Abstract

Functional magnetic resonance imaging (fMRI) is a powerful tool for diagnosing neurological disorders. However, accurately distinguishing disease-related features from confounding covariates (e.g., age, gender, site) and individual variability remains a challenge. To tackle this problem, we propose a novel graph disentanglement learning (GDL) framework that decomposes the latent features from fMRI images into 3 components: disease-related features, covariate-related features, and individual variations. The covariate-related features are learned by aligning 2 subject similarity matrices between the features and the true covariates. The disease-related features are guided by a classification loss. We validate our method on 3 fMRI datasets: ADHD-200, schizophrenia (SCZ), and Presbycusis. The method outperforms existing approaches by an average of 0.5%, 1.7%, and 2.1% in accuracy on the 3 datasets respectively. Ablation studies confirm that our model is robust to hyperparameter selection. The disease-associated regions identified by our model align with established clinical findings. These results suggest that GDL is a promising tool for fMRI-based disease diagnosis and biomarker discovery. The code is publicly available at https://github.com/perpetualmachine/GDL_MICCAI.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2897_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/perpetualmachine/GDL_MICCAI

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaShe_Graph_MICCAI2025,
        author = { Zhang, Shengjie and Jiang, Zhuangzhuang and Shen, Xin and Yu, Ziqi and Chen, Xiang and Zhang, Xiao-Yong and Zhou, Yuan},
        title = { { Graph Disentanglement Learning for fMRI Analysis: Decoupling Disease, Covariates, and Individual Variability } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {352 -- 362}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes a novel Graph Disentanglement Learning (GDL) framework for fMRI-based neurological disorder diagnosis, aiming to disentangle disease-related features from covariate effects (e.g., age, gender, site) and individual variability. The method introduces three feature components—disease, covariate, and individual variation—and employs a GNN-based architecture combined with similarity alignment and classification objectives. The model is validated on three datasets (ADHD-200, SCZ, and Presbycusis) and shows improved performance over baseline 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.
    1. The disentanglement of feature sources (disease, covariates, individual) is a meaningful contribution that addresses a critical challenge in fMRI analysis.

    2. The method is clearly described and experimentally validated on multiple datasets, demonstrating improved classification accuracy.

    3. The framework is general and can potentially be extended to other neuroimaging-based diagnosis tasks.

  • 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 proposed disentanglement of feature sources—specifically the separation of disease-related, covariate-related, and individual-specific components—bears notable conceptual similarity to prior work, particularly the 2022 MICCAI paper by Yu et al. (Longitudinal Infant Functional Connectivity Prediction via Conditional Intensive Triplet Network). Although the application domains differ (their work addresses a regression task rather than classification), both approaches aim to decompose latent representations to isolate individual variability and other contributing factors. However, the current paper does not acknowledge this earlier work or clarify how its proposed method diverges from it, raising concerns about the novelty and originality of the contribution.

  • 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.

    (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?

    While the paper tackles an important and timely problem in fMRI-based disease diagnosis—specifically, the disentanglement of disease-related features from confounding covariates and individual variability—the novelty of the proposed approach is notably limited. The core idea of separating latent representations into disease-, covariate-, and individual-specific components bears strong resemblance to prior work, particularly the 2022 MICCAI paper by Yu et al. Although that work focused on a regression setting involving infant brain development and this study targets classification of neurological disorders, both share a similar underlying architecture and disentanglement strategy. However, the current paper neither cites this prior work nor provides a discussion to differentiate its approach. This lack of acknowledgment and comparative analysis significantly undermines the paper’s originality and weakens its overall contribution.

  • 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 introduces a novel Graph Disentanglement Learning (GDL) framework that addresses a key limitation of existing GNN-based fMRI methods, the inability to disentangle disease-related features from confounding covariates and individual variability. Unlike traditional graph-level or node-level classification methods, GDL decomposes latent features into three distinct components: disease-related, covariate-related, and individual-specific. By incorporating dedicated heads for each component and leveraging contrastive variational autoencoder-inspired design, GDL improves feature interpretability and enhances classification performance. Evaluations on three fMRI datasets (ADHD-200, SCZ, Presbycusis) demonstrate the framework’s robustness and superior accuracy, making it a promising tool for fMRI-based disease diagnosis and biomarker discovery.

  • 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 innovation of the described GDL architecture lies in its disentangled representation learning framework, which separates the learned features into three distinct components: an individual head, a covariate head, and a disease head. This approach addresses the challenge of distinguishing disease-related features from covariates (such as age, gender, and site) and individual variability. The disease head is specifically designed to learn features related to the disease, guided by a classification loss, which enhances disease classification accuracy. The covariate head incorporates a novel technique where it computes a subject feature similarity matrix, which is enforced to align with the true covariates’ similarity matrix.

  • 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 experiment is relatively complete, and I understand that due to space limitations, some details could be further refined or expanded.

  • 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.

    (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 innovation points of the article are relatively clear, and the experimental design is complete. Although space limitations make it difficult to elaborate fully, future work can easily build upon the foundation laid in this paper.

  • 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

    The paper introduces a novel graph representation learning framework that disentangles disease-related, covariate-related, and individual-specific variations.

  • 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.

    Here are the major strenghts of the paper :

    1 ) The proposed approach for disentangling disease-related, covariate-related, and individual-specific features is novel, well-motivated, and clearly defined.

    2 ) The model architecture and associated loss functions are thoroughly explained, with each component clearly described.

    3 ) The experimental results effectively demonstrate the strength of the proposed method. Both the comparisons with existing models and the ablation study are comprehensive and convincing.

  • 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 main concern is that the individual-specific head does not appear to contribute in a distinct or specialized manner compared to the other components—it primarily impacts the reconstruction loss, similar to the disease-related and covariate-related heads. The motivation for introducing an individual-specific head with a Gaussian distribution prior should be clarified and better justified.

  • 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

    The resulting functional networks are fully connected. It is unclear whether any sparsification or thresholding is applied prior to using them in the GNN architectures. While models such as GAT, GAN, and GraphSAGE can operate on fully connected graphs, GCN involves the inversion of the degree matrix, which may become problematic if negative edge values are present. Therefore, the construction and characteristics of the underlying graph should be explained more clearly.

  • 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 novelty and clarity of the paper, along with the well-conducted experimental results across various datasets and models, contributed to my overall positive evaluation. These aspects collectively demonstrate the robustness of the proposed method, and I believe the paper has the potential to make a meaningful contribution to the field.

  • 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.

    I stand by my decision.




Author Feedback

We want to thank the reviewers for their valuable comments to improve our paper.

Reviewer 1: The conceptual similarity with the MICCAI 2022 paper by Yu et al (Longitudinal Infant Functional Connectivity Prediction via Conditional Intensive Triplet Network) limits the originality and novelty.

Response: Besides the application domain, our method has two critical differences compared to the previous work by Yu et al. First, Yu et al. focus exclusively on age regression within a longitudinal setting while our work considers all the covariates, age, gender, education, site, etc. in a general setting. In the previous work by Yu et al., the objective is to predict developmental trajectories, hence age is the only covariate considered. In contrast, our framework considers all the factors contributing to the extracted features in an encoder-decoder architecture. These factors include disease-related factors, covariate factors, and individual factors. This difference makes our framework more broadly applicable to a diverse range of applications, including classification and subtyping.

Even if only ages are considered in the covariates, a second and potentially more important difference is that Yu et al. directly regress ages onto the extracted features while our framework asks subjects with similar ages to have similar age-related features (Fig. 1). The benefit of our strategy is that the features do not have to be linearly related to ages. Instead, they can lie in any nonlinear 1-dimensional manifold in a high dimensional space, as long as the proximity of ages between subjects is preserved in the age-related features in the high dimensional space. This difference is similar to the difference between linear regression and Gaussian process. Hence, our framework can extract more flexible covariate features without imposing a strong linear relationship between the features and the covariates.

We will cite Yu et al. in the final version and clarify these distinctions in the related work section if the paper is accepted.

Reviewer 3: Motivation for introducing an individual head with a Gaussian prior could be better clarified.

Response: The motivation of introducing an individual head could be understood by considering all the factors (disease, covariate, individual) contributing to the extracted features of an image. Suppose we restrict the subjects of any dataset to healthy controls at the same age, with the same gender, at the same site, etc. There will still be a difference among their images and extracted features. These individual differences should be considered in the reconstruction. Therefore, we introduce an additional branch to model this difference. By assuming that the individual features are sampled from a Gaussian distribution, we provide a simple way to model this variability without introducing additional hyperparameters. The benefit is empirically validated in the ablation studies (Table 1).

We will further clarify this motivation in the final version if the paper is accepted.

Reviewer 3: Unclear whether thresholding is applied to the functional connectivity network for GCN.

Response: For all the GNN encoders, we use a threshold of zero to set all negative values in the adjacency matrix to zero, following prior practices [1]. This ensures that the degree matrix remains well-defined and that GCN can perform stable normalization and inverse operations. We will include this clarification in the final version if the paper is accepted.

References:

[1] Zalesky A, Fornito A, Bullmore E. On the use of correlation as a measure of network connectivity[J]. Neuroimage, 2012, 60(4): 2096-2106.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top