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) is essential for understanding and diagnosing brain disorders. However, the challenge of small sample sizes, due to high acquisition costs and low annotation efficiency, hinders deeper exploration of the mechanisms underlying brain diseases. Recently, generative diffusion models have shown great potential for time series data generation, but directly using them for fMRI generation still has some issues. Firstly, most of them are designed for single time series, ignoring the significant dependency information between multiple time series when applied to fMRI. Since fMRI time series from different brain regions exhibit correlations, it is necessary to consider this characteristic when generating fMRI. Secondly, the generation process often lacks the involvement of label information, which limits their applicability in facilitating classification tasks. Thirdly, the alignment between the generated data and the target tasks is often insufficient, limiting its effectiveness for brain disorder diagnosis. To address these issues, we propose a novel task-aligned fMRI generation method based on the diffusion model. Specifically, a functional brain network (FBN) is incorporated into the diffusion model as prior knowledge to guide and constrain the data generation process, ensuring that the generated fMRI respects the functional connectivity characteristics observed in actual fMRI. To effectively and flexibly generate class-specific fMRI, a representative class-wise FBN is utilized as the prior FBN. Meanwhile, the proposed method ensures that the generated fMRI is well aligned with target brain disorder classification tasks. Extensive experiments are conducted on three datasets, consistently demonstrating the superior performance of the proposed method.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4178_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)

ADNI dataset: https://adni.loni.usc.edu/

BibTex

@InProceedings{LiYif_Taskaligned_MICCAI2025,
        author = { Li, Yifan AND Wu, Xiaotong AND Zhang, Xiaocai AND Jiang, Haiteng AND Wu, Weiwen AND Shen, Dinggang AND Zhang, Jianjia},
        title = { { Task-aligned fMRI Generation Model 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 = {638 -- 648}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a diffusion model framework that incorporates class-wise Functional Brain Network (FBN) features as prior knowledge during the reverse process to guide fMRI generation. The method is evaluated on ADNI and two private datasets, showing improved classification performance (ACC/AUC) when using generated data with BoIT and GCN classifiers.

  • 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. Practical Methodology: Leverages readily available FBN features as low-cost prior knowledge, enhancing clinical applicability. Demonstrates good extensibility to multiple diseases through class-wise FBN conditioning.
    2. Extensive Experimental Validation: Comprehensive comparisons against 6 baseline methods across 3 datasets. Thorough ablation studies on FBN integration intervals.
  • 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. Limited Novelty: The core concept of using FBN to guide generation lacks sufficient differentiation from existing Classifier Guidance Diffusion frameworks.
    2. Limitations of Using Classification Metrics Alone: The authors evaluate the generated data solely through classification performance (e.g., accuracy of a downstream classifier). While this provides some insight into feature discriminability, it fails to comprehensively assess the generative model’s core capabilities, such as diversity (e.g., mode coverage), realism (e.g., visual or statistical fidelity), and generalization (e.g., robustness to unseen data distributions).”
  • 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

    I recommend supplementing with evaluations of generation quality (e.g., Fréchet Inception Distance (FID) or Inception Score (IS) for images, or domain-specific metrics for non-image data, along with visual/qualitative comparisons such as t-SNE plots or expert assessment), diversity (e.g., intra-class metrics like LPIPS to detect mode collapse or latent space coverage via Maximum Mean Discrepancy (MMD)), and downstream utility beyond classification (e.g., reconstruction error such as MSE for VAEs or adversarial robustness if applicable). Without these additional evaluations, claims about the model’s superiority may be overstated, as classification accuracy alone cannot distinguish between high-fidelity generation and overfitting to trivial features.

  • 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?

    While this work applies diffusion models to fMRI generation, the core methodology remains largely unchanged from existing Classifier Guidance Diffusion frameworks. The application to neuroimaging is certainly interesting and could prove valuable to the field. However, the current presentation does not sufficiently highlight conceptual or technical advances that would distinguish it from prior work.

  • 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

    The authors propose a novel diffusion model-based framework for generating fMRI time series that are guided by class-wise functional brain networks. By incorporating prior FBNs into the generative process, the proposed method ensures that the generated fMRI data preserves the functional connectivity characteristics inherent in real fMRI signals. This approach also enhances the applicability of the generated data for downstream classification tasks in diagnosing brain disorders.

  • 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 use of representative class-wise FBNs to guide the data generation process is quite innovative, thereby ensuring that generated fMRI signals accurately reflect functional connectivity patterns.
    2. By preserving the dynamic information inherent in fMRI data, the proposed method addresses a key limitation in current generative methods that typically lose temporal context.
    3. Experimental results across multiple datasets validate that the generated data not only resemble real fMRI in their functional connectivity but also improve downstream classification performance.
  • 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.

    Missing negative sign in Equation (6): The gradient-based update in Equation (6) appears to be missing a negative sign if the goal is to bring the generated correlation matrix C_t closer to the prior C_p. Typically, minimizing |C_t – C_p|_F would involve subtracting (rather than adding) the gradient. Although one could assume that a negative sign is absorbed into the hyperparameter lambda, the paper’s implementation states that lambda is fixed at 1, leaving it unclear how the sign is handled in practice. This raises questions about whether the update enforces alignment with the prior correlation structure.

  • 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.

    (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 presents a novel framework for generating fMRI data that is both innovative and well-motivated, with a clear presentation of ideas and thorough experimental validation. The integration of class-wise FBN guidance and a diffusion model framework addresses critical challenges in preserving functional connectivity and temporal dynamics, making it a valuable contribution to the MICCAI community. However, a key implementation detail regarding the handling of the negative sign in Equation (6) remains ambiguous - specifically, the mechanism used to align the generated and prior correlation matrices is unclear since the designated lambda is set to 1. This unresolved issue was a significant factor in arriving at an overall score of 5.

  • 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

    This paper describes a novel application of diffusion model to generate ROI time series extracted from fMRI data. It enables task-aware multi-class data generation through a novel loss constraint enforcing similar FC between group average and generated data.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1. Generating 4D fMRI sequence is a difficult and computational intensive task. Generating ROI time series, on the other hand, is an efficient way to bypass the computational complexity for a wide range of ROI-based analytical tasks on fMRI.

    2. The FC connectivity matrix is an important component for analyzing brain functional dynamics. Enforcing constraint on FC is reasonable and effective as shown in Table 1.

  • 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. While the reviewer believes that generating ROI time series and effective, it still limits the application in ROI-based analysis.

    2. The constraint on FC is only applied on the entire sequence without considering potential temporal dynamics. The experiments would be more comprehensive if ablations like applying the FC constraint on different-sized or multi-scale sliding window is compared.

    3. The experiments only compare performances under two classifiers: BoIT and GCN. The authors might consider adding experiments using simpler classifiers like SVM.

    4. The authors mainly referenced works in the general machine learning community for time series data generation. There are existing works in fMRI synthetic augmentation that can potentially be added into reference.

  • 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 provide sufficient information for 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.

    (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?

    Generation of 4D spatiotemporal data is a challenging and difficult task. This paper describes a simple but widely applicable idea to generate ROI time series rather than the entire image. While the experiments can still be extended, the application value and consistent performances on three datasets justify the rating.

  • 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

N/A



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”.

    The reviewers agree on the interest in the methodology, which applies diffusion models for generating ROI-based fMRI data with consideration for functional brain networks, and strong experimental validation and results. However, in addition to some suggestions for added metrics/ablations, one reviewer had concerns regarding the novelty of the approach, which may not distinguish the presented method from prior diffusion-based work, and another had concerns regarding positioning this work among other generative methods for fMRI. Still, the new adaptation of diffusion-based models to consider brain-domain specific information has the potential to make an impact in the field of fMRI analysis. Therefore, I recommend provisional accept for this work.



back to top