Abstract

Brain tumor segmentation and detection have advanced significantly with the introduction of multimodal magnetic resonance imaging. However, data privacy concerns restrict most studies to centralized environments, limiting their real-world applicability. While federated learning (FL) offers a privacy-preserving solution for cross-institutional brain tumor research, existing multimodal FL approaches primarily address scenarios wherein clients possess either a single modality or complete missing modality data. These methods fail to account for the modality heterogeneity caused by arbitrary missing modalities, a frequent challenge in clinical practice. To address this issue, we propose FedAMM, a novel FL framework designed for brain tumor segmentation under arbitrary missing modalities. FedAMM incorporates multiple strategies to mitigate discrepancies arising from varying modality combinations across clients. First, FedAMM introduces a unimodal prototype distillation technique during local training to balance the contributions of different modalities. Additionally, the server aggregates multimodal prototypes uploaded by clients to generate cluster centers that represent the global modality distribution, thereby guiding local training toward global optimality. Furthermore, we implement a weighted aggregation strategy based on modality proportions. Experimental results on the BraTS2020 dataset demonstrate that FedAMM outperforms existing methods in handling arbitrary missing modalities, highlighting its strong adaptability to imbalanced and heterogeneous federated systems. The code is available at https://github.com/13sky/FedAMM.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1764_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)

N/A

BibTex

@InProceedings{ShiYuk_FedAMM_MICCAI2025,
        author = { Shi, Yukun and Xue, Meiting and Zeng, Yan and Zhang, Jilin and Wan, Jian and Zhou, Ye},
        title = { { FedAMM: Federated Learning for Brain Tumor Segmentation with Arbitrary Missing Modalities } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {203 -- 213}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper presents an FL framework for multimodal brain tumor segmentation that effectively addresses heterogeneity caused by arbitrary missing modalities. The proposed method balances intra-client modality variations while ensuring model consistency across clients. Additionally, it proposes a modality-weighted aggregation strategy that enhances global performance.

  • 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. Propose a novel FL framework (FedAMM), designed for brain tumor segmentation under arbitrary missing modalities
    2. clear description of the method
  • 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 paper compares against only a few baseline methods, and does not include comparisons with recent state-of-the-art (SOTA) techniques. Without such comparisons, it is difficult to assess how competitive the proposed method truly is in the current research landscape.
    2. The evaluation is conducted on a single dataset, which significantly limits the generalizability of the results. Given the proposed method’s potential applicability across various domains, it would have been more compelling to demonstrate its effectiveness on multiple, diverse datasets.
  • 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?

    1.Limited datasets

    1. Lack of strong baselines
  • 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

    The paper proposed the prototype based method to handle arbitrary missing modalities in Multimodal Federated Learning which aligns in real world scenarios. The proposed methods shows promising results compared to different baselines.

  • 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 scenario of arbitrary missing modalities—where each sample across clients has a different combination of modalities—in multi-modal Federated Learning is both interesting and practica in clinical scenario. There is limited prior work addressing this setting.

    2) The experimental evaluation covers 3 modality heterogeneity settings and includes comparisons with multiple baselines. The results demonstrate that the proposed method outperforms existing 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 methodology section is unclear, especially Equations 6 and 8. In the given setup, out of the four modalities, some samples may contain only three modalities. However, Equations 6 and 8 are defined for two modalities, which limits their generalizability to cases with more than two modalities.

    2) The study is limited to only four clients. It appears that the global prototype creation depends on both the number of clients and the number of samples per client. This raises an unanswered question: how scalable is the approach?

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

  • 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

    There are existing prototype-based methods, such as [1], for handling missing modalities in FL. How does the proposed method differ from such methods?

    [1] Bao, Guangyin, et al. “Multimodal federated learning with missing modality via prototype mask and contrast.” arXiv preprint arXiv:2312.13508 (2023).

  • 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?

    The problem this paper is trying to address is interesting and more aligned with real-world scenarios, especially since very few works exists. The proposed method shows good results. However, the given setup—with only a few clients—may not serve as a suitable benchmark for this task.

  • Reviewer confidence

    Somewhat confident (2)

  • [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.

    My previous concerns, especially regarding scalability, remain unaddressed. However, given the limited amount of work on this problem, the availability of code to support further research and experiments, and the strong alignment of this task with real-world scenarios, I believe this paper deserves acceptance. Please note that PmcmFL does consider arbitrary missing modalities (see Fig. 1a). It is indeed designed for classification problems.



Review #3

  • Please describe the contribution of the paper

    The authors introduce a prototype-based strategy to federated learning for brain tumor segmentation with missing modalities and the proposed method outperforms previous approaches on a public brain tumor segmentation dataset.

  • 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 authors present the first study on arbitrary missing modalities in multimodal federated learning, addressing a more realistic and practical challenge.
    • The authors introduce a prototype-based strategy to balance cross-modality discrepancies and mitigate client heterogeneity, further alleviating distribution mismatches through modality-specific encoder aggregation.
    • The proposed method outperforms previous approaches on a public brain tumor segmentation dataset under diverse missing-modality conditions.
  • 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.
    • Only one dataset is used for validation.
    • The citations and references in LaTeX should be associated via ~\cite{}.
    • The source code cannot be validated during review. Could the authors provide an anonymized link to the source code for review?
  • 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.

  • 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?

    The proposed method is innovative for brain tumor segmentation with missing modalities. However, the authors provided no code in the repository of anonymous GitHub.

  • 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.

    My concerns are resolved after reading the authors’ rebuttal.




Author Feedback

We sincerely thank all reviewers for their time and constructive feedback. We appreciate the recognition of our work’s novelty (R2, R3, R4) and practical value (R2, R4). Below, we provide detailed responses to the main concerns raised. Q1: Limited datasets (R3 & R4). A1: Our method targets multimodal brain tumor segmentation, where BraTS2018 and BraTS2020 are the most widely adopted benchmarks. While earlier studies (e.g., [4,28]) used BraTS2018, recent works (e.g., [5,13,31]) have shifted to BraTS2020 due to its larger size and improved data quality. In line with the latest SOTA method FedMAME [31], we followed the common practice of using BraTS2020 exclusively and simulated 15 modality combinations under 3 levels of missing severity to reflect real-world clinical scenarios. We acknowledge the value of using multiple datasets and will explore this in future work. Q2: Lack of strong baselines (R3). A2: To our knowledge, FedAvg and FedProx are widely used as standard baselines in the FL field. In addition, we compared with FedNorm, FedMM (ICASSP 2024), and FedMEMA (AAAI 2024), which represent the latest multimodal FL methods specifically developed for medical imaging. Our method demonstrates consistent performance advantages over these SOTA approaches, validating its effectiveness in handling arbitrary missing modalities. Q3: Clarification on Equations (6) and (8) (R2). A3: We apologize for the confusion. Equation (6) is used to calculate the imbalance ratio between any two available modalities within a sample. Equation (8) defines the corresponding loss function by traversing all such modality pairs. This formulation is not limited to two modalities. In practice, for samples with ≥3 modalities, we compute the loss across all valid modality pairs. We will revise the text to explicitly clarify this flexibility. Q4: Scalability concern due to the small number of clients (R2). A4: We acknowledge that the main experiments were conducted using four clients, limited by GPU resources. However, our proposed aggregation strategy is inherently agnostic to the number of clients. Both the local distillation process and the global prototype clustering are naturally extendable to more clients. Moreover, the aggregation weights are computed based on modality frequency rather than client ID, ensuring strong scalability. Q5: Difference from existing prototype-based methods (R2). A5: Unlike PmcmFL [1], our method assumes arbitrary missing modalities at the sample level across all clients, leading to intra-sample modality imbalance that significantly affects segmentation tasks. We address this by combining prototype-based learning with local knowledge distillation. In contrast, PmcmFL assumes consistent missing-modality distributions across clients and targets classification tasks, primarily using prototypes to replace missing modalities during training and inference. While both methods adopt prototype strategies, they differ fundamentally in assumptions, tasks, and methodology. Q6: Reproducibility (R3 & R4). A6: We have now included the full source code in the anonymous repository linked in the submission. We ensure complete reproducibility of all experiments. The code will be made publicly available on GitHub upon acceptance. Q7: Citation formatting issues (R4). A7: Thank you for pointing out the citation format issue. We will correct the citation style in the final version. Q8: Misclassification of paper track (MIC vs CAI) (R3). A8: We respectfully clarify that this paper belongs to the MIC track. Our work addresses the realistic and challenging problem of arbitrary missing modalities in multimodal federated brain tumor segmentation. We propose novel methodologies, including a prototype-based strategy and a modality-aware client aggregation mechanism. These contributions are aligned with the core research topics of the MIC track.




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’

    Most of the key concerns presented by reviewers are addressed. This work presents enough technical contributions and meets the bar of MICCAI.



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