Abstract

Multi-center fMRI data analysis faces significant challenges such as data privacy concerns and data integration issues. Federated learning, as an innovative distributed machine learning approach, enables cross-center collaboration by sharing model parameters instead of raw data. However, existing methods often struggle with improving the robustness and inference efficiency of multi-center fMRI data processing. To address these challenges, we propose a novel hypergraph-guided federated distillation framework(HGFD) for multi-center fMRI data analysis. HGFD utilizes a hypergraph structure to model the spatiotemporal features of brain activity, capturing high-order correlations across brain regions. Furthermore, a hypergraph-based knowledge distillation approach is utilized to transfer high-order structural representations into shallow neural networks, thereby preserving their ability for complex relational inference and significantly enhancing computational efficiency. In the federated learning process, participating centers only need to share the parameters of their shallow neural networks to a central server. Through parameter aggregation, each center’s shallow network can learn the high-order structural information of other centers. Experiments on multi-center fMRI dataset demonstrate that the proposed method not only improves the robustness and consistency of fMRI-based prediction tasks but also achieves efficient and accurate predictions while ensuring data privacy.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2786_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{JinTao_HypergraphGuided_MICCAI2025,
        author = { Jin, Tao and Xu, Yidan and Gao, Yuhan and Sheng, Xichun and Yan, Chenggang and Sun, Yaoqi and Han, Xiangmin and Gao, Yue},
        title = { { Hypergraph-Guided Federated Distillation Learning for Efficient and Robust Multi-Center fMRI Data Analysis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15970},
        month = {September},
        page = {299 -- 308}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a hypergraph-guided federated distillation framework(HGFD) for multi-center fMRI data analysis, which uses a hypergraph structure to model the spatiotemporal features of brain activity, capturing high-order correlations across brain regions.

  • 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 proposes a hypergraph-guided federated distillation framework (HGFD), which uses BrainHGNN as the teacher to guide a lightweight MLP student model. This design effectively transfers high-order brain connectivity knowledge across centers without sharing raw data.
    2. The method achieves good results on the ABIDE dataset (ACC = 0.70 ± 0.05) and outperforms other baseline models. They also provided ablation studies to validate the contribution of each component of HGFD.
  • 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 authors should more clearly explain the process of constructing the hypergraph from functional brain network data. Key steps such as how nodes and hyperedges are defined or selected remain unclear, making it difficult to assess the validity of this component.
    2. Important hyperparameters such as the loss weight (α), temperature (T) in distillation, and the number of clusters (k) used in hypergraph construction are mentioned but not specified. It should be clarified that how these parameters were chosen.
    3. How the ABIDE dataset was divided into different centers for federated learning?
    4. ABIDE is a relatively small dataset for training a complex model involving hypergraph convolutions and distillation. So, could you report the number of trainable parameters or measures taken to avoid overfitting? .
    5. The authors mentioned that the model captures high-order information related to ASD. Could you provide more concrete visualizations or analyses to support this? for example, showing brain maps, attention scores, or key hyperedges would improve interpretability.
  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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?

    This paper presents a framework combining hypergraph modeling, knowledge distillation, and federated learning, but several important issues limit its overall contribution. The construction process of the hypergraph is not clearly explained, and key hyperparameters are unspecified, making the method difficult to interpret or reproduce. The claimed interpretability benefits are not well supported by visualizations or quantitative analyses. Given these concerns, I recommend weak rejection.

  • 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.

    The rebuttal has addressed most of my concerns. I think the paper could be accepted.



Review #2

  • Please describe the contribution of the paper

    The manuscript proposes HGFD, a hypergraph-guided federated distillation framework for multi-center fMRI analysis, addressing data privacy and computational efficiency. It employs a teacher-student architecture where a hypergraph convolutional network (BrainHGNN-Teacher) captures high-order brain interactions, while a lightweight MLP (MLP-Student) distills this knowledge for federated aggregation. Experiments on the ABIDE dataset demonstrate good performance over existing federated and centralized 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.

    This manuscript has following advantages,

    1. Federated learning used for FMRI dataset avoids raw data sharing, critical for sensitive medical data.
    2. Hypergraphs capture multi-region brain interactions would be better compared to traditional graphs.
    3. Distillation used in the manuscript reduces computational/communication costs by transferring knowledge to a shallow MLP (student model).
    4. Performs better compared to the 8 baselines models used for comparison, in terms of accuracy (70%), AUC (0.69), and specificity (73%).
  • 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.

    Few Questions / Comments:

    1. Information regarding dataset distribution (from the AIBDE dataset) between institution was not provided in the manuscript. How was the dataset distributed between institution? Was any specific strategy / approach used for distributing the dataset between institutions?
    2. The dataset distribution was not mentioned in the manuscript. The data split used for training, validation and testing performed for each institution, was not mentioned. It would be good if the authors can add this information to the manuscript.
    3. The ABIDE dataset has a near (1:1) Normal to ASD ratio (468:403). How did the author address potential class imbalance during federated training, especially if individual centers had skewed distributions?
    4. The results in Table 1, F-1 Score show the MLP trained on centralized server achieves better performance than the proposed model. The highlighted results appear to show the proposed model as the best F-1 score whereas MLP model’s F1-score appears to be the highest. Can the author mention, why was the F-1 score from the proposed model highlighted?
    5. Can the authors mention, which federated framework was used for Hypergraph Guided Federated Distillation learning? Was any of the off the shelf frameworks used (such as, Nvidia Flare or Flower) or was the framework developed by the authors for this research study?
    6. The author mentioned, preprocessing details (motion correction, denoising). What were the software tools used for performing the preprocessing of the input images? It would be good to mention specific algorithm and software packages used for each of the pre-processing step?
    7. Can the author explain, why was the AAL template (116 ROIs) chosen over alternatives like Harvard-Oxford, FreeSurfer or other brain atlas?
    8. The author mentioned that hyperedges (each node) are constructed via K-NN on Pearson correlation matrix. What criteria determined k, and how does sensitivity to k impact model performance?
    9. The manuscript does not mention whether the federated learning was conducted on simulated framework or on (real) multiple institutions?
    10. Can the author mention, if the federated learning was conducted in a simulated framework, how many centers were simulated for federated learning, and what distribution was used for data partitioning across centers?
    11. The manuscript doesn’t mention about any differential privacy or encryption used during student model aggregation and for distilling information from teacher model to student model?
  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure 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

    NA

  • 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?
    1. The manuscript’s hypergraph-driven feature extraction demonstrates an advancement over traditional graph methods, validated by an improved Accuracy / AUC gain over GNNs in ablation studies.
    2. As mentioned above there are few points / questions that needs to be addressed by the authors for a strong acceptance of the manuscript.
  • 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.

    Thank you to the authors, comprehensive responses. The clarifications on the ABIDE dataset partitioning, the rationale for K=10 in hypergraph construction, and the distillation parameter settings are well-received. I also appreciate the explanation regarding the FedAvg framework and how privacy is maintained by sharing only student model weights. These details effectively address my previous queries.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is a new framework called HGFD (Hypergraph Guided Federated Distillation) designed for analyzing fMRI data from multiple research centers without sharing the raw, sensitive patient data. It tackles the challenges of data privacy and efficiency in multi-center studies. The core idea is to combine three techniques: 1) Hypergraphs; 2) Federated Learning; 3) Knowledge Distillation.

  • 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 integration of hypergraphs, federated learning, and knowledge distillation specifically for multi-center fMRI analysis is a novel approach. Using hypergraphs to capture higher-order brain connectivity patterns is well-justified, as brain networks often involve multi-region interactions not captured by standard graphs. Combining this with federated distillation addresses both privacy and the computational/communication burden often seen in federated learning.
    2. The method directly targets critical issues in multi-center medical imaging: data privacy (via FL), modeling complex biological patterns (via hypergraphs), and improving efficiency (via distillation to a simpler model for communication).
    3. The paper compares HGFD against a range of relevant baselines, including standard FL methods (FedAvg, FedProx, etc.) and more specialized FL methods for brain imaging (FedNI, FedBrain, etc.). They also perform ablation studies to show the benefit of the hypergraph teacher and the distillation process itself, strengthening their claims about the contribution of each component.
  • 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. While distillation aims to transfer knowledge, using a simple MLP as the student model might still lead to a loss of some complex information captured by the more sophisticated hypergraph teacher model. The ablation study shows the combined system is better than either alone, but doesn’t fully quantify if the distilled MLP truly captures all the essential higher-order information identified by the hypergraph. Is there a trade-off between communication efficiency and model expressiveness here?
    2. The hypergraph is constructed based on K-Nearest Neighbors (KNN) applied to Pearson correlation features. KNN is a standard technique but can be sensitive to the choice of ‘k’ and the distance metric. It’s not guaranteed that KNN based on feature similarity always identifies the most functionally relevant higher-order groupings of brain regions. Is there any exploration of sensitivity to this construction method?
    3. The experiments appear to be solely based on the ABIDE dataset for ASD classification.
  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    This paper tackles an important problem (privacy-preserving, efficient, and robust multi-center fMRI analysis) with an innovative combination of techniques.

    The core idea of using hypergraph distillation within a federated learning framework is novel and well-motivated for brain network analysis. The method directly addresses practical challenges like privacy and communication overhead.

    Potential concerns remain about whether the simple student model fully captures the teacher’s complexity, the sensitivity of the hypergraph construction method, and the validation being limited to a single dataset/condition.

  • 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

Thanks for all the valuable comments. R1, R2, R3: Dataset Partitioning and Preprocessing Pipeline (1) The ABIDE dataset is a publicly available resource for autism spectrum disorder (ASD), collected from 16 international imaging centers, making it a representative dataset for real-world multi-center federated learning research. We partition the ABIDE dataset into three sub-centers, comprising 340 (149 ASD / 191 normal controls), 254 (130 ASD / 124 NC), and 252 (112 ASD / 140 NC) samples, respectively. The data is split into training, validation, and test sets (8:1:1), with five-fold cross-validation using stratified sampling to preserve class balance. Our model achieves consistent performance under this strategy, with an ACC of 0.70 and an F1 score of 0.67, demonstrating classification reliability and class balance (see Table 1, p.7). (2) Data preprocessing follows the pipeline described in I2HBN (MICCAI, 2024), as detailed in Section 2.1 (p.4) and the experimental setup in Section 3 (p.5). We utilize the DPARSF toolbox (Frontiers in Systems Neuroscience, 2010) for preprocessing the raw ABIDE data and adopt the AAL atlas for brain parcellation, a widely accepted template in fMRI analysis(A-GCL, MIA, 2023). Due to space limits, this paper reports only ABIDE experiments; other multi-center validations will be addressed later.

R1, R2, R3: KNN for Hypergraph Construction HGFD employs hypergraph modeling to capture high-order relationships between brain regions. Brain regions are represented as nodes, with the Pearson correlation matrix serving as node features. Each node is connected to its top-K most feature-similar nodes to form hyperedges, thereby constructing the hypergraph structure. The choice of K is closely related to the number of nodes. Based on empirical evidence and extensive experimental validation across a range of candidate values K = {5, 8, 10, 12, 15, 20}, we set that K = 10 offers the optimal balance between capturing sufficient inter-regional interactions and avoiding excessive redundancy in the constructed hypergraph.

R2, R3: Hypergraph Distillation: Parameter Settings and Overfitting Avoidance BrainHGNN consists of two layers of hypergraph convolution (input dimension 116; intermediate dimensions 696 and 232), enabling the model to capture complex functional relationships between brain regions. We distill BrainHGNN into an MLP-Student to balance high-order representation and communication efficiency. The distillation phase uses T=3.0, alpha=0.5, and 50 epochs. To mitigate overfitting given the limited size of the ABIDE dataset, we employed dropout (with a rate of 0.3), weight decay, and an early stopping strategy. Although the MLP-Student may not fully match the teacher’s expressiveness, it greatly enhances cross-center transmission efficiency. Table 2 (p.8) shows that the distilled MLP-Student outperforms the basic MLP, improving accuracy from 0.59 to 0.70, demonstrating the effectiveness of hypergraph distillation learning.

R1: Selection of Federated Framework and Privacy Protection We propose a hypergraph-guided federated distillation framework based on FedAvg (PMLR, 2017), as detailed in Section 2.3. Each center preprocesses data, builds hypergraphs, and distills high-order features via the teacher model, uploading only MLP-Student weights to the server, thus preserving data privacy without sharing raw data.

R1:Misunderstanding of F1 Score Centralized and federated training have different goals. Since the focus of our work is to optimize performance within a federated learning framework, the highlighted results reflect the optimal performance under distributed training conditions.

R3:Visualization or Analysis Our work focuses on improving model representation and lightweighting to enhance multi-center fMRI analysis efficiency. Due to space limits, interpretability and visualization are omitted but will be explored in future studies.




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