Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Functional Magnetic Resonance Imaging (fMRI) provides crucial insights into brain activity but presents challenges due to its high-dimensional, dynamic, and noisy nature. Traditional graph-based approaches for fMRI analysis often rely on predefined correlation structures, which may not accurately reflect the true underlying functional connectivity. To address this limitation, we propose a graph learning framework that dynamically constructs brain graphs and leverages Spline Convolutional Neural Networks (SplineCNN) for localized spatial feature extraction. Our model introduces a Learner Graph module, which infers graph structures in a data-driven manner, mitigating the reliance on predefined connectivity measures. The SplineCNN and Multi-Graph Convolution modules capture fine-grained spatial dependencies, offering improved adaptability to the heterogeneous nature of fMRI data. Additionally, we incorporate contrastive learning to align learned representations with domain-specific priors to improve generalization. Experimental results demonstrate that our approach outperforms traditional correlation-based methods in neurological disorder classification. The proposed framework provides a principled, adaptive solution for learning graph representations from fMRI, enhancing generalizability and robustness in brain network analysis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4507_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/MNFuad/AGMGC

Link to the Dataset(s)

ABIDE dataset: http://preprocessed-connectomes-project.org/abide/ Rest-Meta MDD dataset: https://rfmri.org/REST-meta-MDD

BibTex

@InProceedings{NomFua_Adaptive_MICCAI2025,
        author = { Noman, Fuad AND Phan, Raphaël C.-W. AND Ombao, Hernando AND Ting, Chee-Ming},
        title = { { Adaptive Graph Learning with Multi-Graph Convolutions for Brain Disorder Classification } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {56 -- 65}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a graph learning framework that dynamically constructs brain graphs. The method consists of three main modules: the Graph Learner module, the SplineCNN and Multi-Graph Convolution modules, and a contrastive learning component along with additional loss functions. The authors claim that the proposed framework provides a principled and adaptive approach for learning graph representations from fMRI data, enhancing both interpretability and robustness in brain network analysis.

    The claimed contributions are as follows:

    • A novel graph learning framework.
    • Integration of SplineCNN and multi-graph convolutions to improve diagnostic accuracy.
    • A contrastive learning strategy to enhance the robustness and interpretability of the learned representations.
  • 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 is well written and follows the defined structure; it is easy to read and follow.
    • Good competitive methods comparison based on the provided code, nice ablation experiments, showing the influence of each part.
    • The idea of augmenting the learned graph with a graph constructed based on the correlations looks interesting and meaningful, and according to the experiments, improves the results.
    • It’s an interesting idea to integrate contrastive loss to match the distributions of both graphs.
  • 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.
    • Novelty: The graph learning module is not particularly novel; it closely follows the ideas presented in Cosmo et al. (Latent-Graph Learning for Disease Prediction), with some minor modifications such as kNN-based thresholding and application in a different setup. Could the authors clarify the major methodological differences between the mentioned work and the proposed module?

    The overall setup with different losses is not entirely clear:

    • Why is it necessary to align both the learned and augmented graphs? It feels like this pushes the learned graph toward the correlation-based graph, yet the results and ablation tables suggest that they still preserve complementary information. This is quite confusing.
    • What is the influence of the clustering loss? It would be interesting to see some ablation results with and without this component in future.
    • In Fig. 1, it is labeled as “Graph Learner” followed by “KNN Graph,” which is a bit confusing. The KNN graph is a predefined graph and not a learned one. It seems the authors refer to kNN thresholding, which retains the top 15% strongest connections. It might be helpful to revise the figure to reduce confusion.
    • Also, what about the choice of k? How does it influence the graph? And how can one be sure that important connections are not being discarded? In this case, it appears to be a form of hard thresholding — clarification would be appreciated.

    Minor comments:

    • “These methods commonly rely on population-level graph structures that incorporate information from both training and test sets, introducing a bias that inflates performance metrics and undermines fair evaluation [13].”
    • “Such dependence on test-time access to training data distributions is impractical for real-world clinical applications, where unseen data must be processed independently, and increases the risk of overfitting to specific training cohorts [16].” - This is not always true. There are inductive and transductive methods, and the former can operate without relying on training data at test time. The authors might want to revisit population-based methods introduced after the Parisot et al. paper.
    • It is worth revising some of the wording, as there is no clear evidence provided to support the claim of enhanced interpretability — the current phrasing feels too strong.
  • 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

    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?

    It feels more like an engineering-focused paper, though it does incorporate some methodologically interesting ideas, such as data-driven augmentation and feels like novel combinations of loss functions.

  • 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



Review #2

  • Please describe the contribution of the paper

    This paper presents a novel graph-based deep learning framework, AGMGC, for fMRI-based brain disorder classification. The method features a dynamic graph construction module (Learner Graph) that learns connectivity directly from raw fMRI time series, overcoming the limitations of static, pre-defined correlation-based graphs. Additionally, the architecture integrates SplineCNN and a Multi-Graph Convolutional Network (MGCNet) to capture localized and global structural dependencies. A contrastive loss aligns learned graphs with domain priors to improve generalization. The framework demonstrates superior performance on two datasets (ABIDE and REST-meta-MDD) for ASD and MDD classification, outperforming multiple state-of-the-art 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.

    Adaptive Graph Learning: The Learner Graph module dynamically infers subject-specific adjacency matrices, avoiding dependence on fixed or population-level graphs. Multi-Graph Integration: The framework combines data-driven and knowledge-based graphs, leveraging both learned structure and prior neurobiological connectivity. SplineCNN-based Local Feature Extraction: The use of spline-based convolutional filters enables spatially localized yet flexible learning on graph data.

  • 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.

    Generalization Across Modalities: The model is designed solely for fMRI; integration of multimodal data (e.g., sMRI, DTI) could further strengthen the clinical applicability.

  • 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 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

    It would be helpful to visualize learned graph structures and compare them with standard FC matrices for better interpretability.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

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

    The proposed AGMGC framework addresses key limitations of previous graph-based models in neuroimaging analysis by offering adaptive graph construction, multi-graph integration, and a biologically grounded contrastive learning component. The comprehensive evaluation on ASD and MDD datasets demonstrates state-of-the-art performance. While some limitations (e.g., dynamic modeling, multimodal extension) are acknowledged, the current contributions are methodologically sound and practically impactful.

  • 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



Review #3

  • Please describe the contribution of the paper

    This work proposes a novel framework called Adaptive Graph Learning with Multi-Graph Convolutions (AGMGC) for fMRI-based brain disorder diagnosis. This framework addresses the limitations of traditional graph-based methods that often rely on predefined and static correlation structures, which may not accurately represent the underlying functional connectivity in the brain. AGMGC introduces a data-driven Learner Graph module that dynamically constructs brain graphs directly from raw fMRI time series, reducing the reliance on fixed connectivity measures. This adaptively learned graph is then combined with a data-driven graph based on partial correlation to form a multi-graph representation. The framework further integrates a Spline Convolutional Neural Network (SplineCNN) module to extract localized spatial features from these graphs and a Multi-Graph Convolution (MGCNet) module to capture fine-grained dependencies for robust classification. To improve generalization, the AGMGC framework incorporates contrastive learning to align the learned representations with domain-specific priors. The fusion of features from SplineCNN and MGCNet is achieved through a dynamic gated Mixture of Experts (MoE) approach, allowing for adaptive combination of their contributions. The paper demonstrates that this approach outperforms traditional correlation-based methods and other state-of-the-art graph neural networks in classifying neurological disorders like Autism Spectrum Disorder (ASD) and Major Depressive Disorder (MDD) using the ABIDE and REST-meta-MDD datasets. The authors emphasize that their framework provides a principled and adaptive solution for learning graph representations from fMRI data, enhancing interpretability and robustness in brain network analysis.

  • 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.

    Adaptive Graph Learning: A key strength is the introduction of the Learner Graph module that dynamically infers brain graph structures directly from fMRI time series in a data-driven manner. This overcomes the limitations of traditional methods that rely on predefined correlation measures, which might not accurately capture subject-specific topological structures inherent in brain networks.

    Multi-Graph Integration: The framework effectively utilizes multi-graph convolutions by combining the adaptively learned graph with a data-driven graph based on partial correlation. This allows the model to leverage different perspectives on functional connectivity, potentially capturing a more comprehensive representation of brain networks.

    • Localized and Global Feature Extraction: The integration of SplineCNN and MGCNet modules enables the capture of both localized spatial features and global dependencies within the brain networks. SplineCNN focuses on localized features, while MGCNet processes the multi-graph representations to capture broader interactions.
    • Dynamic Feature Fusion: The use of a dynamic gated Mixture of Experts (MoE) for fusing the embeddings from SplineCNN and MGCNet is a significant strength. This allows the model to adaptively combine the contributions of these two modules based on the input data, leading to a more robust and informative final representation.
    • Contrastive Learning for Enhanced Generalization: The incorporation of a contrastive loss helps align the learned representations with domain-specific priors, which is shown to enhance generalization across datasets. This addresses a critical limitation of population-based GCN approaches that can suffer from biases due to test-time access to training data distributions.
    • Superior Empirical Performance: The experimental results on two large and widely used fMRI datasets (ABIDE and REST-meta-MDD) demonstrate that the proposed AGMGC framework outperforms traditional machine learning methods and several state-of-the-art graph neural network models in neurological disorder classification. The consistently high accuracy, F1-score, and AUC values across both datasets highlight the effectiveness of the approach
  • 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.

    Complexity and Computational Cost: The AGMGC framework involves several complex modules, including the Learner Graph, SplineCNN, MGCNet, and MoE fusion. While the paper demonstrates strong performance, the computational cost and training time compared to simpler methods are not explicitly discussed. This could be a practical limitation for real-world deployment.

    • Hyperparameter Sensitivity: The performance of deep learning models is often sensitive to the choice of hyperparameters (e.g., learning rate, weight decay, kNN parameter, contrastive loss weight). The paper mentions the specific hyperparameters used, but a detailed analysis of the sensitivity of the results to these choices is not provided.
    • Interpretability Depth: While the conclusion mentions interpretability analysis, the excerpts do not provide specific details or examples of these findings. A deeper dive into how the model’s learned graphs and feature importance scores relate to known neurobiological mechanisms would strengthen the paper.
    • Generalizability Across Diverse Datasets: The model is evaluated on ABIDE and REST-meta-MDD. While these are prominent datasets, the generalizability to other fMRI datasets with different acquisition parameters, preprocessing pipelines, and subject demographics needs further investigation.
    • Handling of Temporal Dynamics: The current framework constructs static graphs from fMRI time series. The conclusion suggests extending AGMGC to dynamic fMRI data to capture temporal connectivity fluctuations as future research. The inability to directly model temporal dynamics in the current version could be considered a limitation.
    • Reliance on Preprocessing: The paper mentions standardizing the BOLD time series. The performance of graph-based methods can be influenced by the fMRI preprocessing steps. The robustness of AGMGC to different preprocessing choices is not explored in detail.
    • Biological Validation: While the interpretability analysis hinted at biologically plausible patterns, a more thorough validation of the learned connectivity patterns with established neuroscience literature or independent biological markers would be beneficial.
    • Writing structure: The reader comprehension of this manuscript could be improved by representing text into figures or tables where necessary e.g section 3.1 could be represented as a simple table for better comprehension.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

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

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    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?

    The paper addresses a significant challenge in fMRI analysis – the limitations of static and predefined connectivity measures for brain disorder diagnosis – by proposing a novel and adaptive graph learning framework. The main contribution of the AGMGC framework, with its dynamically learned graphs, multi-graph convolutions, SplineCNN for localized feature extraction, dynamic feature fusion, and contrastive learning, represents a significant advancement in the field. The major strengths of the paper are compelling, including the data-driven approach to graph construction, the effective integration of multiple graph representations and convolutional techniques, the use of contrastive learning for better generalization, and the consistently superior performance demonstrated on two large, real-world fMRI datasets. The ablation studies further support the importance of the individual components of the framework and their synergistic effect. The indication of improved stability compared to baselines is also a valuable characteristic for clinical applications. While there are potential weaknesses, such as the complexity of the model and the need for further investigation into hyperparameter sensitivity, interpretability depth, and generalizability across more diverse datasets [see section 3], these do not overshadow the significant contributions and promising results presented in the paper. The authors themselves acknowledge the potential for future work, such as extending the framework to dynamic fMRI data and exploring multimodal integration, indicating an awareness of ongoing research directions.

  • 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



Author Feedback

  1. Interpretability Concerns (R1, R2, R3): We acknowledge that Section 4’s (Conclusion) claim of biologically plausible ROI importance scores lacks specific evidence in the current manuscript, and references to enhanced interpretability and visualization of learned graphs versus FC matrices were premature. We will revise Section 4 to remove these claims and reframe interpretability analysis, including ROI validation and graph visualization, as future work to align with neuroscience findings, ensuring claims match presented results.
  2. Generalization and Modalities (R1, R2): Our focus on fMRI addressed challenges in functional connectivity, achieving strong performance on diverse datasets (ABIDE and REST-meta-MDD). Their varied acquisition and demographics support AGMGC’s robustness. Multimodal integration (e.g., sMRI, DTI) is noted as future work in Section 4, aligning with reviewers’ suggestions.
  3. Computational and Practical Concerns (R2, R2): AGMGC’s complexity is justified by its adaptive modules, yielding superior results. While not explicitly quantified, training is feasible on standard hardware (Section 3.2), comparable to baselines like PopulationGCN. Hyperparameters (lr=0.000761, k=0.15N) were empirically tuned, with low metric variance (e.g., F1-score: ±2.86) indicating robustness, addressing practical deployment concerns.
  4. Temporal Dynamics and Preprocessing (R2, R2): Static graphs were chosen for computational efficiency (Section 2), with dynamic modeling noted as future work (Section 4). Standardized preprocessing (Section 2.1) ensures consistency, and strong performance across datasets suggests robustness, though varied pipelines are a future direction (Section 4).
  5. Novelty and Comparison (R3): Our Learner Graph module extends Cosmo et al.’s Latent-Graph Learning by using a convolutional encoder to process fMRI BOLD time series, capturing temporal dynamics, and applying kNN thresholding for sparse, subject-specific graphs suited to brain connectivity (Section 2.2). Unlike their single-graph MLP-based embedding with a global threshold, we integrate a learned graph ( \mathbf{A}_L ) with a partial correlation graph ( \mathbf{A}_K ), fused via a Mixture of Experts (Section 2.4), and use contrastive and cluster losses to enhance domain alignment and robustness (Section 2.5). These adaptations address fMRI-specific challenges (e.g., noisy BOLD signals) for ASD/MDD classification, distinct from Cosmo et al.’s general CADx focus. Also, our critique of transductive methods like [13] (Section 1) targets test-time bias, not inductive approaches. We will clarify this distinction to avoid overgeneralization.
  6. Loss Functions (R3): Aligning A_L with A_K balances data-driven and prior-driven features, with ablation studies (Tables 1, 2) showing complementary benefits. The cluster loss enhances topology consistency, reflected in AGMGC’s superior performance.
  7. Clarity and Structure (R2, R3): Section 3.1’s text clearly describes dataset characteristics, complementing Tables 1 and 2. In Fig. 1, “KNN Graph” refers to kNN-thresholded A_L (Section 2.2); we clarify this in the text. k=0.15N balances sparsity and connectivity, validated by stable performance (Tables 1, 2).



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”.

    This paper presents a framework for fMRI-based brain disorder classification, which advances the state-of-the-art in several meaningful ways. The key strengths include a novel dynamic graph learning module that adapts connectivity structures directly from time series data, an effective integration of SplineCNN and multi-graph convolutions for capturing both localized and global dependencies, and the application of contrastive learning to improve generalization and robustness. Experiments on two large and relevant datasets (ABIDE and REST-meta-MDD) show consistent improvements over strong baselines, supported by ablation studies that validate the contribution of each module. While certain components build upon prior work, the combination demonstrate innovation and practical relevance, particularly for applications in clinical neuroimaging.



back to top