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

Abstract

Multimodal neuroimaging grounded in standardized brain atlases enables precise decoding of Alzheimer’s progression by capturing both structural atrophy and functional decline across neural circuits. Current methods compromise anatomical fidelity in whole-brain modeling while generating biologically inconsistent cross-modal interactions. To address these dual challenges, we develop a graph learning framework that integrates three synergistic components: anatomically constrained feature extraction preserving region-specific biomarkers through spatial priors, channel-wise attention mechanisms for discriminative pattern refinement, and bidirectional cross-modal adaptation governed by alternating attention to enforce neuropathological consistency. This unified architecture processes sMRI and PET data through sequential stages of anatomical feature preservation, noise-robust feature enhancement, and dynamic modality fusion, ultimately mapping neurodegeneration patterns across scales. Evaluated on ADNI, our framework achieves superior classification accuracy while graph topology analysis reveals clinically significant hub reorganization within the default mode network, directly correlating with progressive connectivity deterioration. The method’s capacity to reconcile localized biomarker specificity with systemic network dynamics establishes new standards for computational neuropathology.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2738_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{HuWen_AnatomyGuided_MICCAI2025,
        author = { Hu, Wenzheng and Guan, Zhenghua and Yang, Peng and Li, Jiaqiang and Liu, Yi and Gan, Shushen and Cai, Tuo and Zhang, Ao and Zhang, Tengda and Qu, Junlong and Wang, Shaolong and Cai, Gege and Dong, Xiang and Wang, Tianfu and Lei, Baiying},
        title = { { Anatomy-Guided Multimodal Graph Networks for Alzheimer’s Disease: Integrative Analysis of Cross-Modal Brain Connectivity Signatures } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {65 -- 74}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes an architecture for Alzheimer’s Disease diagnosis that processes region-specific features from multi-modal neuroimaging data using an anatomically guided graph structure. It incorporates channel-wise feature enhancement to refine discriminative features and employs cross-modal attention to adaptively fuse complementary information from different modalities.

  • 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.

    Major strengths of the paper : 1) The paper presents extensive experimental results and comprehensive ablation studies, demonstrating the effectiveness of the proposed 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.

    Although the experimental results demonstrate the effectiveness of the proposed architecture, there are a few key concerns: 1 ) Anatomy-guided graph construction is introduced by comparing the normalized volume values of brain regions. While the use of anatomical guidance is well-motivated, the reasoning behind why this specific graph construction method yields anatomically meaningful patterns is not clearly explained. A clearer justification is needed—specifically, why regions with similar volumes are assumed to be connected, and what insights this volume-based similarity graph provides in the context of disease progression. 2 ) The formulation of the modules is somewhat difficult to follow. Some components appear disconnected, and it is unclear where certain elements, such as Eq. 3, are used in the overall framework. Clarification of these modules would improve readability and understanding.

  • 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 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

    (1) It is unclear whether the absolute or relative volume values are used to form the normalized volume vectors in Equation 1. Please specify which one is applied. (2) In the first paragraph of Section 2.2, the underscore notation for “enc” appears to be incorrect. (3) The algorithm is currently provided as a figure with low resolution. For better clarity, you may include the algorithm as formatted LaTeX code

  • 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 the experimental results and discussion are well-presented, the novelty of the proposed architecture—particularly the graph structure—requires further justification. A more thorough motivation for its design choice would strengthen the paper’s contribution. A strong rebuttal addressing these concerns would be highly beneficial.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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.

    While the authors addressed the volume-based graph construction and provided their motivation, I believe their approach should be more prominently emphasized in the Introduction. Additionally, its advantages and limitations, especially in comparison to white matter connectivity-based approaches, should be discussed more clearly. Therefore, I stand by my decision.



Review #2

  • Please describe the contribution of the paper

    Authors propose a multimodal graph learning framework with three key components: anatomy-guided feature extraction to capture regional information from sMRI and PET; channel-wise feature enhancement to refine discriminative features; and dynamic cross-modal fusion to adaptively integrate two modality features. Authors evaluate the proposed graph learning framework on ADNI dataset and the results show that this model achieves superior AD classification accuracy.

  • 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 proposed graph learning framework integrates spatial, structural, and functional insights to improve AD classification task. Specifically, authors design a dynamic fusion mechanism within anatomical units, named as the Cross-Graph Co-Attention (CGCA) module, to calculate mutual information between sMRI/PET nodes and dynamically adjust cross-modal weights, thus effectively integrating multimodal features. [2]. The proposed anatomy-guided graph learning framework identifies the most important regions and their connections associated with AD patients.

  • 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]. Table 1 shows the performance comparison for classification tasks. How were the results in Table 1 derived? How were the training and testing sets partitioned? It would be preferable to have a clear description of the training and test set partitions. [2]. Lack of clarity: the demographic information about the ADNI dataset is incomplete, such as age and gender. In addition, authors state “The data include sMRI and longitudinal follow-up information”. What is the specific longitudinal follow-up information? A brief description would be helpful. What type of PET data is used?

  • 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

    [1]. There appear to be several typos. On page 2: “xare”? In Equation 1, what does “Di(m,c)” mean? On page 4: the section of 2.2 Anatomical Brain Region Rncoding, “Rncoding”? Section 2.3: “The key component is a is a ….” [2]. On page 2: Authors state “Graph-based models such as BrainMAE (Yang et al., 2024 [6]) capture inter-region relations but focus mainly on functional connectivity.” The content described in this sentence may not be entirely accurate. Many recent publications also applied graph-based models to brain structural connectivity or morphometric similarity network. [3]. All tables and figures must be referenced in the text.

  • 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 framework exhibits good innovation in data fusion and cross-modal analysis. But some descriptions lack clarity, such as the training and test set partitions, and the demographic information. There are still some minor issues in terms of writing.

  • 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.

    Some of the issues has been addressed. Specific data details remain insufficient. The elaboration or citation of literature content is inadequately comprehensive. This study demonstrates good innovation.



Review #3

  • Please describe the contribution of the paper

    This paper introduces a novel graph-learning framework for the fusion of multimodal neuroimaging data. The proposed approach is applied to amyloid-PET and structural MRI data to classify individuals as cognitively normal, mildly cognitively impaired, or diagnosed with Alzheimer’s disease.

    The framework combines: anatomically guided feature extraction via an atlas; construction of a graph for each modality based on volumetric similarity; adaptive channel-wise graph refinement using the graphs from the previous step as an anatomical constraint; and a cross-graph co-attention process to fuse the information from the two modalities.

  • 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 key strength of the paper is the novel model architecture, which generates and aggregates graphs from the different imaging modalities to provide highly accurate disease classification. Each of the components of the architecture are well motivated and their contributions are assessed using an ablation study. The graph-based framework provides some interpretability, as the hubs of the derived networks can be identified. The proposed method is also benchmarked against many other state-of-the-art methods.

  • 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) There is a lack of clarity about the PET data used. For instance, the introduction states that “glucose metabolism” is a region-specific biomarker that can be extracted from the data. However, in section 3.1 it states briefly that the PET data used was amyloid-PET, which doesn’t measure metabolism. In addition, it would help to have more details about the PET data, such as whether all participants had both sMRI and PET, and the specific tracer.

    2) There are further key details missing about the experimental methods:

    • How was the model trained?
    • How was the data split between training and testing?
    • In the comparative experiments, are the same MRI and PET scans used as inputs to all the comparison methods?

    3) I also find some of the language around brain connectivity to be slightly misleading. For example in section 2.1:

    “An anatomy-driven sparse graph construction strategy retains only region-to-region associations supported by white matter fibre tracts” – unless I have misunderstood, the anatomical constraints for this process come from the volume-based graphs, which encode regional similarity in brain volumes across subjects. These types of networks partly reflect white matter connectivity but are not a direct measurement of this: (e.g. https://www.sciencedirect.com/science/article/pii/S105381191830435X). There should be some discussion added about the interpretation of these networks and their links to brain connectivity.

  • 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 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

    I think it would make more sense move the description of the Volume-based graph builder so that it comes after the Anatomical Brain Region encoding methods in section 2.2.

    Also, for future work, I don’t think it makes sense to build a volume-based graph using PET data, which typically is of low resolution and has poor contrast for segmenting different structures. Regional SUVR measures would be more appropriate for deriving the network.

  • 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?

    I think that the work is sufficiently interesting and novel to be accepted at MICCAI. However, I would like to see some additional details added to the methods and some more careful interpretation of the results.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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 have addressed my concerns satisfactorily, assuming that the clarifications are added to the camera-ready version should the paper be accepted.



Author Feedback

Thanks for the rebuttal invitation. We itemize our responses to major points as follows: (1) Training design and data details(R1+R3): In the experiments, datasets were split into training/test sets (8:2) using a fixed random seed and subject IDs to prevent data leakage of the same subject across sets. We analyzed FDG-PET data to quantify glucose metabolism in brain regions reflecting functional neurodegeneration. Future work will incorporate PIB-PET (measuring amyloid plaques) combined with FDG-PET and SUVR analysis to validate structural-pathological correlations in AD. (2) Description of Eq.1(R1+R2):The symbol Di(m,c) denotes structural distance (volume difference) between the ith brain region and others under modality m and contrast c. We compute Euclidean squared distances of volume vectors to form matrix B(m,c), then apply TopK on Di to select K-nearest nodes for constructing sparse adjacency matrix A(m,c). As defined in Section 2.1, the normalized volume vector v uses [0,1]-scaled relative values, eliminating absolute volume disparities and ensuring scale consistency to avoid model bias from numerical magnitude variations. (3) Description of Eq.3(R2): In the adaptive channel-wise graph refinement block, F1 and F2 are dual-path modules that model channel relationships and higher-order interactions, to dynamically optimise the topology. After dimensionality reduction of brain region feature pairs (xi,xj) by linear projection, F1 explicitly computes the normalized differences reflecting volume-driven channel relations (anatomical prior), while F2 uses an MLP to implicitly learn nonlinear channel interactions, enhancing discrimination.Their outputs are weighted by channel weights Alpha and fused with shared anatomical topology T(from volume ratios) to form the final channel-aware topology. (4) Biological Basis of Anatomically Significant Patterns(R2+R3): The biological rationale for volume similarity-based brain connectivity maps lies in synchronized developmental/degenerative changes and functional correlations among structurally co-varying regions, reflecting shared genetic-environmental regulation (e.g., tau protein diffusion). In AD, combined atrophy of the entorhinal cortex and hippocampus captures tau spread, where volumetric covariation networks detect early changes better than single-region volumes. Similarly, hippocampus-amygdala structural covariance abnormalities in depression precede global atrophy, supporting such metrics for early disease recognition[1]. This validated approach quantifies volumetric covariation, offering network-level insights for early AD diagnosis and pathology studies. (5) Volume Similarity and White Matter Connectivity(R3):The graph construction method that we employed is based on volume similarity and essentially falls under the category of structural covariance networks (SCNs), rather than structural connectivity from white matter fibre tracing. SCNs reflect synergistic changes between morphological features (e.g., volume or cortical thickness) across individual brain regions. These covariance patterns are co-regulated by multiple factors such as genetics, development and experience. They also have a partially overlap with white matter connectivity[2]. Therefore, although volumetric similarity networks cannot directly represent white matter fibre connectivity, they have been shown to serve as an indirect proxy for it, offering good interpretability and applicability, when anatomical connectivity is not easily observable or data access is limited. In AD, SCNs have been used to reveal early patterns of network-level atrophy, showing better sensitivity than single-region metrics. 1.Zhu et al.,2021,Structural covariance network of the hippocampus amygdala complex in medication-naïve patients with first-episode major depressive disorder 2.Seidlitz et al.,2018,Morphometric similarity networks detect microscale cortical organization and predict inter-individual cognitive variation.



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