Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Histopathological region segmentation faces two main challenges: catastrophic forgetting and the high cost of pixel-level annotations. Recent studies have focused on incremental learning of new categories using low-cost image-level labels. However, the limitations of multiple instance learning (MIL) in modeling instance relationships hinder further improvement in segmentation performance. To address these challenges, we propose the Dual-branch Dynamic Coupling (DDCWISS) network for weakly supervised class-incremental learning in histopathological region segmentation. Our architecture overcomes the limitations of isolated local feature computation in traditional MIL by enabling complementary feature extraction through parallel local representation and global modeling branches. Additionally, we propose a learnable coupling module to ensure effective multi-scale feature fusion, while the dual-path supervision mechanism simultaneously enhances segmentation accuracy. Experiments on the CPATH dataset demonstrate that our method significantly reduces reliance on costly pixel-level annotations for histopathological region segmentation, while effectively alleviating the catastrophic forgetting problem during incremental learning. These results highlight the potential of DDCWISS as a scalable, weakly supervised Class-Incremental paradigm for medical image analysis. The source code is publicly available at: https://github.com/XiaoyanHong24/DDCWISS

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2287_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/XiaoyanHong24/DDCWISS

Link to the Dataset(s)

BCSS dataset: https://drive.google.com/drive/folders/1iS2Z0DsbACqGp7m6VDJbAcgzeXNEFr77 LUAD-HistoSeg dataset: https://drive.google.com/drive/folders/1E3Yei3Or3xJXukHIybZAgochxfn6FJpr WSSS4LUAD dataset: https://drive.google.com/drive/folders/1qTTTaHAp8HOnvxnKi1RXp-bC7sito9DF

BibTex

@InProceedings{HonXia_DualBranch_MICCAI2025,
        author = { Hong, Xiaoyan AND Fan, Jiansong AND Deng, Zhaohong AND Pan, Xiang},
        title = { { Dual-Branch Dynamic Coupling Weakly Supervised Learning for Class-Incremental Histopathological Region Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15969},
        month = {September},
        page = {149 -- 158}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a novel framework for weakly supervised, class-incremental segmentation of histopathological whole slide images (WSIs). It introduces a dual-branch architecture that combines local feature extraction with global context modeling to improve segmentation performance using only image-level labels. Additionally, the method incorporates a learnable coupling module for multi-scale feature fusion and a dual-path supervision mechanism to enhance learning effectiveness while mitigating catastrophic forgetting during incremental learning.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The proposed method effectively models both local features and global contextual dependencies through a dual-branch architecture, which enhances segmentation accuracy under weak supervision.
    • The framework addresses the challenge of catastrophic forgetting in incremental learning, allowing the model to incorporate new classes without retraining from scratch.
    • By relying only on image-level labels for training, the approach significantly reduces the need for expensive and time-consuming pixel-level annotations, which typically require expert pathologists.
    • The study addresses a clinically relevant and scalable problem in digital pathology, aligning well with real-world workflows where WSIs are standard practice.
    • The method is evaluated on a public dataset (CPATH), demonstrating practical applicability and potential for real-world deployment.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The motivation for using class-incremental learning in the context of histopathological segmentation is not well justified. It is unclear whether new tissue classes are commonly introduced in clinical practice, and why catastrophic forgetting is a significant concern in this setting. While incremental learning is a logical need in natural image segmentation, its relevance in histology-based segmentation needs better explanation.
    • The definition and role of classes in this study are not clearly explained. The goal appears to be learning to segment new classes over time, but the nature of these classes and whether they represent distinct tissue types or pathologies is not discussed.
    • The connection between catastrophic forgetting and centralized prediction features is not clearly described.
    • Reference [14], cited to support segmentation using unsupervised methods, seems irrelevant or inadequately justified in the context of this work.
    • The phrase “isolated computation of local features” is vague, and its purported contribution to false positive segmentation errors is not well elaborated.
    • The term “semantic gap” is used but not clearly defined or contextualized.
    • The mechanism by which the DFFC module enforces complementarity and orthogonality between features is not explained. The underlying theory and implementation details are missing.
    • The application of multiple instance learning (MIL) at the pixel level is unclear, as MIL is traditionally used for bag-level or instance-level supervision, not dense prediction tasks.
    • In Equation 1, the functions G and H are undefined, and it is unclear how they facilitate alignment of local and global representations.
    • In Equations 2 and 3, the variable P is not defined, and the function φ is not explained. The computation and role of the mᵢ values (pixel relevance scores) are unclear. The rationale for using Equation 3 as a regularization term is not sufficiently justified.
    • In Equation 5, the same scalar r is applied across all pixels, but it is unclear how this controls individual pixel-level probability contributions.
    • The distinction between zᵢ and f^(i,j) is not clear. If j represents pixel coordinates, it should also be consistently used in the formulation of z.
    • Equation 6 includes pixel-level ground truth, which contradicts the claim of weak supervision (i.e., training with only image-level labels).
    • The background confidence map is mentioned without explanation of how it is generated or used.
    • The process of generating new pseudo-labels y at iteration t is not described in detail, raising concerns about training stability and performance.
    • Equation 8 seems tailored for binary segmentation, yet the paper suggests multiclass segmentation. Also, the meaning of aˢ in this equation is undefined.
    • It is unclear whether the model operates on full WSIs or on patch-wise processing. WSIs are extremely large (e.g., 20,000×20,000 pixels), and this crucial detail should be clarified.
    • Table 1 lacks sufficient details: the evaluation metric is not defined, abbreviations like “Sup,” “Joint,” and “FT”are unexplained, and column headings are ambiguous.
    • The experimental design is limited, with insufficient ablation studies or comparisons to alternative weakly supervised incremental learning approaches.
    • Several cited references are preprints from arXiv. For some peer-reviewed versions being available. For example, Reference 1 has been officially published in CVF Open Access and IEEE Xplore, yet it is cited using its arXiv version.
  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper presents a promising approach that models both local and global features under weak supervision and addresses catastrophic forgetting in incremental learning. It also reduces annotation costs and is evaluated on a relevant public dataset. However, the motivation for incremental learning in histopathology is unclear, and many aspects of the method—such as key definitions, equations, and implementation details—lack clarity. The experimental design is limited, with insufficient comparisons and missing details. These issues make it difficult to fully assess the novelty and practicality of the work.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    Most of my comments will be addressed in the revised version.



Review #2

  • Please describe the contribution of the paper

    This paper proposes DDCWISS, a dual-branch dynamic coupling network for weakly supervised class-incremental segmentation of histopathological regions. The method integrates: A Dual-Branch Encoder (DBE) that extracts both local (MIL-based) and global (Transformer-based) features; A Dynamic Feature Coupler (DFC) for feature alignment and fusion; A Dual-path supervision mechanism combining pseudo-label learning and feature consistency constraints. The method aims to address two challenges: reducing dependence on pixel-level annotations and mitigating catastrophic forgetting during incremental learning.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper addresses an important and underexplored problem—weakly supervised class-incremental learning for histopathology, where both catastrophic forgetting and high annotation costs are critical challenges. The combination of dual-branch feature extraction and dynamic feature coupling is a thoughtful architectural design that aims to simultaneously capture local and global dependencies, particularly suited for the complex structures in whole slide images (WSIs). Evaluation on both overlapping and disjoint protocols using a reasonably sized dataset (CPATH) enhances the credibility and robustness of the experimental validation. A meaningful ablation study is conducted, clearly demonstrating the contribution of each core module (DBE, DFC, and dual-path supervision) to the overall performance.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Limited methodological novelty: While the dual-branch idea is well-motivated, the individual components (MIL extension, sliding window Transformer, feature fusion) are adaptations of existing techniques rather than fundamentally new mechanisms. Insufficient justification for module design choices: Although the proposed architecture combines several modules, the paper lacks a detailed explanation of how each component (e.g., dual-branch feature extraction, dynamic coupling, dual-path supervision) specifically addresses the challenges of weakly supervised class-incremental learning. A stronger connection between problem formulation and architectural design would improve the clarity and impact of the work. Relatively narrow evaluation: Experiments are conducted only on the CPATH dataset. Testing on additional histopathology datasets (e.g., CAMELYON, CRAG) would strengthen claims of generalizability. Clarity of writing: The method description is sometimes verbose, making it difficult to quickly understand the role and interaction of different modules. Additionally, minor writing issues are present, such as spelling issue in “semantic segmatation” (I think it should be “segmentation”), which slightly detract from the paper’s presentation quality.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper addresses an important and underexplored problem—weakly supervised class-incremental learning for histopathological image segmentation—where both catastrophic forgetting and the high cost of pixel-level annotations are critical challenges. The proposed DDCWISS framework thoughtfully combines dual-branch feature extraction, dynamic feature coupling, and dual-path supervision to enhance segmentation accuracy under weak supervision. The evaluation follows standard protocols (overlapping and disjoint) on a reasonably sized CPATH dataset, and meaningful ablation studies are included to demonstrate the contribution of each module. However, the methodological novelty is relatively limited, as the core components (MIL extension, sliding window Transformer, feature fusion) are largely adapted from existing techniques rather than introducing fundamentally new mechanisms. Additionally, the paper lacks a detailed explanation of how the individual modules specifically address the identified challenges, and the evaluation is restricted to a single dataset without broader validation. Minor writing issues slightly detract from the overall presentation quality. Despite these limitations, the paper makes a valuable contribution toward scalable weakly supervised learning in medical imaging. Therefore, I recommend Weak Reject.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After carefully reading the authors’ rebuttal, I find that most major concerns raised by the reviewers—particularly those related to the motivation for class-incremental learning in histopathology, the design rationale for the network components, and the clarity of experimental settings—have been adequately addressed.

    The authors clarified the clinical relevance of introducing new tissue classes over time and justified the need for continual learning strategies in this domain. They also elaborated on how the dual-branch architecture and dynamic feature coupling are designed to mitigate limitations in both local and global feature modeling under weak supervision. While the individual modules are built on existing concepts, the overall integration into a coherent and clinically relevant framework reflects solid engineering novelty.

    Furthermore, the experimental protocol—though still limited to the CPATH dataset—includes both overlapping and disjoint setups, and the ablation study supports the efficacy of the proposed modules. The authors also committed to open-sourcing the code, which would significantly improve reproducibility. Minor concerns regarding notation, phrasing, and clarity were acknowledged and will be revised.

    Given the originality in problem formulation, the practicality of the solution, and the adequacy of the rebuttal, I am inclined to recommend acceptance, contingent on the authors delivering a revised version that addresses clarity and reproducibility as promised.



Review #3

  • Please describe the contribution of the paper

    In this manuscript, the authors tackles two major problems in histopathological datasets, high labeling cost and catastrophic forgetting problems. To resolve the catastrophic forgetting problem, they propose using incremental learning to indicate the new category with the previous known features for old information. Second, to overcome the limitations of the traditional MIL approaches regarding high FP segmentation errors, they adopt weakly supervised continual learning.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The major merits of this paper contains two parts. First, the proposed dual-branch feature extractor aims at capturing cross-regional dependencies to obtain the widen receptive field while capturing local context in histopathology images. Second, the proposed DDCWISS promotes a progressive knowledge integration mechanism via pseudo-supervision distillation.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The overall contributions of the proposed method appear relatively limited within the inadequate comparative studies and ablation studies. While the authors adopt a dual-channel confidence simulation approach, it remains unclear how robust this is to label noise compared to existing pseudo-supervision signal-based methods. Further, the effectiveness of the proposed dynamic label correction mechanism for multi-label learning is difficult to assess based solely on the provided formulations and explanations. Second, to support the clams that the proposed method enables fine-grained feature representation learning for unknown class observations, it would be essential to explicitly evaluate how much forgetting for known classes and how much intransigence for unknown classes. These metrics would provide a more direct understanding of the model’s ability to handle evolving label spaces.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, the proposed method A presents an interesting approach to incremental weakly supervised continual learning. However, the limited explanation and insufficient experimental validation suggest that further clarification and thorough evaluation are necessary to fully substantiate the effectiveness of the proposed approach.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all three reviewers for their constructive comments and appreciations of our strengths such as ‘the method presents a promising approach’ (R1), ‘the method presents an interesting approach’ (R2), and ‘addresses an important and underexplored problem’ (R3). Below, we respond to the main concerns and weaknesses raised.

Reviewer#1 Q7.1: In clinical practice, pathologists need to distinguish different types of pathological tissues at various stages of cancer diagnosis. As the disease progresses and diagnostic criteria become more refined, increasingly subtle tissue types must be identified (e.g., lymphatic infiltration, vascular invasion). Re-annotating all slides and retraining models from scratch is costly. Moreover, focusing only on new classes can cause the model to forget previously learned ones, which severely compromises its clinical utility. Therefore, a joint incremental strategy is essential to mitigate such forgetting caused by the shift between old and new classes. We will refine the corresponding textual descriptions in the revision. Q7.2/9/13/15-16/17-18: We agree on the necessity of adding more details on image processing, category definitions, and abbreviations in Table 1. We will revise and supplement the explanation of formula parameters and unify all notation accordingly. Q7.3/5-8/14: We agree on the importance of phrase clarification in improving the quality of our paper and will include these clarifications in the revision. Q7.19: We sincerely appreciate your valuable feedback. We compared a fully supervised method without pseudo-supervision or incremental learning (Joint) to the WILSON method, and showed that our model, using only image-level labels, achieves performance close to full supervision. We also validated each component through ablation studies. Further textual details will be included in the revision. Q7.4/20: We will revise and update the references accordingly.

Reviewer#3 Q7.1/2: We thank you for the insightful comments. Our study is the first to address weakly supervised incremental learning for histopathological tissue segmentation. We validate the feasibility of our model, especially under clinically relevant conditions where pixel-level labels are unavailable and new classes are introduced. The effectiveness of label smoothing in reducing noise has been discussed in reference 1; we will include relevant details and citations. We will also further elaborate on the role of dynamic label correction. Q7.3: We appreciate your suggestion. We currently evaluate performance by averaging across tasks. To present the results more clearly, we will report per-task, per-class mIOU results instead of the mean values shown in Table 1. This will better reflect the model’s ability to retain previously learned classes and learn new ones.

Reviewer#4 Q7.1/2: While individual components may not seem novel, our contribution is the first systematic framework for weakly supervised incremental learning in histopathological tissue segmentation. To address the challenges of catastrophic forgetting and multi-scale modeling in pathology images, we propose a dual-branch architecture with a dynamic coupling mechanism. To compensate for insufficient pixel-level supervision, we integrate image-level labels and dynamically generated pseudo-labels into a dual weak supervision path that enhances the model’s learning capability. We agree that a clearer motivation would improve the clarity and impact of our work, and we will supplement the textual explanation accordingly in the revision. Q7.3: The CPATH dataset is composed of three different datasets and includes multiple tissue categories. However, CAMELYON and CRAG have limited class annotations. We plan to further annotate these datasets in future work to enhance the generalizability of our model. Q7.4: This will be corrected in the revision. Code Availability(Q9) Our code is open-sourced on GitHub, due to rebuttal limits, the link will be provided in the revision.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top