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

Abstract

Thanks to advances in neuroimaging, graph neural networks (GNNs) have emerged as a powerful tool for learning brain graph representations to identify Alzheimer’s Disease (AD). However, existing methods often overlook the brain’s hemispherical lateralization, enforcing homogeneous information propagation between hemispheres, which limits their learning capabilities. In this study, we propose a novel dissociative brain graph learning framework (LG-DBGL) guided by brain lateralization to enhance AD identification. Specifically, the Lateralized Decoupling (LD) module partitions brain networks into left/right hemispheric and cross-hemispheric sub-networks. The Dissociative Graph Encoder (DGE) module then independently learns representations for each sub-network, preserving lateralized functional features and avoiding feature confusion. Finally, the Multi-Source Fusion Mechanism (MSFM) dynamically quantifies the contribution of each sub-network to AD-related pathological features, enabling lateralization-guided multi-source feature fusion. Comprehensive experiments conducted on a real-world dataset demonstrate the effectiveness of our LG-DBGL. Our code is publicly available at \text{https://github.com/ilove-gh/LG-DBGL.}

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2711_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{YeJia_LGDBGL_MICCAI2025,
        author = { Ye, Jiazhen and Yuan, Manman and Li, Junlin and Jia, Weiming and Wang, Jiacheng and Li, Jiapei},
        title = { { LG-DBGL: Lateralization-Guided Dissociative Brain Graph Learning for Alzheimer’s Disease Identification } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {449 -- 459}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposes a lateralization-guided decoupling method to partition the brain network into three distinct subnetworks: two inter-hemispheric (left and right) and one intra-hemispheric. These subnetworks are processed independently, and an attention-based module is employed to highlight their individual contributions to the final graph representation. This approach aims to improve the representational capacity of brain networks for the identification of Alzheimer’s Disease.

  • 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 main strengths of the paper are: 1 ) The lateralization-guided decoupling method is well-motivated, enabling the model to capture distinct brain patterns effectively. 2 ) The experimental results demonstrate the model’s effectiveness, with comparisons to state-of-the-art models yielding satisfactory outcomes. 3 ) The ablation study is thorough, offering valuable insights into the individual components, the contributions of the subnetworks, and the diffusion process.

  • 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 weaknesses of the paper are: 1 ) The authors do not provide sufficient details about the brain network construction process, such as which atlas is used for parcellating the brain, and what features are extracted from the fMRI data. Specifically, it is unclear whether the features represent common correlations between regions or the BOLD signal itself. 2 ) While the lateralization-guided decoupling method is well-motivated, the rationale for processing the inter-hemispheric subnetwork with a separate Graph Neural Network, while not applying the same approach to the intra-hemispheric subnetworks, is not clearly explained.

  • 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 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

    (1) The inter-hemispheric subnetworks are processed using a heat-kernel-based diffusion process. The paper cites [4] for the overall diffusion process. However, authors of [4] modify the graph structure by sparsifying the resulting diffusion graph, a step that is not explicitly described in the paper. As a result, it is assumed that the heat kernel is applied directly without any further modifications. To provide clarity to the readers, it would be beneficial to cite Ref1 (F. Chung, “The heat kernel as the pagerank of a graph,” Proceedings of the National Academy of Sciences, 104(50):19735–19740, 2007) if no sparsification is applied. (2) In the ablation study, it is stated that “w/o LD will input the entire brain network into the GCN encoder without lateralization-based decoupling.” Therefore, the results of “w/o LD” should not be directly comparable to the GCN model performance presented in Table 1. If this is not the case, further clarification is needed to explain why. (3) In Table 2, the term “w/o DL” should be corrected to “w/o LD” for consistency. (4) It is interesting that the NC vs AD performance is lower than the NC vs MCI performance, especially considering that the latter is assumed to be the more challenging classification task. Could the authors provide further insight into this observation? (5) In Figure 3, the diffused networks appear sparse. However, as the heat kernel 𝑒^−𝑡𝐴 results in a dense network even with very small values, it would be helpful to clarify whether sparsification was applied or if the sparsity observed in the figure is for visual clarity.

  • 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?

    Although the lateralization-guided decoupling approach has the potential to enhance diagnostic processes for Alzheimer’s Disease, and the paper presents a novel methodology, I would recommend a weak reject due to the technical concerns highlighted above. A strong rebuttal addressing these issues could 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.

    Accept

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

    Thank you to the authors for addressing my concerns. The explanations clarified my questions regarding network generation and model architecture. Accordingly, I have raised my score.



Review #2

  • Please describe the contribution of the paper

    This study introduces LG-DBGL, a novel graph neural network framework designed for AD identification by explicitly modeling brain hemispherical lateralization. The proposed framework has three main contributions: 1) LD is conducted to split brain networks into left and right hemispheric and cross-hemispheric sub-networks to squuze specific characteristics; 2) three DGEs independently learns representations for each sub-network; 3) MSFM dynamically quantifies sub-network contributions via attention-based weighting.

  • 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 study innovatively introduces hemispherical lateralization as a critical factor for identifying and analyzing Alzheimer’s Disease (AD), and constructs a Dissociative Graph Encoder (DGE) architecture comprising three sub-encoders, learning the lateralized functional features for different sub-networks. 2) This work quantifies the contributions of left/right hemispheres and cross-hemispheric networks to AD prediction by introducing the Multi-Source Fusion Mechanism (MSFM) with an attention mechanism, and reveals that left hemisphere lateralization is more strongly associated with AD.

  • 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) The dataset used in the study has a relatively small sample size, comprising only 120 subjects (28 NC, 51 MCI, and 41 AD), and lacks validation on other AD-related public datasets or clinical datasets, leaving the model’s generalization capability uncertain. 2) Insufficient discussion on the relationship between AD/MCI and brain lateralization. Though the authors mentioned “In patients with AD, regions associated with language in the left hemisphere often exhibit significant changes due to cognitive decline”, more sentences and references are still needed to establish a causal relationship between AD and brain lateralization.

  • 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 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.

    (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 authors provided a well-written manuscript which proposed LG-DBGL, a hemispheric lateralization-guided graph neural network framework that decouples brain networks into left/right and cross-hemispheric sub-networks, independently learns their representations to preserve functional specificity, and dynamically fuses multi-source features via attention mechanisms, achieving superior Alzheimer’s Disease identification accuracy on the ADNI dataset. The proposed LG-DBGL framework is rationally designed, and demonstrates a relatively significant improvement over other GCN-based networks in the AD prediction task. The proposed model is able to quantify the contributions from different sub-networks to AD identification, providing novel insights for AD and lateralization-related research. However, there are still several issues in this work that need to be discussed or improved: 1) the generalization capability of the proposed model still requires further investigation. 2) more references are needed to support the study’s findings on the association between AD and brain lateralization. 3) More details of the data processing part are needed, e.g Does the fMRI data include only resting-state data, or are there also task-based scans?Which atlas do the authors use for the connectivity matrix construction? Is any additional pre-processing applied on the ADNI dataset? 4) In Fig. 2(a) right hemisphere part, shouldn’t the annotation be GR = (XR, AR) to be distinguished from GL?

  • 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.

    My major concerns have been adequately addressed. The clarifications regarding data preprocessing, network construction, and the rationale for encoder choices enhance the methodological transparency, and the authors’ justification for dataset usage, along with validation on an independent dataset (PPMI), strengthens the generalizability of their approach.



Review #3

  • Please describe the contribution of the paper

    This paper proposes a novel, clinically-relevant, biologically-informed approach to improving the classification of MCI and AD patients. The authors take advantage of hemispheric lateralization to better inform the features fed into a graph encoder. Results show superior performance compared to a wide range of benchmark models.

  • 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 is a strong submission as the work is a solid contribution to the field and the authors address many of the initial questions proactively. The primary strength of the work is that it uses biology to inform and improve deep/graph learning classification, making the results clinically-relevant. Second, the authors perform several ablation studies to show where their approach fails and make a strong argument for the combination of the 3 innovations they propose.

  • 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. Introduction is well written overall, but contains many grammatical errors. Please revise.
    2. More details must be provided for the graph construction. How many regions? Preprocessing? Functional connectivity metric that was used? How were structural and functional connectivity integrated? How were subjects chosen? Quality control? etc. Please be comprehensive.
    3. I do not see a limitations section.
  • 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 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?

    This paper is very well written and nearly comprehensive. The main criticism I have is that there is not enough detail about the imaging data itself and how the graphs were constructed. The method itself is innovative, well-reasoned, and well-tested.

  • 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

We thank in-depth reviews and appreciate your affirmation of our contributions. The main concerns are addressed below. [R1&R2&R3]-Regarding the detailed procedures for neuroimaging data processing and brain network construction. We preprocessed the resting-state fMRI and DTI data using the DPARSF and PANDA toolkits, respectively. During that process, the AAL template parcellated the human brain into 116 regions (including both 90 cerebral and 26 cerebellar regions), and we used the 90 cerebral regions for our Alzheimer’s disease (AD) research. The node features were the raw fMRI BOLD signals, and the edges between nodes were quantified by the fractional anisotropy (FA) values obtained from DTI. Moreover, detailed information, including data processing and subject characteristics, has been documented in the anonymous GitHub repository mentioned in the abstract. [R2]-Regarding using different encoders in hemispherical networks and cross-hemispherical networks. This design is based on the distinct topological structures and information propagation needs of these two types of networks. Hemispheric networks, with dense and rich interconnections including local and long-range links, utilize a heat diffusion-based encoder to enhance critical pathways and capture both local and global structural features. In contrast, cross-hemispherical networks have a bipartite topology with sparse connections mainly linking nodes across hemispheres, prompting us to use a GIN encoder that excels at learning discriminative graph structures. These encoder distinctions will be clearly outlined in the manuscript. [R2]-The lower performance for NC vs. AD compared to NC vs. MCI is noteworthy. It can be attributed to two aspects of data heterogeneity:1) Biomarker Consistency. The AD group generally shows higher clinical variability (e.g., mixed pathologies and medication effects), introducing confounding factors. In contrast, MCI subjects exhibit more consistent early biomarkers, which aligns with studies showing higher biomarker consistency in the prodromal stage (Reference: Revised criteria for diagnosis and staging of Alzheimer’s disease: Alzheimer’s Association Workgroup). 2) GNN Limitations Due to Heterogeneity. Since the AD group generally has greater brain network complexity and heterogeneity than the MCI group, posing a challenge to the ability of GNN-based methods to handle heterogeneous structures. Several literatures have demonstrated similar performance. (Reference: Multi-view Brain Networks Construction for Alzheimer’s Disease Diagnosis, A novel graph neural network method for Alzheimer’s disease classification). [R2]-Regarding the sparsification of the heat kernel. For code implementation, we used a threshold of 0.0001 to exclude extremely weak connections in the experiment, which corresponds to Eq. (3). For visualization effect, a larger threshold was adopted to enhance visual clarity, as shown in Fig. 3. We will detail the threshold setting in the manuscript. [R2]-The results of “w/o LD” should not be directly compared with the performance of the GCN model shown in Table 1. In the ablation study, “w/o LD” involves inputting the entire brain network into the GCN encoder, which is the proposed Hemispheric Encoder (i.e., Eq. (3)), not a vanilla GCN. So the comparison methods in Table 1 include vanilla GCN. Thanks for your correction, and we will explain it clearly in the manuscript. [R3]-Regarding the limited sample size and few datasets. Due to the complexity of AD diagnosis, high data acquisition costs, and ethical constraints, such datasets are scarce. In this study, we employ the ADNI dataset, the most authoritative and widely used public resource for AD’s research. Additionally, we validate our method on the PPMI dataset for Parkinson’s disease recognition, where it consistently outperforms baseline methods. These results further support the effectiveness of our method and are detailed in the future supplementary materials.



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”.

    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