Abstract

Resting-state functional MRI (rs-fMRI) has been increasingly employed to aid in brain disorder diagnosis and reveal the pathological mechanisms underlying neurological diseases. However, clinical applications of current automated diagnosing techniques remain constrained by the complexity of brain topology structures and the high costs associated with expert-derived biomarkers. Recent advancements in research have shown that Graph Contrastive Learning (GCL) holds substantial potential for overcoming these challenges and improving diagnosis accuracy. Nevertheless, existing GCL-based methods predominantly generate a static augmented brain network during graph augmentation and primarily focus on the semantic differences between the original and augmented views. To address above issues, we introduce MGCL-DA (Multi-view Graph Contrastive Learning with Dynamic Self-aware and Cross-sample Topology Augmentation), a novel framework aimed at generating two complementary augmentations of brain networks that account for both individual-specific and inter-subject functional heterogeneity, as well as dynamically regulating the update of augmented views to optimize the transmission of discriminative features. Furthermore, we incorporate multi-view graph contrastive learning with min-max constraints, applying distinct contrastive constraints based on specific augmentation semantics to enable pairwise comparisons between the original network and its two augmented views. Extensive experiments on the MDD dataset demonstrate the superior classification performance of MGCL-DA over several state-of-the-arts.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2205_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{ZhaHao_Multiview_MICCAI2025,
        author = { Zhang, Hao and Liu, Xiaoyun and Huang, Shuo and Yuan, Yonggui and Zhang, Daoqiang and Zhang, Li},
        title = { { Multi-view Graph Contrastive Learning with Dynamic Self-aware and Cross-sample Topology Augmentation for Brain Disorder Diagnosis } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {530 -- 540}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a framework called MGCL-DA, which combines dynamic self-awareness and cross-sample topology augmentation with multi-view graph contrastive learning for the diagnosis of brain diseases.

  • 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. This paper addresses the limitations of existing graph contrastive learning methods, which rely on static augmentations and single-view contrast, by introducing a dynamically adjusted topological enhancement mechanism. This mechanism adaptively adjusts the updates of augmented views using dynamic weight factors based on changes in loss and accuracy during training. In theory, this approach can better capture the dynamic evolution characteristics of brain networks.
    2. The paper proposes a dynamic adjustment topological enhancement mechanism to overcome the limitations of static augmentations and single-view contrast in existing graph contrastive learning methods. This mechanism adapts the updates of augmented views using dynamic weight factors, based on the changes in loss and accuracy during training, which theoretically enhances the ability to capture the dynamic evolution of brain networks.
    3. The experimental section demonstrates the superior performance of MGCL-DA on the MDD dataset.
  • 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 proposes dynamically adjusting the weight factor ωt (Equation 3) by modifying the update weight of augmented views based on the rate of change of loss and accuracy. However, the update of ωt depends on the objective loss function of MGCL-DA and the training accuracy for MDD classification, which have inconsistent dimensions and numerical ranges (e.g., as shown in Figure 2(a), the loss changes within the range [0.2], while accuracy ranges from [0, 1]). Directly adding these and mapping through a Sigmoid function may lead to unstable weight updates, especially during the early stages of training when the loss fluctuates significantly. This rapid change in weights could disrupt the model’s convergence, as observed in Figure 2(a), where the weight factor ωt decreases quickly during the first 10 iterations, but the validity of this phenomenon has not been verified.
    2. Additionally, the adjustment of the weight factor ωt based on the rate of change in both loss and accuracy leads to another potential issue. From my observation, the loss term seems to gradually decrease with iteration t, while the accuracy term increases with iteration t. This could potentially cancel out the effects of each other. Moreover, Figure 2(a) shows a rapid decline in the weight factor between iteration 0 and 10, followed by a steady trend with some sharp fluctuations, which may indicate that the effects of the loss and accuracy terms are counteracting each other.
    3. The paper proposes three contrastive loss terms (Equations 4, 5, and 6) targeting the original view (the initial brain network directly constructed through rs-fMRI, i.e., functional connectivity matrix), self-perceived augmented view (Z-pooling operation on the subject’s brain network, i.e., combining MaxPool and AvgPool to suppress noise and redundant information), and cross-sample augmented view (spatial and temporal rotation of the original brain network combined with equal-weight fusion of cross-sample features to capture functional connectivity differences between subjects). However, their combination might introduce optimization conflicts. For instance, enforcing differences between the self-perceived and cross-sample augmented views (Equation 6) could be problematic, as these two views are generated through different mechanisms, and such enforced differences might disrupt their semantic information. The authors did not experimentally validate the necessity of this approach.
    4. Contrastive learning models typically require large datasets, and 171 samples may be insufficient to train a complex dynamic augmentation mechanism. As shown in Table 1, the accuracy standard deviation of the proposed MGCL-DA is as high as ±3.98%, suggesting significant model performance fluctuations, which may be due to overfitting or insufficient data.
    5. I am particularly curious whether this iterative updating brings information gain with each iteration, especially during early iterations when the augmented view noise is high. Is the dynamic mechanism effectively filtering out this noise? The authors only validate the final classification performance but do not provide a quantitative analysis of intermediate results. For example, methods like significance maps or gradient backpropagation visualizations could be used to show which brain regions or connections are emphasized in the augmented views.
  • 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
    1. It is recommended to further explore how to unify the scales of loss and accuracy, or use other mechanisms to balance them, in order to avoid instability during the early stages of training. Additionally, the authors should further analyze the relationship between loss and accuracy changes and explore how to adjust the mechanisms to prevent any counteracting effects between them.
    2. The three contrastive loss terms (original view, self-perceived augmented view, and cross-sample augmented view) may lead to optimization conflicts. In particular, enforcing a difference between the self-perceived and cross-sample views (Equation 6) may not be reasonable, as these two views are generated through different mechanisms, and forcing a discrepancy could undermine their respective semantic information. It is recommended that the authors further validate whether the combination of these three contrastive views introduces unnecessary conflicts, and consider whether the loss weights need to be adjusted accordingly.
    3. It would be beneficial to include a visual analysis of intermediate results to provide further insight into the model’s performance and behavior.
  • 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?

    Although the proposed framework is innovative, it suffers from issues such as unstable weight factors, optimization conflicts, and insufficient data, which affect the reliability and persuasiveness of the paper.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 proposes MGCL-DA, a multi-view graph contrastive learning framework with dynamic self-aware and cross-sample topology augmentation for rs-fMRI-based brain disorder 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.Dynamic Topology Augmentation is proposed to iteratively adjust self-aware (individual-specific) and cross-sample (inter-subject) augmented brain networks using an adaptive weight factor (ωt​) tied to training loss and accuracy, enabling dynamic refinement of discriminative features. 2.The Multi-view Graph Contrastive Learning with Min-Max Constraints enforces distinct semantic comparisons between original and augmented views—minimizing divergence for self-aware augmentation while maximizing diversity for cross-sample augmentation—to capture both intra- and inter-subject heterogeneity.

  • 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 MDD dataset comprises only 171 subjects, raising concerns about generalizability. Without validation on external datasets, it is difficult to assess whether the learned representations can generalize well to other populations or clinical settings. 2.While performance metrics are strong, the paper does not elucidate how the model identifies MDD-related biomarkers (e.g., connectivity patterns or ROIs), limiting clinical utility.

  • 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?

    Good methodological contribution and paper organization.

  • 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 #3

  • Please describe the contribution of the paper

    The Authors developed a new deep learning neural network MGCL-DA (Multi- view Graph Contrastive Learning with Dynamic Self-aware and Cross- sample Topology Augmentation). This method was applied on rs-fMRI sequences using a dynamic, dual augmentation strategies moving beyond static graph augmentations. The applicability of the method was applied by demonstrating improved diagnostic accuracy on the MDD (Major Depressive Disorder) dataset, outperforming several state-of-the-art 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.
    • Strong methodological analysis, with appropriate description of the methodology
    • Comparison between the proposed method with other state-of-the-art methods
    • Ablation study
  • 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.
    • Limited applicability of the method in the clinics. Even though the method provide high accuracy, no explaination has been extracted from the graphs to drive clinical research to the analysis of specific neural connections.
  • 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.

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

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    • Strong methodological analysis, with appropriate description of the methodology
    • Comparison between the proposed method with other state-of-the-art methods
    • Ablation study
  • 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

We sincerely thank the Area Chairs and Reviewers (R1, R2, and R4) for their detailed feedback, and are honored to receive the “provisional acceptance.” The constructive comments greatly improved the clarity and quality of our manuscript. In response, we provide the following clarifications:

Q1: The fluctuation of the iterative update factor ωₜ during training affects the stability of the augmented views (R1). A1: As shown in Fig. 2(a), both the objective loss and the dynamic weight factor ωₜ stabilize at low values in the later training stages, indicating that the overall optimization is stable and well-converged.

Q2: Elucidate the design issues of the iterative update mechanism of the update factor ωₜ (R1). A2: To address potential instability from the scale mismatch between loss and accuracy, the dynamic factor ωₜ (Eq. 3) is computed based on the relative change rates of these metrics. Specifically, we compute the relative changes in both loss and accuracy between adjacent epochs, and then average them to represent the dynamic performance shift. This combined indicator is further passed through a Sigmoid function, which naturally maps the value into a bounded range, ensuring smooth and stable adjustment of the weight factor during training. To regulate the influence of this dynamic weight, we introduce an upper bound ωₘₐₓ to limit excessive fluctuation in early training due to unstable model initialization, and a lower bound ωₘᵢₙ to avoid feature degradation caused by diminishing weight updates in later stages. This design ensures that the feature contributions from all epochs are appropriately preserved and balanced throughout the training process.

Q3: The validity and effectiveness of the loss term introduced in Eq. 6 for improving the model’s performance (R1). A3: Regarding the contrastive loss defined in Eq. 6, our model performs topology augmentation of brain networks from two distinct perspectives: individual specificity and cross-subject variability. These two augmented views inherently capture semantically different functional patterns. Given this semantic divergence, introducing a constraint to maximize their discrepancy is both reasonable and necessary, as it enhances the model’s ability to capture heterogeneity in brain connectivity. To verify its effectiveness, we have included an ablation model (MGCL-DA-III) in Table 2, where the loss component is removed by setting β = 0. The results show a noticeable drop in diagnostic performance across multiple metrics, confirming the contribution of the discrepancy-based contrastive loss to model optimization and its necessity within our framework.

Q4: The concern regarding the model’s generalization ability and the identification of discriminative functional connectivity (R1,R2,R4). A4: Due to space constraints, we are unable to include additional disorder-specific biomarkers or extended generalization results. Nevertheless, our model demonstrates consistently strong performance across cross-validation folds, indicating stable optimization and robustness. In addition, discriminative functional connectivity can be identified by computing Pearson correlation coefficients between brain region pairs in the augmented networks, enabling the localization of disorder-relevant patterns.

We will further extend our work in the journal version to address these concerns more fully by: (1) providing a rigorous mathematical proof for the multi-view contrastive constraint using mutual information theory; (2) performing external validation on two brain disorder datasets to evaluate model generalization; and (3) visualizing discriminative functional connectivity in the augmented networks to enhance interpretability and clinical relevance. We kindly invite the reviewers to refer to that version for more detailed insights.

Once again, we express our gratitude to the Area Chairs and the Reviewers for their invaluable feedback.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top