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

Abstract

Federated learning (FL) on gigapixel whole slide images (WSIs) remains challenged by severe cross-center heterogeneity arising from inconsistent staining protocols, scanning optics, and slide digitization settings. Existing FL approaches typically rely on coarse feature statistics or require iterative transmission of synthetic images, making it difficult to capture domain-specific pathology priors while preserving privacy. To solve these problems, we propose FedGMM-Bag, a generative federated learning framework that performs cross-center distribution alignment through attribute-aware pseudo-bag generation without sharing raw WSIs. Each center constructs a compact multi-attribute Gaussian Mixture Model (GMM) group of its local data, which is communicated once and used to synthesize cross-domain pseudo-bags for augmenting local MIL training. FedGMM-Bag introduces two complementary mechanisms: distribution replay to introduce external morphological diversity and covariance-aligned feature mapping to achieve fine-grained cross-domain alignment. Comprehensive experiments on three heterogeneous multi-center datasets demonstrate that FedGMM-Bag consistently outperforms existing federated and pseudo-bag methods on both seen and unseen centers. These results demonstrate its strong cross-domain generalization and its suitability for privacy-preserving federated pathology under real-world heterogeneity.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2550_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

CAMELYON17 dataset: https://camelyon17.grand-challenge.org/Data/

BibTex

@InProceedings{DenYin_AttributeAware_MICCAI2026,
        author = { Deng, Yingjiao AND Zhang, Qing AND Tang, Hui AND Zhou, Chunhua AND Wang, Jiansheng AND Sun, Li AND Wang, Yan AND Zou, Duowu AND Xiong, Hongkai AND Li, Qingli},
        title = { { Attribute-Aware Federated Learning with Pseudo-Bag Augmentation for Multi-Center WSI Classification } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16891},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This study introduces a generative federated learning (FL) framework for whole slide image classification that synthesizes pseudo-bag across institutions based on attribute aware gaussian mixture model (GMM) summaries coined FedGMM. As opposed to prior works focused on stain-style transfer and harmonization, including synthetic slide distillation; here, the authors posit that GMMs can capture key statistical descriptors for effective transfer across institutions. This includes modeling the variability in patch counts across slides, class wise densities, intra-bag compositions and top-class wise patches, coupled together each local model can sample from the other center’s distributions for augmentation in what is termed distribution replay, and features are further aligned with covariance-based mappings. To validate this approach, an inhouse pancreatic cohorts was used, including public TCGA renal cell carcinoma subtype cohort, and the Camelyon17 dataset. Experiments reveal that FedGMM improves over existing approaches with notable gains, and lower entropy discrepancy across institutions.

  • 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 is well written with a lot of details on the proposed attribute aware GMM construction. The motivations are equally reasonable as FL is a timely and practical framework to address the challenge of privacy in medical AI.
    • The idea of using a generative approach to sampling is intuitive and interesting. The different level of attributes modeled in this work seem reasonable and capture the core of sources of heterogeneity across institutions.
    • The authors perform several experiments to validate their approach i. e. , a private cohort and two public datasets. This supports the possibility of reproducibility. In addition, several recent state of the art approaches were compared, which clearly shows the benefits of FedGMM.
  • 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.
    • While I appreciate the use of GMMs as presented in this work. It is not entirely clear why this is a more principled approach compared to other generative approaches. The authors do not sufficiently motivate this choice.
    • This framework has several moving parts and the selection of components in the GMM is not quantitatively validated or discussed. For instance, for each GMM group 4 components for patch count, 8 for class specific modeling, and so on without any justification. Thus, it is unclear to the reader if there was hyperparameter search or some heuristic for balancing these components.
    • The authors leverage ImageNet pretrained ResNet50 to obtain features from the datasets. While this is reasonable, it is unclear if the results presented are a result of the backbone limited feature representation. Ideally, including recent pathology foundation specific models can clearly contextualize the FL task since these models have been trained on diverse data and less sensitive to the existing FL challenges. This should have been validated.
    • It is unclear if only using pseudo bags-only would provide reasonable gains. This evaluation is missing from the study, as the authors only augment the same number of pseudo bags as the initial training set. Thus, it is unclear how effective sensitive the approach is to the number of bags generated.
  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/en/REVIEWER-GUIDELINES.html
    • In Introduction, L4.”analisys” – > “analysis”
    • The methodological design of this work is similar to Fedbip [3][Chen et al. CVPR25], a discussion on the differences would help clarify differences if any. Including [17] Pettersson et al. Federated Gaussian Mixture Models.
    • In Methods 2.2 Class-Specific Attribute Modeling subsection: [last sentence] Why do the authors describe ‘class specific attribute modeling’ as capturing ‘multi-modal morphological patterns’ ? I do not see how this is multi-modal.
    • In Methods 2.2 Bag-Composition Attribute Modeling subsection: [Last sentence] the authors state that ‘synthetic bags preserve realistic intra-bag composition’; was this validated quantitatively?
    • In implementation details, How did the authors arrive at the number of components 4/8/4/4, no justification about this.
    • There is a discrepancy on the reported scores of HistoFS in Table 1 [RCC] from that of the original paper, this is unclear especially if the same splits were used.
    • The discussion on Fig2.(d,e) is unclear. From a reader perspective, it is hard to understand how to interpret this particular analysis since we do not know which Institutes are annotated and how discrepancy and cross domain heteregeinty is resolved from one cohort to the other.
  • 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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    Overall, this is an interesting study that address a timely challenge in FL for computational pathology. There are merits to the presented methodology, but some of the claims in the paper are not supported by the results and lack some clear discussions on choice of hyperparameters and positioning of their work in the literature. One of the larger issues is the lack of ablations on the existing pathology foundation models. This is important, because if the main concern is heterogeneity and cross cohort variability, it’s only natural to base all evaluations on a strong foundation model rather than a ResNet50.In turn, this will ultimately inform readers whether this problem is still largely unsolved despite having models trained on millions of data points from diverse cohorts. The current score is a reflection of the aforementioned issues and would benefit from further clarifications.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    I thank the authors for the detailed rebuttal and for running the foundation model experiment I requested.

    However, the new CONCH results actually reinforce my concerns rather than resolve it. This makes the overall story weaker in my view. On RCC the gap is essentially saturated; this suggests much of the reported gain in the main paper is tied to using ResNet50 features, and the cross-center heterogeneity problem is largely already addressed by a modern foundation model backbone. The GMM grid search and pseudo-bag ratio sweep are useful additions, and the HistoFS discrepancy clarification is appreciated.



Review #2

  • Please describe the contribution of the paper

    This paper addresses the problem of federated learning for multi-center whole slide image (WSI) classification under strong data heterogeneity and privacy constraints. The authors propose a framework termed FedGMM-Bag, which models each center’s data distribution using an attribute-aware Gaussian Mixture Model (GMM) and leverages these compact statistical representations to generate cross-domain pseudo-bags for augmenting local multiple instance learning (MIL) training.

    Specifically, each center constructs a multi-attribute GMM capturing patch-level distributions, class-conditional features, intra-bag composition, and top-K discriminative regions. These GMMs are shared across centers instead of raw data. Two pseudo-bag generation strategies are introduced: (1) distribution replay via GMM sampling to inject external domain diversity, and (2) covariance-aligned feature mapping to explicitly align feature statistics across centers. The generated pseudo-bags are then used to enhance federated MIL training under a standard FedAvg paradigm.

    The method is evaluated on multiple multi-center WSI datasets, demonstrating improved cross-center generalization compared to several existing federated and pseudo-bag-based methods.

  • 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 tackles federated learning under cross-center heterogeneity for WSI classification, which is a highly relevant and realistic problem. Variations in staining, scanning, and acquisition protocols are well-known challenges in pathology AI, and the inability to share raw WSIs further complicates model training. The problem setting is therefore well-motivated and of clear clinical relevance.

    Proposes a privacy-preserving and communication-efficient framework: Instead of sharing raw data or high-dimensional features, the method communicates compact GMM-based statistical summaries. This is an interesting design choice, as it reduces communication overhead while preserving privacy to a certain extent. Compared to approaches that rely on synthetic image sharing or repeated feature exchange, this formulation is relatively lightweight and potentially more practical.

    Introduces a structured way to model cross-center distribution via multiple attributes: The use of an attribute-aware GMM group (patch count, class distribution, composition, top-K features) is a structured attempt to capture different aspects of WSI variability. This multi-level modeling of distribution (instance-level and bag-level) is conceptually reasonable and aligns with the hierarchical nature of WSI data.

    Combines two complementary pseudo-bag generation strategies: The combination of distribution replay (sampling-based augmentation) and covariance-aligned mapping (feature-level alignment) is a sensible design. The former increases diversity, while the latter attempts to explicitly reduce domain gaps. This dual strategy is a strength compared to methods that rely solely on one type of augmentation.

    Relatively comprehensive experimental validation: The method is evaluated on multiple multi-center datasets, including unseen centers, which is important for demonstrating generalization. Reporting both AUC and entropy (for cross-center consistency) is also a positive aspect, as it reflects robustness rather than only accuracy.

  • 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.

    Major Comments

    1.Weak Methodological Innovation; More Like a Combination of Existing Ideas

    Structurally, this work mainly integrates several existing directions: federated learning (FedAvg), the MIL framework (CLAM), GMM distribution modeling, and pseudo-bag data augmentation. These components themselves are not new; the authors’ main contribution lies in combining them and designing an attribute-aware GMM system and two pseudo-bag generation strategies. This combination is more of a “reasonable engineering design” than a clear methodological breakthrough. For example, GMM is used to model feature distribution, pseudo-bags for enhancement, and covariance alignment for alignment—similar approaches exist in different fields. It is suggested that the authors further clarify: what is the truly new core idea compared to existing federated WSI methods, rather than just improvements on modules?

    2.The rationale and necessity of GMM modeling are not adequately explained.

    The paper’s key assumption is to use a multi-attribute GMM to characterize the central feature distribution of WSI and to perform cross-domain generation based on this. However, several questions remain:

    Is GMM sufficient to represent the complex distribution of high-dimensional WSI features?

    Why choose GMM instead of a stronger generative model (such as VAE, diffusion, etc.)?

    Is there redundancy or repetitive modeling among multiple attributes (patch count, class, composition, top-K)?

    3.Performance Improvement Exists, but Its Persuasiveness is Limited

    Experimental results show that the method does outperform the baseline on multiple datasets, but the overall improvement is small, mostly around 1%. Considering the additional modeling, generation, and alignment steps introduced by the model, this gain is somewhat limited. More importantly, there is a lack of in-depth analysis of the “source of improvement.” For example, which part (GMM modeling, replay, mapping) truly brings about the improvement? Under which data distributions or centrality differences is the effect more significant? It is recommended to add more granular analysis, rather than just reporting overall metrics.

    4.Insufficient Statistical Analysis

    Although the paper reports the mean and standard deviation of multiple experiments, it does not perform significance testing. This is crucial given the current small performance improvement. If the difference is not statistically significant, then the method’s advantage needs to be stated more cautiously. It is recommended to supplement with statistical significance analysis, or at least to discuss the reliability of the results more objectively in the paper.

    5.Comparison Methods Still Have Room for Improvement

    The current baseline includes some classic MIL methods and existing federated learning methods, but overall it is still biased towards traditional methods. Given that this work emphasizes cross-center generalization and distribution modeling, it is recommended to add some more recent and stronger comparative methods, especially those from the last two years in the areas of federated pathology or self-supervised methods. Otherwise, while the current comparison is reasonable, it is insufficient to fully demonstrate the advantages of the method.

    6.The authors demonstrated the alignment effect of pseudo-bag on distributions using t-SNE, but this is more qualitative. Several questions remain:

    Does pseudo-bag introduce noise or misdistribution?

    Could it lead to overfitting the model to a “synthetic distribution”?

    Are the benefits of replay and mapping stable?

  • 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 mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    Overall innovations and improvements in experimental results

  • 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.

    The author answered some of my questions, but due to space limitations, they didn’t address all of them in detail. However, I believe this article meets MICCAI’s requirements, and I recommend accepting it.



Review #3

  • Please describe the contribution of the paper

    This paper proposes FedGMM-Bag, a federated learning framework for multi-center whole slide image (WSI) classification under strong data heterogeneity. The key idea is to represent each center’s data distribution using an attribute-aware Gaussian Mixture Model. These compact GMM representations are shared across centers and used to generate cross-domain pseudo-bags to augment local MIL training and improve cross-center generalization without sharing raw data. Experiments on multiple datasets show improved performance over existing federated and pseudo-bag methods

  • 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.

    1.Representing each center’s data distribution with an attribute-aware GMM is a neat and conceptually sound idea. 2.The paper clearly addresses a realistic challenge in federated pathology: cross-center heterogeneity under strict privacy constraints. 3.The integration of GMM-based distribution modeling with pseudo-bag augmentation is reasonably novel and technically meaningful. 4.The method is evaluated on multiple datasets (PANC, RCC, CAMELYON17), and shows consistent improvements over strong baselines

  • 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.

    1.The number of GMM components is a critical hyperparameter that directly affects distribution modeling quality. However, the paper does not provide ablation or analysis on this factor, making it unclear how robust or sensitive the method is to this choice. 2.Although communication cost is low, the cost of fitting multiple GMMs and generating pseudo-bags across centers is not thoroughly analyzed. 3.The experimental evaluation is conducted on datasets with relatively few centers (typically 2–3 centers) . While the results are promising, this setup may not fully reflect the level of heterogeneity encountered in real-world federated pathology settings.

  • Please rate the clarity and organization of this paper

    Good

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    I lean towards a weak accept for this paper. The work addresses an important and practical challenge in federated computational pathology and introduces an interesting and intuitive GMM-based framework for modeling cross-center heterogeneity. However, the paper has several limitations. The novelty is moderate, and key design choices, particularly the number of GMM components, are not sufficiently analyzed. In addition, the evaluation is conducted on datasets with relatively few centers, which limits the strength of claims regarding scalability and robustness under real-world heterogeneity. Some modeling assumptions (e.g., Gaussian mixtures) may also restrict expressiveness. Finally, the code is not publicly available, which limits reproducibility and makes it harder to verify the reported results and adopt the method in practice.

  • 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.

    I want to thank the author for their effort in preparing the response, i think most of my concerns were alleviated.



Author Feedback

We thank reviewers and the AC for constructive feedback. We address the concerns below. 1@R1R2 Novelty: We respectfully clarify that our contribution is not the isolated use of FL, MIL and GMM, but a WSI-specific federated formulation for hierarchical, bag-structured pathology data. Each center is summarized by an attribute-aware GMM group modeling patch-count variability, class-conditional feature density, intra-bag composition, and top-K discriminative patterns, which enables DR for diversity injection and CM for second-order alignment. Thus, the novelty lies in turning cross-center WSI heterogeneity into a controllable generative representation for privacy-preserving federated MIL. 2@R1R2 GMM: We do not claim GMM is universally stronger than VAE/GAN/diffusion; it better fits our federated WSI-MIL objective: feature-space, class-conditional pseudo-bag synthesis/alignment rather than pixel-level generation. GMM compactly estimates local bag-level feature distributions via mixture weights, means, and covariances, enabling closed-form density estimation, interpretable components, low communication, and component-wise sampling. VAE/GAN/diffusion would require heavier training and extra conditioning to control structured bag attributes. Thus, GMM is a principled choice for efficient bag-level modeling. R2-2 The attributes are sequentially dependent rather than redundant, GMM groups preserve multi-level WSI heterogeneity and naturally support DR/CM. R2-3 Tab.3 already shows progressive gains. Since GMM and DR CM are sequentially coupled in the generation pipeline, they cannot be fully separated into independent ablations. R1-2R4-1 GMM component: 4/8/4/4 was selected by grid search. On RCC, it gives the best average AUC/entropy 0.9821/0.0765, outperforming 2/4/2/2 0.9564/0.1139, 4/8/8/8 0.9812/0.0867, and 4/12/8/8 0.9807/0.0862.These results were omitted due to space limits. 3@R1R2R4 Experimental evidence: R1-3 Foundation model: With CONCH features, FedGMM-Bag still improves over HistoFS (PANC 0.9532 vs. 0.9485; RCC 0.9981 vs. 0.9978), showing the gain is not tied to a weak backbone. We used ResNet50 in the main paper to better isolate the contribution of federated distribution modeling. R1-4 PseMix in Tab1 already serves as a pseudo-bag-only baseline. We also evaluated different pseudo-bag ratios 0.5x/1x/1.5x/2x/2.5x. On RCC, average AUCs are 0.9673/0.9821/0.9818/0.9782/0.9756; on PANC, the best result is 0.8516 at 1.5x. Since larger ratios do not consistently improve performance, we use 1x as the best performance-overhead trade-off. R1-minor: We will clarify these points in the revision. The discrepancy mainly arises from different feature extractors (HistoFS: ViT; ours: ResNet50). R2-3 Gains are consistent across 3 datasets, with lower entropy and one-shot GMM transmission (~0.28 MB/center in RCC), indicating a favorable robustness-efficiency trade-off compared with HistoFS. R2-4 Paired testing against HistoFS gives p=0.04782, indicating significance. R2-5 We already include recent strong FL WSI baselines, especially FedWSIDD(MICCAI 2025) and HistoFS(CVPR 2025), together with representative general FL and MIL pseudo-bag methods. To the best of our knowledge, there are very limited additional recent federated WSI pathology methods that are both directly comparable to our setting and reproducible under the same protocol. R2-6 We agree that t-SNE is only qualitative. Tab3 already shows that DR-only improves robustness quantitatively, CM-only is weaker alone, and the full model performs best, indicating that replay/mapping are stable and complementary rather than causing harmful drift to a synthetic distribution. R3-2 The extra cost mainly comes from one-time GMM fitting and lightweight pseudo-bag sampling on pre-extracted features, rather than a heavy per-round burden. R3-3 We follows standard Fed WSI benchmarks, which are limited in center number; we will clarify this scope and note larger-scale validation as future work.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper addresses an important and practical problem in federated learning for computational pathology, and the proposed GMM-based framework is conceptually interesting with promising empirical results across multiple datasets. However, the main issue is the unclear methodological novelty and justification, as the approach appears to combine existing components (FL, MIL, GMM, pseudo-bag augmentation) without sufficiently clarifying the core innovation or why GMM is the most appropriate modeling choice compared to alternative generative approaches. In addition, the experimental validation is not fully convincing, due to limited analysis of key design choices (e.g., GMM components), lack of statistical significance testing, and insufficient investigation into the source of performance gains. There are also concerns regarding baseline comparisons and evaluation settings, including the absence of stronger or more recent methods (e.g., foundation models), and limited validation under realistic multi-center heterogeneity. The final decision will depend on how convincingly these concerns are addressed.

  • 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.

    The paper addresses an important and practically relevant problem in federated computational pathology: robust WSI classification under severe cross-center heterogeneity and privacy constraints. Reviewers appreciated the clear motivation, communication-efficient framework, and the integration of attribute-aware GMM modeling with pseudo-bag augmentation. The extensive evaluation across multiple datasets, including unseen-center settings, further supports the practical value of the proposed approach. While concerns were raised regarding the degree of methodological novelty, the reliance on an ImageNet-pretrained ResNet50 backbone, and the lack of evaluation with recent pathology foundation models, I do not believe these issues outweigh the overall strengths of the work. The proposed framework provides a practical and effective solution to a challenging real-world problem and demonstrates consistent improvements under heterogeneous federated settings. The authors should nevertheless address the reviewers’ comments in the final version, particularly by clarifying the methodological novelty, discussing the relationship to recent pathology foundation models, and better highlighting the limitations and scope of the proposed approach. Overall, I believe the paper makes a meaningful contribution and merits acceptance.



Meta-review #2

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

    Reject

  • Please justify your recommendation.

    The paper got two accepts and one reject after rebuttal, and I tend to reject this paper. The main reason is that as the fast development of pathology models, a new and powerful foundation model should be used for patch feature extraction. Many problems that exist when using traditional feature extractors (such as ResNet50) no longer exist when using foundation model feature extractors. The results in the rebutal also support this point. When a foundation model is used for the CAMELYON dataset, the genalization performance whill be much higher than the results listed in this paper.



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.

    The paper proposes FedGMM-Bag, a federated learning framework for multi-center whole slide image classification. Based on the reviewers’ feedback and the authors’ rebuttal, I recommend accepting this manuscript. Two out of three reviewers moved to accept post-rebuttal, agree that the major concerns have been addressed, recognizing the method design and evaluation.



back to top