Fully Automatic Binary Glioma Grading based on Pre-Therapy MRI using 3D Convolutional Neural Networks

Milan Decuyper; Roel Van Holen

REVIEW 4 major objections 6 minor 13 references

Fully Automatic Binary Glioma Grading based on Pre-Therapy MRI using 3D Convolutional Neural Networks

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Fully automatic MRI pipeline grades gliomas with 91% accuracy

desk verdict A clinically plausible but incremental two-stage 3D CNN for glioma grading whose external validation is undermined by an untested assumption that segmentation errors don't affect classification. read the letter →

arxiv 1908.01506 v1 pith:SJR2VDZH submitted 2019-08-05 eess.IV

Milan Decuyper , Roel Van Holen This is my paper · ORCID

classification eess.IV

keywords gliomagradingglioblastomalower-grade3DconvolutionalneuralnetworkMRIsegmentationU-Netcomputer-aideddiagnosispre-therapy

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that glioma grading—the decision between high-grade glioblastoma and lower-grade glioma—can be made non-invasively and without any manual steps from standard pre-therapy MRI. The proposed system is a two-stage neural network: a 3D U-Net first segments the whole tumour, and a 3D residual classifier then labels the extracted tumour region. On a public multi-centre test cohort the system reaches 91.25% accuracy, and on independent data from a second hospital it reaches 91.96%. A sympathetic reader would care because manual MRI grading is unreliable, while biopsy needed for tissue diagnosis can shorten survival; a dependable automatic grader could guide initial treatment decisions.

What carries the argument

The load-bearing mechanism is the cascade of two networks, with the tumour ROI extracted from the automatic segmentation as the central object. The first stage is a 3D U-Net—a convolutional network with a contracting path and an expanding path that outputs a per-voxel whole-tumour mask—built with 25 features at the highest resolution, instance normalisation and leaky ReLUs, and trained with a soft dice loss. The T1ce volume inside that predicted mask is resized to 112x112x112 and fed to the classifier, a 3D residual network with one 7x7x7 convolution, four residual blocks, instance normalisation and ReLU after every convolution, then average pooling and a two-class output. Randomly zeroing the T1, T2 or FLAIR channels during segmentation training makes the system robust to missing modality sequences. The resizing step is what turns tumours of different shapes and sizes into a fixed input so the classifier only sees the tumour region.

What would settle it

On the same public test patients, feed the classifier ROIs defined by expert manual segmentation and compare accuracy to the 91.25% obtained with automatic masks; a large drop would directly falsify the paper's assertion that segmentation errors do not influence the tumour ROI.

Watch

Extended reading notes

Core claim

The central claim is that this fully automatic two-stage pipeline—segmentation with a 3D U-Net, classification of the tumour region of interest with a 3D residual CNN—achieves state-of-the-art binary glioma grading on a large, heterogeneous, multi-centre dataset. On 80 held-out public test patients the classifier reaches 91.25% accuracy with an AUC of 96.29%, and on 112 independent patients from another hospital it reaches 91.96% accuracy with an AUC of 93.39%. The segmenter itself scores a whole-tumour dice coefficient of 88.53% on a public benchmark test set, which the authors say matches top-performing entries in the most recent challenge. They further assert that small differences between manual and predicted segmentations have little effect on the tumour ROI, so the segmentation quality is sufficient for grading. Their conclusion is that state-of-the-art performance is achievable with a system trained on a large heterogeneous dataset and that the method generalises to different imaging protocols and centres.

Load-bearing premise

The pipeline assumes that the tumour region automatically cut out by the 3D U-Net and squeezed into a fixed cube still contains all the image clues needed to distinguish glioblastoma from lower-grade glioma; if common segmentation mistakes leave out or add in the wrong tissue, the reported accuracy would not transfer.

Editorial extensions

If this is right

If the accuracy holds in clinical use, initial management decisions—watch-and-wait versus aggressive resection with chemoradiotherapy—could be made from MRI alone, without biopsy or manual tumour grading.
Because the segmentation network was trained to tolerate missing input channels, the full pipeline could be used at centres that do not acquire all four standard MRI sequences.
The strong result on a second hospital's data implies the method is not tied to one scanner or protocol, provided the preprocessing steps are reproduced.
A fully automatic grader could be run at the time of the first scan, giving an immediate malignancy estimate while histopathology is pending or unavailable.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper leaves implicit that the fixed-cube resize makes tumour size a potentially usable signal, since larger lesions are more often glioblastoma; a volume-matched control experiment would reveal whether the classifier learns size or tissue-level patterns.
An extension the authors do not attempt is applying the same cascade to finer distinctions—WHO grade II versus III, or IDH-mutant versus IDH-wild-type disease—so it is unknown how far the learned features generalise beyond the coarse binary split.
A single end-to-end network trained jointly on segmentation and classification could plausibly outperform the two-stage design, and the reported accuracy gives a concrete baseline for such a comparison.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a solid extended abstract that does one thing better than most papers in this niche — it validates the full pipeline on a genuinely held-out external hospital dataset. The architecture is standard (3D U-Net plus 3D residual classifier), but the scale is not: 549 patients for training, 112 patients from Ghent University Hospital for final testing, with explicit separation between segmentation training and test data. The reported accuracies (91.25% on TCIA, 91.96% on Ghent) are plausible and the AUC/MCC numbers help. The authors also avoid self-citation and are honest about the data sources. That is real credit where it's due.

The soft spots are real but not disqualifying. The stress-test note is on target: the paper asserts in Section 3 that small variations between manual and predicted segmentations won't strongly influence the ROI, but never tests this directly. The classifier sees only the automatically segmented, resized 112x112x112 ROI. If the U-Net systematically under-segments non-enhancing tumor or over-includes edema on the Ghent data, the input distribution shifts, and the 91.96% accuracy becomes harder to interpret. We don't even get U-Net Dice on the Ghent set, which is the only basis for the cross-center robustness claim. Also missing: confidence intervals and comparisons to other methods on the same test sets, so "state-of-the-art" is not established. The single external center is a limitation but a minor one for a four-page abstract.

The paper does not do anything circular or incoherent. The problem is an untested assumption, not a misrepresented result. For an MIDL extended abstract, this is a legitimate feasibility study. I would not call the robustness claim proven, but I would call it a reasonable hypothesis that deserves a proper test.

Who is this for? People building CAD pipelines for brain tumors and referees who want a clean example of external validation design. I'd send this to peer review rather than desk reject — it's a legitimate empirical study with a clear gap that a full version could address (ROI sensitivity analysis, segmentation Dice on the external set, confidence intervals). The stress-test note identifies the paper's actual weak spot, not a manufactured one. My bottom line: conditionally accept, with that missing analysis as the main condition.

Referee Report

4 major / 6 minor

Summary. The paper proposes a fully automatic two-stage pipeline for binary glioma grading (glioblastoma versus lower-grade glioma, WHO grades II/III) from pre-therapy structural MRI. The first stage segments the whole tumor with a 3D U-Net; the second extracts a 112x112x112 ROI from the T1ce image inside the predicted mask and classifies it with a 3D residual CNN. The pipeline is trained on 549 patients from public datasets (BraTS 2018 training data, TCGA-GBM, TCGA-LGG, LGG-1p19qDeletion) and evaluated on 80 held-out TCIA patients and on 112 patients from Ghent University Hospital. Reported performance includes a whole-tumor Dice of 88.53% on 76 BraTS 2018 test cases, and classification accuracy of 91.25% on the TCIA test set and 91.96% on the Ghent set. The authors conclude that state-of-the-art grading performance is achieved and that the system is robust to data from different centres.

Significance. If the reported results hold, the paper makes a useful empirical contribution: it demonstrates a fully automatic, two-stage deep-learning system for glioma grading on a relatively large multi-source dataset, with an independent external evaluation. The use of public BraTS/TCIA data and an unseen single-center dataset is a strength, as are the explicit train/validation/test splits and the statement that test patients were excluded from segmentation training. The paper is an extended abstract and does not claim a mathematical derivation, so the main scientific value is the measured accuracy on held-out data. However, the strength of the cross-center robustness claim depends on an untested assumption about the interaction between segmentation errors and classification accuracy, and the reported performance metrics lack uncertainty quantification. These issues are addressable and do not require a change in the core methodology.

major comments (4)

[Section 3, first paragraph] The assertion that 'small variations between manual and predicted segmentations won't have a strong influence on the tumour ROI' is load-bearing for the fully automatic claim, but it is never tested. The paper reports segmentation Dice and classification accuracy as separate numbers, but does not compare classification accuracy using manual ROIs versus automatic ROIs on the same patients. Because the 112x112x112 ROI is the only input to the classifier, systematic under-segmentation or over-inclusion of tissue could shift the classifier's input distribution. Please provide such a comparison, or clearly restrict the robustness claim to the specific automatic pipeline as evaluated.
[Section 2.3] The paper does not specify whether the ROIs used to train the classification network were generated by the automatic U-Net, by manual segmentations, or by a mixture. This is critical for interpreting the reported accuracies: if the classifier was trained on manual ROIs and tested on automatic ROIs, a train/test mismatch exists; if it was trained on automatic ROIs, the network may have absorbed the segmentation bias of the public data. Please state explicitly how the training ROIs were obtained.
[Sections 2.2 and 3] The U-Net Dice score is reported only for the 76 BraTS 2018 test patients, not for the Ghent University Hospital data. The Ghent accuracy of 91.96% is the sole basis for the conclusion of robustness to different centres and imaging protocols, but without any measure of segmentation quality or input-distribution shift on the external data, this conclusion is not fully supported. Please report segmentation performance on the Ghent data or otherwise analyze the effect of cross-center variation on the classifier inputs.
[Table 1] No confidence intervals or statistical significance tests are provided for accuracy, AUC, MCC, sensitivity, or specificity. With 80 and 112 test patients, the difference between 91.25% and 91.96% accuracy is within sampling noise, and sensitivity/specificity estimates have wide intervals. Please add bootstrap confidence intervals or another form of uncertainty quantification for the headline metrics.

minor comments (6)

[Author affiliations] The author names appear as 'Roel V an Holen' with spacing irregularities; please correct the typesetting.
[Table 1 caption] The table header text 'Ghent Univeristy Hospital' contains a typo; it should be 'Ghent University Hospital'.
[Section 2.2] The term 'dice score' should be capitalized as 'Dice score' for consistency with standard usage.
[Section 2.2] The hyperparameters 'lrinit = 1·10−4' and 'lrinit = 1 · 10−3' would be clearer as 'initial learning rate' spelled out, and the multiplication dot should be typeset consistently.
[Figure 1] The figure caption contains a garbled substring ('7x7x7 conv, 64 /23x3x3 conv...') that appears to be a layout artifact; please regenerate the figure and caption so that the architecture is readable.
[Section 3] The claim that performance 'matches state-of-the-art algorithms in the most recent BraTS 2018 challenge' is not accompanied by a direct comparison table or citation to specific challenge results; please clarify the basis of this comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the accuracy claims are direct measurements on held-out data, with no fitted input renamed as a prediction.

full rationale

This paper is an empirical engineering study, not a mathematical derivation, so the standard circularity patterns do not apply. The claimed results are direct measurements on held-out test sets: the segmentation Dice score is measured on 76 BraTS 2018 test patients (Section 3), and the grading accuracy of 91.25% is measured on 80 TCIA test patients that were explicitly excluded from training and validation (Sections 2.3 and 3). The independent evaluation on 112 patients from Ghent University Hospital is likewise a held-out external dataset. There is no equation that defines one reported quantity in terms of another reported quantity, and no fitted parameter is subsequently relabeled as a prediction. The only arguable weakness is the untested assertion that 'small variations between manual and predicted segmentations won't have a strong influence on the tumour ROI' (Section 3), but this is an empirical assumption about robustness, not a circular definition or a self-citation chain. The paper contains no self-citations that bear on the central claim, and the BraTS 2018 benchmark comparison is an independent external reference point. Therefore the derivation chain, such as it is, is self-contained with respect to circularity, and the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

This is an empirical deep learning paper, so no new physical entities or adjusted scientific constants are introduced. The central claim is a performance measurement, not a derived law. The main assumptions are that the MRI sequences carry grading information, that the histology labels are accurate, and that the public and hospital test sets represent the target population. Hyperparameters such as learning rate and ROI size are standard training choices, with the ROI size listed as a hand-chosen free parameter.

free parameters (3)

ROI resize dimension = 112x112x112
The tumor ROI is resized to a fixed cube of 112 voxels per side before classification; this hand-chosen size affects the information available to the classifier.
Initial learning rate (classification network) = 1e-3
Set manually for SGD training of the residual network; affects convergence and final accuracy.
Initial learning rate (segmentation network) = 1e-4
Set manually for ADAM training of the 3D U-Net; affects the segmentation quality.

assumptions (4)

domain assumption MRI intensity patterns in T1ce, T2, and FLAIR sequences are sufficient to discriminate glioblastoma from lower-grade glioma.
The entire classification rests on the premise that these structural sequences carry the grading information; without this, the neural network would have no signal.
domain assumption Histologically confirmed WHO grade labels used for training and evaluation are correct and consistent across the public datasets and the Ghent cohort.
Ground truth comes from histology at each center; any label noise or protocol differences could bias the reported accuracy.
domain assumption The BraTS 2018 test set and the Ghent hospital data are representative of the general clinical population for glioma grading.
The authors generalize from two test sets to a claim of robustness across centers.
domain assumption Co-registration, interpolation to 1 mm3, bias correction, and skull-stripping do not remove or distort the grading-relevant signal.
All scans are preprocessed with SPM12; if preprocessing introduces artifacts, the results would change.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Fully Automatic Binary Glioma Grading based on Pre-Therapy MRI using 3D Convolutional Neural Networks." pith.science (2026). https://pith.science/paper/SJR2VDZH

@misc{pith2026190801506,
  author       = {Pith},
  title        = {Pith review of: Fully Automatic Binary Glioma Grading based on Pre-Therapy MRI using 3D Convolutional Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJR2VDZH}},
  note         = {Machine review of arXiv:1908.01506}
}

read the original abstract

The optimal treatment strategy of newly diagnosed glioma is strongly influenced by tumour malignancy. Manual non-invasive grading based on MRI is not always accurate and biopsies to verify diagnosis negatively impact overall survival. In this paper, we propose a fully automatic 3D computer-aided diagnosis (CAD) system to non-invasively differentiate high-grade glioblastoma from lower-grade glioma. The approach consists of an automatic segmentation step to extract the tumour ROI followed by classification using a 3D convolutional neural network. Segmentation was performed using a 3D U-Net achieving a dice score of 88.53% which matches top performing algorithms in the BraTS 2018 challenge. The classification network was trained and evaluated on a large heterogeneous dataset of 549 patients reaching an accuracy of 91%. Additionally, the CAD system was evaluated on data from the Ghent University Hospital and achieved an accuracy of 92% which shows that the algorithm is robust to data from different centres.

Figures

Figures reproduced from arXiv: 1908.01506 by the authors.

**Figure 1.** Architecture used to classify a tumour ROI as GBM or LGG. Every convolutional layer is succeeded with instance normalisation and a ReLU activation. 3. Results and Discussion The obtained whole tumour dice scores on the 76 BraTS 2018 test patients were 88.53%, 86.38% and 84.62% when providing all four modalities, only T1ce and FLAIR and only T1ce and T2 sequences as input respectively. These scores match the performa… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages

[1]

Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features

Spyridon Bakas, Hamed Akbari, et al. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features . Scientific Data, 4: 0 170117, sep 2017 a . ISSN 2052-4463. doi:10.1038/sdata.2017.117

doi:10.1038/sdata.2017.117 2017
[2]

Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection

Spyridon Bakas, Hamed Akbari, et al. Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection . The Cancer Imaging Archive, 2017 b . doi:10.7937/K9/TCIA.2017.KLXWJJ1Q

doi:10.7937/k9/tcia.2017.klxwjj1q 2017
[3]

The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository

Kenneth Clark, Bruce Vendt, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository . Journal of Digital Imaging, 26 0 (6): 0 1045--1057, dec 2013. ISSN 0897-1889. doi:10.1007/s10278-013-9622-7

doi:10.1007/s10278-013-9622-7 2013
[4]

Data From LGG-1p19qDeletion

Bradley Erickson, Zeynettin Akkus, et al. Data From LGG-1p19qDeletion . The Cancer Imaging Archive, 2017. doi:10.7937/K9/TCIA.2017.dwehtz9v

work page doi:10.7937/k9/tcia.2017.dwehtz9v 2017
[5]

No new-net

Fabian Isensee, Philipp Kickingereder, et al. No new-net. CoRR, abs/1809.10483, 2018

arXiv 2018
[6]

Jansen, Vera Graute, et al

Nathalie L. Jansen, Vera Graute, et al. MRI-suspected low-grade glioma: is there a need to perform dynamic FET PET? European Journal of Nuclear Medicine and Molecular Imaging, 39 0 (6): 0 1021--1029, jun 2012. ISSN 1619-7070. doi:10.1007/s00259-012-2109-9

work page doi:10.1007/s00259-012-2109-9 2012
[7]

Menze, Andras Jakab, et al

Bjoern H. Menze, Andras Jakab, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) . IEEE Transactions on Medical Imaging, 34 0 (10): 0 1993--2024, oct 2015. ISSN 0278-0062. doi:10.1109/TMI.2014.2377694

arXiv 1993
[8]

3d MRI brain tumor segmentation using autoencoder regularization

Andriy Myronenko. 3d MRI brain tumor segmentation using autoencoder regularization. CoRR, abs/1810.11654, 2018

arXiv 2018

Show all 13 references

[9]

Flanders, et al

Nancy Pedano, Adam E. Flanders, et al. Radiology Data from The Cancer Genome Atlas Low Grade Glioma [TCGA-LGG] collection . jan 2016. doi:10.7937/K9/TCIA.2016.L4LTD3TK

2016 doi
[10]

Radiology Data from The Cancer Genome Atlas Glioblastoma Multiforme [TCGA-GBM] collection

Lisa Scarpace, Tom Mikkelsen, and others. Radiology Data from The Cancer Genome Atlas Glioblastoma Multiforme [TCGA-GBM] collection . jan 2016. doi:10.7937/K9/TCIA.2016.RNYFUYE9

2016 doi
[11]

European Association for Neuro-Oncology (EANO) guideline on the diagnosis and treatment of adult astrocytic and oligodendroglial gliomas

Michael Weller, Martin van den Bent, et al. European Association for Neuro-Oncology (EANO) guideline on the diagnosis and treatment of adult astrocytic and oligodendroglial gliomas . The Lancet Oncology, 18 0 (6): 0 e315--e329, jun 2017. ISSN 1470-2045. doi:10.1016/S1470-2045(...

2017 doi
[12]

Does early resection of presumed low-grade glioma improve survival? A clinical perspective

Maarten M J Wijnenga, Tariq Mattni, et al. Does early resection of presumed low-grade glioma improve survival? A clinical perspective. Journal of neuro-oncology, 133 0 (1): 0 137--146, may 2017. ISSN 1573-7373. doi:10.1007/s11060-017-2418-8

2017 doi
[13]

Glioma Grading on Conventional MR Images: A Deep Learning Study With Transfer Learning

Yang Yang, Lin-Feng Yan, et al. Glioma Grading on Conventional MR Images: A Deep Learning Study With Transfer Learning. Frontiers in neuroscience, 12: 0 804, 2018. ISSN 1662-4548. doi:10.3389/fnins.2018.00804

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features

Spyridon Bakas, Hamed Akbari, et al. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features . Scientific Data, 4: 0 170117, sep 2017 a . ISSN 2052-4463. doi:10.1038/sdata.2017.117

doi:10.1038/sdata.2017.117 2017

[2] [2]

Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection

Spyridon Bakas, Hamed Akbari, et al. Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection . The Cancer Imaging Archive, 2017 b . doi:10.7937/K9/TCIA.2017.KLXWJJ1Q

doi:10.7937/k9/tcia.2017.klxwjj1q 2017

[3] [3]

The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository

Kenneth Clark, Bruce Vendt, et al. The Cancer Imaging Archive (TCIA): Maintaining and Operating a Public Information Repository . Journal of Digital Imaging, 26 0 (6): 0 1045--1057, dec 2013. ISSN 0897-1889. doi:10.1007/s10278-013-9622-7

doi:10.1007/s10278-013-9622-7 2013

[4] [4]

Data From LGG-1p19qDeletion

Bradley Erickson, Zeynettin Akkus, et al. Data From LGG-1p19qDeletion . The Cancer Imaging Archive, 2017. doi:10.7937/K9/TCIA.2017.dwehtz9v

work page doi:10.7937/k9/tcia.2017.dwehtz9v 2017

[5] [5]

No new-net

Fabian Isensee, Philipp Kickingereder, et al. No new-net. CoRR, abs/1809.10483, 2018

arXiv 2018

[6] [6]

Jansen, Vera Graute, et al

Nathalie L. Jansen, Vera Graute, et al. MRI-suspected low-grade glioma: is there a need to perform dynamic FET PET? European Journal of Nuclear Medicine and Molecular Imaging, 39 0 (6): 0 1021--1029, jun 2012. ISSN 1619-7070. doi:10.1007/s00259-012-2109-9

work page doi:10.1007/s00259-012-2109-9 2012

[7] [7]

Menze, Andras Jakab, et al

Bjoern H. Menze, Andras Jakab, et al. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) . IEEE Transactions on Medical Imaging, 34 0 (10): 0 1993--2024, oct 2015. ISSN 0278-0062. doi:10.1109/TMI.2014.2377694

arXiv 1993

[8] [8]

3d MRI brain tumor segmentation using autoencoder regularization

Andriy Myronenko. 3d MRI brain tumor segmentation using autoencoder regularization. CoRR, abs/1810.11654, 2018

arXiv 2018