Abstract

Deciphering dynamic brain states from functional magnetic resonance imaging (fMRI) requires disentangling temporal network fluctuations from intrinsic subject-level heterogeneity. However, most existing graph learning methods assume an individual-based or homogeneous population, failing to account for latent subject subpopulations that exhibit distinct node-level community structures. To address this, we propose a Multilayer Modularity-aware Graph Autoencoder (MMGAE) utilizing a hierarchical two-stage approach. First, we identify latent subpopulations by clustering Weighted Stochastic Block Model (WSBM) parameters derived from shared prior node communities. Second, these subpopulation labels guide a novel contrastive coupling mechanism within the MMGAE. This mechanism enforces community-constrained edges to preserve local node topology (modular organization) while utilizing a contrastive coupling objective to explicitly differentiate between disparate subject subgroups across network layers. The resulting community partition from each modularity-aware subject embedding is subsequently integrated into a Markov-switching WSBM to identify time-resolved brain states. This pipeline effectively uncovers the specific modular organization defining each subpopulation before analyzing temporal dynamics. We validated the framework on Human Connectome Project task fMRI data. The MMGAE achieved superior consensus clustering performance compared to baseline methods, validating its ability to resolve subpopulation-specific patterns. Furthermore, the model shows significant improvement in identifying distinct brain states characterized by varying levels of network segregation and integration between subgroups. The code is available at https://github.com/KaiJunSee/MMGAE.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3586_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/KaiJunSee/MMGAE

Link to the Dataset(s)

HCP dataset: https://db.humanconnectome.org/data/projects/HCP_1200

BibTex

@InProceedings{SeeKai_Multilayer_MICCAI2026,
        author = { See, Kai-Jun AND Noman, Fuad AND Phan, Raphaël C.-W. AND Ombao, Hernando AND Ting, Chee-Ming},
        title = { { Multilayer Modularity-Aware Graph Autoencoder for Detecting Subpopulation Community Structure in Dynamic Brain Networks } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16894},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This is a strong paper that tackles a genuine methodological gap in dynamic brain network analysis. Most existing methods either assume a homogeneous population or ignore individual heterogeneity, leading to misleading group-averaged community structures. The proposed hierarchical approach—first discovering subpopulations, then using them to guide contrastive learning in a multilayer GAE—is both novel and well-motivated. The integration of modularity maximization with graph autoencoders is clever, and the contrastive coupling across subgroups (Equation 2) is a notable contribution. The experiments are thorough, with clear comparisons to spectral clustering, MGAE, and GenLouvain. The results are striking: perfect consensus clustering within subgroups (Table 1) suggests that the method successfully identifies highly coherent subpopulations. The dynamic state identification also shows improvements, though less dramatic. The paper is well-written, the figures are informative (especially the connectome plots in Fig. 3), and the method is clearly described.

  • 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. Explicitly disentangling subject-level heterogeneity (subpopulations) from temporal dynamics is an important step beyond current multilayer methods that assume a single common community structure across all subjects.
    2. The use of the supra modularity matrix (P) combined with a contrastive node similarity matrix (S) that penalizes embeddings from different subgroups is elegant.
    3. First clustering WSBM parameters to discover subpopulations, then using those labels to guide embedding learning, is logical and effective.
  • 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 reports mean and standard deviation but does not provide p-values or statistical tests (e. g. , paired t-test or Wilcoxon) to confirm that the improvements of MMGAE over GenLouvain (second best) are significant.
    2. Section 2.2 states that “the optimal number of subgroups is selected empirically based on the best state identification performance. ” This is somewhat circular (state identification depends on subpopulations). The authors should provide a more principled criterion (e. g. , silhouette score on WSBM parameters, or cross-validation) or at least show a sensitivity analysis.
    3. How exactly are the WSBM parameters (average and standard deviation of community connectivity magnitudes) computed?
    4. The contrastive coupling uses δ(g_r, g_s) which is ±1 depending on whether subjects belong to the same subgroup. This is derived from the first-stage clustering. There is a risk of double-dipping (using the same data to define subgroups and then to enforce separation). The authors should discuss this and perhaps propose a cross-validation scheme or an alternative (e. g. , using a held-out set for subgroup discovery).
    5. In Table 2, the improvements of MMGAE over baselines in the motor task are small (e. g. , RI 0.7330 vs. 0.7334? Actually MMGAE is 0.7330 vs. GenLouvain 0.7242, so a small gain). In the language task, MMGAE does not outperform MGAE in some metrics (e. g. , ACC: 0.5885 vs. 0.5894 for MGAE). The authors acknowledge that “unsynchronized states across subjects” may cause this. This is honest, but the paper would benefit from a deeper analysis of why the method works better for motor than language.
  • 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.

  • 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 is a strong paper that addresses an important problem in network neuroscience with a novel and sophisticated framework. The methodological innovation is clear, the experiments are extensive, and the results are promising.

  • 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 presents a complex pipeline for identifying latent subpopulations and dynamic brain states from multi-subject fMRI data. They practically propose a novel multi layer modularity based on Graph Autoencoder.

  • 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 method is an integration of classical and deep models, since it combines probabilistic models (WSBM, Markov-switching) with deep learning (graph autoencoder). The idea of modularity-aware representation learning is probably new. The authors use multilayer modularity constraints to ensure that embeddings preserve community structure, which is central in brain network analysis.

  • 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.

    PERFECT RESULTS (1.0000 ± 0.0000) across all metrics, this is too good to be true. Table 1 shows the proposed MMGAE achieving perfect consensus clustering (AMI=1.0, ARI=1.0, ACC=1.0, F1=1.0) with zero standard deviation on both motor and language tasks. But then Table 2 shows close to chance performance. Something is really off. This is statistically impossible for real fMRI data with 496/485 subjects and 116 ROIs unless either the method is trivial (e.g., assigning all nodes to the same community), or there is data leakage between training and evaluation Biology is noisy, I have never seen anything working 100% unless it was so simple that did not require the described method.

    I might have missed, but how many communities were detected?

    There is a circular reasoning which might be a bit speculative: The adjacency matrix for each subject is constructed as Ar = ϕ(Ac ⊙ Xr) where Ac is from subgroup-level community detection. This means community structure determines the graph topology rather than the data determining communities. I don’t see relevant aspects for the ML side: How was the train/test split implemetned? Was the k-means initialization done with fix seeds?

    There is sometimes a redundancy issue with multilayer clustering versus a unique layer clustering on the same data. An ablation-like analysis like in Blanco et al. Journal of Computational Science 2024 would be useful.

  • 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.

    (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?

    Results are suspicious and contradictory. There is a lack of details for reproducibility.

  • 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 paper still needs improvements, but it can be accepted for MICCAI, there are some valid contributions



Review #3

  • Please describe the contribution of the paper

    The paper proposes a unified framework (MMGAE-WSBM) for analyzing dynamic functional brain networks that explicitly accounts for inter-subject heterogeneity by first identifying latent subject subpopulations and subsequently incorporating these subgroup assignments into a multilayer graph autoencoder with a contrastive coupling mechanism. This approach aims to preserve community-level modular structure while enhancing separability across subject groups, and integrates the learned representations into a Markov-switching WSBM to model time-resolved brain states. The framework combines elements from graph representation learning, community detection, and probabilistic temporal modeling to jointly address subject variability and dynamic connectivity in fMRI data.

  • 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 presents a technically novel integration of multilayer graph autoencoding, contrastive subgroup-aware learning, and probabilistic temporal modeling to jointly address subject heterogeneity and dynamic brain network analysis.

  • 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 experimental validation raises several concerns: (1) the reported perfect clustering performance (e.g., 1.0000±0.0000 in Table 1) lacks sufficient explanation or analysis, making it unclear whether this reflects meaningful structure or potential methodological bias (e.g., circularity in subgroup definition or evaluation); (2) there is a notable inconsistency between strong clustering results and relatively modest gains in downstream state identification (Table 2), with overlapping variability suggesting limited practical improvement over simpler baselines; (3) the practical advantage of the proposed complex framework over existing, simpler methods is not clearly demonstrated; (4) hyperparameter tuning appears insufficiently rigorous for both the proposed method and baselines, raising concerns about fairness of comparison and sensitivity of results; and (5) clarity issues in presentation (e.g., illegible text in Fig. 1) hinder 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 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    The methodological idea is interesting but the experimental validation is insufficiently justified, with unclear interpretation of perfect clustering results, limited demonstrated advantage over simpler baselines, and lack of rigorous hyperparameter analysis.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    Thank you for the detailed rebuttal and clarifications. I appreciate the additional explanations, particularly regarding the transductive setting and the distinction between clustering consensus and downstream state identification. However, I will maintain my original decision, as several concerns remain only partially addressed. In particular, I still find the experimental validation and practical significance insufficiently convincing, including the interpretation of the perfect clustering results, the relatively modest downstream gains despite strong clustering performance, the unclear advantage over simpler alternatives, the limited rigor and detail of the hyperparameter tuning and comparison procedures, and presentation issues affecting clarity and interpretability.



Author Feedback

Transductive Learning & Cross-Validation (R1, R2): We clarify that the proposed MMGAE-WSBM is an unsupervised, transductive learning framework. The goal is to discover the optimal latent representation and modular organization for a specific population rather than generalizing to unseen subjects. In neuroimaging, community detection typically relies on the full dataset to find a robust consensus via single-subject or group-average analysis, neither of which relies on supervised train/test splits. Our framework follows this transductive standard but moves beyond the homogeneous population assumption by identifying latent subpopulations with distinct topologies, preventing the group average from obscuring meaningful inter-subject heterogeneity.

Informational Independence & Leakage (R1, R3): To prevent information leakage, our hierarchical architecture maintains informational independence between stages. Stage A identifies latent subpopulations using time-averaged WSBM parameters (global statistical profiles) and learns modular embeddings. Conversely, Stage B uses time-resolved sliding-window data to identify dynamic states. Because Stage A utilizes the temporal mean while Stage B utilizes transient windowed dynamics, the subgroup assignments do not leak the temporal regime changes evaluated in the state identification task.

Cluster Consensus & State Identification Results (R2, R3): Table 1 reports intra-group cluster consensus, not a comparison against external ground-truth labels. A score of 1.0 indicates that for every subject r assigned to a group, the model’s learned embeddings produced the exact same node-to-community assignments Ĉ_r as the group consensus partition Ĉ. This perfection is a mathematical result of the contrastive coupling mechanism (δ(g_r, g_s)), which minimizes intra-group variance in the embedding space to find stable modular organizations. This proves the coupling successfully denoises individual embeddings into the coherent group representation. While raw fMRI is noisy, MMGAE uses community-constrained edges to filter this noise. Notably, the state identification results in Table 2 are not perfect, accurately reflecting the expected biological noise and asynchrony in dynamic states.

Adjacency Matrix & Graph Topology Circularity (R2): Adjacency construction is a form of informed regularization, not circularity. A_c provides a structural prior (subpopulation community edges), while X_r contains subject-specific functional connectivity. The operator ⊙ ensures we only retain edges that are both statistically strong in the subject’s data (X_r) and consistent with the subgroup’s modular organization (A_c). This suppresses idiosyncratic noise that often plagues individual fMRI graphs.

Methodological Contribution & Effect Size (R1, R3): While we report mean and standard deviation per standard requirements, the consistent performance advantage of MMGAE-WSBM over baseline methods across a large cohort (R=496/485) demonstrates stable methodological gains. As a methodological work, our primary contribution is the innovative ability to disentangle subject-level heterogeneity and dynamic states. We believe the evaluation serves as a rigorous proof of concept sufficient to validate this innovation. Consistent with conference focuses, we reserve extensive post-hoc modeling for future work.

Implementation & Reproducibility (R2, R3): To ensure stability, we utilized k-means++ initialization with 20 replicates, ensuring results represent global optima for each subject. Hyperparameters for MMGAE-WSBM were empirically tuned by monitoring objective loss convergence and maximizing cluster consensus. We provided all critical hyperparameters (γ, ξ, η, β) and observed the model to be robust to variations. Similarly, all baseline methods were rigorously tuned following their original implementations to ensure a fair comparison. Code and preprocessed data will be made available after acceptance.




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.

    Overall, the reviewers appreciated the novel methodology that combines classical ML and DL techniques. However, there are concerns with the experimental results. Specifically, the cross-validation setup is not adequately explained, and the perfect results suggest either overfitting or information leakage. I tend to agree with Reviewer 2 that perfect performance is highly suspicious, particularly given the limitations of fMRI data. There is also some inconsistency between the perfect clustering performance and relatively modest gains in the downstream tasks. The authors should clearly address these concerns in their rebuttal.

  • 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.

    While the authors speak to the major reviewer critiques, their method seems to parse and use the same fMRI data in multiple ways. For example, the claim that using static measures in Stage A is “informationally independent” of the sliding window analysis in Stage B requires either empirical or theoretical validation. On the surface, there is most certainly an indirect relationship between time-averaged and sliding-window metrics. A similar validation is required for the claim that the group-level A_c is not circularly related to subject-specific functional connectivity, as the former can be derived from the collection of latter measures. I am also unconvinced about the explanation for the perfect results. Overall, I believe that this paper requires further analyses to strengthen its underlying claims.



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.

    The paper propose a novel hierarchical framework for dynamic brain network analysis that disentangles subject-level heterogeneity from temporal dynamics. In the rebuttal, the authors clarified that the framework operates in a transductive setting. Although R3 maintained a Reject due to lingering doubts about practical significance over simpler baselines, R1 and R2 found the clarifications satisfactory. Ultimately, the methodological innovation of integrating graph autoencoders with modularity constraints is deemed valuable and warrants acceptance.



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.

    After carefully reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors directly addressed key concerns including clarifying the transductive learning setting, explaining the informational independence between stages to prevent leakage, distinguishing intra-group cluster consensus (Table 1) from noisy dynamic state identification (Table 2), justifying adjacency construction as informed regularization rather than circularity, and providing implementation details for reproducibility.



back to top