List of Papers Browse by Subject Areas Author List
Abstract
Brain networks exhibit a modular community structure that varies across individuals and neurological conditions. However, existing self-supervised learning (SSL) methods often overlook this heterogeneity, relying on generic masking strategies that fail to capture subject-specific functional organization. We propose BrainPICM, a self-supervised framework for brain network analysis via progressive individualized community-aware masking. BrainPICM formulates ROI-to-community mapping as a progressive unbalanced optimal transport process, yielding soft assignments and per-ROI confidence scores. Guided by these confidence estimates, a curriculum-style masking strategy gradually incorporates low-confidence, potentially pathological regions into training, enabling the model to learn both stable modular structures and individual variations. Additionally, a deviation-aware aggregation module quantifies functional reorganization by measuring mass redistribution relative to a population template, enhancing interpretability and downstream prediction. Experiments on three fMRI datasets (ABIDE-I, ADHD-200, ADNI) show that BrainPICM consistently outperforms state-of-the-art supervised and SSL methods in diagnostic accuracy, indicating that explicitly injecting modular community structure into masked modeling yields more functionally consistent and generalizable representations. The source code for this approach will be released at https://github.com/Hrychen7/BrainPICM.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3365_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Hrychen7/BrainPICM
Link to the Dataset(s)
ABIDE-I dataset: https://fcon_1000.projects.nitrc.org/indi/abide/abide_I.html
ADHD-200 dataset: https://fcon_1000.projects.nitrc.org/indi/adhd200/
ADNI dataset: https://adni.loni.usc.edu/data-samples/adni-data/
BibTex
@InProceedings{CheHai_Progressive_MICCAI2026,
author = { Chen, Hairui AND Yang, Yanwu AND Cao, Jianfeng AND Peng, Hanyang AND Ye, Chenfei AND Ma, Ting},
title = { { Progressive Self-supervised Learning with Individualized Community Assignment for Brain Network Analysis } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes BrainPICM, a self-supervised learning framework for brain network analysis that incorporates individualized community structure into masked modeling. The method formulates ROI-to-community assignment as a progressive unbalanced optimal transport process, yielding soft assignments and ROI-level confidence scores. Based on these, a curriculum-style community-aware masking strategy is designed. Experiments on three datasets demonstrate improvements over both supervised and self-supervised 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 paper addresses an important limitation of existing self-supervised learning methods, namely the neglect of individualized community structure in brain networks. 2.Incorporating soft community assignment into the masking process is a reasonable and intuitive design, aligning well with the modular organization of brain networks. 3.The method is evaluated on multiple datasets with a range of 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.Although the integration of UOT-based assignment and community-aware masking is interesting, the overall framework remains an extension of existing masked modeling paradigms. The contribution is incremental rather than fundamentally novel. 2.The improvements over strong baselines are relatively modest, which limits the practical impact of the proposed method. 3.The paper lacks deeper analysis of key design choices, for example, the advantage of UOT over simpler clustering methods.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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 presents a technically sound and reasonably motivated extension of self-supervised learning for brain networks by incorporating individualized community structure. While the performance gains are limited, the proposed framework is well-designed and may inspire further work in community-aware representation learning.
- 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 author appropriately addressed the concerns.
Review #2
- Please describe the contribution of the paper
The paper proposes a self-supervised framework for brain network analysis that explicitly incorporates individualized functional community structure into masked modeling. Specifically, it introduces a progressive, confidence-aware masking strategy guided by soft ROI-to-community assignments obtained via an unbalanced optimal transport formulation. In addition, it models subject-specific functional reorganization by aggregating deviations from a population-level community template, and integrates these features with ROI-level representations for downstream diagnosis.
- 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.Well-motivated problem setting addressing inter-subject heterogeneity in functional brain networks, which is an important and realistic challenge in fMRI-based disease analysis. 2.Proposes a community-aware self-supervised learning framework that explicitly incorporates functional network structure into representation learning. 3.Introduces a confidence-aware, progressive masking strategy, which leverages soft ROI-to-community assignments to guide masking in a curriculum learning manner, rather than using random masking. 4.Employs an unbalanced optimal transport formulation to obtain soft and individualized ROI-to-community assignments, allowing modeling of uncertainty and potential outlier regions. 5.Designs a deviation-based aggregation module to capture subject-specific functional reorganization relative to a population template, providing complementary information to ROI-level features. 6.Effectively combines local (ROI-level connectivity) and global (community-level deviation) features for downstream classification. 7.Demonstrates consistent improvements over a range of supervised and self-supervised baselines across multiple public datasets (ABIDE, ADHD-200, ADNI). 8.Includes ablation studies validating the contribution of key components (pretraining, assignment, deviation), supporting the overall framework design. 9.Evaluates robustness under different atlas choices and analyzes the impact of the masking schedule, showing reasonable stability 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 ROI-to-community assignment (Q) is not directly supervised or evaluated. It is unclear whether the learned assignments are meaningful or better than simpler alternatives such as k-means or spectral clustering. 2.The method relies on a fixed atlas-based community template (e.g., Yeo networks), which may limit flexibility and introduce bias. The approach does not demonstrate robustness to different or data-driven community definitions. 3.Pretraining and downstream evaluation are performed on the same dataset, so the method does not demonstrate transferability or generalization across datasets. Modest performance gains: The improvements over strong baselines are relatively small, and the paper does not provide sufficient analysis to explain when and why the method works better. 4.While the masking design is a key contribution, the paper does not provide detailed analysis showing how it changes learned representations compared to random or attention-based masking.
- 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.
(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?
weakness: 1.The ROI-to-community assignment (Q) is not explicitly validated, and it remains unclear whether it captures meaningful or biologically relevant functional structures. 2.The use of unbalanced optimal transport lacks sufficient justification, as the paper does not demonstrate why it is preferable to simpler alternatives such as softmax-based or clustering-based assignments. 3.The proposed masking strategy lacks theoretical support, particularly the design choice of prioritizing high-confidence ROIs for masking, which is counterintuitive and not well justified. 4.Pretraining and downstream evaluation are conducted on the same dataset, without cross-dataset validation, limiting the evidence for generalization and making it closer to in-dataset optimization rather than a transferable foundation model. 5.The baseline comparisons are insufficient, as the method is not compared against simpler assignment strategies (e.g., k-means or direct similarity-based assignment), making it difficult to assess the necessity of the proposed components. 6.The selection of the initial parameter ρ₀ relies on dataset-specific tuning, which raises concerns about potential overfitting and lack of robustness across datasets. 7.The deviation matrix D is claimed to reflect functional reorganization, but this interpretation is weakly supported; it remains unclear how well D alone performs for classification or whether it provides meaningful biological insights. Advantage: 1.The use of predefined functional communities (e.g., Yeo networks) as prior knowledge provides a structured way to incorporate domain-specific information into the learning process, which is meaningful for brain network analysis. 2.Leveraging community-level priors to guide representation learning helps introduce biologically informed inductive bias, potentially improving robustness compared to purely data-driven approaches. 3.The deviation matrix (D) offers a novel way to quantify subject-specific differences relative to a population template, enabling the modeling of functional reorganization at the network level. 4.The idea of aggregating ROI-level deviations into community-level interactions provides an intuitive and interpretable feature representation for downstream tasks. 5.Combining ROI-level features with deviation-based global features introduces a multi-scale perspective, which is a reasonable and potentially effective design choice.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes BrainPICM, a self-supervised framework for functional connectivity learning. Instead of adopting an ROI-wise masked autoencoder, BrainPICM employs a cluster-based masking strategy, in which the clusters are dynamically updated through progressive unbalanced optimal transport. A Brain Transformer is then trained via masked functional connectivity reconstruction. For downstream prediction tasks, BrainPICM leverages both the pretrained Transformer representations and the subnetwork deviation map.
- 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 idea of using subnetwork-level masking for self-supervised learning is interesting and novel. The introduction of the subnetwork deviation map for both interpretability and downstream prediction is valuable and well-motivated.
- 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.Functional connectivity matrices computed using Pearson correlation are symmetric, meaning that the i-th column vector is identical to the i-th row vector. Therefore, masking only at the row level may lead to information leakage. The authors should clearly state whether symmetric masking or another leakage-prevention strategy is used to address this issue. 2.The cost function or assignment rule for the virtual communities is not sufficiently clarified. It is currently unclear whether the assignment is determined by a thresholding mechanism or by another objective function, as in conventional UOT-based formulations. This part should be described more explicitly. 3.The Brain Transformer appears to be optimized jointly with the clustering module, while the initial ROI features may be noisy or randomly initialized. The authors should provide evidence that such randomness does not adversely affect the self-supervised learning dynamics. For example, it would be useful to examine whether the clustering assignments converge consistently under different random initializations. 4.The optimization of Equation 1 appears to be highly nontrivial, yet the implementation details are largely omitted. Given that this optimization is central to the proposed method, the paper should provide a more complete description of how it is carried out in practice.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The paper may suffer from a serious information leakage issue in the self-supervised learning setup. In addition, several key aspects of Section 2.3 remain unclear, even though this section contains the main technical contribution of the work. These missing details make it difficult to assess both the validity and reproducibility of the method.
- 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 authors have satisfactorily addressed my concerns, and I appreciate their comprehensive response.
Author Feedback
We thank all reviewers for their valuable comments. For R1: Q: the advantage of UOT. A: We additionally replaced UOT with a simpler clustering-based assignment strategy. Although the clustering variant still outperformed several baselines (64.5% ACC on ADHD), our full UOT-based framework achieved the best performance (67.3%), demonstrating the benefit of modeling unmatched transport mass and ROI-level uncertainty for progressive masking.
For R2: Q: The assignment Q is not explicitly validated A: The learned assignment matrix Q achieved a higher silhouette coefficient (0.39) than the original prior partition (0.10), indicating better subnetwork separation. Q: The role of the deviation matrix A: The deviation matrix D can be interpreted as a functional reassignment pattern, characterizing how ROI assignments shift from the population-level template to individualized subnetworks. In ADHD, we observed noticeable reassignment deviations among the DMN, DAN, and SMN, suggesting abnormal cross-network reorganization. Q: Dataset-specific tuning of ρ₀ A: Model performance remains stable with ρ₀∈[0.6,0.9]. Since ρ₀ controls transported ROI mass, larger ρ₀ benefits neurodevelopmental studies, while smaller ρ₀ benefits neurodegenerative studies. Q: lacks comparison to simpler assignment strategies A: We compared our method with a standard OT assignment strategy without progressive masking pretraining (“w/o pretrain”). We further evaluated MAE-style masking (62.61%), EvolvedMask-style masking (63.39%), and direct hard-label partition masking (65.15%). Our full method achieved the best performance (67.37%), validating the proposed progressive confidence-aware masking strategy. Q: Why does masking high-confidence ROIs first help learning A: Our masking strategy follows curriculum learning. High-confidence ROIs correspond to stable community structures and provide easier reconstruction targets during early training. As ρ increases, more uncertain ROIs are progressively incorporated, leading to smoother optimization. Empirically, the opposite strategy that masks low-confidence ROIs first achieved only 64.1% ACC on ADHD, lower than our high-confidence-first strategy (67.37%). Similar progressive masking designs were also adopted in TARDRL[26] and EvolvedMask[7]
For R3: Q: Potential leakage from symmetric FC matrices A: Thank you for raising this important concern. This issue is related to the structured dependency of symmetric FC matrices rather than label leakage, which is a common structured dependency strategy in masking-based connectivity frameworks. We adopts a row-wise masking strategy, where each row x_i is treated as an ROI-wise connectivity profile and reconstructed at the whole-profile level. Similar ROI-wise connection profile modeling is also used in BrainNetTF[12] and BrainMass[23]. We further tested stricter symmetric masking by zeroing corresponding column positions and observed only marginal performance changes (67.37% → 66.9%). Q: The virtual community formulation is unclear A: The virtual community is implicitly modeled by the unmatched transport mass in UOT rather than thresholding. The transported mass to K subnetworks is constrained by ρ(t), while the remaining mass 1−ρ(t) is assigned to the virtual community. The residual δ_i reflects how poorly ROI i matches predefined subnetworks, and the confidence score is computed as q_i=1−δ_i Q: Optimization details of Eq(1) A: Eq. (1) is optimized using an entropy-regularized UOT solver. After introducing the virtual community, the soft assignment matrix is iteratively updated via Sinkhorn-style matrix scaling: Q̂ = diag(a) exp(−C/ε) diag(b), where C = −log softmax(P). Q: Random initialization may destabilize assignment learning A: To improve stability, we maintain an EMA feature bank instead of directly using instantaneous encoder outputs. This reduces assignment fluctuation during early training. Stable convergence was observed across runs with low variance in Tab1
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.
Reviewers have highlighted several positives with the paper, such as clear motivation, interesting and novel solution design.
In the rebuttal, I would suggest the authors focus on providing clarification to the main criticism’s raised (1) the optimization/assignment rules they use for Eq.1 (2) potential for data-leakage in the masking strategy and justifying the design (3) interpretation of the deviation matrix as functional reassignment and (4) parameter choices made.
- 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.
All reviewer concerns have been sufficiently addressed in the rebuttal and all reviewers are leaning towards accepting the paper.
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.
While the authors respond to each concern, the rebuttal includes several performance metrics. As far as I can tell, many of these metrics are derived from new ablation experiments that were conducted during the rebuttal process. Such experiments are not allowed at MICCAI. Removing all of this information substantially weakens the rebuttal, as many points remain unaddressed.
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.
This paper contributes a compelling self-supervised framework for brain network analysis by coupling progressive representation learning with individualized community assignment, and the contribution feels both technically meaningful and well aligned with the domain. The rebuttal is concrete, offering targeted clarifications and improving confidence that the proposed components matter. With strong overall reviewer support after rebuttal and a convincing methodology, this paper should be accepted
