1 -1
BraTS CBICA_ASR -1 -1 -1
BraTS CBICA_ASU -1 -1 -1
BraTS CBICA_ASV -1 -1 -1
BraTS CBICA_ASW -1 -1 -1
BraTS CBICA_ASY -1 -1 -1
BraTS CBICA_ATB -1 -1 -1
BraTS CBICA_ATD -1 -1 -1
BraTS CBICA_ATF -1 -1 -1
BraTS CBICA_ATN -1 -1 -1
BraTS CBICA_ATP -1 -1 -1
BraTS CBICA_ATV -1 -1 -1
BraTS CBICA_ATX -1 -1 -1
BraTS CBICA_AUA -1 -1 -1
BraTS CBICA_AUN -1 -1 -1
BraTS CBICA_AUQ -1 -1 -1
BraTS CBICA_AUR -1 -1 -1
BraTS CBICA_AUW -1 -1 -1
BraTS CBICA_AUX -1 -1 -1
BraTS CBICA_AVB -1 -1 -1
BraTS CBICA_AVF -1 -1 -1
BraTS CBICA_AVG -1 -1 -1
BraTS CBICA_AVJ -1 -1 -1
BraTS CBICA_AVT -1 -1 -1
BraTS CBICA_AVV -1 -1 -1
BraTS CBICA_AWG -1 -1 -1
BraTS CBICA_AWH -1 -1 -1
BraTS CBICA_AWI -1 -1 -1
BraTS CBICA_AWV -1 -1 -1
BraTS CBICA_AWX -1 -1 -1
BraTS CBICA_AXJ -1 -1 -1
BraTS CBICA_AXL -1 -1 -1
BraTS CBICA_AXM -1 -1 -1
BraTS CBICA_AXN -1 -1 -1
BraTS CBICA_AXO -1 -1 -1
BraTS CBICA_AXQ -1 -1 -1
BraTS CBICA_AXW -1 -1 -1
BraTS CBICA_AYA -1 -1 -1
BraTS CBICA_AYC -1 -1 -1
BraTS CBICA_AYG -1 -1 -1
BraTS CBICA_AYI -1 -1 -1
BraTS CBICA_AYU -1 -1 -1
BraTS CBICA_AYW -1 -1 -1
BraTS CBICA_AZD -1 -1 -1
BraTS CBICA_AZH -1 -1 -1
BraTS CBICA_BAN -1 -1 -1
BraTS CBICA_BAP -1 -1 -1
BraTS CBICA_BAX -1 -1 -1
BraTS CBICA_BBG -1 -1 -1
BraTS CBICA_BCF -1 -1 -1
BraTS CBICA_BCL -1 -1 -1
BraTS CBICA_BDK -1 -1 -1
BraTS CBICA_BEM -1 -1 -1
BraTS CBICA_BFB -1 -1 -1
BraTS CBICA_BFP -1 -1 -1
BraTS CBICA_BGE -1 -1 -1
BraTS CBICA_BGG -1 -1 -1
BraTS CBICA_BGN -1 -1 -1
BraTS CBICA_BGO -1 -1 -1
BraTS CBICA_BGR -1 -1 -1
BraTS CBICA_BGT -1 -1 -1
BraTS CBICA_BGW -1 -1 -1
BraTS CBICA_BGX -1 -1 -1
BraTS CBICA_BHB -1 -1 -1
BraTS CBICA_BHK -1 -1 -1
BraTS CBICA_BHM -1 -1 -1
BraTS CBICA_BHQ -1 -1 -1
BraTS CBICA_BHV -1 -1 -1
BraTS CBICA_BHZ -1 -1 -1
BraTS CBICA_BIC -1 -1 -1
BraTS CBICA_BJY -1 -1 -1
BraTS CBICA_BKV -1 -1 -1
BraTS CBICA_BLJ -1 -1 -1
BraTS CBICA_BNR -1 -1 -1
BraTS TMC_6290 -1 -1 -1
BraTS TMC_6643 -1 -1 -1
BraTS TMC_9043 -1 -1 -1
BraTS TMC_11964 -1 -1 -1
BraTS TMC_12866 -1 -1 -1
BraTS TMC_15477 -1 -1 -1
BraTS TMC_21360 -1 -1 -1
BraTS TMC_27374 -1 -1 -1
BraTS TMC_30014 -1 -1 -1
CPTAC-GBM C3L-00016 -1 -1 4
CPTAC-GBM C3L-00019 -1 -1 4
CPTAC-GBM C3L-00265 -1 -1 4
CPTAC-GBM C3L-00278 -1 -1 4
CPTAC-GBM C3L-00349 -1 -1 4
CPTAC-GBM C3L-00424 -1 -1 4
CPTAC-GBM C3L-00429 -1 -1 4
CPTAC-GBM C3L-00506 -1 -1 4
CPTAC-GBM C3L-00528 -1 -1 4
CPTAC-GBM C3L-00591 -1 -1 4
CPTAC-GBM C3L-00631 -1 -1 4
CPTAC-GBM C3L-00636 -1 -1 4
CPTAC-GBM C3L-00671 -1 -1 4
CPTAC-GBM C3L-00674 -1 -1 4
CPTAC-GBM C3L-00677 -1 -1 4
CPTAC-GBM C3L-01045 -1 -1 4
CPTAC-GBM C3L-01046 -1 -1 4
CPTAC-GBM C3L-01142 -1 -1 4
CPTAC-GBM C3L-01156 -1 -1 4
CPTAC-GBM C3L-01327 -1 -1 4
CPTAC-GBM C3L-02041 -1 -1 4
CPTAC-GBM C3L-02465 -1 -1 4
CPTAC-GBM C3L-02504 -1 -1 4
CPTAC-GBM C3L-02704 -1 -1 4
CPTAC-GBM C3L-02706 -1 -1 4
CPTAC-GBM C3L-02707 -1 -1 4
CPTAC-GBM C3L-02708 -1 -1 4
CPTAC-GBM C3L-03260 -1 -1 4
CPTAC-GBM C3L-03266 -1 -1 4
CPTAC-GBM C3L-03727 -1 -1 4
CPTAC-GBM C3L-03728 -1 -1 4
CPTAC-GBM C3L-03747 -1 -1 4
CPTAC-GBM C3L-03748 -1 -1 4
CPTAC-GBM C3L-04084 -1 -1 4
CPTAC-GBM C3N-00661 -1 -1 4
CPTAC-GBM C3N-00662 -1 -1 4
CPTAC-GBM C3N-00663 -1 -1 4
CPTAC-GBM C3N-00665 -1 -1 4
CPTAC-GBM C3N-01192 -1 -1 4
CPTAC-GBM C3N-01196 -1 -1 4
CPTAC-GBM C3N-01505 -1 -1 4
CPTAC-GBM C3N-01849 -1 -1 4
CPTAC-GBM C3N-01851 -1 -1 4
CPTAC-GBM C3N-01852 -1 -1 4
CPTAC-GBM C3N-02255 -1 -1 4
CPTAC-GBM C3N-02256 -1 -1 4
CPTAC-GBM C3N-02286 -1 -1 4
CPTAC-GBM C3N-03001 -1 -1 4
CPTAC-GBM C3N-03003 -1 -1 4
CPTAC-GBM C3N-03755 -1 -1 4
CPTAC-GBM C3N-04686 -1 -1 4
IvyGAP W10 1 1 4
IvyGAP W11 0 0 4
IvyGAP W12 0 0 4
IvyGAP W13 0 0 4
IvyGAP W16 0 0 4
IvyGAP W18 0 0 4
IvyGAP W19 0 0 4
IvyGAP W1 0 0 4
IvyGAP W20 0 0 4
IvyGAP W21 0 0 4
IvyGAP W22 0 0 -1
IvyGAP W26 0 -1 4
IvyGAP W29 0 0 4
IvyGAP W2 0 1 4
IvyGAP W30 0 0 4
IvyGAP W31 1 1 4
IvyGAP W32 0 0 4
IvyGAP W33 0 0 4
IvyGAP W34 0 0 4
IvyGAP W35 1 0 3
IvyGAP W36 0 0 4
IvyGAP W38 0 0 4
IvyGAP W39 0 0 4
IvyGAP W3 1 0 4
IvyGAP W40 0 0 4
IvyGAP W42 0 -1 4
IvyGAP W43 0 -1 4
IvyGAP W45 -1 -1 4
IvyGAP W48 0 -1 4
IvyGAP W4 1 0 4
IvyGAP W50 0 -1 3
IvyGAP W53 1 -1 4
IvyGAP W54 0 -1 4
IvyGAP W55 0 -1 4
IvyGAP W5 0 0 4
IvyGAP W6 0 0 4
IvyGAP W7 0 0 4
IvyGAP W8 0 0 4
IvyGAP W9 0 0 4
REMBRANDT 900-00-5299 -1 -1 4
REMBRANDT 900-00-5303 -1 -1 4
REMBRANDT 900-00-5308 -1 -1 3
REMBRANDT 900-00-5316 -1 -1 4
REMBRANDT 900-00-5317 -1 -1 4
REMBRANDT 900-00-5332 -1 -1 4
REMBRANDT 900-00-5339 -1 -1 4
REMBRANDT 900-00-5341 -1 -1 -1
REMBRANDT 900-00-5342 -1 -1 4
REMBRANDT 900-00-5346 -1 -1 4
REMBRANDT 900-00-5380 -1 -1 -1
REMBRANDT 900-00-5381 -1 -1 4
REMBRANDT 900-00-5382 -1 -1 2
REMBRANDT 900-00-5385 -1 -1 3
REMBRANDT 900-00-5396 -1 -1 4
REMBRANDT 900-00-5404 -1 -1 4
REMBRANDT 900-00-5412 -1 -1 -1
REMBRANDT 900-00-5414 -1 -1 4
REMBRANDT 900-00-5458 -1 -1 4
REMBRANDT 900-00-5459 -1 -1 3
REMBRANDT 900-00-5462 -1 -1 4
REMBRANDT 900-00-5468 -1 -1 2
REMBRANDT 900-00-5476 -1 -1 2
REMBRANDT 900-00-5477 -1 -1 2
REMBRANDT HF0763 -1 -1 -1
REMBRANDT HF0828 -1 -1 3
REMBRANDT HF0835 -1 -1 2
REMBRANDT HF0855 -1 -1 2
REMBRANDT HF0868 -1 -1 -1
REMBRANDT HF0883 -1 -1 -1
REMBRANDT HF0899 -1 -1 2
REMBRANDT HF0920 -1 -1 2
REMBRANDT HF0931 -1 -1 2
REMBRANDT HF0953 -1 -1 2
REMBRANDT HF0960 -1 -1 2
REMBRANDT HF0966 -1 -1 3
REMBRANDT HF0986 -1 -1 4
REMBRANDT HF0990 -1 -1 4
REMBRANDT HF1000 -1 -1 2
REMBRANDT HF1058 -1 -1 4
REMBRANDT HF1059 -1 -1 3
REMBRANDT HF1071 -1 -1 4
REMBRANDT HF1077 -1 -1 4
REMBRANDT HF1078 -1 -1 4
REMBRANDT HF1097 -1 -1 4
REMBRANDT HF1113 -1 -1 -1
REMBRANDT HF1122 -1 -1 4
REMBRANDT HF1136 -1 -1 3
REMBRANDT HF1139 -1 -1 4
REMBRANDT HF1150 -1 -1 3
REMBRANDT HF1156 -1 -1 2
REMBRANDT HF1167 -1 -1 2
REMBRANDT HF1185 -1 -1 3
REMBRANDT HF1191 -1 -1 4
REMBRANDT HF1199 -1 -1 -1
REMBRANDT HF1219 -1 -1 3
REMBRANDT HF1227 -1 -1 2
REMBRANDT HF1232 -1 -1 3
REMBRANDT HF1235 -1 -1 2
REMBRANDT HF1242 -1 -1 3
REMBRANDT HF1246 -1 -1 2
REMBRANDT HF1264 -1 -1 2
REMBRANDT HF1269 -1 -1 4
REMBRANDT HF1280 -1 -1 3
REMBRANDT HF1292 -1 -1 4
REMBRANDT HF1293 -1 -1 -1
REMBRANDT HF1297 -1 -1 4
REMBRANDT HF1300 -1 -1 -1
REMBRANDT HF1307 -1 -1 -1
REMBRANDT HF1316 -1 -1 2
REMBRANDT HF1318 -1 -1 -1
REMBRANDT HF1325 -1 -1 2
REMBRANDT HF1331 -1 -1 -1
REMBRANDT HF1334 -1 -1 2
REMBRANDT HF1344 -1 -1 2
REMBRANDT HF1345 -1 -1 2
REMBRANDT HF1357 -1 -1 3
REMBRANDT HF1381 -1 -1 2
REMBRANDT HF1397 -1 -1 4
REMBRANDT HF1398 -1 -1 3
REMBRANDT HF1407 -1 -1 2
REMBRANDT HF1409 -1 -1 3
REMBRANDT HF1420 -1 -1 -1
REMBRANDT HF1429 -1 -1 -1
REMBRANDT HF1433 -1 -1 2
REMBRANDT HF1437 -1 -1 -1
REMBRANDT HF1442 -1 -1 2
REMBRANDT HF1458 -1 -1 3
REMBRANDT HF1463 -1 -1 2
REMBRANDT HF1489 -1 -1 2
REMBRANDT HF1490 -1 -1 3
REMBRANDT HF1493 -1 -1 -1
REMBRANDT HF1510 -1 -1 -1
REMBRANDT HF1511 -1 -1 2
REMBRANDT HF1517 -1 -1 4
REMBRANDT HF1538 -1 -1 4
REMBRANDT HF1551 -1 -1 2
REMBRANDT HF1553 -1 -1 2
REMBRANDT HF1560 -1 -1 4
REMBRANDT HF1568 -1 -1 2
REMBRANDT HF1587 -1 -1 3
REMBRANDT HF1588 -1 -1 2
REMBRANDT HF1606 -1 -1 2
REMBRANDT HF1613 -1 -1 3
REMBRANDT HF1628 -1 -1 4
REMBRANDT HF1652 -1 -1 -1
REMBRANDT HF1677 -1 -1 2
REMBRANDT HF1702 -1 -1 3
REMBRANDT HF1708 -1 -1 2
TCGA-GBM TCGA-02-0003 0 0 4
TCGA-GBM TCGA-02-0006 0 0 4
TCGA-GBM TCGA-02-0009 0 0 4
TCGA-GBM TCGA-02-0011 0 0 4
TCGA-GBM TCGA-02-0027 0 0 4
TCGA-GBM TCGA-02-0033 0 0 4
TCGA-GBM TCGA-02-0034 0 0 4
TCGA-GBM TCGA-02-0037 0 0 4
TCGA-GBM TCGA-02-0046 0 0 4
TCGA-GBM TCGA-02-0047 0 0 4
TCGA-GBM TCGA-02-0048 0 0 4
TCGA-GBM TCGA-02-0054 0 0 4
TCGA-GBM TCGA-02-0059 -1 0 4
TCGA-GBM TCGA-02-0060 0 0 4
TCGA-GBM TCGA-02-0064 0 0 4
TCGA-GBM TCGA-02-0068 0 0 4
TCGA-GBM TCGA-02-0069 0 0 4
TCGA-GBM TCGA-02-0070 0 0 4
TCGA-GBM TCGA-02-0075 0 0 4
TCGA-GBM TCGA-02-0085 0 0 4
TCGA-GBM TCGA-02-0086 0 0 4
TCGA-GBM TCGA-02-0087 -1 0 4
TCGA-GBM TCGA-02-0102 0 0 4
TCGA-GBM TCGA-02-0106 -1 0 4
TCGA-GBM TCGA-02-0116 0 0 4
TCGA-GBM TCGA-06-0119 0 0 4
TCGA-GBM TCGA-06-0122 0 0 4
TCGA-GBM TCGA-06-0128 1 0 4
TCGA-GBM TCGA-06-0130 0 0 4
TCGA-GBM TCGA-06-0132 0 0 4
TCGA-GBM TCGA-06-0133 0 0 4
TCGA-GBM TCGA-06-0137 0 0 4
TCGA-GBM TCGA-06-0138 0 0 4
TCGA-GBM TCGA-06-0139 0 0 4
TCGA-GBM TCGA-06-0142 0 0 4
TCGA-GBM TCGA-06-0145 0 0 4
TCGA-GBM TCGA-06-0149 -1 0 4
TCGA-GBM TCGA-06-0154 0 0 4
TCGA-GBM TCGA-06-0158 0 0 4
TCGA-GBM TCGA-06-0162 -1 0 4
TCGA-GBM TCGA-06-0164 -1 0 4
TCGA-GBM TCGA-06-0166 0 0 4
TCGA-GBM TCGA-06-0168 0 0 4
TCGA-GBM TCGA-06-0175 -1 0 4
TCGA-GBM TCGA-06-0176 0 0 4
TCGA-GBM TCGA-06-0177 -1 0 4
TCGA-GBM TCGA-06-0179 -1 0 4
TCGA-GBM TCGA-06-0182 -1 0 4
TCGA-GBM TCGA-06-0184 0 0 4
TCGA-GBM TCGA-06-0185 0 0 4
TCGA-GBM TCGA-06-0187 0 0 4
TCGA-GBM TCGA-06-0188 0 0 4
TCGA-GBM TCGA-06-0189 0 0 4
TCGA-GBM TCGA-06-0190 0 0 4
TCGA-GBM TCGA-06-0192 0 0 4
TCGA-GBM TCGA-06-0213 0 0 4
TCGA-GBM TCGA-06-0238 0 0 4
TCGA-GBM TCGA-06-0240 0 0 4
TCGA-GBM TCGA-06-0241 0 0 4
TCGA-GBM TCGA-06-0644 0 0 4
TCGA-GBM TCGA-06-0646 0 0 4
TCGA-GBM TCGA-06-0648 0 0 4
TCGA-GBM TCGA-06-0649 0 0 4
TCGA-GBM TCGA-06-1084 0 0 4
TCGA-GBM TCGA-06-1802 -1 0 4
TCGA-GBM TCGA-06-2570 1 0 4
TCGA-GBM TCGA-06-5408 0 0 4
TCGA-GBM TCGA-06-5412 0 0 4
TCGA-GBM TCGA-06-5413 0 0 4
TCGA-GBM TCGA-06-5417 1 -1 4
TCGA-GBM TCGA-06-6389 1 0 4
TCGA-GBM TCGA-08-0350 0 0 4
TCGA-GBM TCGA-08-0352 0 0 4
TCGA-GBM TCGA-08-0353 0 0 4
TCGA-GBM TCGA-08-0354 0 0 4
TCGA-GBM TCGA-08-0355 0 0 4
TCGA-GBM TCGA-08-0356 0 0 4
TCGA-GBM TCGA-08-0357 0 0 4
TCGA-GBM TCGA-08-0358 0 0 4
TCGA-GBM TCGA-08-0359 0 0 4
TCGA-GBM TCGA-08-0360 0 -1 4
TCGA-GBM TCGA-08-0385 0 -1 4
TCGA-GBM TCGA-08-0389 0 0 4
TCGA-GBM TCGA-08-0390 0 0 4
TCGA-GBM TCGA-08-0392 0 0 4
TCGA-GBM TCGA-08-0512 -1 0 4
TCGA-GBM TCGA-08-0520 -1 0 4
TCGA-GBM TCGA-08-0521 -1 0 4
TCGA-GBM TCGA-08-0522 -1 -1 4
TCGA-GBM TCGA-08-0524 -1 0 4
TCGA-GBM TCGA-08-0529 -1 0 4
TCGA-GBM TCGA-12-0616 0 0 4
TCGA-GBM TCGA-12-0776 -1 0 4
TCGA-GBM TCGA-12-0829 0 0 4
TCGA-GBM TCGA-12-1093 0 0 4
TCGA-GBM TCGA-12-1094 -1 0 4
TCGA-GBM TCGA-12-1098 -1 0 4
TCGA-GBM TCGA-12-1598 0 0 4
TCGA-GBM TCGA-12-1601 0 -1 -1
TCGA-GBM TCGA-12-1602 0 0 4
TCGA-GBM TCGA-12-3650 0 0 4
TCGA-GBM TCGA-14-0789 0 0 4
TCGA-GBM TCGA-14-1456 1 0 4
TCGA-GBM TCGA-14-1794 0 0 4
TCGA-GBM TCGA-14-1825 0 0 4
TCGA-GBM TCGA-14-1829 0 0 4
TCGA-GBM TCGA-14-3477 0 0 4
TCGA-GBM TCGA-19-0963 -1 0 4
TCGA-GBM TCGA-19-1390 0 0 4
TCGA-GBM TCGA-19-1789 0 0 4
TCGA-GBM TCGA-19-2624 0 0 4
TCGA-GBM TCGA-19-2631 0 0 4
TCGA-GBM TCGA-19-5951 0 0 4
TCGA-GBM TCGA-19-5954 0 0 4
TCGA-GBM TCGA-19-5958 0 0 4
TCGA-GBM TCGA-19-5960 0 0 4
TCGA-GBM TCGA-27-1834 0 0 4
TCGA-GBM TCGA-27-1838 0 0 4
TCGA-GBM TCGA-27-2526 0 0 4
TCGA-GBM TCGA-76-4932 0 -1 4
TCGA-GBM TCGA-76-4934 0 0 4
TCGA-GBM TCGA-76-4935 0 0 4
TCGA-GBM TCGA-76-6191 0 0 4
TCGA-GBM TCGA-76-6193 0 0 4
TCGA-GBM TCGA-76-6280 0 0 4
TCGA-GBM TCGA-76-6282 0 0 4
TCGA-GBM TCGA-76-6285 0 0 4
TCGA-GBM TCGA-76-6656 0 0 4
TCGA-GBM TCGA-76-6657 0 0 4
TCGA-GBM TCGA-76-6661 0 0 4
TCGA-GBM TCGA-76-6662 0 0 4
TCGA-GBM TCGA-76-6663 0 0 4
TCGA-GBM TCGA-76-6664 0 0 4
TCGA-LGG TCGA-CS-4941 0 0 3
TCGA-LGG TCGA-CS-4942 1 0 3
TCGA-LGG TCGA-CS-4943 1 0 3
TCGA-LGG TCGA-CS-4944 1 0 2
TCGA-LGG TCGA-CS-5393 1 0 3
TCGA-LGG TCGA-CS-5395 0 0 2
TCGA-LGG TCGA-CS-5396 1 1 3
TCGA-LGG TCGA-CS-5397 0 0 3
TCGA-LGG TCGA-CS-6186 0 0 3
TCGA-LGG TCGA-CS-6188 0 0 3
TCGA-LGG TCGA-CS-6290 1 0 3
TCGA-LGG TCGA-CS-6665 1 0 3
TCGA-LGG TCGA-CS-6666 1 0 3
TCGA-LGG TCGA-CS-6667 1 0 2
TCGA-LGG TCGA-CS-6668 1 1 2
TCGA-LGG TCGA-CS-6669 0 0 2
TCGA-LGG TCGA-DU-5849 1 1 2
TCGA-LGG TCGA-DU-5851 1 0 3
TCGA-LGG TCGA-DU-5852 0 0 3
TCGA-LGG TCGA-DU-5853 1 0 2
TCGA-LGG TCGA-DU-5854 0 0 3
TCGA-LGG TCGA-DU-5855 1 0 3
TCGA-LGG TCGA-DU-5871 1 0 2
TCGA-LGG TCGA-DU-5872 1 0 2
TCGA-LGG TCGA-DU-5874 1 1 2
TCGA-LGG TCGA-DU-6397 1 1 3
TCGA-LGG TCGA-DU-6399 1 0 2
TCGA-LGG TCGA-DU-6400 1 1 2
TCGA-LGG TCGA-DU-6401 1 0 2
TCGA-LGG TCGA-DU-6404 0 0 3
TCGA-LGG TCGA-DU-6405 0 0 3
TCGA-LGG TCGA-DU-6407 1 0 2
TCGA-LGG TCGA-DU-6408 1 0 3
TCGA-LGG TCGA-DU-6410 1 1 3
TCGA-LGG TCGA-DU-6542 1 0 3
TCGA-LGG TCGA-DU-7008 1 0 2
TCGA-LGG TCGA-DU-7010 1 0 3
TCGA-LGG TCGA-DU-7014 -1 0 2
TCGA-LGG TCGA-DU-7015 1 0 2
TCGA-LGG TCGA-DU-7018 1 1 3
TCGA-LGG TCGA-DU-7019 1 0 3
TCGA-LGG TCGA-DU-7294 1 1 2
TCGA-LGG TCGA-DU-7298 1 0 3
TCGA-LGG TCGA-DU-7299 1 0 3
TCGA-LGG TCGA-DU-7300 1 1 3
TCGA-LGG TCGA-DU-7301 1 0 2
TCGA-LGG TCGA-DU-7302 1 1 3
TCGA-LGG TCGA-DU-7304 1 0 3
TCGA-LGG TCGA-DU-7306 1 0 2
TCGA-LGG TCGA-DU-7309 1 0 3
TCGA-LGG TCGA-DU-8162 0 0 3
TCGA-LGG TCGA-DU-8164 1 1 2
TCGA-LGG TCGA-DU-8165 0 0 3
TCGA-LGG TCGA-DU-8166 1 0 2
TCGA-LGG TCGA-DU-8167 1 0 2
TCGA-LGG TCGA-DU-8168 1 1 3
TCGA-LGG TCGA-DU-A5TP 1 0 3
TCGA-LGG TCGA-DU-A5TR 1 0 2
TCGA-LGG TCGA-DU-A5TS 1 0 2
TCGA-LGG TCGA-DU-A5TT 0 0 3
TCGA-LGG TCGA-DU-A5TU 1 0 2
TCGA-LGG TCGA-DU-A5TW 1 0 3
TCGA-LGG TCGA-DU-A5TY 0 0 3
TCGA-LGG TCGA-DU-A6S2 1 1 2
TCGA-LGG TCGA-DU-A6S3 1 1 2
TCGA-LGG TCGA-DU-A6S6 1 1 2
TCGA-LGG TCGA-DU-A6S7 1 0 3
TCGA-LGG TCGA-DU-A6S8 1 1 3
TCGA-LGG TCGA-EZ-7265A -1 -1 -1
TCGA-LGG TCGA-FG-5964 1 1 2
TCGA-LGG TCGA-FG-6688 0 0 3
TCGA-LGG TCGA-FG-6689 1 0 2
TCGA-LGG TCGA-FG-6691 1 0 2
TCGA-LGG TCGA-FG-6692 0 0 3
TCGA-LGG TCGA-FG-7643 0 0 2
TCGA-LGG TCGA-FG-A4MT 1 0 2
TCGA-LGG TCGA-FG-A6IZ 1 1 2
TCGA-LGG TCGA-FG-A713 1 1 2
TCGA-LGG TCGA-HT-7473 1 0 2
TCGA-LGG TCGA-HT-7475 1 0 3
TCGA-LGG TCGA-HT-7602 1 0 2
TCGA-LGG TCGA-HT-7616 1 1 3
TCGA-LGG TCGA-HT-7680 0 0 2
TCGA-LGG TCGA-HT-7684 1 0 3
TCGA-LGG TCGA-HT-7686 1 0 3
TCGA-LGG TCGA-HT-7690 1 0 3
TCGA-LGG TCGA-HT-7692 1 1 2
TCGA-LGG TCGA-HT-7693 1 0 2
TCGA-LGG TCGA-HT-7694 1 1 3
TCGA-LGG TCGA-HT-7855 1 0 3
TCGA-LGG TCGA-HT-7856 1 1 3
TCGA-LGG TCGA-HT-7860 0 0 3
TCGA-LGG TCGA-HT-7874 1 1 3
TCGA-LGG TCGA-HT-7879 1 0 3
TCGA-LGG TCGA-HT-7882 0 0 3
TCGA-LGG TCGA-HT-7884 1 0 2
TCGA-LGG TCGA-HT-8018 1 0 2
TCGA-LGG TCGA-HT-8105 1 1 3
TCGA-LGG TCGA-HT-8106 1 0 3
TCGA-LGG TCGA-HT-8107 0 0 2
TCGA-LGG TCGA-HT-8111 1 0 3
TCGA-LGG TCGA-HT-8113 1 0 2
TCGA-LGG TCGA-HT-8114 1 0 3
TCGA-LGG TCGA-HT-8563 1 0 3
TCGA-LGG TCGA-HT-A5RC 0 0 3
TCGA-LGG TCGA-HT-A614 1 0 2
TCGA-LGG TCGA-HT-A61A 1 0 2
Data_collection Patient IDH_mutated Prediction_score_IDH_wildtype Prediction_score_IDH_mutated 1p19q_codeleted Prediction_score_1p19q_codeleted Prediction_score_1p19q_intact Grade Prediction_score_grade_2 Prediction_score_grade_3 Prediction_score_grade_4
TCGA-GBM TCGA-02-0003 0 099998915 10867886E-05 0 099996686 3308471E-05 4 7377526E-05 000074111245 099918514
TCGA-GBM TCGA-02-0006 0 042321962 05767803 0 068791837 031208166 4 060229343 026596427 013174225
TCGA-GBM TCGA-02-0009 0 099306935 0006930672 0 09906961 0009303949 4 0056565534 010282235 08406121
TCGA-GBM TCGA-02-0011 0 013531776 08646823 0 085318035 01468197 4 0015055533 092510724 005983725
TCGA-GBM TCGA-02-0027 0 09997279 000027212297 0 09986827 00013172914 4 00016104137 00038575265 0994532
TCGA-GBM TCGA-02-0033 0 099974436 000025564007 0 099940693 0000593021 4 00020670628 0003761288 09941717
TCGA-GBM TCGA-02-0034 0 091404164 008595832 0 089209336 01079066 4 00116944825 0061110377 092719513
TCGA-GBM TCGA-02-0037 0 09999577 42315594E-05 0 099992716 72827526E-05 4 82080274E-05 0009249337 09906686
TCGA-GBM TCGA-02-0046 0 0999129 00008710656 0 09989637 00010362669 4 0004290756 0022799779 097290945
TCGA-GBM TCGA-02-0047 0 099991703 83008505E-05 0 09999292 70863265E-05 4 000016252015 0040118434 095971906
TCGA-GBM TCGA-02-0048 0 09998785 000012148175 0 099959475 000040527192 4 00002215901 000039696065 09993814
TCGA-GBM TCGA-02-0054 0 09999831 1689829E-05 0 09999442 5583975E-05 4 00010063206 0060579527 093841416
TCGA-GBM TCGA-02-0059 -1 09993749 000062511285 0 09996424 00003576683 4 00007046657 0010920537 09883748
TCGA-GBM TCGA-02-0060 0 07197039 028029615 0 09016612 009833879 4 017739706 03728545 04497484
TCGA-GBM TCGA-02-0064 0 09999083 9170197E-05 0 09995073 000049264234 4 000043781495 00028024286 099675983
TCGA-GBM TCGA-02-0068 0 099187535 0008124709 0 099528164 00047183693 4 00030539853 059695286 039999318
TCGA-GBM TCGA-02-0069 0 09890871 0010912909 0 099704784 0002952148 4 00057067247 0061368063 09329252
TCGA-GBM TCGA-02-0070 0 09940659 00059340666 0 0957794 0042206 4 0008216515 003556913 09562143
TCGA-GBM TCGA-02-0075 0 099933076 00006693099 0 099735296 00026470982 4 000044697264 00035736929 09959793
TCGA-GBM TCGA-02-0085 0 099114406 0008855922 0 09756698 002433019 4 00065203947 0035171553 095830804
TCGA-GBM TCGA-02-0086 0 099965334 000034666777 0 0998698 00013019645 4 000032699382 00018025768 099787045
TCGA-GBM TCGA-02-0087 -1 09974885 00025114634 0 09990638 000093628286 4 0007505083 0008562708 098393226
TCGA-GBM TCGA-02-0102 0 09797647 0020235319 0 098292196 0017078074 4 003512482 03901857 05746895
TCGA-GBM TCGA-02-0106 -1 099993694 6302759E-05 0 099980897 000019110431 4 60797247E-05 00008735659 09990657
TCGA-GBM TCGA-02-0116 0 09999778 22125667E-05 0 09996886 00003113695 4 000015498884 000051770627 09993273
TCGA-GBM TCGA-06-0119 0 09999362 63770494E-05 0 09999355 6452215E-05 4 000013225728 00028902534 099697745
TCGA-GBM TCGA-06-0122 0 09915298 0008470196 0 09859093 00140907345 4 00121390615 027333176 071452916
TCGA-GBM TCGA-06-0128 1 099988174 000011820537 0 099980634 000019373452 4 000016409029 0007865882 099197
TCGA-GBM TCGA-06-0130 0 099998784 12123987E-05 0 09999323 6775062E-05 4 80872844E-05 00026260202 099729306
TCGA-GBM TCGA-06-0132 0 09998566 000014341719 0 099988496 000011501736 4 000072843547 0005115947 099415565
TCGA-GBM TCGA-06-0133 0 097782 002218004 0 0993807 00061929906 4 0026753133 004659919 09266477
TCGA-GBM TCGA-06-0137 0 096448904 003551094 0 099403125 0005968731 4 000649511 038909483 060441005
TCGA-GBM TCGA-06-0138 0 09977743 00022256707 0 099736834 0002631674 4 00032954598 0011606657 09850979
TCGA-GBM TCGA-06-0139 0 09992649 00007350447 0 099898964 00010103129 4 00021781863 00069256434 099089617
TCGA-GBM TCGA-06-0142 0 099909425 00009057334 0 09985896 00014103584 4 0002598974 0046451908 09509491
TCGA-GBM TCGA-06-0145 0 099964654 000035350278 0 0999652 000034802416 4 00009068022 0021991275 0977102
TCGA-GBM TCGA-06-0149 -1 09992161 00007839425 0 09981067 00018932257 4 00057726577 0013888515 09803388
TCGA-GBM TCGA-06-0154 0 099968064 000031937403 0 0999729 000027106237 4 000041507537 023430935 07652756
TCGA-GBM TCGA-06-0158 0 09999199 8014118E-05 0 099992514 74846226E-05 4 00026547876 020762624 0789719
TCGA-GBM TCGA-06-0162 -1 099964297 00003569706 0 09997459 0000254147 4 000033955855 004318936 09564711
TCGA-GBM TCGA-06-0164 -1 09983991 00016009645 0 09873262 0012673735 4 00016517473 00048346478 09935136
TCGA-GBM TCGA-06-0166 0 099991715 82846556E-05 0 0999554 00004459562 4 000013499439 0011635037 098823
TCGA-GBM TCGA-06-0168 0 09975561 00024438864 0 09964825 00035174883 4 0004766434 010448053 089075303
TCGA-GBM TCGA-06-0175 -1 09996252 000037482675 0 09988098 00011902251 4 00026097735 004992068 094746953
TCGA-GBM TCGA-06-0176 0 099550986 00044901576 0 09998872 000011279297 4 0032868527 036690876 06002227
TCGA-GBM TCGA-06-0177 -1 081774735 018225263 0 09946464 00053536464 4 0026683953 013013016 08431859
TCGA-GBM TCGA-06-0179 -1 09997508 000024923254 0 09989778 00010222099 4 0002628482 0004127114 099324447
TCGA-GBM TCGA-06-0182 -1 099999547 45838406E-06 0 099998736 12656287E-05 4 00002591103 000018499703 09995559
TCGA-GBM TCGA-06-0184 0 09935369 00064631375 0 099458355 00054164114 4 0023110552 0017436244 09594532
TCGA-GBM TCGA-06-0185 0 09999337 66310655E-05 0 099986255 000013738607 4 7657532E-05 0016089642 098383385
TCGA-GBM TCGA-06-0187 0 09991689 00008312097 0 099700147 00029984985 4 00020616595 0033111423 096482694
TCGA-GBM TCGA-06-0188 0 09883802 0011619771 0 09826743 0017325714 4 0013776424 0112841725 087338185
TCGA-GBM TCGA-06-0189 0 099906737 0000932636 0 09983865 00016135005 4 00022760795 00106745735 09870494
TCGA-GBM TCGA-06-0190 0 099954176 000045831292 0 09967013 00032986512 4 000040555766 0001246768 099834764
TCGA-GBM TCGA-06-0192 0 09997876 00002123566 0 09992735 00007264875 4 00004505576 00014473333 09981021
TCGA-GBM TCGA-06-0213 0 099986935 00001305845 0 099971646 000028351307 4 8755587E-05 00013480412 09985644
TCGA-GBM TCGA-06-0238 0 09999982 17603431E-06 0 09999894 10616134E-05 4 8076515E-05 56053756E-05 099986315
TCGA-GBM TCGA-06-0240 0 09989956 00010044163 0 099948466 00005152657 4 00016040986 021931975 077907616
TCGA-GBM TCGA-06-0241 0 099959785 000040211933 0 099910825 00008917038 4 00023411359 0007850656 098980826
TCGA-GBM TCGA-06-0644 0 09871044 0012895588 0 09859228 00140771745 4 0013671214 009819665 088813215
TCGA-GBM TCGA-06-0646 0 099959 00004100472 0 099936503 000063495064 4 00019223108 0040443853 095763385
TCGA-GBM TCGA-06-0648 0 09999709 29083441E-05 0 099982435 000017571273 4 000077678583 000038868992 099883455
TCGA-GBM TCGA-06-0649 0 09997805 000021952427 0 099951684 000048311835 4 0042641632 00058432207 095151514
TCGA-GBM TCGA-06-1084 0 099985826 000014174655 0 099968565 00003144242 4 00002676724 020492287 079480946
TCGA-GBM TCGA-06-1802 -1 09991928 00008072305 0 09956176 0004382337 4 000043478087 00019495043 09976157
TCGA-GBM TCGA-06-2570 1 096841115 0031588882 0 09842457 0015754245 4 0015369608 0030956635 09536738
TCGA-GBM TCGA-06-5408 0 099857306 00014269598 0 09962638 00037362208 4 00027690146 0016195394 098103565
TCGA-GBM TCGA-06-5412 0 099366105 0006338921 0 099193794 0008061992 4 0011476759 006606435 09224589
TCGA-GBM TCGA-06-5413 0 09994105 000058955856 0 09983026 00016974095 4 00027100197 0021083053 097620696
TCGA-GBM TCGA-06-5417 1 01521267 08478733 -1 03064492 06935508 4 013736826 037757674 048505494
TCGA-GBM TCGA-06-6389 1 099987435 000012558252 0 09997017 000029827762 4 00014020519 00020044278 099659353
TCGA-GBM TCGA-08-0350 0 019229275 08077072 0 0033211168 09667888 4 0051619414 022280572 072557485
TCGA-GBM TCGA-08-0352 0 099997497 25071595E-05 0 099992514 74846226E-05 4 000024192198 000048111935 099927694
TCGA-GBM TCGA-08-0353 0 09901496 0009850325 0 09967775 0003222484 4 00053748637 0004291497 09903336
TCGA-GBM TCGA-08-0354 0 076413894 023586108 0 07554566 024454337 4 008784444 02004897 071166587
TCGA-GBM TCGA-08-0355 0 09998349 000016506859 0 099984336 000015659066 4 000076689845 0023648744 09755844
TCGA-GBM TCGA-08-0356 0 097673583 0023264103 0 097773504 0022264915 4 001175834 0031075679 095716596
TCGA-GBM TCGA-08-0357 0 099509466 0004905406 0 099300176 00069982093 4 0005191745 0038681854 095612645
TCGA-GBM TCGA-08-0358 0 099999785 2199356E-06 0 099999034 9628425E-06 4 6113315E-06 00011283219 09988656
TCGA-GBM TCGA-08-0359 0 097885466 0021145396 0 09956006 00043994132 4 0009885523 0066605434 092350906
TCGA-GBM TCGA-08-0360 0 09922444 00077555366 -1 09948704 00051296344 4 0013318472 003317344 095350814
TCGA-GBM TCGA-08-0385 0 099605453 00039454065 -1 099686414 0003135836 4 00050293226 0029977333 096499336
TCGA-GBM TCGA-08-0389 0 099964714 000035281325 0 09991272 000087276706 4 00017554013 00024730961 099577147
TCGA-GBM TCGA-08-0390 0 099945146 000054847915 0 099936 00006399274 4 00036811908 00050958768 0991223
TCGA-GBM TCGA-08-0392 0 099962366 000037629317 0 09993575 00006424303 4 000036593352 0010291994 09893421
TCGA-GBM TCGA-08-0512 -1 09982893 00017106998 0 099193794 0008061992 4 00016200381 00027773918 09956026
TCGA-GBM TCGA-08-0520 -1 099603915 00039607873 0 09981933 00018066854 4 00007140295 0019064669 09802213
TCGA-GBM TCGA-08-0521 -1 09975274 0002472623 0 099490017 00050998176 4 0001514669 0020103427 09783819
TCGA-GBM TCGA-08-0522 -1 099960107 000039899128 -1 09992053 00007947255 4 0000269389 0006173321 09935573
TCGA-GBM TCGA-08-0524 -1 09964619 0003538086 0 099620515 0003794834 4 000019140428 0010096702 09897119
TCGA-GBM TCGA-08-0529 -1 09996567 000034329997 0 099952066 00004793605 4 000032077235 0035970636 09637086
TCGA-GBM TCGA-12-0616 0 098521465 0014785408 0 098704207 001295789 4 001592791 012875569 08553164
TCGA-GBM TCGA-12-0776 -1 099899167 00010083434 0 09987031 00012968953 4 0019219175 00637484 09170324
TCGA-GBM TCGA-12-0829 0 099913067 00008693674 0 099821776 00017821962 4 00021031094 0055067167 09428297
TCGA-GBM TCGA-12-1093 0 099992585 7411892E-05 0 09999448 5518923E-05 4 000046803855 0012115157 098741674
TCGA-GBM TCGA-12-1094 -1 09980045 00019955388 0 09866105 0013389497 4 00053194338 001599471 097868586
TCGA-GBM TCGA-12-1098 -1 09998406 000015936712 0 09977216 00022783307 4 000010218692 0035607774 09642901
TCGA-GBM TCGA-12-1598 0 096309197 0036908068 0 097933435 0020665688 4 0012952217 052912676 045792103
TCGA-GBM TCGA-12-1601 0 09875683 0012431651 -1 0991891 0008108984 -1 00118053425 0105477065 088271755
TCGA-GBM TCGA-12-1602 0 099830914 00016908031 0 099858415 00014158705 4 0008427611 0025996923 09655755
TCGA-GBM TCGA-12-3650 0 09761519 0023848088 0 097467697 0025323058 4 0010450666 043705726 05524921
TCGA-GBM TCGA-14-0789 0 099856466 00014353332 0 099666256 00033374047 4 0001406897 0008273975 099031913
TCGA-GBM TCGA-14-1456 1 006299064 093700933 0 08656222 013437784 4 016490369 047177824 036331803
TCGA-GBM TCGA-14-1794 0 08579393 014206071 0 09850429 0014957087 4 0023009384 009868736 08783033
TCGA-GBM TCGA-14-1825 0 099960107 000039899128 0 099968123 00003187511 4 0008552247 0010156045 09812918
TCGA-GBM TCGA-14-1829 0 090690076 009309922 0 09907856 0009214366 4 0008461936 0102735735 088880235
TCGA-GBM TCGA-14-3477 0 099796116 0002038787 0 09990728 00009271923 4 00032272525 0021644868 09751279
TCGA-GBM TCGA-19-0963 -1 099876726 00012327607 0 09983612 00016388679 4 00031698826 013153598 086529416
TCGA-GBM TCGA-19-1390 0 099913234 00008676725 0 099703634 00029636684 4 00015592943 0026028048 097241265
TCGA-GBM TCGA-19-1789 0 09809491 00190509 0 09915216 0008478402 4 0038703684 014341596 08178804
TCGA-GBM TCGA-19-2624 0 07535573 024644265 0 09816127 0018387254 4 012311598 012769651 07491875
TCGA-GBM TCGA-19-2631 0 099860877 0001391234 0 09981178 00018821858 4 00009839778 001843531 09805807
TCGA-GBM TCGA-19-5951 0 09999031 9685608E-05 0 099977034 000022960825 4 00020246736 0004014765 09939606
TCGA-GBM TCGA-19-5954 0 099456257 00054374957 0 09968273 00031726828 4 00073725334 006310084 09295266
TCGA-GBM TCGA-19-5958 0 099999475 5234907E-06 0 0999941 58978338E-05 4 35422294E-05 86819025E-05 09998777
TCGA-GBM TCGA-19-5960 0 09683962 003160382 0 09013577 009864227 4 0011394806 018114014 08074651
TCGA-GBM TCGA-27-1834 0 099998164 18342893E-05 0 09999685 31446623E-05 4 8611921E-05 000031686216 0999597
TCGA-GBM TCGA-27-1838 0 09993625 000063743413 0 09940428 0005957154 4 00006736379 0007191195 099213517
TCGA-GBM TCGA-27-2526 0 099983776 000016219281 0 09996898 000031015594 4 000016658282 00006323714 09992011
TCGA-GBM TCGA-76-4932 0 09867389 0013261103 -1 09949397 0005060332 4 00007321126 0003016794 099625117
TCGA-GBM TCGA-76-4934 0 099318933 0006810731 0 09995073 000049264234 4 00061555947 00070025027 098684186
TCGA-GBM TCGA-76-4935 0 074562997 025437003 0 098242307 001757688 4 076535034 006437644 017027317
TCGA-GBM TCGA-76-6191 0 09981067 00018932257 0 09970879 00029121784 4 00044340584 00096095055 098595643
TCGA-GBM TCGA-76-6193 0 09966168 0003383191 0 099850464 00014953383 4 00037061477 007873953 09175543
TCGA-GBM TCGA-76-6280 0 099948776 00005122569 0 099908185 000091819017 4 00001475792 000846075 099139166
TCGA-GBM TCGA-76-6282 0 0995906 0004093958 0 099861956 00013804223 4 00006694951 0009437619 09898929
TCGA-GBM TCGA-76-6285 0 099949074 00005092657 0 09971661 00028338495 4 00031175872 004005614 095682627
TCGA-GBM TCGA-76-6656 0 09996917 000030834455 0 09983897 00016103574 4 002648366 00017969633 09717193
TCGA-GBM TCGA-76-6657 0 099987245 000012755992 0 099951494 00004850083 4 000096620515 0005599633 09934342
TCGA-GBM TCGA-76-6661 0 093211424 006788577 0 09640178 0035982177 4 003490037 0026863772 09382358
TCGA-GBM TCGA-76-6662 0 096425414 0035745807 0 09963924 0003607617 4 002845819 002544755 09460942
TCGA-GBM TCGA-76-6663 0 088664144 0113358565 0 09984207 00015792594 4 0010206689 043740335 05523899
TCGA-GBM TCGA-76-6664 0 011047115 08895289 0 09559813 004401865 4 00049677677 08806894 011434281
TCGA-LGG TCGA-CS-4941 0 088931274 011068726 0 087037706 012962292 3 002865127 0048591908 092275685
TCGA-LGG TCGA-CS-4942 1 00031327847 099686724 0 096309197 0036908068 3 096261597 00148612335 0022522787
TCGA-LGG TCGA-CS-4943 1 0005265965 099473405 0 09940544 00059455987 3 09439103 0023049146 003304057
TCGA-LGG TCGA-CS-4944 1 009363656 09063635 0 08755211 0124478824 2 034047556 033881712 03207073
TCGA-LGG TCGA-CS-5393 1 009623762 09037624 0 098178816 001821182 3 014111634 042021698 043866673
TCGA-LGG TCGA-CS-5395 0 08502822 014971776 0 09932025 00067975316 2 0052374925 018397054 076365453
TCGA-LGG TCGA-CS-5396 1 099839586 00016040892 1 099967945 000032062363 3 00016345463 029090768 07074577
TCGA-LGG TCGA-CS-5397 0 049304244 050695753 0 08829839 0117016025 3 038702008 021211159 040086827
TCGA-LGG TCGA-CS-6186 0 099913234 00008676725 0 099956185 000043818905 3 00008662089 016898473 083014905
TCGA-LGG TCGA-CS-6188 0 052768165 047231838 0 08584221 014157787 3 019437431 047675493 03288707
TCGA-LGG TCGA-CS-6290 1 09102666 008973339 0 09462997 0053700306 3 0104100704 025633416 06395651
TCGA-LGG TCGA-CS-6665 1 099600047 0003999501 0 099756086 00024391294 3 0011873978 001634113 097178483
TCGA-LGG TCGA-CS-6666 1 021655986 07834402 0 09327296 0067270435 3 017667453 036334327 045998225
TCGA-LGG TCGA-CS-6667 1 012061995 087938 0 095699733 0043002643 2 063733935 019323014 016943048
TCGA-LGG TCGA-CS-6668 1 0076787576 09232124 1 04240933 057590663 2 06810894 013706882 018184178
TCGA-LGG TCGA-CS-6669 0 08488156 01511844 0 094018847 005981148 2 0037862387 002352077 09386168
TCGA-LGG TCGA-DU-5849 1 005773187 094226813 1 08664153 013358466 2 072753835 015028271 012217898
TCGA-LGG TCGA-DU-5851 1 09963994 0003600603 0 099808073 00019192374 3 00060602655 012558761 08683521
TCGA-LGG TCGA-DU-5852 0 09998591 000014091856 0 099954873 000045121062 3 0002267452 00038046916 099392784
TCGA-LGG TCGA-DU-5853 1 0010986943 09890131 0 09549844 0045015533 2 08603989 0077804394 0061796777
TCGA-LGG TCGA-DU-5854 0 09567354 0043264627 0 098768765 0012312326 3 01194655 027027336 06102612
TCGA-LGG TCGA-DU-5855 1 0009312956 09906871 0 046602532 053397465 3 0008289882 097042197 0021288157
TCGA-LGG TCGA-DU-5871 1 005623634 09437636 0 09449439 005505607 2 042517176 020180763 037302068
TCGA-LGG TCGA-DU-5872 1 0062359583 09376405 0 015278916 08472108 2 012133307 048199505 039667192
TCGA-LGG TCGA-DU-5874 1 022858672 077141327 1 06457066 03542934 2 058503634 020639434 02085693
TCGA-LGG TCGA-DU-6397 1 097691274 002308724 1 09908213 0009178773 3 00048094327 00412339 09539566
TCGA-LGG TCGA-DU-6399 1 00023920655 099760795 0 09970073 00029926652 2 098691386 0007037292 0006048777
TCGA-LGG TCGA-DU-6400 1 0030923586 09690764 1 037771282 06222872 2 09710506 0015339471 001360994
TCGA-LGG TCGA-DU-6401 1 0014545513 098545444 0 045332992 054667014 2 0878585 006398724 005742785
TCGA-LGG TCGA-DU-6404 0 08563024 014369765 0 09857318 00142681915 3 0012578745 08931047 009431658
TCGA-LGG TCGA-DU-6405 0 094122344 0058776554 0 09657707 0034229323 3 0015099723 0858934 012596628
TCGA-LGG TCGA-DU-6407 1 00046772743 099532276 0 095787287 0042127114 2 095650303 0019410672 0024086302
TCGA-LGG TCGA-DU-6408 1 0032852467 09671475 0 02978783 070212173 3 046377006 04552443 008098562
TCGA-LGG TCGA-DU-6410 1 084198 015801999 1 09610981 0038901985 3 0029748935 0547783 042246798
TCGA-LGG TCGA-DU-6542 1 099541724 0004582765 0 099690056 00030994152 3 00036504513 0033356518 0962993
TCGA-LGG TCGA-DU-7008 1 00027017966 09972982 0 09924154 0007584589 2 0945233 0033200152 0021566862
TCGA-LGG TCGA-DU-7010 1 09090629 0090937115 0 083999664 016000335 3 0011747591 011156695 08766855
TCGA-LGG TCGA-DU-7014 -1 00067384504 09932615 0 09144437 008555635 2 090214694 005846623 003938676
TCGA-LGG TCGA-DU-7015 1 011059116 08894088 0 09457512 005424881 2 04990067 023008518 027090812
TCGA-LGG TCGA-DU-7018 1 06190684 038093168 1 09720721 0027927874 3 002608347 03462771 06276394
TCGA-LGG TCGA-DU-7019 1 006866228 09313377 0 068647516 031352484 3 06280373 02546188 011734395
TCGA-LGG TCGA-DU-7294 1 039513415 06048658 1 04910898 05089102 2 044678423 011827048 04349453
TCGA-LGG TCGA-DU-7298 1 002178117 097821885 0 058896303 041103697 3 04621931 040058115 013722575
TCGA-LGG TCGA-DU-7299 1 0050494254 094950575 0 09805993 0019400762 3 088520575 003754964 0077244624
TCGA-LGG TCGA-DU-7300 1 020334144 07966585 1 021174264 078825736 3 06957292 014594184 015832895
TCGA-LGG TCGA-DU-7301 1 0028517082 09714829 0 07594931 024050693 2 07559878 013617343 010783881
TCGA-LGG TCGA-DU-7302 1 007878401 092121595 1 097414124 0025858777 3 059945434 013100924 02695364
TCGA-LGG TCGA-DU-7304 1 0049359404 09506406 0 09947084 0005291605 3 05746174 017312215 025226048
TCGA-LGG TCGA-DU-7306 1 0774658 022534202 0 09720191 0027980946 2 007909051 04979186 042299092
TCGA-LGG TCGA-DU-7309 1 002068546 097931457 0 091696864 008303132 3 091011685 0041825026 0048058107
TCGA-LGG TCGA-DU-8162 0 019030987 08096902 0 084344435 015655571 3 06724078 015660264 017098951
TCGA-LGG TCGA-DU-8164 1 0026989132 09730109 1 06119184 038808158 2 078654927 011851947 00949313
TCGA-LGG TCGA-DU-8165 0 099918324 000081673806 0 09982692 00017308301 3 00077142627 001586733 097641844
TCGA-LGG TCGA-DU-8166 1 0062617026 093738294 0 052265906 047734097 2 0571523 027175376 015672325
TCGA-LGG TCGA-DU-8167 1 008068282 09193171 0 08626991 013730097 2 07117111 014616342 014212546
TCGA-LGG TCGA-DU-8168 1 04501781 05498219 1 09405718 005942822 3 028535154 039651006 031813842
TCGA-LGG TCGA-DU-A5TP 1 013576113 08642388 0 098667485 0013325148 3 06805368 011191124 020755199
TCGA-LGG TCGA-DU-A5TR 1 0038810804 09611892 0 094154674 005845324 2 07418394 01198958 013826479
TCGA-LGG TCGA-DU-A5TS 1 036534345 06346565 0 097664696 0023353029 2 0076500095 07058904 021760948
TCGA-LGG TCGA-DU-A5TT 0 057493186 042506814 0 08586593 014134066 3 024835269 018135522 05702921
TCGA-LGG TCGA-DU-A5TU 1 017411166 082588834 0 08903419 0109658085 2 026840523 031951824 041207647
TCGA-LGG TCGA-DU-A5TW 1 00015382263 099846184 0 09784259 0021574067 3 099424005 00014788082 0004281163
TCGA-LGG TCGA-DU-A5TY 0 099497885 000502115 0 09904406 0009559399 3 00076062134 003340487 09589889
TCGA-LGG TCGA-DU-A6S2 1 01338958 08661042 1 010181248 08981875 2 08703488 0033631936 0096019216
TCGA-LGG TCGA-DU-A6S3 1 007097701 092902297 1 0049773447 09502266 2 08236395 0043779366 013258114
TCGA-LGG TCGA-DU-A6S6 1 00054852334 09945148 1 00052813343 09947187 2 095740056 0030734295 0011865048
TCGA-LGG TCGA-DU-A6S7 1 00015218158 099847823 0 09977216 00022783307 3 097611564 0011231668 0012652792
TCGA-LGG TCGA-DU-A6S8 1 090418625 009581377 1 09320215 006797852 3 015433969 006605734 0779603
TCGA-LGG TCGA-EZ-7265A -1 001654544 09834546 -1 092290026 0077099696 -1 091443384 0045349486 0040216673
TCGA-LGG TCGA-FG-5964 1 095945925 004054074 1 09480585 0051941562 2 0052469887 018844457 075908554
TCGA-LGG TCGA-FG-6688 0 041685596 0583144 0 0400786 0599214 3 032869554 028211078 03891937
TCGA-LGG TCGA-FG-6689 1 0040960647 09590394 0 088871056 01112895 2 078484637 01247657 009038795
TCGA-LGG TCGA-FG-6691 1 00066411127 09933589 0 09705485 002945148 2 082394814 01353293 004072261
TCGA-LGG TCGA-FG-6692 0 099044985 0009550158 0 098370695 0016293105 3 002482948 023811981 07370507
TCGA-LGG TCGA-FG-7643 0 067991304 032008696 0 094600123 0053998843 2 032237333 020420441 047342223
TCGA-LGG TCGA-FG-A4MT 1 00037180893 09962819 0 098237246 0017627545 2 09685786 0019877713 0011543726
TCGA-LGG TCGA-FG-A6IZ 1 0023916386 097608364 1 003330537 096669465 2 016134319 0751304 0087352775
TCGA-LGG TCGA-FG-A713 1 020932822 07906717 1 053740746 046259254 2 068370515 013241291 018388201
TCGA-LGG TCGA-HT-7473 1 026437023 073562974 0 09914391 0008560891 2 009070598 05834457 032584828
TCGA-LGG TCGA-HT-7475 1 0014885316 09851147 0 093397486 006602513 3 09713343 0009202645 0019462984
TCGA-LGG TCGA-HT-7602 1 0078306936 09216931 0 044295275 055704725 2 06683338 024550638 00861598
TCGA-LGG TCGA-HT-7616 1 0994089 0005911069 1 08912444 010875558 3 00015109215 00081261955 09903628
TCGA-LGG TCGA-HT-7680 0 01775255 08224745 0 079779327 020220678 2 06160002 021518312 016881672
TCGA-LGG TCGA-HT-7684 1 099250317 00074968883 0 09977216 00022783307 3 0001585032 0011880362 09865346
TCGA-LGG TCGA-HT-7686 1 043986762 05601324 0 09985134 00014866153 3 08800356 0017893802 010207067
TCGA-LGG TCGA-HT-7690 1 0508178 049182203 0 09986749 00013250223 3 007351807 06479417 027854022
TCGA-LGG TCGA-HT-7692 1 0006764646 09932354 1 00017718028 099822825 2 084003216 009709251 00628754
TCGA-LGG TCGA-HT-7693 1 08835126 011648734 0 098880965 0011190402 2 00389769 060259813 035842496
TCGA-LGG TCGA-HT-7694 1 006299064 093700933 1 06663645 03336355 3 06368097 02515797 011161056
TCGA-LGG TCGA-HT-7855 1 013434944 086565053 0 06805072 031949285 3 03900612 035394293 02559958
TCGA-LGG TCGA-HT-7856 1 0037151825 09628482 1 045108467 05489153 3 0024809493 094240344 0032787096
TCGA-LGG TCGA-HT-7860 0 09996338 000036614697 0 099890125 00010987312 3 00023981468 004139189 095620996
TCGA-LGG TCGA-HT-7874 1 027373514 07262649 1 061277324 03872268 3 030690825 04321765 026091516
TCGA-LGG TCGA-HT-7879 1 006545533 09345446 0 07643643 023563562 3 07316188 01349482 0133433
TCGA-LGG TCGA-HT-7882 0 099826247 00017375927 0 099920684 0000793176 3 00026636408 0011237644 098609877
TCGA-LGG TCGA-HT-7884 1 0045437213 09545628 0 09804874 0019512545 2 067944294 020575646 011480064
TCGA-LGG TCGA-HT-8018 1 0090937115 09090629 0 08061669 019383314 2 069696444 017501967 012801588
TCGA-LGG TCGA-HT-8105 1 09291196 0070880495 1 09865976 0013402403 3 036353382 007970196 055676425
TCGA-LGG TCGA-HT-8106 1 09987081 00012918457 0 099922514 00007748164 3 0019909225 0057560045 09225307
TCGA-LGG TCGA-HT-8107 0 006150854 09384914 0 018944609 08105539 2 071847403 015055439 013097167
TCGA-LGG TCGA-HT-8111 1 062096643 037903354 0 09220272 007797278 3 00015017459 090417147 009432678
TCGA-LGG TCGA-HT-8113 1 0003941571 099605846 0 00025608707 099743915 2 088384247 009021307 002594447
TCGA-LGG TCGA-HT-8114 1 09404078 005959219 0 09970708 00029292419 3 0015292862 028022403 07044831
TCGA-LGG TCGA-HT-8563 1 099999154 843094E-06 0 099999607 39515203E-06 3 32918017E-06 031742522 068257153
TCGA-LGG TCGA-HT-A5RC 0 06915494 030845058 0 045883363 054116637 3 012257236 02777765 059965116
TCGA-LGG TCGA-HT_A614 1 08180474 018195263 0 09584989 00415011 2 0067164555 0059489973 087334543
TCGA-LGG TCGA-HT-A61A 1 0035779487 09642206 0 07277821 0272218 2 07607039 014429174 009500429
Page 9: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 10: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 11: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 12: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 13: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 14: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 15: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 16: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 17: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 18: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 19: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 20: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 21: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 22: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 23: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 24: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 25: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 26: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 27: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 28: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 29: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 30: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 31: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 32: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 33: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 34: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 35: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 36: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 37: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 38: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 39: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 40: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 41: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 42: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 43: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 44: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 45: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 46: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 47: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 48: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin
Page 49: arXiv:2010.04425v1 [eess.IV] 9 Oct 2020 · 2020. 10. 12. · De Witt Hamer 7, Roelant S Eijgelaar , Pim J French4, Hendrikus J Dubbink8, Arnaud JPE Vincent3, Wiro J Niessen1,9, Martin