pith. sign in

arxiv: 2603.15941 · v2 · submitted 2026-03-16 · 💻 cs.CV

Towards Fair and Robust Volumetric CT Classification via KL-Regularised Group Distributionally Robust Optimisation

Samuel Johnny , Blessed Guda , Goodness Obasi , Aaron Emmanuel , Moise Busogi This is my paper

Pith reviewed 2026-05-15 09:39 UTC · model grok-4.3

classification 💻 cs.CV
keywords volumetric CT classificationGroup DROKL regularizationfairnessrobustnessCOVID-19 detectionlung pathologymulti-site imaging
0
0 comments X

The pith

KL-regularised Group DRO with a MobileViT encoder raises both worst-group F1 and average performance in volumetric CT classification.

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

The paper demonstrates that adding a KL penalty to Group Distributionally Robust Optimisation prevents the group weights from collapsing to a single worst case while still upweighting underperforming acquisition sites and demographic subgroups. The resulting training objective is applied to a lightweight MobileViT-XXS slice encoder followed by a SliceTransformer aggregator for two CT tasks: binary COVID-19 detection across multiple sites and four-class lung pathology recognition with explicit gender fairness constraints. On the first task the method reaches a challenge F1 of 0.835; on the second it reaches a mean per-gender macro F1 of 0.815, lifting the most disadvantaged subgroup (female squamous cell carcinoma) by 17.4 points over a focal-loss baseline.

Core claim

KL-regularised Group DRO adaptively reweights training batches according to current per-group loss while the KL term keeps the group distribution from degenerating, allowing a single model to improve both robustness to site shifts and fairness across gender-class combinations without separate per-site retraining.

What carries the argument

KL-regularised Group Distributionally Robust Optimisation that upweights underperforming groups during optimisation while the KL divergence penalty on the group weighting distribution prevents collapse.

If this is right

  • A single set of hyperparameters can be used across multiple acquisition centres without site-specific retuning.
  • Direct definition of groups at the gender-class granularity lifts accuracy on severely underrepresented combinations such as female squamous cell carcinoma.
  • The same lightweight volumetric architecture works for both binary and multi-class CT tasks while satisfying the robustness and fairness objectives.
  • Worst-group performance improves without a proportional drop in average performance.

Where Pith is reading between the lines

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

  • The same regularisation pattern could be tested on other medical imaging tasks where scanner vendor and patient demographics create distribution shifts.
  • If the optimal KL coefficient proves stable across new datasets, the method would reduce the engineering cost of deploying models in new hospitals.
  • The approach suggests that explicit regularisation on the robustness objective itself can shrink the usual gap between average and worst-group accuracy.

Load-bearing premise

The KL penalty will keep group weights from collapsing to one or two sites or subgroups while still delivering meaningful worst-case protection across all centres and genders.

What would settle it

Train the identical MobileViT-plus-SliceTransformer architecture with standard Group DRO (KL coefficient set to zero) on the same data splits and measure whether the group weights concentrate on a single centre or gender class and whether the reported per-group F1 scores fall below the KL-regularised figures.

Figures

Figures reproduced from arXiv: 2603.15941 by Aaron Emmanuel, Blessed Guda, Goodness Obasi, Moise Busogi, Samuel Johnny.

Figure 1
Figure 1. Figure 1: Representative axial CT slices at five depth lev [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed pipeline. A 3D CT scan is split into 64 slices, each encoded independently by a shared MobileViT [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of KL regularisation strength α on Task 2 vali￾dation performance, reported separately for male and female sub￾groups. Group DRO with α = 0.5 achieves the best mean F1 of 0.815 and the smallest gender gap, outperforming Focal Loss (0.777) and the best challenge entry (0.704 [17]). At α = 1.0, male macro rises while female macro falls sharply, indicating that forcing uniform weights over-regularises … view at source ↗
Figure 4
Figure 4. Figure 4: Effect of KL regularisation strength α on Task 1 val￾idation performance. Group DRO with α = 0.5 achieves the best mean F1 of 0.835, surpassing both the weighted CE baseline (0.804) and the best published challenge entry (0.776 [17]). Large α forces uniform group weights, collapsing toward ERM and de￾grading performance (α= 0.3, F1 = 0.726). proves over both baselines, confirming the value of group￾aware t… view at source ↗
read the original abstract

Automated diagnosis from chest computed tomography (CT) scans faces two persistent challenges in clinical deployment: distribution shift across acquisition sites and performance disparity across demographic subgroups. We address both simultaneously across two complementary tasks: binary COVID-19 classification from multi-site CT volumes (Task 1) and four-class lung pathology recognition with gender-based fairness constraints (Task 2). Our framework combines a lightweight MobileViT-XXS slice encoder with a two-layer SliceTransformer aggregator for volumetric reasoning, and trains with a KL-regularised Group Distributionally Robust Optimisation (Group DRO) objective that adaptively upweights underperforming acquisition centres and demographic subgroups. Unlike standard Group DRO, the KL penalty prevents group weight collapse, providing a stable balance between worst-case protection and average performance. For Task 2, we define groups at the granularity of gender class, directly targeting severely underrepresented combinations such as female Squamous cell carcinoma. On Task 1, our best configuration achieves a challenge F1 of 0.835, surpassing the best published challenge entry by +5.9. On Task 2, Group DRO with {\alpha} = 0.5 achieves a mean per-gender macro F1 of 0.815, outperforming the best challenge entry by +11.1 pp and improving Female Squamous F1 by +17.4 over the Focal Loss baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents a KL-regularised Group Distributionally Robust Optimisation (Group DRO) method integrated with a MobileViT-XXS slice encoder and SliceTransformer for volumetric CT scan classification. It targets distribution shifts across acquisition sites and fairness across gender subgroups in two tasks: binary COVID-19 classification (Task 1) and multi-class lung pathology recognition (Task 2). The key innovation is the KL penalty to prevent group weight collapse in Group DRO, leading to reported F1 improvements of 0.835 on Task 1 (+5.9 over best challenge entry) and 0.815 mean per-gender macro F1 on Task 2 (+11.1 pp, with +17.4 on Female Squamous).

Significance. If the performance gains can be robustly attributed to the KL-regularised Group DRO rather than confounding factors like architecture choices, this work offers a promising direction for developing fair and robust models in medical imaging that balance worst-group performance with overall accuracy, potentially aiding clinical deployment across diverse sites and demographics.

major comments (2)
  1. [Abstract] The central claim that the KL penalty prevents group weight collapse while delivering worst-case protection is not supported by any reported analysis of group weights, ablation studies on the regularisation parameter α, or direct comparisons to unregularised Group DRO using the same MobileViT+SliceTransformer backbone. This verification is load-bearing for the attribution of the +17.4 pp Female Squamous F1 improvement and the overall +11.1 pp gain.
  2. [Experimental evaluation] The reported F1 scores (e.g., 0.835 on Task 1 and 0.815 on Task 2) are presented without error bars, statistical significance tests, or detailed experimental protocols including data splits and hyperparameter tuning procedures, making it impossible to assess the reliability and reproducibility of the claimed improvements over challenge baselines.
minor comments (1)
  1. [Abstract] The notation for the regularisation parameter is introduced as α = 0.5 without prior definition in the abstract, which could be clarified for readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas where additional evidence and details will strengthen the manuscript. We address each major comment below and will revise the paper to incorporate the requested analyses and protocols.

read point-by-point responses

  1. Referee: [Abstract] The central claim that the KL penalty prevents group weight collapse while delivering worst-case protection is not supported by any reported analysis of group weights, ablation studies on the regularisation parameter α, or direct comparisons to unregularised Group DRO using the same MobileViT+SliceTransformer backbone. This verification is load-bearing for the attribution of the +17.4 pp Female Squamous F1 improvement and the overall +11.1 pp gain.

    Authors: We agree that the current manuscript lacks explicit verification of the KL penalty's effect on group weights and direct ablations. In the revised version we will add: (i) training curves and histograms of group weights with and without the KL term to demonstrate prevention of collapse; (ii) a full ablation table varying α ∈ {0, 0.1, 0.5, 1.0} on both tasks using the identical MobileViT-XXS + SliceTransformer backbone; and (iii) side-by-side results against unregularised Group DRO. These additions will directly support attribution of the reported gains, especially on the Female Squamous subgroup. revision: yes

  2. Referee: [Experimental evaluation] The reported F1 scores (e.g., 0.835 on Task 1 and 0.815 on Task 2) are presented without error bars, statistical significance tests, or detailed experimental protocols including data splits and hyperparameter tuning procedures, making it impossible to assess the reliability and reproducibility of the claimed improvements over challenge baselines.

    Authors: We acknowledge that the absence of variability measures and protocol details limits reproducibility assessment. The revised manuscript will include: standard deviations from five independent runs with different random seeds, statistical significance tests (paired t-tests) against the challenge baselines, and an expanded experimental section detailing site-stratified splits for Task 1, gender-and-class stratified splits for Task 2, the full hyperparameter search grid, and training schedules. These will appear in the main text with additional tables in the supplement. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results on held-out data independent of training objective

full rationale

The manuscript presents an empirical ML pipeline (MobileViT + SliceTransformer trained with KL-regularised Group DRO) and reports F1 metrics on two separate challenge tasks using held-out test data. No derivation, equation, or self-citation chain is shown that reduces the reported gains (+5.9 F1 on Task 1, +11.1 pp macro F1 on Task 2) to a quantity defined by the fitted α or by the objective itself. The KL penalty is described as preventing collapse, but its effect is measured externally rather than being tautological. All performance numbers are falsifiable on external benchmarks and do not collapse to the training inputs by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions plus the domain-specific premise that a KL penalty on group weights yields a stable worst-case/average trade-off; no new physical entities are introduced.

free parameters (1)
  • α = 0.5
    KL regularisation coefficient set to 0.5 for Task 2 to obtain the reported per-gender macro F1.
axioms (1)
  • domain assumption Group DRO with an added KL penalty on group weights prevents collapse while still up-weighting under-performing acquisition sites and demographic subgroups.
    Invoked to justify the training objective for both tasks.

pith-pipeline@v0.9.0 · 5555 in / 1396 out tokens · 31658 ms · 2026-05-15T09:39:15.966434+00:00 · methodology

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

30 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    Abdel-Khalek

    Ebtesam Al-Mansor, Mohammed Al-Jabbar, Anis Ben Ishak, and S. Abdel-Khalek. Medical image edge detection in the framework of quantum representations.Alexandria Engi- neering Journal, 81:234–242, 2023. 2

    work page 2023

  2. [2]

    A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis

    Anastasios Arsenos, Dimitrios Kollias, and Stefanos Kol- lias. A large imaging database and novel deep neural ar- chitecture for covid-19 diagnosis. In2022 IEEE 14th Im- age, Video, and Multidimensional Signal Processing Work- shop (IVMSP), page 1–5. IEEE, 2022. 2

    work page 2022

  3. [3]

    Light-weight vision transformer-based semantic segmentation for medical im- ages

    Wen-Ling Chou, Guo-Shiang Lin, Ku-Yaw Chang, Sheng- Lei Yan, and Wei-Cheng Yeh. Light-weight vision transformer-based semantic segmentation for medical im- ages. In2025 IEEE International Conference on Advanced Visual and Signal-Based Systems (AVSS), pages 1–4, 2025. 3

    work page 2025

  4. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representa- tions (ICLR), 2021. 2

    work page 2021

  5. [5]

    Data-driven distributionally robust optimization using the wasserstein metric.Mathematical Programming, 171:115–166, 2018

    Peyman Mohajerin Esfahani and Daniel Kuhn. Data-driven distributionally robust optimization using the wasserstein metric.Mathematical Programming, 171:115–166, 2018. 2

    work page 2018

  6. [6]

    Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system

    Demetris Gerogiannis, Anastasios Arsenos, Dimitrios Kol- lias, Dimitris Nikitopoulos, and Stefanos Kollias. Covid- 19 computer-aided diagnosis through ai-assisted ct imaging analysis: Deploying a medical ai system. In2024 IEEE In- ternational Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2024. 1, 2

    work page 2024

  7. [7]

    Domain adaptation for medical image analysis: A survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2022

    Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: A survey.IEEE Transactions on Biomedical Engineering, 69(3):1173–1185, 2022. 1, 2

    work page 2022

  8. [8]

    Huang, Y

    B. Huang, Y . Liu, B. Tang, et al. Inceptionmamba: A lightweight and effective model for medical image classifi- cation revealing mamba’s low-frequency bias.Neural Pro- cessing Letters, 58(15), 2026. 3

    work page 2026

  9. [9]

    Deep neural archi- tectures for prediction in healthcare.Complex & Intelligent Systems, 4(2):119–131, 2018

    Dimitrios Kollias, Athanasios Tagaris, Andreas Stafylopatis, Stefanos Kollias, and Georgios Tagaris. Deep neural archi- tectures for prediction in healthcare.Complex & Intelligent Systems, 4(2):119–131, 2018. 2

    work page 2018

  10. [10]

    Deep transparent prediction through latent repre- sentation analysis.arXiv preprint arXiv:2009.07044, 2020

    Dimitrios Kollias, N Bouas, Y Vlaxos, V Brillakis, M Se- feris, Ilianna Kollia, Levon Sukissian, James Wingate, and S Kollias. Deep transparent prediction through latent repre- sentation analysis.arXiv preprint arXiv:2009.07044, 2020. 3

    work page arXiv 2009

  11. [11]

    Transpar- ent adaptation in deep medical image diagnosis

    Dimitris Kollias, Y Vlaxos, M Seferis, Ilianna Kollia, Levon Sukissian, James Wingate, and Stefanos D Kollias. Transpar- ent adaptation in deep medical image diagnosis. InTAILOR, page 251–267, 2020. 3

    work page 2020

  12. [12]

    Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis

    Dimitrios Kollias, Anastasios Arsenos, Levon Soukissian, and Stefanos Kollias. Mia-cov19d: Covid-19 detection through 3-d chest ct image analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, page 537–544, 2021. 1, 2

    work page 2021

  13. [13]

    Ai-mia: Covid-19 detection and severity analysis through medical imaging

    Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-mia: Covid-19 detection and severity analysis through medical imaging. InEuropean Conference on Computer Vi- sion, page 677–690. Springer, 2022. 2

    work page 2022

  14. [14]

    Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity

    Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Ai-enabled analysis of 3-d ct scans for diagnosis of covid-19 & its severity. In2023 IEEE International Conference on Acoustics, Speech, and Signal Processing Workshops (ICAS- SPW), page 1–5. IEEE, 2023. 1, 2

    work page 2023

  15. [15]

    A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging.Neuro- computing, 542:126244, 2023

    Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. A deep neural architecture for harmonizing 3-d input data analysis and decision making in medical imaging.Neuro- computing, 542:126244, 2023. 1, 3

    work page 2023

  16. [16]

    Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection.arXiv preprint arXiv:2407.15728, 2024

    Dimitrios Kollias, Anastasios Arsenos, James Wingate, and Stefanos Kollias. Sam2clip2sam: Vision language model for segmentation of 3d ct scans for covid-19 detection.arXiv preprint arXiv:2407.15728, 2024. 2, 3

    work page arXiv 2024

  17. [17]

    Pharos-afe-aimi: Multi-source & fair disease diagnosis

    Dimitrios Kollias, Anastasios Arsenos, and Stefanos Kollias. Pharos-afe-aimi: Multi-source & fair disease diagnosis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7265–7273, 2025. 1, 2, 5, 6

    work page 2025

  18. [18]

    Multi-source covid-19 detec- tion via kernel-density-based slice sampling.arXiv preprint arXiv:2507.01564, 2025

    Chia-Ming Lee, Bo-Cheng Qiu, Ting-Yao Chen, Ming-Han Sun, Fang-Ying Lin, Jung-Tse Tsai, I-An Tsai, Yu-Fan Lin, and Chih-Chung Hsu. Multi-source covid-19 detec- tion via kernel-density-based slice sampling.arXiv preprint arXiv:2507.01564, 2025. 3, 5

    work page arXiv 2025

  19. [19]

    Advancing lung disease diagnosis in 3d ct scans.arXiv preprint arXiv:2507.00993,

    Qingqiu Li, Runtian Yuan, Junlin Hou, Jilan Xu, Yuejie Zhang, Rui Feng, and Hao Chen. Advancing lung disease diagnosis in 3d ct scans.arXiv preprint arXiv:2507.00993,

    work page arXiv

  20. [20]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. InPro- ceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2980–2988, 2017. 2

    work page 2017

  21. [21]

    MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer

    Sachin Mehta and Mohammad Rastegari. Mobilevit: Light- weight, general-purpose, and mobile-friendly vision trans- former. InInternational Conference on Learning Represen- tations (ICLR), 2022. Also available as arXiv:2110.02178. 1, 3

    work page internal anchor Pith review arXiv 2022

  22. [22]

    Hashimoto, and Percy Liang

    Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst- case generalization. InInternational Conference on Learn- ing Representations (ICLR), 2020. 1, 2, 3, 4

    work page 2020

  23. [23]

    An investigation of why overparameterization exacerbates spurious correlations

    Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. InInternational Confer- ence on Machine Learning (ICML), 2020. 2

    work page 2020

  24. [24]

    Q. Wang, F. Liu, R. Zou, et al. Enhancing medical im- age object detection with collaborative multi-agent deep q- networks and multi-scale representation.EURASIP Journal on Advances in Signal Processing, 2023(132):1–18, 2023. 2

    work page 2023

  25. [25]

    3d-2d medical image registration technology and its application development: a survey

    Handan Xiao. 3d-2d medical image registration technology and its application development: a survey. InProceedings of the 2023 4th International Symposium on Artificial Intel- ligence for Medicine Science, page 95–100, New York, NY , USA, 2024. Association for Computing Machinery. 2

    work page 2023

  26. [26]

    Mazurowski, and Heung-Il Suk

    Jee Seok Yoon, Kwanseok Oh, Yooseung Shin, Maciej A. Mazurowski, and Heung-Il Suk. Domain generalization for medical image analysis: A review.Proceedings of the IEEE, 112(10):1583–1609, 2024. 2

    work page 2024

  27. [27]

    Multi-source covid-19 detection via variance risk extrapolation.arXiv preprint arXiv:2506.23208, 2025

    Runtian Yuan, Qingqiu Li, Junlin Hou, Jilan Xu, Yuejie Zhang, Rui Feng, and Hao Chen. Multi-source covid-19 detection via variance risk extrapolation.arXiv preprint arXiv:2506.23208, 2025. 3, 5

    work page arXiv 2025

  28. [28]

    Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation.IEEE Transactions on Medical Imaging, 39(7):2531–2540, 2020

    Li Zhang, Xiaosong Wang, Dong Yang, Thomas Sanford, Stephanie Harmon, Baris Turkbey, Bradford J Wood, Hol- ger Roth, Berengere Aubert-Broche, D Louis Collins, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation.IEEE Transactions on Medical Imaging, 39(7):2531–2540, 2020. 1, 2

    work page 2020

  29. [29]

    Zhang, P

    Y . Zhang, P. Gu, N. Sapkota, and D. Z. Chen. Swipe: Ef- ficient and robust medical image segmentation with implicit patch embeddings. InMedical Image Computing and Com- puter Assisted Intervention – MICCAI 2023, page 315–326, Cham, 2023. Springer. 3

    work page 2023

  30. [30]

    Adverin: Monotonic adversarial in- tensity attack for domain generalization in medical image segmentation.Medical Image Analysis, 107:103848, 2026

    Zheyuan Zhang, Bin Wang, Lanhong Yao, Elif Keles, Debesh Jha, Matthew Antalek, Gorkem Durak, Alpay Mede- talibeyoglu, Concetto Spampinato, Baris Turkbey, Boqing Gong, and Ulas Bagci. Adverin: Monotonic adversarial in- tensity attack for domain generalization in medical image segmentation.Medical Image Analysis, 107:103848, 2026. 2

    work page 2026