Pith. sign in

REVIEW 4 major objections 5 minor 12 references

Predicting Brain Tumor Response to Therapy using a Hybrid Deep Learning and Radiomics Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that predicting glioblastoma therapy response from longitudinal MRI works best when deep learning features and handcrafted radiomics are fused, reaching a mean ROC AUC of 0.81 in four-class RANO response prediction.

desk verdict The longitudinal radiomics (growth/shrinkage masks, centroid shift) are genuinely new and the evaluation setup is sound, but the paper's central claim that fusion beats single-feature sources is not statistically supported—the gaps are within one standard deviation over only 5 folds. read the letter →

arxiv 2509.06511 v1 pith:EZ3QWYEZ submitted 2025-09-08 cs.CV

Daniil Tikhonov , Matheus Scatolin , Mohor Banerjee , Qiankun Ji , Ahmed Jaheen , Mostafa Salem , Abdelrahman Elsayed , Hu Wang
show 2 more authors
Sarim Hashmi Mohammad Yaqub
This is my paper
classification cs.CV
keywords glioblastomatreatmentresponsepredictionRANOcriteriaradiomicsdeeplearningfeaturefusionlongitudinalMRICatBoost
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

The paper tries to establish that automated RANO response classification, into Complete Response, Partial Response, Stable Disease, and Progressive Disease, is improved by combining learned deep features with a large set of handcrafted radiomics features. It argues that neither deep learning nor radiomics alone is sufficient; the fusion of both, fed into a CatBoost classifier, yields the strongest predictive performance. This matters because manual RANO assessment is time-consuming and observer-dependent, and an automated, objective tool could standardize treatment-response evaluation in neuro-oncology. The evidence comes from a patient-wise 5-fold cross-validation on the LUMIERE dataset, where the hybrid model reaches a mean ROC AUC of 0.81, Macro F1 of 0.50, and accuracy of 0.72, outperforming every single feature source. The key clinical promise is that dynamic tumor changes captured by growth and shrinkage masks, together with deep image representations, give a more reliable signal than volume measurements alone.

What carries the argument

The central object is the growth and shrinkage mask pair: the follow-up-only tumor region and the baseline-only tumor region derived from automatic segmentations. Radiomic features computed separately on these masks make dynamic tumor change explicit, and when combined with centroid shift and compartment volume changes, they carry the temporal signal that volumes alone miss. The second load-bearing component is feature fusion itself: concatenating deep features from four modality-specific ResNet-18 encoders with the full radiomics suite and feeding the combined vector to gradient boosting (CatBoost) is what produces the reported performance gain.

What would settle it

Recompute the radiomics and hybrid features on the same LUMIERE cases using two independent segmentations, for example expert manual masks and a different automatic tool, and compare the patient-wise 5-fold ROC AUC. If the fused model's AUC drops to the deep-only level of 0.74 or shifts more than the reported plus or minus 0.08 when masks change, the claimed hybrid gain would be shown to depend on the specific segmentation rather than on true tumor changes.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid feature vector, built by concatenating four fine-tuned ResNet-18 encoders (one per MRI modality) with more than 4,800 radiomic features computed on original, growth, and shrinkage tumor masks plus engineered temporal features, lets a CatBoost classifier predict the four RANO response classes with a mean ROC AUC of 0.81, a Macro F1 of 0.50, and an accuracy of 0.72. The ablation study shows incremental gains: volumes alone score 0.59 AUC, adding full radiomics raises it to 0.70, growth/shrinkage masks bring it to 0.73, 2D radiomics and centroid shift to 0.78, and the deep-only encoder reaches 0.74, while the fused model reaches 0.81. The paper interprets this

Load-bearing premise

The entire feature set, all 4,896 radiomic values and the engineered volume and centroid features, is only as good as the automatic tumor masks they are computed from, and the paper does not measure how segmentation errors affect prediction.

Editorial extensions

If this is right

  • Automated RANO classification becomes feasible without manual feature selection, achieving 0.81 AUC on a standardized challenge dataset.
  • Explicitly modeling newly grown and shrunken tumor regions contributes more predictive value than tracking overall volume, with AUC improving from 0.59 to 0.73 when such masks are added.
  • Deep features and handcrafted radiomics are complementary: the deep-only model reaches 0.74 AUC while the fused model reaches 0.81, so neither feature family saturates the available information.
  • The pipeline's clinical utility depends on segmentation quality and cross-institutional validation, which the paper identifies as open limitations.
  • The computational cost of preprocessing and radiomics extraction, roughly 5 minutes and 40 seconds per case, currently limits real-time clinical deployment.

Reading between the lines

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

  • A natural next test the paper does not run: feature selection on the roughly 5,000 features could reveal whether most predictive power concentrates in a small subset of growth and shrinkage texture features, potentially cutting computational cost sharply.
  • The 2D largest-slice ResNet-18 encoder discards most of the volume; a 3D or transformer-based longitudinal encoder might capture the same temporal and spatial information that the engineered 3D radiomics currently provide, possibly making handcrafted features redundant.
  • The same fusion recipe could transfer to other oncology response criteria, such as RECIST for solid tumors, wherever two-time-point masks are available.
  • Because the masks are assumed given, the hybrid gain may partly reflect segmentation-quality artifacts; a segmentation-quality-aware version or uncertainty estimation would test whether the reported advantage persists under imperfect input.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes an automated pipeline for predicting 4-class RANO treatment response from longitudinal glioblastoma MRI pairs in the BraTS 2025 / LUMIERE setting. The method preprocesses registered, bias-corrected, intensity-normalized scans; extracts deep features from the largest-tumor-area 2D axial slice for each of four MRI modalities using fine-tuned ResNet-18 encoders; computes 4,896 PyRadiomics features from original, growth, and shrinkage masks at both timepoints; adds engineered volumetric/centroid-shift/time-gap features; concatenates all features into a single vector and trains CatBoost. Evaluation is patient-wise stratified 5-fold cross-validation over 616 scan-pairs, with macro F1, ROC AUC, and accuracy. Ablations (Table 1) show the fused model best (F1 0.50±0.08, AUC 0.81±0.08, accuracy 0.72±0.05), and the paper concludes that fusing deep and radiomics features is synergistic and robust.

Significance. If the results are valid, the paper provides a useful engineering benchmark for automated RANO response prediction and demonstrates a plausible fusion architecture. Strengths include use of a public dataset, patient-wise and stratified splits, a stepwise ablation, and transparent runtime reporting. However, the central claim of synergistic superiority rests on differences that are within one standard deviation across only five folds, and the high-dimensional feature set with unspecified feature selection and nested evaluation leaves room for overfitting. The segmentation-dependence and lack of sensitivity analysis further temper the 'robust' characterization. As a challenge-style methods paper the contribution is of interest, but the statistical and evaluation-integrity questions need to be resolved before the central claim can be accepted.

major comments (4)
  1. [§3, Table 1] The paper's central claim that 'integrating learned deep features with domain-specific handcrafted features provides the most robust and accurate framework' is supported only by small, noisy differences. Approach 6 reaches AUC 0.81±0.08 vs. 0.78±0.06 for Approach 4 and Macro F1 0.50±0.08 vs. 0.45±0.08 for Approach 3. With n=5 folds, these differences are within one standard deviation, and no paired significance test, bootstrap confidence interval, or per-fold table is provided. Please add paired tests (e.g., Wilcoxon signed-rank across folds or patient-level bootstrap), report confidence intervals and effect sizes, and revise the 'clearly demonstrate' language in §3 and the abstract if significance is not established.
  2. [§2.2, §2.3] The evaluation does not describe how the ResNet-18 encoders were fine-tuned relative to the cross-validation folds, nor how CatBoost and its hyperparameters were selected ('through our experimentation'). If encoder fine-tuning or classifier selection used the same folds that are later used for evaluation, the reported numbers are optimistically biased. In addition, the final feature vector contains 4,896 radiomics features plus deep features for only 616 scan-pairs with no feature selection or dimensionality reduction. Please specify the exact training/validation protocol for every fitted component, run nested or repeated cross-validation, and add feature selection or a regularization analysis (e.g., number of selected features vs. performance).
  3. [§2.1] The preprocessing description says 'histogram matching to standardize image intensities across all patient scans.' Please clarify whether the reference histogram is derived from the full dataset including the validation folds. If so, this is a form of test-data leakage and would invalidate the cross-validation estimates; if not, specify the reference source. This point is load-bearing for the validity of all reported metrics.
  4. [§5] The conclusion concedes that 'reliance on pre-defined tumor masks may introduce variability due to segmentation quality,' but the paper does not quantify segmentation accuracy or test sensitivity to mask errors. Since all 4,896 radiomics features and the growth/shrinkage masks are derived from automatic segmentations, the 'robust' claim in §4/§5 is not yet supported. Please add a sensitivity analysis (e.g., comparing the two segmentation tools, or perturbing masks) or weaken the robustness claim to 'robust with respect to the provided segmentations.'
minor comments (5)
  1. [Table 1] 'CV (ResNet-18)' is not defined; if it means concatenated deep feature vector across modalities/timepoints, spell that out. Also clarify whether the same ResNet features are used in Approach 5 and Approach 6.
  2. [§2.2] Specify the dimensionality of the deep feature vector and how the four modality features are combined. The text says four independent ResNet-18 models but does not state whether their outputs are concatenated, averaged, or otherwise fused.
  3. [§2.2] Clarify why weighted random sampling is used alongside Focal Loss without class-specific alpha; state whether both are applied and how the inverse class-frequency weights are computed.
  4. [References] Reference [9] (MedPerf) appears in the bibliography but is not cited in the text; either cite it where federated benchmarking is relevant or remove it.
  5. [Fig. 2] The text says 'as shown in Figure 2' but the figure caption does not specify the dimensions of the final feature vector; consider making the vector lengths and the concatenation order explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: features are derived from imaging and labels are external expert RANO annotations; no equation or fitted parameter reduces to the prediction.

full rationale

The paper's derivation chain is an empirical supervised pipeline: it extracts deep features (via fine-tuned ResNet-18 encoders) and handcrafted radiomic/volumetric/engineered features from MRI, concatenates them, and trains a CatBoost classifier to predict expert-annotated RANO response labels. The labels come from the external LUMIERE dataset, not from the paper's own feature computation. The features are not fitted to the labels by construction, and no equation in the paper defines a feature in terms of the outcome or defines the outcome in terms of a fitted parameter. The ablation study (Table 1) compares different feature sets under the same cross-validation scheme, so the final model is an empirical comparison rather than a tautological derivation. The only conceptual overlap is that RANO labels are themselves based on imaging criteria and some features (volumetric changes, growth/shrinkage masks) are correlated with those criteria; however, the paper does not compute the labels from the features, and the volume-only baseline AUC of 0.59 confirms that the mapping is not a trivial identity. Model-selection optimism from choosing CatBoost on the same folds is a statistical validity concern, not circular reasoning. There is no self-citation load-bearing, no imported uniqueness theorem, and no ansatz smuggled via citation. Therefore, no significant circularity.

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

The central claim rests primarily on the LUMIERE labels and automatic segmentations, on the assumption that a 2D largest slice is sufficient, and on various undisclosed hyperparameter choices. No new physical or conceptual entities are introduced.

free parameters (4)
  • ResNet-18 fine-tuning schedule = LR warm-up to 1e-4, cosine decay to 1e-6, 150 epochs
    Chosen by hand to optimize validation performance; no sensitivity analysis is reported, so the central result could depend on this schedule.
  • Focal Loss gamma = 2.0
    Set to 2.0 with no class alpha; this choice affects how hard examples are weighted and therefore classifier behavior.
  • ROI crop size = 128x128 pixels
    The largest tumor slice is cropped to 128x128; alternative crop sizes or multi-slice strategies are not tested.
  • CatBoost hyperparameters = not reported
    CatBoost was selected as best among five classifiers, but its hyperparameters are not disclosed, making the reported metrics dependent on undisclosed tuning.
assumptions (4)
  • domain assumption LUMIERE expert RANO annotations are reliable ground truth.
    The model is trained and evaluated against these labels; if expert labels contain significant observer variability, reported AUC/F1 are against a noisy target (Section 2.1 and Section 3).
  • domain assumption Provided automatic tumor segmentations are accurate enough for radiomics and mask computation.
    All radiomic, volume, growth, and shrinkage features are computed from these masks; no segmentation quality control is reported (Section 2.1 and Section 5).
  • ad hoc to paper The single axial slice with the largest tumor area contains sufficient information for response prediction.
    This 2D cropping choice discards 3D context; the paper does not compare with full-volume input or multi-slice selection (Section 2.2).
  • domain assumption Preprocessing, including MNI registration, N4 bias correction, and histogram matching, preserves clinically relevant signal.
    Any distortion from registration and intensity standardization could affect radiomics; no validation is provided that preprocessing does not remove signal (Section 2.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Brain Tumor Response to Therapy using a Hybrid Deep Learning and Radiomics Approach." pith.science (2026). https://pith.science/paper/EZ3QWYEZ

@misc{pith2026250906511,
  author       = {Pith},
  title        = {Pith review of: Predicting Brain Tumor Response to Therapy using a Hybrid Deep Learning and Radiomics Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EZ3QWYEZ}},
  note         = {Machine review of arXiv:2509.06511}
}
read the original abstract

Accurate evaluation of the response of glioblastoma to therapy is crucial for clinical decision-making and patient management. The Response Assessment in Neuro-Oncology (RANO) criteria provide a standardized framework to assess patients' clinical response, but their application can be complex and subject to observer variability. This paper presents an automated method for classifying the intervention response from longitudinal MRI scans, developed to predict tumor response during therapy as part of the BraTS 2025 challenge. We propose a novel hybrid framework that combines deep learning derived feature extraction and an extensive set of radiomics and clinically chosen features. Our approach utilizes a fine-tuned ResNet-18 model to extract features from 2D regions of interest across four MRI modalities. These deep features are then fused with a rich set of more than 4800 radiomic and clinically driven features, including 3D radiomics of tumor growth and shrinkage masks, volumetric changes relative to the nadir, and tumor centroid shift. Using the fused feature set, a CatBoost classifier achieves a mean ROC AUC of 0.81 and a Macro F1 score of 0.50 in the 4-class response prediction task (Complete Response, Partial Response, Stable Disease, Progressive Disease). Our results highlight that synergizing learned image representations with domain-targeted radiomic features provides a robust and effective solution for automated treatment response assessment in neuro-oncology.

Figures

Figures reproduced from arXiv: 2509.06511 by the authors.

Figure 1
Figure 1. Overview of the preprocessing pipeline applied to the raw MRI data. The central diagram represents the sequential transformation of the original scans, aiming to ensure consistency across all samples. The four surrounding boxes illustrate key normalization steps: (1) spatial registration to MNI152 space, (2) N4 bias field correction, (3) inten￾sity normalization via histogram matching and (4) z-scoring. These steps … view at source ↗
Figure 2
Figure 2. Illustration of how The final feature vector was created and used as input to CatBoost. 3 Results The performance of our proposed methodology was evaluated through a series of experiments designed to assess the contribution of each major component of our feature set. The results, averaged over the 5-fold patient-wise cross-validation, are presented in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the dataset’s label distribution and the stratified 5-fold patient￾wise splitting strategy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 5 canonical work pages

  1. [1]

    The Lancet Oncology.20(5), 728–740 (2019)

    Kickingereder, P., Isensee, F., Tursunova, I., et al.: Automated quantitative tu- mor response assessment of MRI in neuro-oncology with artificial neural networks: a multicentre, retrospective study. The Lancet Oncology.20(5), 728–740 (2019). https://doi.org/10.1016/S1470-2045(19)30098-1

    doi:10.1016/s1470-2045(19)30098-1 2019

  2. [2]

    Scientific Data.9(1), 768 (2022)

    Suter, Y., Knecht, U., Valenzuela, W., et al.: The LUMIERE dataset: longitudinal glioblastoma MRI with expert RANO evaluation. Scientific Data.9(1), 768 (2022). https://doi.org/10.1038/s41597-022-01881-7

    doi:10.1038/s41597-022-01881-7 2022

  3. [3]

    Journal of Clinical Oncology.28(11), 1963–1972 (2010)

    Wen, P.Y., Macdonald, D.R., Reardon, D.A., et al.: Updated response assess- ment criteria for high-grade gliomas: Response Assessment in Neuro-Oncology (RANO) working group. Journal of Clinical Oncology.28(11), 1963–1972 (2010). https://doi.org/10.1200/JCO.2009.26.3541

    work page doi:10.1200/jco.2009.26.3541 1963

  4. [4]

    NeuroImage.54(1), 313–327 (2011)

    Fonov, V., Evans, A.C., Botteron, K., Almli, C.R., McKinstry, R.C., Collins, D.L.: Unbiased average age-appropriate atlases for pediatric studies. NeuroImage.54(1), 313–327 (2011). https://doi.org/10.1016/j.neuroimage.2010.07.033

    doi:10.1016/j.neuroimage.2010.07.033 2011

  5. [5]

    Tustison, Nicholas & Avants, Brian & Cook, Philip & Gee, James. (2010). N4ITK: Improved N3 bias correction with robust B-spline approximation. 708-711. 10.1109/ISBI.2010.5490078

    work page arXiv 2010

  6. [6]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR)

    He, K., Zhang, X., Ren, S., Sun, J.: Deep Residual Learning for Image Recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR). pp. 770–778 (2016)

    work page 2016

  7. [7]

    In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R

    Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A.V., Gulin, A.: Cat- Boost: unbiased boosting with categorical features. In: Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., Garnett, R. (eds.) Advances in Neural Information Processing Systems, vol. 31. Curran Associates, Inc. (2018). https://proceedings.neurips.cc/paper_files/...

    work page 2018

  8. [8]

    https://doi.org/10.1158/0008-5472.CAN-17-0339

    van Griethuysen, J.J.M., Fedorov, A., Parmar, C., et al.: Computational Radiomics SystemtoDecodetheRadiographicPhenotype.CancerResearch.77(21),e104–e107 (2017). https://doi.org/10.1158/0008-5472.CAN-17-0339

    doi:10.1158/0008-5472.can-17-0339 2017

Show all 12 references
  1. [9]

    https://doi.org/10.1038/s42256-023-00652-2

    Karargyris, A., Umeton, R., Sheller, M.J., Aristizabal, A., George, J., Wuest, A., Pati, S., et al.: Federated benchmarking of medical artificial intelligence with MedPerf.Nature Machine Intelligence5, 799–810 (2023). https://doi.org/10.1038/s42256-023-00652-2

    2023 doi

  2. [10]

    Human Brain Mapping

    Isensee, F., Schell, M., Tursunova, I., Brugnara, G., Bonekamp, D., Neuberger, U., Wick, A., Schlemmer, H.P., Heiland, S., Wick, W., Bendszus, M., Maier- Hein, K.H., Kickingereder, P.: Automated brain extraction of multi-sequence MRI using artificial neural networks. Human Bra...

    2019 doi

  3. [11]

    arXiv preprint arXiv:1904.08128 (2019)

    Isensee, F., Petersen, J., Kohl, S.A.A., Jaeger, P.F., Maier-Hein, K.H.: nnU-Net: Breaking the Spell on Successful Medical Image Segmentation. arXiv preprint arXiv:1904.08128 (2019). https://arxiv.org/abs/1904.08128

    1904 arXiv

  4. [12]

    Neuro-Oncology Advances.5(1), vdad089 (2023)

    Grün, D.A., Kerl, H.U., Bette, S., Weng, A.M., Sommer, W.H., Meyer, B., Zimmer, C., Pyka, T., Unterrainer, M.: DeepBraTumIA: Automatic segmentation of glioma and brain tissues using deep learning in routine clinical workflows. Neuro-Oncology Advances.5(1), vdad089 (2023). http...

    2023 doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.