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

Abstract

By leveraging complementary Euclidean and graph-based spatial information from structural Magnetic Resonance Imaging (sMRI), the effective fusion of multi-spatial brain features holds the potential to enhance the classification accuracy for Alzheimer’s Disease (AD). Existing deep learning models often rely on simplistic methods such as concatenation, weighted summation, and self-attention to integrate Euclidean and graph spatial features. However, these models neglect the causal relationships between feature domains and labels, resulting in redundancies and limiting the classification accuracy. In this study, we propose a Multi-Spatial Granger Causality Features Fusion Network (MSGCFNet). Specifically, the MSGCFNet consists of a Multi-Spatial Features Encoder (MSFEN) module that extracts Euclidean and graph spatial features, a Multi-Spatial Granger Causality Features Disentanglement (MSGCFD) module that uses Granger causality-based learning to disentangle the causal dependencies within Euclidean and graph spatial features, and a Multi-Spatial Features Fusion Classification (MSFFC) module that employs a bidirectional cross-attention mechanism to robustly fuse the disentangled features from the two spatial features. Additionally, we design a multi-spatial Granger causal contrast disentanglement loss function that effectively minimizes the bias and redundancy of the disentangled features. Experimental results demonstrate that MSGCFNet achieves classification accuracies of 93.6% for Alzheimer’s Disease (AD) vs. Normal Controls (NC) and 83.4% for Early Mild Cognitive Impairment (EMCI) vs. Late Mild Cognitive Impairment (LMCI) tasks, highlighting its superior classification performance. The code is available at https://github.com/FindBrain/MSGCFNet.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0927_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/FindBrain/MSGCFNet

Link to the Dataset(s)

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

BibTex

@InProceedings{SonZhi_MultiSpatial_MICCAI2025,
        author = { Song, Zhiwei and Li, Jingming and Yu, Hu and Guo, Xiaojuan},
        title = { { Multi-Spatial Granger Causality Features Fusion Network for Alzheimer’s Disease Classification } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15967},
        month = {September},
        page = {374 -- 383}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a novel architecture, Multi-Spatial Granger Causality Feature Fusion Network (MSGCFNet), which integrates multi-spatial brain features derived from both Euclidean and graph spatial domains using structural MRI (sMRI). MSGCFNet further disentangles causal dependencies across different feature domains and employs a bidirectional cross-attention mechanism to achieve robust feature fusion. Experimental results demonstrate that MSGCFNet outperforms existing models in Alzheimer’s Disease (AD) classification.

  • 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 ablation study is comprehensive and includes informative visualizations, which help to validate the effectiveness of each proposed component.
    • The comparison with other state-of-the-art methods is appropriate and supports the superiority of the proposed framework.
  • 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.
    • There are several typographical and formatting errors in the methodology section, such as in Equation (1). These issues should be carefully addressed and corrected prior to submission.
    • The overall writing lacks clarity and coherence. The manuscript would benefit greatly from thorough polishing and reorganization to improve readability.
    • The novelty of the proposed method is limited, as it primarily consists of combining an existing Granger Causality Feature Disentanglement (MSGCFD) module and cross-attention mechanism. The current design does not fully exploit the potential interactions between the two feature branches. A more tightly coupled or co-adaptive design might enhance both effectiveness and originality.
  • Please rate the clarity and organization of this paper

    Poor

  • 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.

    (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?

    The proposed approach has potential, but the current presentation lacks sufficient innovation and clarity. The authors are encouraged to revise the architecture by more thoughtfully incorporating both feature branches into the MSGCFD module and to improve the manuscript’s organization and precision. Additionally, the authors should clarify whether the baseline results were reproduced by themselves.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The novelty is the major concern which is still not being addressed



Review #2

  • Please describe the contribution of the paper

    Authors introduce MSGCFNet, a multi-spatial brain feature integration deep learning framework designed to fuse Euclidean space and graph-based spatial features extracted from structural magnetic resonance imaging (sMRI), thereby enhancing the classification performance for Alzheimer’s disease (AD). This framework employs Granger causality-based learning to preserve the causal dependence of multi-spatial features. Additionally, a bidirectional cross-attention mechanism is utilized to capture intricate cross-domain interactions. Authors validate the method on two AD-related tasks, outperforming 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.

    Strength: The authors introduce causality into multi-space feature decoupling and combine cross-attention to solve the limitations of traditional fusion methods (such as concatenation and weighted summation) that ignore causality. At the same time, the authors use causal contrast disentanglement loss and causal disentanglement loss to separate causal features from non-causal features to reduce redundant information interference.

  • 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. Insufficient Methodological Innovation: The authors’ primary contributions appear to be a causal analysis method for feature-label relationships in dual-feature fusion and a contrastive learning approach to decouple task-independent and task-correlated features. However, the causal analysis has already been proposed in prior work (Reference [17]). While the authors use it across two feature spaces, the manuscript does not sufficiently clarify how this differs from or advances existing approaches.

    2. Insufficient Experimental Comparison and Analysis: The paper presents a feature fusion framework but evaluates it against only two fusion methods. A more comprehensive comparison with state-of-the-art feature fusion techniques would strengthen the validity of the claims. Additionally, the experimental analysis is insufficient. The authors list comparative results without delving into the underlying reasons for performance differences. For instance, while the proposed contrastive loss for feature decomposition is a claimed innovation, no comparison is made with alternative feature disentanglement methods (e.g., adversarial loss or other contrastive learning variants). A deeper discussion of methodological and performance distinctions is necessary.

    3. Organizational and Clarity are Weak: The overall paper structure is logical, but the methodology section (Section 2) suffers from disorganization. Section 2.2 introduces four loss functions but only elaborates on one (L_Mgcausal). The multi-spatial contrastive loss (a core component of this module) is deferred to Section 2.3, while the Binary Cross-Entropy loss is misplaced, as it does not belong to this module. The explanations of L_MG (Equations 3–5) and Equation 7 are unclear, potentially confusing readers. The conclusion lacks a discussion on future research plan.

    4. Technical Errors and Ambiguities: In Section 2.1, there is a symbol inconsistency (text uses A, while the equation uses Â), with no definition provided. In Section 2.2:, “Loss_Mgccl” should likely be L_Mgccl, and the variable I in Equation 3 is undefined. In Section 2.3, “bidirectional cross-attention” should be capitalized (“Bidirectional cross-attention”). The σ function in Equation 7 is missing parentheses. The phrase “where S^i S represents the cosine similarity calculation of sample i” contains an extraneous “S” (it should be “where S^i represents…”). Figure references should follow standard formatting (e.g., Fig. X).

    5. Clinical Interpretability not Verified: Although the feature decoupling module isolated causal correlation features (α), the specific biological significance corresponding to these features was not explicitly analyzed (such as whether they were associated with known pathological brain regions or biomarkers of AD).

  • 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.

    (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?
    1. The proposed causal analysis method is repeated with the existing work [17], while authors applied it in two feature spaces, the novelty may be under explained, and the feature decoupling lacks comparison with mainstream methods; 2. Insufficient interpretation of experimental results and lack of in-depth analysis of performance differences; 3. In the methodology section, the logic is loose, the introduction of loss functions is unclear, and the formula symbols are undefined or wrong in many places.
  • 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’ response partially addresses my concerns, but critical issues remain unresolved. However, the authors admit to addressing the following in the revised manuscript: Ablation studies on dual-space causal disentanglement; Comparisons with feature disentanglement methods; Preliminary analysis linking causal features to AD pathology.



Review #3

  • Please describe the contribution of the paper

    The key idea of the paper is to is to integrates the granger causality features from Euclidean and graph spatial features for AD diagnosis. Specifically, each type of features is disentangled into related and unrelated features to label using granger causal disentangle loss. Then the bidirectional cross-attention module is designed to fuse the label-related features of both spatial types for subsequent classification. The dependency of the label-unrelated features with other label-related features are reduced through contrastive loss function. The results show that the proposed method achieves better performance compared with other methods for AD classification with good interpretability.

  • 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.

    This paper decouples each category of features into disease-related features and disease-unrelated features, and fuses the disease-related features from two perspectives to enhance the performance of AD identification. Specifically, Granger causality is employed to retain the disease-related features within each category. Meanwhile, a contrastive loss is introduced to reduce the correlation between the disease-unrelated features and the disease-related features. The significance test between causal and non-causal features is provided for two classification task.

  • 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 comparative methods are relatively limited. In the Violin Plot of feature disentanglement analysis, no illustration is provided for Violin shapes. In Section 2.1 “module” is incorrectly written as “Moule” The dimension size is not provided for 𝑍𝐸 and 𝑍G . In Fig.1, the input of bidirectional cross-attention module from the bottom is Q𝐸, K𝐸, V𝐸, which should be QG, KG, VG ？

  • 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?

    Using Granger causality to retain the disease-related features within each category of features.

  • 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 gratefully acknowledge all reviewers for their invaluable time and insightful feedback, which significantly enhanced the quality of this work. Below, we respond to each major concern. @R2,R3-Novelty & Contribution Clarifications (a) We propose MSGCFNet, an innovative framework that seamlessly integrates multi-spatial feature encoding, causal disentanglement, and feature fusion to tackle the critical challenges of feature redundancy and cross-spatial dependency modeling in AD classification task. Specifically:1)we propose a multi-spatial Granger causality disentanglement module, which utilizes conditional mutual information and contrastive learning to disentangle label-relevant causal features from both Euclidean and graph spaces;2)we design a bidirectional cross-attention module with contrastive learning-enhanced fusion to strengthen cross-spatial feature interactions;3)we propose a multi-spatial Granger causal contrastive disentanglement loss, a novel formulation designed to explicitly disentangle AD-related causal representations. (b) While inspired by reference [17], our approach differs in both methodology and objective. Reference [17] proposes a Granger causality-inspired graph neural network to identify the most influential subgraph that is causally related to psychiatric diseases. In contrast, the pivotal innovation of our MSGCFNet resides in not only leveraging Granger causality to disentangle AD-related causal features from multi-spatial inputs, but also using causal and non-causal features to guide cross-space feature fusion of the model. @R1,R3- Experimental Comparison & Analysis (a) Our experimental comparison includes a comprehensive set of eleven representative baselines, including Euclidean-based, graph-based, and recent multi-spatial fusion models. In contrast to early and intermediate fusion strategies, the proposed method adopts late fusion strategy. Therefore, to ensure a fair comparison, we focused on two recent and representative models with late fusion strategy (s2MRI-ADNet,MSRNet), which are most relevant to our approach in terms of model structure. Due to rebuttal policy, we will compare a broader range of fusion strategies in future work. (b) Our proposed multi-spatial Granger causal contrastive disentanglement loss focuses on disentangling causal features across Euclidean and graph spaces. As the resulting causal features (ZEα,ZGα) are not high-dimensional, we employ a lightweight contrastive learning strategy to effectively capture task-relevant causal information. Due to rebuttal policy, we will consider comparisons with state-of-the-art contrastive strategies in future work. @R3-Clinical Interpretability We wish to clarify that our framework can provide interpretability. Specifically, the causal features ZEα and ZGα obtained via Equation (5) were fused through the bidirectional cross-attention mechanism in Equation (6) to produce ZEGα, which can be used to perform statistical correlation with the biomarkers of AD to explore their clinical relevance. Due to space limitations, we will further report these interpretability analyses in future work. @R1,R2,R3-Organizational Clarity & Technical Ambiguities (a) The total loss L_Mgccl consists of three components in our model: L_Mgcausal, L_Mcontrastive, and L_BCE. Section 2.2 focuses on disentangling AD-related features in the Euclidean and graph spaces, thus we describe L_Mgcausal here. Section 2.3 introduces the fusion of the disentangled features ZEα and ZGα for classification, so the L_Mcontrastive and L_BCE are described in this section. (b) We thank all R1,R2, and R3 for the manuscript corrections/suggestions. We will correct all identified errors and revise the architecture by more thoughtfully incorporating both feature branches into the MSGCFD module to improve the manuscript’s organization and precision. @R2-Others We reproduced all baseline results based on the original paper and code.



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”.

    This paper presents MSGCFNet, a multi - spatial brain feature integration deep learning framework for Alzheimer’s disease (AD) classification. It combines Euclidean and graph - based spatial features from sMRI, using Granger causality learning and a bidirectional cross - attention mechanism to capture cross - domain interactions. Despite outperforming baselines in two AD - related tasks, it has several drawbacks: insufficient innovation, inadequate experimental comparison and analysis, poor organizational clarity with technical errors and ambiguities, and lack of verification of clinical interpretability. The author needs to explain the shortcomings mentioned above.

  • 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.

    Reject

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

    This paper integrates the granger causality features from Euclidean and graph spatial features for AD diagnosis. However, the main concern is that the novelty of the proposed method is limited, as it primarily consists of combining an existing MSGCFD module and cross-attention mechanism. Besides, the experimental comparison with sota methods is very insufficient. Moreover, the overall writing lacks clarity and coherence. Based these concerns, the paper cannot be accepted.



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’

    The rebuttal addressed several concerns and got acceptance from one of the reviewers with initial negative rating. While novelty is limited, the proposed method yields meaningful contribution and results. The camera-ready should incorporate suggestions from R3.



back to top