Pith. sign in

REVIEW 3 major objections 36 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · grok-4.3

Four training enhancements to the WT-PSE framework raise optic-disc Dice score from 0.939 to 0.956

2026-06-28 11:05 UTC pith:N3GKM7WT

load-bearing objection The paper adds four standard training tweaks to WT-PSE and reports a Dice gain, but compares the new final-epoch result only to the baseline at epoch 5 with no ablations or variance shown. the 3 major comments →

arxiv 2606.03069 v1 pith:N3GKM7WT submitted 2026-06-02 cs.CV cs.AIcs.LG

ROBUST-WT: Robust Uncertainty-aware Segmentation Transform via Whitening and Training Enhancements

Aqsa Naseer , Maryam Bibi , Syeda Samiya Urooj , Muhammad Khurram Shahzad This is my paper
classification cs.CV cs.AIcs.LG
keywords medical image segmentationwhitening transformdomain adaptationdata augmentationloss schedulingoptic disc segmentationcross-domain segmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper identifies four limitations in the original WT-PSE implementation for cross-domain medical image segmentation: insufficient augmentations to mimic scanner differences, per-pixel binary cross-entropy loss vulnerable to edge noise, missing scheduled loss weighting that can destabilize training, and no built-in ablation controls. It proposes four direct fixes consisting of domain-adaptive augmentations, a hybrid BCE-Dice loss, curriculum-based Dice weight scheduling, and command-line flags for systematic ablations. Experiments on the fundus optic disc benchmark show the updated pipeline reaching a final Dice score of 0.956 and ASD of 13.31. A sympathetic reader would care because these changes target practical training bottlenecks that affect reliable segmentation when imaging devices and protocols vary.

Core claim

The improved pipeline that applies domain-adaptive augmentation including random erasing, gamma correction and salt-and-pepper noise, replaces per-pixel BCE with a hybrid BCE-Dice loss, introduces curriculum-based Dice weight scheduling, and adds command-line ablation flags achieves a final-epoch optic-disc Dice score of 0.956 and an ASD score of 13.31 on the fundus benchmark, outperforming the baseline epoch-5 Dice score of 0.939, while leaving the underlying WT-PSE architecture unchanged.

What carries the argument

The four training enhancements (domain-adaptive augmentation, hybrid BCE-Dice loss, curriculum Dice-weight scheduling, and command-line ablation flags) that directly target the four listed limitations in the original WT-PSE learning framework.

Load-bearing premise

The four listed limitations are the primary causes of any performance gap and the proposed fixes address them without introducing new uncontrolled variables.

What would settle it

A controlled re-run of the fundus optic-disc experiments in which the four enhancements are applied yet the Dice score remains at or below the baseline value of 0.939 would falsify the central claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Performance gains are obtainable solely through training-level changes without any modification to the WT-PSE architecture.
  • Domain-adaptive augmentations can simulate real scanner variations more effectively than the original limited set.
  • The hybrid loss improves edge-aware segmentation under noisy conditions compared with per-pixel BCE alone.
  • Curriculum scheduling of the Dice term stabilizes early training phases that would otherwise be destabilized.
  • Command-line ablation flags enable systematic scientific comparison of each enhancement.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same four enhancements could be tested on segmentation tasks outside fundus imaging to check whether the gains generalize.
  • The approach suggests that many existing whitening-based frameworks might benefit from similar training-level adjustments rather than architectural redesign.
  • Further runs with varied random seeds or additional clinical datasets would test whether the reported Dice and ASD improvements remain stable.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. The paper identifies four limitations in the original WT-PSE framework (limited augmentations, per-pixel BCE loss, unscheduled loss weighting, and lack of ablation controls) and proposes four corresponding enhancements (domain-adaptive augmentations including random erasing/gamma/salt-and-pepper, hybrid BCE+Dice loss, curriculum Dice weight scheduling, and command-line ablation flags). On the fundus optic disc segmentation benchmark, the full enhanced pipeline is reported to reach a final-epoch Dice of 0.956 and ASD of 13.31, outperforming the baseline's epoch-5 Dice of 0.939; the central claim is that these training-level changes produce the gains without modifying the WT-PSE architecture.

Significance. If the attribution of gains to the specific enhancements were supported by ablations, variance estimates, and fair epoch-matched comparisons, the work would demonstrate that modest training modifications can improve cross-domain robustness in medical segmentation pipelines. The manuscript provides no machine-checked proofs or parameter-free derivations; its value would rest entirely on the strength of the empirical controls.

major comments (3)
  1. [Abstract / Experiments] Abstract and results: the reported comparison evaluates the proposed pipeline at its final training epoch but the baseline only at epoch 5. This epoch mismatch leaves open the possibility that the 0.017 Dice improvement is driven by additional training steps rather than the four listed enhancements.
  2. [Experiments] Experiments section: despite the introduction of command-line ablation flags, no per-enhancement ablation results, incremental addition tables, or leave-one-out controls are presented. Without these, the claim that the four specific fixes address the four listed limitations cannot be isolated from confounders such as longer training or altered optimization.
  3. [Results] Results: the manuscript reports point estimates (Dice 0.956, ASD 13.31) with neither run-to-run standard deviations, error bars, nor statistical significance tests. Given the empirical nature of the work and known sensitivity of segmentation metrics to random seeds, this omission weakens the reliability of the performance claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper to improve the experimental rigor and clarity of our claims.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and results: the reported comparison evaluates the proposed pipeline at its final training epoch but the baseline only at epoch 5. This epoch mismatch leaves open the possibility that the 0.017 Dice improvement is driven by additional training steps rather than the four listed enhancements.

    Authors: We agree that the epoch mismatch is a valid concern that could confound attribution of the observed gains. The original WT-PSE baseline was reported at epoch 5 per its published protocol, while our enhanced pipeline was evaluated at convergence. In the revised manuscript, we will add epoch-matched comparisons by reporting the baseline performance at its final epoch as well and include learning curves for both pipelines to demonstrate that improvements are not solely attributable to additional training steps. revision: yes

  2. Referee: [Experiments] Experiments section: despite the introduction of command-line ablation flags, no per-enhancement ablation results, incremental addition tables, or leave-one-out controls are presented. Without these, the claim that the four specific fixes address the four listed limitations cannot be isolated from confounders such as longer training or altered optimization.

    Authors: We acknowledge that although command-line ablation flags were added to support controlled experiments, the manuscript does not present the corresponding ablation results. This limits isolation of each enhancement's contribution. In the revision, we will include incremental addition tables and leave-one-out ablation studies showing the effect of each component (augmentations, hybrid loss, scheduling) when added sequentially to the baseline. revision: yes

  3. Referee: [Results] Results: the manuscript reports point estimates (Dice 0.956, ASD 13.31) with neither run-to-run standard deviations, error bars, nor statistical significance tests. Given the empirical nature of the work and known sensitivity of segmentation metrics to random seeds, this omission weakens the reliability of the performance claim.

    Authors: We agree that the absence of variability measures weakens the empirical claims. In the revised manuscript, we will rerun the experiments with multiple random seeds, report mean and standard deviation for Dice and ASD, and include error bars in the results tables. Statistical significance testing will be added where appropriate to support the comparisons. revision: yes

Circularity Check

0 steps flagged

No significant circularity; purely empirical study with no derivation chain

full rationale

The paper identifies four limitations in the prior WT-PSE work and proposes four training enhancements, then reports benchmark results (Dice 0.956 vs baseline 0.939). No mathematical derivation, first-principles prediction, or equation chain exists that could reduce to fitted inputs or self-citations by construction. The cited WT-PSE is external prior publication; enhancements are described as independent training changes. Attribution of gains rests on a single full-pipeline comparison rather than ablations, but this is an evidence-strength issue, not circularity. The work is self-contained as an empirical pipeline improvement.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the empirical assertion that the listed training changes produce the observed Dice improvement; no new theoretical entities or derivations are introduced, but the assumption that the four limitations are the dominant bottlenecks is taken as given without further justification.

free parameters (1)
  • Dice loss weight schedule
    Curriculum-based weighting strategy is introduced; specific ramp parameters are not stated in the abstract.
axioms (1)
  • domain assumption The four identified limitations are the main factors limiting WT-PSE performance across domains
    Abstract states these limitations without providing evidence or citations that they are primary.

pith-pipeline@v0.9.1-grok · 5819 in / 1334 out tokens · 37931 ms · 2026-06-28T11:05:48.298013+00:00 · methodology

0 comments
read the original abstract

Generalized segmentation of medical images prevents performance degradation when different imaging devices and clinical protocols are used across multiple domains. The Whitening Transform-based Probabilistic Shape Regularization Extractor (WT-PSE), published in IEEE Transactions on Medical Imaging in 2024, addresses this challenge by employing feature decorrelation and Wasserstein distance-based knowledge distillation to achieve robust cross-domain segmentation. This study systematically examines improvements to the WT-PSE learning framework. Four limitations in the original implementation are identified: limited training augmentations that fail to simulate real scanner variations, reliance on per-pixel binary cross-entropy loss that is sensitive to edge noise, the absence of a scheduled loss weighting strategy that may destabilize early training, and the lack of ablation switches for controlled scientific comparison. To address these issues, we propose four enhancements: (1) domain-adaptive augmentation including random erasing, gamma correction, and salt-and-pepper noise; (2) a hybrid BCE and Dice loss function for improved edge-aware segmentation under noisy conditions; (3) a curriculum-based Dice weight scheduling strategy; and (4) command-line control flags for systematic ablation studies. Experiments on the fundus optic disc segmentation benchmark demonstrate that the improved pipeline achieves a final epoch optic-disc Dice score of 0.956 and an ASD score of 13.31, outperforming the baseline epoch-5 Dice score of 0.939. These results indicate that training-level improvements can provide consistent performance gains without modifying the underlying WT-PSE architecture.

Figures

Figures reproduced from arXiv: 2606.03069 by Aqsa Naseer, Maryam Bibi, Muhammad Khurram Shahzad, Syeda Samiya Urooj.

Figure 1
Figure 1. Figure 1: Method overview of the WT-PSE inference pipeline used [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cross-domain segmentation performance comparison [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pixel-level confusion matrix for OD segmentation on the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization of Domain 3 fundus images. Top: input images. Bottom: corresponding OD segmentation overlays. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: False-positive and false-negative overlays for representa [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Threshold analysis on the held-out target domain (Domain [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 6 canonical work pages · 1 internal anchor

  1. [1]

    Enhancing Medical Image Segmentation with Whitening Transform- Based Probabilistic Shape Regularization,

    C. Chen et al., "Enhancing Medical Image Segmentation with Whitening Transform- Based Probabilistic Shape Regularization," IEEE Transactions on Medical Imaging, vol. 43, no. 7, pp. 2693–2706, Jul. 2024. doi: 10.1109/TMI.2024.3371987

    work page doi:10.1109/tmi.2024.3371987 2024

  2. [2]

    DoFE: Domain-oriented Feature EmbeODing for Generalizable Fundus Image Seg- mentation on Unseen Datasets,

    Y . Wang et al., "DoFE: Domain-oriented Feature EmbeODing for Generalizable Fundus Image Seg- mentation on Unseen Datasets,"IEEE Transac- tions on Medical Imaging, vol. 39, no. 12, pp. 4237–4248, 2020

    2020

  3. [3]

    Shape-Aware Meta-Learning for Generalizable Medical Image Segmentation,

    X. Liu et al., "Shape-Aware Meta-Learning for Generalizable Medical Image Segmentation," in Proc. MICCAI, 2020, pp. 493–502. 10

    2020

  4. [4]

    ISD: Self-Supervised Learning with Instance-Semantic Distinction for Medical Image Segmentation,

    X. Liu et al., "ISD: Self-Supervised Learning with Instance-Semantic Distinction for Medical Image Segmentation," inProc. MICCAI, 2021

    2021

  5. [5]

    Texture and Shape Biased Two- Stream Image Recognition for Effective Segmen- tation,

    X. Xie et al., "Texture and Shape Biased Two- Stream Image Recognition for Effective Segmen- tation," inProc. CVPR, 2021

    2021

  6. [6]

    Domain Generalization on Medical Imaging Classification Using Linear- Dependency Regularization,

    Q. Zhou et al., "Domain Generalization on Medical Imaging Classification Using Linear- Dependency Regularization," inProc. NeurIPS, 2020

    2020

  7. [7]

    V-Net: Fully Convolutional Neural Networks for V olu- metric Medical Image Segmentation,

    F. Milletari, N. Navab, and S.-A. Ahmadi, "V-Net: Fully Convolutional Neural Networks for V olu- metric Medical Image Segmentation," inProc. 3DV, 2016, pp. 565–571

    2016

  8. [8]

    Evaluating the Repro- ducibility of Deep Learning-Based Image Seg- mentation,

    J. M. Reinhold et al., "Evaluating the Repro- ducibility of Deep Learning-Based Image Seg- mentation," inSPIE Medical Imaging, 2020

    2020

  9. [9]

    Spinal Cord Grey Matter Seg- mentation Challenge,

    A. Prados et al., "Spinal Cord Grey Matter Seg- mentation Challenge,"NeuroImage, vol. 152, pp. 312–329, 2017

    2017

  10. [10]

    Shape-Aware Complementary-Task Learning for Multi-Organ Segmentation of CT Im- ages,

    S. Liu et al., "Shape-Aware Complementary-Task Learning for Multi-Organ Segmentation of CT Im- ages," inProc. MICCAI, 2019

    2019

  11. [11]

    A Computational Approach to Edge Detection,

    J. E. Canny, "A Computational Approach to Edge Detection,"IEEE Transactions on Pattern Anal- ysis and Machine Intelligence, vol. 8, no. 6, pp. 679–698, 1986

    1986

  12. [12]

    Knowledge-Grounded Attention-Based Neural Machine Translation Model,

    H. Israr, H. Israr, S. A. Khan, M. A. Tahir, M. K. Shahzad, M. Ahmad, and J. M. Zain, "Knowledge-Grounded Attention-Based Neural Machine Translation Model," inApplied Compu- tational Intelligence and Soft Computing, 2025

    2025

  13. [13]

    Neural Machine Translation Models with Attention-Based Dropout Layer,

    H. Israr, S. A. Khan, M. A. Tahir, M. K. Shahzad, M. Ahmad, and J. M. Zain, "Neural Machine Translation Models with Attention-Based Dropout Layer," inComputers, Materials&Continua, vol. 75, no. 2, 2023

    2023

  14. [14]

    mixup: Beyond Empirical Risk Minimization,

    H. Zhang et al., "mixup: Beyond Empirical Risk Minimization," inProc. ICLR, 2018

    2018

  15. [15]

    Deep Learn- ing Based Fabric Defect Detection,

    S. R. Arshad and M. K. Shahzad, "Deep Learn- ing Based Fabric Defect Detection," inResearch Reports on Computer Science, vol. 3, no. 1, pp. 1–11, 2024

    2024

  16. [16]

    EPFG: Electricity Price Forecast- ing with Enhanced GANs Neural Network,

    M. Hanif et al., "EPFG: Electricity Price Forecast- ing with Enhanced GANs Neural Network,"IETE Journal of Research, vol. 69, no. 9, pp. 6473– 6482, 2023

    2023

  17. [17]

    LNDIR: A Lightweight Non-Increasing Delivery-Latency Interval-Based Routing for Duty-Cycled Sensor Networks,

    M. K. Shahzad et al., "LNDIR: A Lightweight Non-Increasing Delivery-Latency Interval-Based Routing for Duty-Cycled Sensor Networks,"Inter- national Journal of Distributed Sensor Networks, vol. 14, no. 4, 2018

    2018

  18. [18]

    Two-dimensional whitening reconstruction for enhancing robustness of principal component analysis,

    X. Shi, Z. Guo, F. Nie, L. Yang, J. You, and D. Tao, "Two-dimensional whitening reconstruction for enhancing robustness of principal component analysis," inIEEE Trans. Pattern Anal. Mach. In- tell., vol. 38, no. 10, pp. 2130–2136, 2016

    2016

  19. [19]

    Modeling uncertainty with hedged instance embeODing,

    S. J. Oh, K. Murphy, J. Pan, J. Roth, F. Schroff, and A. Gallagher, "Modeling uncertainty with hedged instance embeODing," 2018, arXiv:1810.00319

    work page arXiv 2018

  20. [20]

    Probabilistic embeODings revisited,

    I. Karpukhin, S. Dereka, and S. Kolesnikov, "Probabilistic embeODings revisited," 2022, arXiv:2202.06768

    work page arXiv 2022

  21. [21]

    Distances between prob- ability distributions of different dimensions,

    Y . Cai and L.-H. Lim, "Distances between prob- ability distributions of different dimensions," in IEEE Trans. Inf. Theory, vol. 68, no. 6, pp. 4020– 4031, 2022

    2022

  22. [22]

    Statistical aspects of Wasserstein distances,

    V . M. Panaretos and Y . Zemel, "Statistical aspects of Wasserstein distances," inAnnu. Rev. Statist. Appl., vol. 6, pp. 405–431, 2019

    2019

  23. [23]

    A new algorithm for computing the square root of a matrix,

    J. Nichols, "A new algorithm for computing the square root of a matrix," 2016

    2016

  24. [24]

    Infor- mation geometry connecting Wasserstein distance and Kullback–Leibler divergence,

    S.-I. Amari, R. Karakida, and M. Oizumi, "Infor- mation geometry connecting Wasserstein distance and Kullback–Leibler divergence," inInf. Geome- try, vol. 1, no. 1, pp. 13–37, 2018

    2018

  25. [25]

    Provable robustness against Wasserstein distri- bution shifts via input randomization,

    A. Kumar, A. Levine, T. Goldstein, and S. Feizi, "Provable robustness against Wasserstein distri- bution shifts via input randomization," inProc. ICLR, 2022

    2022

  26. [26]

    Domain generalization with adversarial feature learning,

    H. Li, S. J. Pan, S. Wang, and A. C. Kot, "Domain generalization with adversarial feature learning," inProc. CVPR, 2018

    2018

  27. [27]

    Prior attention network for multi- lesion segmentation in medical images,

    X. Zhao et al., "Prior attention network for multi- lesion segmentation in medical images," inIEEE Trans. Med. Imag., vol. 41, no. 12, pp. 3812–3823, 2022. 11

    2022

  28. [28]

    Y-Net: A one-to-two deep learning frame- work for digital holographic reconstruction,

    K. Wang, D. J. Dou, Q. Kemao, J. Di, and J. Zhao, "Y-Net: A one-to-two deep learning frame- work for digital holographic reconstruction," in Opt. Lett., vol. 44, pp. 4765–4768, 2019

    2019

  29. [29]

    Learning structured output representation using deep conditional gen- erative models,

    K. Sohn, H. Lee, and X. Yan, "Learning structured output representation using deep conditional gen- erative models," inProc. NeurIPS, 2015

    2015

  30. [30]

    Em- bracing the dark knowledge: Domain generaliza- tion using regularized knowledge distillation,

    Y . Wang, H. Li, L.-P. Chau, and A. C. Kot, "Em- bracing the dark knowledge: Domain generaliza- tion using regularized knowledge distillation," in Proc. ACM MM, 2021

    2021

  31. [31]

    An overview of statistical learning theory,

    V . N. Vapnik, "An overview of statistical learning theory," inIEEE Trans. Neural Netw., vol. 10, no. 5, pp. 988–999, 1999

    1999

  32. [32]

    Uncertainty modeling for out-of- distribution generalization,

    X. Li et al., "Uncertainty modeling for out-of- distribution generalization," inProc. ICLR, 2022

    2022

  33. [33]

    Comparing Kullback–Leibler divergence and mean squared error loss in knowledge distil- lation,

    T. Kim, J. Oh, N. Kim, S. Cho, and S.-Y . Yun, "Comparing Kullback–Leibler divergence and mean squared error loss in knowledge distil- lation," 2021, arXiv:2105.08919

    work page arXiv 2021

  34. [34]

    Benchmarking neural network robustness to common corruptions and perturbations,

    D. Hendrycks and T. Dietterich, "Benchmarking neural network robustness to common corruptions and perturbations," 2019, arXiv:1903.12249

    work page arXiv 2019

  35. [35]

    Mg, Al, Si, Ca, Ti, Fe, and Ni abundance for a sample of solar analogues

    M. Arjovsky, S. Chintala, and L. Bottou, "Wasser- stein GAN," 2017, arXiv:1701.07850

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [36]

    Steel Defect Classification Using Machine Learning,

    S. R. Arshad and M. K. Shahzad, "Steel Defect Classification Using Machine Learning," inProc. IMCOM, 2022. 12

    2022