Pith. sign in

REVIEW 3 major objections 5 minor 9 references

A Voxel-Wise Uncertainty-Guided Framework for Glioma Segmentation Using Spherical Projection-Based U-Net and Localized Refinement in Multi-Parametric MRI

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Targeted 3D refinement, pointed at voxels where spherical projections disagree most, lifts glioma segmentation Dice past both 2D and 3D baselines, with the largest gains in enhancing tumor and tumor core.

desk verdict The central superiority claim is invalid because fusion weights were tuned directly on the test set (B.3) contradicting B.4's training-only statement; the framework idea is coherent, but the paper needs a major fix before the numbers can be believed. read the letter →

arxiv 2507.15622 v1 pith:7TXKGXS4 submitted 2025-07-21 physics.med-ph

Zhenyu Yang , Chen Yang , Rihui Zhang , Minbin Chen , Chunhao Wang , Fang-Fang Yin This is my paper
classification physics.med-ph
keywords gliomasegmentationuncertaintyquantificationsphericalprojectionmulti-parametricMRI3Drefinementkernel-basedregionselectionparticleswarmoptimizationnnU-Net
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 is trying to establish that voxel-wise uncertainty, measured as disagreement among 1,024 spherical projections of every 2D MRI slice, can be used as a routing signal for where to apply 3D segmentation refinement. A 2D U-Net first segments the whole brain and produces an entropy map; a cubic sliding-window search selects the subvolumes with the highest cumulative entropy; a 3D U-Net refines only those subvolumes; and a sigmoid-weighted fusion, tuned by particle swarm optimization, combines the 2D and 3D probability maps. On 69 independent test cases from a 369-patient benchmark dataset, the fused model tops both standalone 2D and standalone 3D models on Dice for all three tumor targets, with the largest gains in enhancing tumor (0.8124 vs 0.7527 and 0.6530) and tumor core (0.7499 vs 0.7002 and 0.7027). If true, this matters because uncertainty stops being a retrospective scorecard and becomes a practical mechanism for spending computation where segmentation is hardest.

What carries the argument

The central object is the voxel-wise uncertainty map from spherical projection. Each 2D axial slice is deformed onto 1,024 virtual spherical surfaces with distinct origins; a shared 2D nnU-Net (a self-configuring U-Net variant) segments every projected view, and Shannon entropy across the 1,024 probability histograms scores how much the predictions disagree at each voxel. That entropy map is the triage mechanism: a cubic sliding-window kernel with hand-set size (32 mm for enhancing tumor and tumor core, 64 mm for whole tumor) sums uncertainty inside each cube, ranks cubes, and discards candidates whose overlap with already-selected cubes exceeds 40%. The 3D nnU-Net then refines only the chosen subvolumes, and Equation (2) fuses the 2D and 3D probabilities with a sigmoid-weighted sum whose weights are optimized by particle swarm optimization. The whole scheme is a resource-allocation rule: entropy decides where 3D context is spent, and 2D global context is kept everywhere else.

What would settle it

Compute, on held-out cases, the overlap (Dice or recall) between the high-entropy kernel regions and the voxels where the 2D model's prediction disagrees with the reference segmentation. If that overlap is no better than chance, the claim that uncertainty localizes the errors refined by the 3D model is falsified; the Dice gains would then have to come from averaging or fusion rather than from targeted refinement.

Watch

Extended reading notes

Core claim

The central claim is that localizing 3D refinement by voxel-wise uncertainty improves brain-tumor subregion segmentation more than either pure slice-wise 2D or full-volume 3D segmentation, while keeping inference cheap. The uncertainty maps are produced by projecting each axial slice onto 1,024 spherical surfaces with different projection origins, passing every projected view through a shared 2D nnU-Net (a self-configuring U-Net variant), averaging the predictions for the mask, and computing Shannon entropy over the discretized prediction histogram at each voxel. A kernel-based search then gathers the highest-entropy 3D subvolumes, a 3D nnU-Net refines them, and the final output is a sigmoid-weighted voxel-wise combination of the 2D and 3D probabilities, with weights optimized by particle swarm optimization. The paper reports statistically significant Dice gains mainly for enhancing tumor and tumor core, where 2D models tend to miss boundaries and fragmented lesions and 3D patch models tend to smooth them over.

Load-bearing premise

The entire method works only if the entropy map computed from 1,024 spherical projections reliably points to the voxels where the 2D model is actually wrong, and if the manually chosen refinement cube sizes (32 mm for enhancing tumor and tumor core, 64 mm for whole tumor) with a 40% overlap limit capture those wrong regions; if a boundary error falls outside a selected cube, the 3D network never sees it and fusion cannot fix it.

Editorial extensions

If this is right

  • The enhancing-tumor Dice gain (0.8124 vs 0.7527 for 2D and 0.6530 for 3D) indicates the uncertainty-guided fusion recovers boundary and fragmented-lesion errors that single-model pipelines miss.
  • The tumor-core sensitivity gain (0.7700 vs 0.6481 for 2D and 0.6914 for 3D) indicates the refinement adds true-positive voxels rather than merely reweighting existing predictions.
  • The whole-tumor gain is small (0.9055 vs 0.9038 for 3D), and the optimized fusion weights favor 3D for whole tumor (0.6796) but 2D for enhancing tumor and tumor core (0.5848 and 0.6980), so the method adapts its trust by subregion.
  • Because the full pipeline runs in under four minutes per patient at inference, the targeted refinement buys accuracy without the cost of full-volume 3D inference.

Reading between the lines

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

  • The paper does not report a direct overlap measurement between high-entropy kernels and voxels where the 2D model is actually wrong, so a natural next experiment is to compute the recall of selected kernels against true 2D error voxels; that would isolate whether the entropy map or the fusion weights are doing the work.
  • The hand-set kernel sizes (32 mm for enhancing tumor and tumor core, 64 mm for whole tumor) and the 40% overlap cap are transfer risks; replacing them with data-driven selection based on uncertainty-map statistics would make the triage adaptive across datasets.
  • Whole tumor benefits little from refinement even though 3D receives the larger fusion weight there, suggesting that for whole tumor the fusion is mostly averaging two good predictions; an ablation that disables 3D refinement for whole tumor would test whether the refinement stage earns its cost in that subregion.
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

3 major / 5 minor

Summary. The manuscript proposes a three-stage glioma segmentation framework on BraTS2020: a 2D nnU-Net with spherical-projection deformation produces both segmentations and voxel-wise Shannon-entropy uncertainty maps; a kernel-based search selects high-uncertainty 3D sub-volumes; a 3D nnU-Net refines those sub-volumes; and a sigmoid-weighted fusion combines the 2D and 3D predictions with weights optimized by Particle Swarm Optimization. The authors report Dice scores of 0.8124 (ET), 0.7499 (TC), and 0.9055 (WT), claiming statistically significant improvements over standalone 2D and 3D baselines, and argue that the uncertainty-guided localization of 3D refinement improves both accuracy and interpretability.

Significance. If the results were valid, the conceptual contribution would be of interest to the medical image segmentation community: the idea of using uncertainty maps to switch from 2D to 3D processing only in ambiguous sub-volumes is a plausible route to combining efficiency with volumetric context, and the modular pipeline is easy to understand and potentially applicable to other modalities. However, the paper's central quantitative claim is compromised by an unresolved contradiction in the experimental protocol (Section B.3 vs. B.4), so the reported superiority over the baselines is not currently established. The paper does not provide code, pretrained models, or a reproducible training/evaluation script, which further limits verification. The strengths are the clarity of the overall architecture and the explicit statement of several limitations in the Discussion, but these do not overcome the protocol problem.

major comments (3)
  1. [B.3 and B.4] These two sections directly contradict each other. Section B.3 states that the PSO objective was 'to maximize the Dice Similarity Coefficient (DSC) across the target tumor subregions (ET, TC, and WT) on the test set,' while Section B.4 states that 'all components of the proposed framework—including ... PSO algorithm execution—were strictly confined to the training data to prevent data leakage.' If PSO tuned the fusion weights on the same 69 test cases later reported in Table I, then the reported Dice gains over the untuned baselines are inflated by selection bias, and the Wilcoxon signed-rank tests do not provide valid evidence of superiority. This is a load-bearing issue for the main claim; the authors must either remove the test-set tuning statement and disclose the exact validation protocol, or rerun the optimization on a training/validation split and re-evaluate on a genuinely held-out test set.
  2. [B.3, Eq. (2), and Table II] The PSO setup is under-specified. The manuscript does not report the number of particles, the number of iterations, the search bounds, or whether the bias term b in Eq. (2) is also optimized. Table II lists only w_2d and w_3d, yet the fusion equation includes b. Without these details, it is impossible to assess the size of the parameter space and the potential for overfitting, even if the protocol were corrected to use a validation split.
  3. [B.2 and Results] The paper does not include an ablation that isolates the contribution of uncertainty-guided kernel selection. The kernel size (d=32 mm for ET/TC, 64 mm for WT), the 40% overlap constraint, and the stopping threshold are all hand-set, and no comparison is made against random or uniform selection of sub-volumes of the same size. As a result, the reported improvements could be attributed to the additional 3D network alone rather than to the uncertainty localization. The authors acknowledge the heuristic nature in the Discussion, but the central 'uncertainty-guided' claim requires such an ablation.
minor comments (5)
  1. [Eq. (1)] Equation (1) contains corrupted notation, including '\(\ln0 f\)' and unclear sub/superscripts; please rewrite the entropy formula cleanly so that the binning and frequency terms are unambiguous.
  2. [Table I] In the ET row, '09991±0.0010' appears to be missing a decimal point and should likely be '0.9991±0.0010'. Please correct this typo and re-check the corresponding value in the text.
  3. [B.3, after Eq. (2)] The notation switch between \(P_{3D\_refined}\) and \(P_{3D\_kernel}\) is confusing; define both symbols before first use and state explicitly how multiple overlapping kernel predictions are averaged.
  4. [Section C and Table I] The statistical comparison is described only as a Wilcoxon signed-rank test with significance level 0.05, but the table marks significance with '*' for many metrics. Please state whether a multiple-comparison correction was applied, and consider reporting effect sizes or confidence intervals for the main Dice comparisons.
  5. [Figure 1] Figure 1 would be clearer if the color scheme for ET, TC, and WT were defined in the caption, since the relationship between the three targets is central to the paper.

Circularity Check

1 steps flagged · score 7.0 of 10

As written, the reported Dice gains are partly fitted: Section B.3 optimizes the fusion weights by PSO to maximize Dice on the same test set later reported in Table I, contradicting B.4's training-only claim.

  1. fitted input called prediction [Section B.3 (fusion weights, Eq. 2), Section B.4 (training-only claim), and Table I (headline Dice results)]
    "The optimization objective was to maximize the Dice Similarity Coefficient (DSC) across the target tumor subregions (ET, TC, and WT) on the test set. ... All components of the proposed framework—including model training, identification of high-uncertainty regions, development of 2D and 3D nnU-Net models, and PSO algorithm execution—were strictly confined to the training data to prevent data leakage."

    The free fusion parameters (w2d, w3d, and possibly b) in Eq. (2) are selected by PSO to maximize DSC on the same 69-case test set whose DSC values are then reported in Table I. The headline numbers 0.8124/0.7499/0.9055 are therefore the value of the fitted objective, not an independent out-of-sample measurement. The 2D and 3D baselines received no equivalent test-set weight tuning, so the claimed superiority and the Wilcoxon comparisons are not valid evidence: an optimized score is being compared against unoptimized scores. The B.4 sentence asserting that PSO was strictly confined to training data directly contradicts the B.3 objective sentence, so as written the evaluation is contaminated.

full rationale

The central quantitative claim is the test-set Dice superiority of the proposed framework. Section B.3 states explicitly that PSO maximizes DSC on the test set, making the fusion weights fitted to the primary outcome variable. Consequently, Table I's 'Ours' column is an in-sample optimized value, whereas the baseline columns are fixed models, so the reported improvements and the associated Wilcoxon tests are statistically forced rather than independent. The contradiction between B.3 ('on the test set') and B.4 ('strictly confined to the training data') means that the paper cannot support an unbiased evaluation as written. This is the clearest and most load-bearing circular step. The spherical-projection uncertainty localizer is inherited from the authors' prior work (references 12 and 13) and is validated only qualitatively in this paper; that is an empirical assumption and a self-citation risk, but not a direct equation-level circularity. Similarly, the kernel sizes chosen from BraTS-wide tumor-volume statistics may include test cases, which is a leakage concern rather than a definitional circularity. Overall the pipeline is not tautological, because the 2D and 3D networks are trained on the training split and the fusion output depends on their predictions, so a score of 8 or 10 would be too high; however, the main numerical evidence reduces in part to a test-set-fitted objective, warranting a score of 7.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework rests on several hand-set parameters (number of projections, kernel size, overlap ratio, stopping threshold) and on the fitted test-set fusion weights. No new physical entities are introduced. The most consequential ledger item is the PSO-fitted weights, which make the reported Dice a fitted quantity.

free parameters (5)
  • Fusion weights w_2d, w_3d per target = ET: 0.5848, 0.4152; TC: 0.6980, 0.3020; WT: 0.3204, 0.6796
    Tuned with PSO to maximize test-set Dice (Section B.3, Table II). These are the main parameters that make the reported Dice a fitted result.
  • Kernel size d = 32 mm for ET/TC, 64 mm for WT
    Chosen empirically from average tumor volumes (Section B.2); affects which subvolumes get 3D refinement.
  • Kernel overlap constraint = 40% of kernel volume
    Set empirically and illustrated in Figure 3; balances redundancy and coverage of high-uncertainty regions.
  • Kernel selection stopping threshold = Unspecified 'predefined minimum threshold'
    Controls how many high-uncertainty kernels are selected; no quantitative value is given in Section B.2.
  • Spherical projection grid and entropy binning = 1024 origins at 8-pixel intervals; 100 bins
    Hand-set choices in Section B.1 that define the uncertainty map's fidelity.
assumptions (4)
  • domain assumption Shannon entropy of discretized prediction probabilities across spherical projections reflects true segmentation error
    Used to identify high-uncertainty subvolumes for refinement; if the entropy map does not localize errors, the 3D refinement is wasted. Mentioned in Section B.1 and visualized in Figures 4-6.
  • domain assumption BraTS2020 manual segmentations are accurate ground truth
    Standard dataset assumption; all metrics are computed against these labels.
  • domain assumption A single random 8:2 split is representative
    No cross-validation; reported confidence assumes one split reflects the population.
  • ad hoc to paper The fusion weights fitted on the test set do not bias evaluation
    This assumption is contradicted by the text itself: PSO maximizes test Dice, so it does bias evaluation. It is a load-bearing premise of the reported improvement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Voxel-Wise Uncertainty-Guided Framework for Glioma Segmentation Using Spherical Projection-Based U-Net and Localized Refinement in Multi-Parametric MRI." pith.science (2026). https://pith.science/paper/7TXKGXS4

@misc{pith2026250715622,
  author       = {Pith},
  title        = {Pith review of: A Voxel-Wise Uncertainty-Guided Framework for Glioma Segmentation Using Spherical Projection-Based U-Net and Localized Refinement in Multi-Parametric MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7TXKGXS4}},
  note         = {Machine review of arXiv:2507.15622}
}
read the original abstract

Purpose: Accurate segmentation of glioma subregions in multi-parametric MRI (MP-MRI) is essential for diagnosis and treatment planning but remains challenging due to tumor heterogeneity and ambiguous boundaries. This study proposes an uncertainty-guided hybrid framework integrating spherical projection-based 2D modeling with targeted 3D refinement to enhance segmentation accuracy and interpretability. Methods: Using the BraTS2020 dataset (369 patients, four-modality MP-MRI), three 2D U-Nets were trained to segment enhancing tumor (ET), tumor core (TC), and whole tumor (WT). Voxel-wise uncertainty was quantified via a spherical projection-based 2D nnU-Net, capturing prediction variance across deformed inputs. A 3D sliding window was used to identify high-uncertainty regions, which were refined using a dedicated 3D nnU-Net. Final outputs combined 2D and 3D predictions through a weighted fusion optimized via Particle Swarm Optimization. Results: The proposed method outperformed standalone 2D and 3D baselines, achieving Dice scores of 0.8124 (ET), 0.7499 (TC), and 0.9055 (WT), with consistent gains in sensitivity and visual coherence. Conclusion: This work presents a novel uncertainty-aware segmentation strategy that adaptively integrates 2D and 3D modeling. By focusing refinement on ambiguous regions, it improves both efficiency and accuracy, offering broad applicability to precision neuro-oncology and other high-stakes medical imaging tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

9 extracted references · 8 canonical work pages

  1. [1]

    Introduction Gliomas represent one of the most aggressive and prevalent forms of primary malignant brain tumors in adults, originating from glial cells within the central nervous system1,2. Glioblastoma multiforme (GBM) is one of its aggressive subtypes with heterogeneous morphology, highly invasive growth patterns, and significant resistance to standard ...

    work page

  2. [2]

    (𝑖,𝑗) is derived by averaging these 1024 predictions, with a mean probability value exceeding 0.5 binarized as the target volume. Segmentation uncertainty 𝑈!

    Materials and Methods A. Patient Data The study employed a total of 369 glioma patients from Brain Tumor Segmentation Challenge 2020 (BraTS 2020) dataset28–30. For each patient, 4 MR images were acquired as an MP-MRI protocol, including T1-weighted, T1-weighted contrast-enhanced (T1ce), T2-weighted, and fluid-attenuated inversion recovery (FLAIR). All MR ...

    work page 2020

  3. [3]

    ,𝑤)"), subject to the non-negativity constraint 𝑤!

    For voxels covered by exactly one 3D kernel, 𝑃)0_23453-67.(𝑖,𝑗,𝑘) is replaced by the single 𝑃)0_83298(𝑖,𝑗,𝑘) prediction from that specific kernel, and fusion Equation (2) is applied. 3. In regions with multiple overlapping 3D kernels, multiple 𝑃)0_83298(𝑖,𝑗,𝑘) predictions are available for a given voxel. These are averaged to form a composite 3D probabili...

    work page 2020

  4. [4]

    =0.5848 and 𝑤)

    Results Figure 4-6 present representative examples from the test set, demonstrating the effectiveness of the proposed uncertainty-guided refinement pipeline across multiple glioma subregions. For each case, the four MP-MRI modalities (FLAIR, T1, T1ce, and T2), manual ground truth, model-predicted uncertainty map, and segmentation results from the proposed...

    work page 2020

  5. [5]

    = 0.5848 > 𝑤!

    Discussion In this study, we propose a hybrid segmentation framework that integrates spherical-projection-based uncertainty estimation with region-specific 3D refinement to enhance glioma segmentation in MP-MRI. While prior approaches have typically treated segmentation and uncertainty quantification as separate objectives, our framework explicitly unifie...

    work page 2020

  6. [6]

    Conclusion In this study, we developed an uncertainty-guided segmentation framework that combines voxel-wise uncertainty estimation with targeted 3D refinement to enhance the accuracy and reliability of glioma segmentation from multi-parametric MRI data. By leveraging a 2D nnU-Net with spherical projection for initial segmentation and uncertainty quantifi...

    arXiv doi:10.48550/arxiv.2405.18368 2015

  7. [14]

    MRI brain tumor segmentation and uncertainty estimation using 3D-UNet architectures

    Yang Z, Hu Z, Ji H, et al. A neural ordinary differential equation model for visualizing deep neural network behaviors in multi‐parametric MRI‐based glioma segmentation. Med Phys. Published online 2023. 15. Bakas S, Akbari H, Sotiras A, et al. Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. S...

    work page Pith review arXiv 2023

  8. [27]

    Brain tumor segmentation with self-ensembled, deeply-supervised 3D U-net neural networks: a BraTS 2020 challenge solution

    Wang G, Li W, Aertsen M, Deprest J, Ourselin S, Vercauteren T. Test-time augmentation with uncertainty estimation for deep learning-based medical image segmentation. Sci Rep. 2020;10:5068. 28. Henry T, Carre A. Top 10 BraTS 2020 challenge solution: Brain tumor segmentation with self-ensembled, deeply-supervised 3D- Unet like neural networks. 29. Henry T, ...

    work page Pith review arXiv 2020

Show all 9 references
  1. [40]

    Optimum weighted multimodal medical image fusion using particle swarm optimization

    Shehanaz S, Daniel E, Guntur SR, Satrasupalli S. Optimum weighted multimodal medical image fusion using particle swarm optimization. Optik. 2021;231:166413. 41. Yiqing L, Xigang Y, Yongjian L. An improved PSO algorithm for solving non-convex NLP/MINLP problems with equality co...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.