pith. sign in

arxiv: 2603.28463 · v2 · submitted 2026-03-30 · 💻 cs.CV

Decoupling Wavelet Sub-bands for Single Source Domain Generalization in Fundus Image Segmentation

Shramana Dey , Varun Ajith , Abhirup Banerjee , Sushmita Mitra This is my paper

Pith reviewed 2026-05-14 21:31 UTC · model grok-4.3

classification 💻 cs.CV
keywords domain generalizationfundus image segmentationwavelet sub-bandsoptic discoptic cupsingle sourcemedical imaging
0
0 comments X

The pith

Wavelet sub-band decoupling enables single-source domain generalization in fundus image segmentation by isolating anatomical structure from appearance variations.

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

The paper presents WaveSDG, a segmentation network for optic disc and cup in fundus images that trains on one source domain and performs on multiple unseen target domains. It decomposes encoder features into wavelet sub-bands and uses a dedicated module to hold global anatomy in low-frequency parts while sharpening edges and removing noise in high-frequency parts. This separation aims to keep structural cues stable even when device or clinic conditions change. Evaluations across five target datasets show higher balanced Dice scores and lower Hausdorff distances than seven prior methods, along with lower result variance.

Core claim

WaveSDG decouples anatomical structure from domain-specific appearance through wavelet sub-band decomposition. The WISER module refines low-frequency components to anchor global anatomy while selectively enhancing directional edges and suppressing noise in high-frequency sub-bands, producing more accurate and stable segmentations on unseen domains.

What carries the argument

The Wavelet-based Invariant Structure Extraction and Refinement (WISER) module, which assigns distinct processing roles to each wavelet sub-band to separate stable anatomical topology from varying appearance features.

If this is right

  • Consistent outperformance on optic cup and optic disc segmentation from one source to five unseen target datasets.
  • Best balanced Dice score and lowest 95th-percentile Hausdorff distance among compared methods.
  • Lower variance across domains, indicating greater cross-domain stability.
  • Ablation studies confirm the contribution of both the sub-band decoupling and the WISER refinement steps.

Where Pith is reading between the lines

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

  • The same frequency separation could be tested on other retinal or medical segmentation tasks that suffer from device-induced domain shifts.
  • If the approach preserves all structural information, it may reduce the need to collect multi-site annotated training data in clinical settings.
  • Extending the method to color fundus images with additional artifacts such as cataracts would test whether the decoupling remains effective under stronger appearance changes.

Load-bearing premise

Wavelet sub-band decomposition combined with the WISER module can reliably separate anatomical topology from domain-specific appearance without discarding needed segmentation cues or creating artifacts on new domains.

What would settle it

Running WaveSDG on an additional unseen fundus dataset and finding that its Dice scores and Hausdorff distances match or fall below those of a plain encoder-decoder network without any wavelet decomposition.

Figures

Figures reproduced from arXiv: 2603.28463 by Abhirup Banerjee, Shramana Dey, Sushmita Mitra, Varun Ajith.

Figure 1
Figure 1. Figure 1: Representative wavelet decomposition of (a) Fundus image, with (b) extracted [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The proposed WaveSDG architecture. The lower and upper rectangles illustrate the overall network and the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of predicted segmentation mask by WaveSDG and seven other competing methods, along with [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the effect of various components in WISER. (a) Feature channel and its wavelet-decomposed [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of LL sub-band decomposition, with row (a) showing column-wise, (i) query image, with its [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
read the original abstract

Domain generalization in fundus imaging is challenging due to variations in acquisition conditions across devices and clinical settings. The inability to adapt to these variations causes performance degradation on unseen domains for deep learning models. Besides, obtaining annotated data across domains is often expensive and privacy constraints restricts their availability. Although single-source domain generalization (SDG) offers a realistic solution to this problem, the existing approaches frequently fail to capture anatomical topology or decouple appearance from anatomical features. This research introduces WaveSDG, a new wavelet-guided segmentation network for SDG. It decouples anatomical structure from domain-specific appearance through a wavelet sub-band decomposition. A novel Wavelet-based Invariant Structure Extraction and Refinement (WISER) module is proposed to process encoder features by leveraging distinct semantic roles of each wavelet sub-band. The module refines low-frequency components to anchor global anatomy, while selectively enhancing directional edges and suppressing noise within the high-frequency sub-bands. Extensive ablation studies validate the effectiveness of the WISER module and its decoupling strategy. Our evaluations on optic cup and optic disc segmentation across one source and five unseen target datasets show that WaveSDG consistently outperforms seven state-of-the-art methods. Notably, it achieves the best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance, indicating improved accuracy, robustness, and cross-domain stability.

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 / 2 minor

Summary. The paper proposes WaveSDG, a wavelet-guided network for single-source domain generalization in fundus image segmentation of optic cup and disc. It decomposes encoder features via wavelet sub-bands and introduces the WISER module, which refines low-frequency components to anchor global anatomy while enhancing directional edges and suppressing noise in high-frequency sub-bands. The central claim is that this decoupling yields consistent outperformance over seven SOTA methods on one source and five unseen target datasets, with the best balanced Dice score, lowest 95th-percentile Hausdorff distance, and reduced variance.

Significance. If the quantitative claims hold under rigorous validation, the work would be significant for medical imaging SDG: it offers an architectural route to separate anatomical topology from acquisition-specific appearance without multi-domain training data, directly addressing privacy and annotation costs in fundus analysis.

major comments (2)
  1. [WISER module] Abstract and WISER module description: the load-bearing assumption that low-frequency (approximation) coefficients remain largely domain-invariant is not supported by explicit normalization or disentanglement steps; fundus images commonly exhibit low-frequency shifts (vignetting, sensor gain, global brightness) that directly alter these coefficients, and the paper must demonstrate via pre/post-WISER statistics or targeted ablations that source-specific bias is removed rather than merely assumed absent.
  2. [Experiments] Evaluation section: the abstract asserts 'best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance' across five targets, yet supplies no numerical tables, per-dataset scores, statistical significance tests, or variance breakdowns; without these, the cross-domain stability claim cannot be verified and the comparison to the seven baselines remains unverifiable.
minor comments (2)
  1. Clarify the exact wavelet family, decomposition level, and sub-band selection criteria used in the decomposition step.
  2. Provide a diagram or pseudocode for the WISER module to make the distinct processing of low- versus high-frequency paths explicit.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which help strengthen the presentation of our work on WaveSDG for single-source domain generalization in fundus segmentation. We address each major point below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: [WISER module] Abstract and WISER module description: the load-bearing assumption that low-frequency (approximation) coefficients remain largely domain-invariant is not supported by explicit normalization or disentanglement steps; fundus images commonly exhibit low-frequency shifts (vignetting, sensor gain, global brightness) that directly alter these coefficients, and the paper must demonstrate via pre/post-WISER statistics or targeted ablations that source-specific bias is removed rather than merely assumed absent.

    Authors: We agree that explicit empirical support for the domain-invariance of low-frequency coefficients is required rather than relying solely on the design rationale of wavelet decomposition. In the revised manuscript we will add (i) quantitative pre- and post-WISER statistics (mean and variance of approximation coefficients) computed on source and target domains, and (ii) targeted ablation results that isolate the effect of the low-frequency refinement branch on cross-domain performance. These additions will demonstrate that source-specific low-frequency bias is actively reduced by the WISER module. revision: yes

  2. Referee: [Experiments] Evaluation section: the abstract asserts 'best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance' across five targets, yet supplies no numerical tables, per-dataset scores, statistical significance tests, or variance breakdowns; without these, the cross-domain stability claim cannot be verified and the comparison to the seven baselines remains unverifiable.

    Authors: We acknowledge that the evaluation section as currently written does not contain the full numerical tables needed to substantiate the abstract claims. In the revision we will expand the evaluation section with complete per-dataset tables reporting Dice, 95th-percentile Hausdorff distance, and standard deviation for WaveSDG and all seven baselines across the five target domains. We will also include paired statistical significance tests (e.g., Wilcoxon signed-rank) and explicit variance breakdowns to allow direct verification of the reported improvements in accuracy, robustness, and cross-domain stability. revision: yes

Circularity Check

0 steps flagged

No circularity: architectural proposal evaluated on held-out targets

full rationale

The paper introduces WaveSDG and the WISER module as a new architectural design for decoupling wavelet sub-bands in single-source domain generalization. No equations, derivations, or fitted parameters are presented that reduce reported performance gains to quantities computed from the same source data by construction. The central claims rest on empirical results across one source and five unseen target datasets plus ablation studies, with no self-citation load-bearing uniqueness theorems, ansatz smuggling, or renaming of known results. The derivation chain is self-contained and does not collapse to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into exact parameter counts or training choices; the central claim rests on the unverified assumption that wavelet sub-bands cleanly separate anatomy from appearance.

axioms (1)
  • domain assumption Wavelet sub-band decomposition can separate anatomical structure from domain-specific appearance in fundus images
    Invoked by the decoupling strategy and WISER module description.
invented entities (1)
  • WISER module no independent evidence
    purpose: Process encoder features by leveraging distinct semantic roles of each wavelet sub-band to refine low-frequency anatomy and enhance/suppress high-frequency components
    Newly proposed component whose effectiveness is asserted via ablation studies mentioned but not shown.

pith-pipeline@v0.9.0 · 5550 in / 1405 out tokens · 49164 ms · 2026-05-14T21:31:59.742547+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

37 extracted references · 37 canonical work pages

  1. [1]

    Fundus imaging-based healthcare: Present and future.ACM Trans

    Vijay Kumar and Kolin Paul. Fundus imaging-based healthcare: Present and future.ACM Trans. on Comp. for Healthcare, 4:1–34, 2023

    work page 2023

  2. [2]

    J. Shi, K. Zhang, et al. A survey of label-noise deep learning for medical image analysis.Medical Image Analysis, 95:103166, 2024

    work page 2024

  3. [3]

    FunOTTA: On-the-fly adaptation on cross-domain fundus image via stable test-time training.IEEE Trans

    Qian Zeng, Le Zhang, et al. FunOTTA: On-the-fly adaptation on cross-domain fundus image via stable test-time training.IEEE Trans. on Medical Imaging, pages 1–1, 2025

    work page 2025

  4. [4]

    Hierarchy-aware and knowledge-guided learning for multi-label classification of retinal diseases from fundus images.IEEE Trans

    Zhaocan Yang, Yan Li, et al. Hierarchy-aware and knowledge-guided learning for multi-label classification of retinal diseases from fundus images.IEEE Trans. on Medical Imaging, pages 1–1, 2026

    work page 2026

  5. [5]

    Domain adaptation for medical image analysis: A survey.IEEE Trans

    Hao Guan and Mingxia Liu. Domain adaptation for medical image analysis: A survey.IEEE Trans. on Biomed. Eng., 69:1173–1185, 2021

    work page 2021

  6. [7]

    FedCA: Federated domain generalization for medical image segmentation via cross-client feature style transfer and adaptive style alignment.Exp

    Yihan Ren, Yanfeng Li, et al. FedCA: Federated domain generalization for medical image segmentation via cross-client feature style transfer and adaptive style alignment.Exp. Sys. with Appl., page 131394, 2026

    work page 2026

  7. [8]

    Domain generalization for medical image analysis: A review.Proc

    Jee Seok Yoon, Kwanseok Oh, et al. Domain generalization for medical image analysis: A review.Proc. of the IEEE, 112:1583–1609, 2024

    work page 2024

  8. [9]

    Treasure in distribution: A domain randomization based multi-source domain generalization for 2d medical image segmentation

    Ziyang Chen, Yongsheng Pan, et al. Treasure in distribution: A domain randomization based multi-source domain generalization for 2d medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 89–99, 2023

    work page 2023

  9. [10]

    Integrated domain augmentation and structural content reconstruction based multi- source domain generalization on medical image segmentation.IEEE Trans

    Yihan Ren, Yanfeng Li, et al. Integrated domain augmentation and structural content reconstruction based multi- source domain generalization on medical image segmentation.IEEE Trans. on Instrumentation and Measurement, 74:1–14, 2025

    work page 2025

  10. [11]

    Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation.IEEE Trans

    Ling Zhang, Xiaosong Wang, et al. Generalizing deep learning for medical image segmentation to unseen domains via deep stacked transformation.IEEE Trans. on Medical Imaging, 39:2531–2540, 2020

    work page 2020

  11. [12]

    Causality-inspired single-source domain generalization for medical image segmentation.IEEE Trans

    Cheng Ouyang, Chen Chen, et al. Causality-inspired single-source domain generalization for medical image segmentation.IEEE Trans. on Medical Imaging, 42:1095–1106, 2022. 12 Decoupling Wavelet Sub-bands for Single Source Domain Generalization in Fundus Image Segmentation

    work page 2022

  12. [13]

    Rethinking data augmentation for single-source domain generalization in medical image segmentation

    Zixian Su, Kai Yao, et al. Rethinking data augmentation for single-source domain generalization in medical image segmentation. InProc. of AAAI Conf. on Artificial Intelligence, volume 37, pages 2366–2374, 2023

    work page 2023

  13. [14]

    ADA: An adaptive augmentation framework for single-source domain generalization in medical image segmentation

    Runlin Huang, Hongmin Cai, et al. ADA: An adaptive augmentation framework for single-source domain generalization in medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer- Assisted Intervention, (MICCAI), pages 45–54, 2025

    work page 2025

  14. [15]

    Teacher-student instance-level adversarial augmentation for single domain generalized medical image segmentation.IEEE Trans

    Zhengshan Wang, Long Chen, et al. Teacher-student instance-level adversarial augmentation for single domain generalized medical image segmentation.IEEE Trans. on Medical Imaging, 45:764–776, 2026

    work page 2026

  15. [16]

    A hybrid dual-augmentation constraint framework for single-source domain generalization in medical image segmentation.Pattern Recognition, 170:112082, 2025

    Ruofan Wang, Jintao Guo, et al. A hybrid dual-augmentation constraint framework for single-source domain generalization in medical image segmentation.Pattern Recognition, 170:112082, 2025

    work page 2025

  16. [17]

    Devil is in channels: Contrastive single domain generalization for medical image segmentation

    Shishuai Hu, Zehui Liao, et al. Devil is in channels: Contrastive single domain generalization for medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 14–23, 2023

    work page 2023

  17. [18]

    Structure-aware single-source generalization with pixel-level disentanglement for joint optic disc and cup segmentation.Biomed

    Jia-Xuan Jiang, Yuee Li, et al. Structure-aware single-source generalization with pixel-level disentanglement for joint optic disc and cup segmentation.Biomed. Signal Proc. Control, 99:106801, 2025

    work page 2025

  18. [19]

    MaxStyle: Adversarial style composition for robust medical image segmentation

    Chen Chen, Zeju Li, et al. MaxStyle: Adversarial style composition for robust medical image segmentation. In Proc. Int. Conf. Med. Img. Comp. and Computer-Asst. Intervention, (MICCAI), pages 151–161, 2022

    work page 2022

  19. [20]

    Adversarial consistency for single domain generalization in medical image segmentation

    Yanwu Xu, Shaoan Xie, et al. Adversarial consistency for single domain generalization in medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 671–681, 2022

    work page 2022

  20. [21]

    MixStyleFlow: Domain generalization in medical image segmentation using normalizing flows

    Reza Safdari, Mohammad-Ali Nikouei Mahani, et al. MixStyleFlow: Domain generalization in medical image segmentation using normalizing flows. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 376–385, 2025

    work page 2025

  21. [22]

    Train once, deploy anywhere: Edge-guided single-source domain generalization for medical image segmentation

    Shi Gu et al. Train once, deploy anywhere: Edge-guided single-source domain generalization for medical image segmentation. InProc. Int. Conf. Med. Img. with Deep Learning, volume 250, pages 722–741, 2023

    work page 2023

  22. [23]

    Boundary loss for highly unbalanced segmentation

    Hoel Kervadec, Jihene Bouchtiba, et al. Boundary loss for highly unbalanced segmentation. InProc. of Int. Conf. on Medical Imaging with Deep Learning, (MIDL), volume 102, pages 285–296, 2019

    work page 2019

  23. [24]

    Frequency-mixed single-source domain generalization for medical image segmentation

    Heng Li, Haojin Li, et al. Frequency-mixed single-source domain generalization for medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 127–136, 2023

    work page 2023

  24. [25]

    Medical image segmentation via single-source domain generalization with random amplitude spectrum synthesis

    Qiang Qiao, Wenyu Wang, et al. Medical image segmentation via single-source domain generalization with random amplitude spectrum synthesis. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 435–445, 2024

    work page 2024

  25. [26]

    MoreStyle: Relax low-frequency constraint of Fourier-based image recon- struction in generalizable medical image segmentation

    Haoyu Zhao, Wenhui Dong, et al. MoreStyle: Relax low-frequency constraint of Fourier-based image recon- struction in generalizable medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 434–444, 2024

    work page 2024

  26. [27]

    Wavelet convolutions for large receptive fields

    Shahaf E Finder, Roy Amoyal, et al. Wavelet convolutions for large receptive fields. InProc. of European Conf. on Computer Vision, pages 363–380. Springer, 2024

    work page 2024

  27. [28]

    UWT-Net: Mining low-frequency feature information for medical image segmentation

    Pengcheng Zhang, Xiaocao Ouyang, et al. UWT-Net: Mining low-frequency feature information for medical image segmentation. InProc. of Int. Conf. on Medical Image Computing and Computer-Assisted Intervention, (MICCAI), pages 615–624, 2025

    work page 2025

  28. [29]

    Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution

    Xin Deng, Ren Yang, et al. Wavelet domain style transfer for an effective perception-distortion tradeoff in single image super-resolution. InProc. of Int. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 3076–3085, 2019

    work page 2019

  29. [30]

    Learning frequency adaptation for cross-domain medical image segmentation

    Xiaoming Wang, Lei Liu, et al. Learning frequency adaptation for cross-domain medical image segmentation. In Proc. of Int. Conf. on Bio-Informatics and Bio-Medicine (BIBM), pages 1684–1689. IEEE, 2024

    work page 2024

  30. [31]

    Ronneberger, P

    O. Ronneberger, P. Fischer, et al. U-Net: Convolutional networks for biomedical image segmentation. InProc. Int. Conf. Med. Img. Comp. and Computer-Asst. Intervention, (MICCAI), pages 234–241, 2015

    work page 2015

  31. [32]

    J. I. Orlando, H. Fu, et al. Refuge challenge: A unified framework for evaluating automated methods for Glaucoma assessment from fundus photographs.Med. Img. Analysis, 59:101570, 2020

    work page 2020

  32. [33]

    Sivaswamy, SR Krishnadas, et al

    J. Sivaswamy, SR Krishnadas, et al. Drishti-GS: Retinal image dataset for Optic Nerve Head (ONH) segmentation. InProc. of Int. Symp. on Biomed. Img. ISBI, pages 53–56. IEEE, 2014. 13 Decoupling Wavelet Sub-bands for Single Source Domain Generalization in Fundus Image Segmentation

    work page 2014

  33. [34]

    Gamma challenge: Glaucoma grading from multi-modality images.Med

    Junde Wu, Huihui Fang, et al. Gamma challenge: Glaucoma grading from multi-modality images.Med. Img. Analysis, 90:102938, 2023

    work page 2023

  34. [35]

    Seelamantula, et al

    JRH Kumar, C.S. Seelamantula, et al. Cháks.u IMAGE: A Glaucoma-specific fundus image database.Scientific Data, 10:70, 2023

    work page 2023

  35. [36]

    Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations

    Carole H Sudre, Wenqi Li, et al. Generalised Dice overlap as a deep learning loss function for highly unbalanced segmentations. InProc. of Int. Workshop on Deep Learning in Med. Img Anal., pages 240–248, 2017

    work page 2017

  36. [37]

    Deep Residual learning for image recognition

    Kaiming He, Xiangyu Zhang, et al. Deep Residual learning for image recognition. InProc. of Int. IEEE/CVF Conf. on Computer Vision and Pattern Recognition, pages 770–778, 2016

    work page 2016

  37. [38]

    Evolutionary Generative Adversarial Networks.IEEE Trans

    Chaoyue Wang, Chang Xu, et al. Evolutionary Generative Adversarial Networks.IEEE Trans. on Evol. Comp., 23:921–934, 2019. 14

    work page 2019