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

Abstract

Whole-Slide Image (WSI)-based prognosis prediction often struggles to scale to rare cancers due to limited patient cohorts, creating a critical generalization bottleneck. Furthermore, directly sharing knowledge across different cancers via multi-task learning frequently causes negative transfer due to significant prognostic heterogeneity. To address these challenges, we propose a novel Federated Expert-Routing Framework (FedERF) tailored for cross-cancer knowledge transfer. During pre-training, we utilize Federated LoRA across diverse source cancer clients to construct a shared cross-cancer expert pool. This extracts generalizable morphological priors from diverse multi-cancer cohorts, providing robust initialization for data-scarce clients. In the target-adaptive fine-tuning phase, we introduce a dynamic routing mechanism. Instead of applying a rigid global model, this mechanism adaptively selects and aggregates the most relevant experts from the pool based on the target domain’s specific semantics. This routing strategy precisely injects complementary knowledge while isolating conflicting features, effectively mitigating negative transfer. Extensive experiments demonstrate that our FedERF successfully overcomes the data scarcity dilemma and achieves superior prognostic performance on target cancers. Code is available at https://github.com/ddayzzz/MICCAI2026-FedERF-public.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3342_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/ddayzzz/MICCAI2026-FedERF-public

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WanShu_CrossCancer_MICCAI2026,
        author = { Wang, Shu AND Li, Junjian AND Wang, Jianxin},
        title = { { Cross-Cancer Expert-Routing Knowledge Transfer in Federated Prognosis Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes FedERF for cross-cancer prognosis prediction with data-scarce target cancers. The method has two stages. First, it performs federated expert learning across multiple source cancer cohorts using LoRA modules to build a pool of cancer-specific experts on top of a shared backbone. Second, during target adaptation, it uses a layer-wise and patch-level routing mechanism to selectively combine the most relevant source experts for the target cancer. The method is evaluated on five TCGA cohorts with three MIL backbones under leave-one-out transfer and few-shot settings.

  • 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. The problem is important and clinically relevant. Prognosis prediction for rare or data-scarce cancers is exactly where transfer learning can help, but negative transfer across cancer types is a real concern.
    2. The combination of parameter-efficient expert learning and target-adaptive routing is well matched to the problem setting and more targeted than learning a single monolithic pan-cancer model.
    3. The empirical study is broad: five target cancers, three MIL backbones, few-shot evaluation, and component ablations.
  • 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.Some of the result claims are too strong. FedERF is the best method on average, but not uniformly across every target and backbone. For example, on ABMIL-LIHC it is slightly below Scratch, and on WiKG-UCEC it does not exceed ROUPKT. Claims such as ‘consistently outperforms’ should therefore be softened. 2.The algorithmic schedule needs to be written more clearly. Eq. (1) reads like a joint optimization of the global backbone and expert parameters, while the text describes first training w_G and then freezing w_G to learn the LoRA experts. The two-stage procedure should be made explicit.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    I lean positive on this paper because the problem is meaningful, the method is technically interesting, and the empirical evaluation is fairly broad.

  • Reviewer confidence

    Very confident (4)

  • [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



Review #2

  • Please describe the contribution of the paper

    This paper proposes FedERF, a federated cross-cancer prognosis framework for WSI-based survival prediction that aims to improve transfer to data-scarce target cancers while mitigating negative transfer from heterogeneous source cancers. The method consists of two stages: federated LoRA-based expert learning to construct a cross-cancer expert pool, followed by target-adaptive layer-wise routing to selectively combine relevant experts during fine-tuning on the target cancer. The main claimed advancement over prior FL-MIL approaches is that FedERF avoids compressing all pan-cancer knowledge into a single monolithic global model and instead dynamically routes complementary cancer-specific priors for the target domain.

  • 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.The paper addresses an important and clinically relevant problem: prognosis prediction for rare or data-scarce cancers from WSIs, where both limited target data and cross-cancer heterogeneity create major challenges for robust modeling. 2.The proposed method is well motivated conceptually. The decomposition into federated expert construction and target-adaptive routing aligns naturally with the stated goal of leveraging transferable pan-cancer priors while reducing negative transfer from irrelevant cancers. 3.FedERF shows strong overall empirical performance, achieving the best reported average C-index across all three backbones, which suggests that the framework is broadly effective rather than tied to a single architecture. 4.The ablation study is informative and supports the proposed design. Removing experts, averaging experts without routing, or replacing pretrained experts with random ones all degrades performance relative to the full model, indicating that both the expert pool and the adaptive routing mechanism contribute meaningfully.

  • 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 evaluation setting and the use of the term “federated” would benefit from clearer explanation. The experiments appear to model federation across cancer cohorts in a leave-one-out cross-cancer transfer setting, but the manuscript does not clearly discuss what kind of federation is being modeled and how this setting should be interpreted relative to more conventional FL scenarios. 2.The novelty relative to prior cross-cancer transfer and routing-based methods could be positioned more sharply. While the paper argues that monolithic global models cause negative transfer and introduces expert routing as the solution, the precise methodological distinction from prior transfer/routing paradigms is not articulated as clearly as it could be. 3.Some baseline definitions are under-explained. In particular, Scratch serves as an important reference in Table 1, but the manuscript does not clearly define whether it means target-only training from random initialization, from pretrained backbone weights, or under exactly matched training budgets, making the comparison harder to interpret precisely. 4.The choice of ROUPKT as a key baseline is not sufficiently motivated in the paper’ s earlier positioning. Although it is included in experiments and adapted to the current setting, it is not introduced in the motivation/background discussion in Sec. 1, so the reader lacks context for why this is a particularly relevant comparison and how it conceptually relates to FedERF. 5.The empirical validation would be stronger with more statistical support and implementation detail. The paper reports means and standard deviations and includes useful ablations, but it does not provide stronger significance analysis or enough architecture-specific details about expert/router insertion and training configuration to fully assess robustness and reproducibility.

  • 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.

  • 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?

    This paper tackles an important problem in computational pathology and proposes a technically reasonable framework that combines federated expert learning with target-adaptive routing to improve cross-cancer prognosis prediction. The empirical results are overall strong, with consistent average gains across multiple cohorts and three different MIL backbones, and the ablation/few-shot results provide additional support for the proposed design. At the same time, the paper would be stronger if it clarified the intended interpretation of its federated setting, sharpened the novelty positioning relative to prior routing-based transfer work, and more clearly explained baseline definitions and choices such as Scratch and ROUPKT. Overall, I view this as a borderline positive submission with meaningful technical merit, but one that still needs clearer positioning and presentation to justify a stronger accept recommendation.

  • 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



Review #3

  • Please describe the contribution of the paper

    The authors propose FedERF, a federated learning MIL framework designed to learn cancer-specific characteristics by leveraging multiple experts. The framework focuses on overcoming the customization limits of traditional federated MIL by enabling more localized and specialized feature representation across different clinical sites.

  • 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.
    • Unlike conventional slide-level routers, the proposed Layer-wise Dynamic Routing Mechanism operates at the patch level across different layers. This allows for a more “target-tailored” expert routing, which is particularly effective for capturing the heterogeneous nature of cancer across different clients in a federated setting.
    • While the individual components may not represent entirely new concepts, the study presents a sophisticated integration of existing principles to significantly advance the performance ceiling of federated MIL. It can be interpreted as a well-executed, progressive approach to local model customization.
    • The authors effectively demonstrate the versatility of FedERF by testing it across various backbone architectures. This empirical validation proves the framework’s utility as a plug-and-play solution that is not restricted to a specific feature extractor.
  • 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 overview of the training workflow in Figure 1 is difficult to follow. The mixture of “Phase 1/2” labels with separate numbered markers (1, 2, 3) and multi-directional connections creates visual confusion. A more linear or modular layout would be necessary for readers to grasp the system architecture clearly.
    • The mathematical formulations, such as Equation (5), suffer from an excessive number of superscripts and subscripts, hindering immediate comprehension. Rather than dedicating space to standard LoRA formulations (Equation 3), the authors should focus on simplifying and clarifying the notations for their core proposed mechanisms.
    • There is a discrepancy in terminology between the “Preliminary” section (Phase 1 & 2) and the rest of the manuscript. This inconsistency, combined with a lack of detailed implementation parameters, makes the study difficult to reproduce without the original source code.
    • The process by which a target client initializes from global weights and retrieves from the expert pool is inadequately explained. Without a clear description of the retrieval logic, the claim that the model becomes “cancer-specific” lacks sufficient justification.
    • Beyond numerical performance gains, the paper provides little evidence of cancer-specific learning. The authors should include qualitative evaluations or case studies where patch-level and layer-level routing show distinct, interpretable differences to support their claims.
    • The experimental results indicate high sensitivity to hyperparameter changes. This suggests that the model may require extensive site-specific optimization, which could limit its practical scalability in diverse federated environments.
  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    While the attempt to refine routing at the patch and layer levels is noteworthy, the paper provides insufficient empirical evidence to support its claims regarding “Cross-cancer Knowledge Transfer.” Furthermore, the poor readability of the overall workflow (Figure 1) and mathematical notations (Equation 5) makes it difficult to grasp the core contribution, suggesting that the manuscript requires significant refinement in both clarity and validation.

  • 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

N/A



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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.

    Both reviewers acknowledged the paper’s strong clinical motivation, solid empirical evaluation, and novel methodology. While reviewer 3’s concerns regarding clarity and reproducibility were valid, these minor issues could be addressed in the final version.



back to top