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

Abstract

Many biomedical data exhibit intrinsic graph-like properties, making graph neural networks (GNNs) widely adopted modeling tools. The brain arterial network (BAN) represents the most complex arterial network in humans, where conventional GNNs struggle to capture critical long-range relationships. Recent graph transformers have enabled modeling of these long-range dependencies through attention mechanisms; however, they face challenges in incorporating hierarchical information, especially when strong anatomical priors exist within the graph structure. While some approaches have attempted to integrate hierarchical information into graph transformers, they primarily focus on node feature aggregation, despite BAN’s most clinically significant features residing in edges rather than nodes. To address these limitations, we propose a hierarchical graph transformer (HGT) with edge-aware structural encoding that better incorporates anatomical and multi-scale structural information. Our approach achieves state-of-the-art performance across all 11 tasks. This work lays the foundation for individualized risk assessment that complements traditional systemic risk evaluation methods.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/4940_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{ZhaKai_EdgeAware_MICCAI2025,
        author = { Zhang, Kaiyu AND Chen, Li AND Liu, Wenjin AND Kim, Taewon AND Wang, Xin AND Guo, Yin AND Tan, Zhiwei AND Chen, Zhensen AND Tang, Angie AND Zhao, Xihai AND Hatsukami, Thomas S. AND Mossa-Basha, Mahmud AND Balu, Niranjan AND Yuan, Chun},
        title = { { Edge-Aware Hierarchical Graph Transformer to Decode Brain Arterial Network } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15971},
        month = {September},
        page = {309 -- 319}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper aims to model the inherent complex graph properties of the brain arterial network (BAN) for individualized risk assessment. To address the challenges posed by conventional methods in integrating hierarchical information, the authors proposed a hierarchical graph transformer (HGT) with edge-aware structural encoding to better integrate anatomical and multi-scale structural information. Experimental results on the BAN dataset show that the proposed HGT achieves state-of-the-art performance in all 11 tasks.

  • 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 authors proposed a dataset, BAN, for classification of FRS levels (<0.1, 0.1-0.2, >0.2), hypertension, diabetes, and smoking status; and regression of FRS, age, SBP, DBP, HDL, LDL, and TC.
    2. A Hierarchical Graph Transformer (HGT) framework with edge-aware structural encoding.
    3. Sufficient experimental setup on the proposed BAN dataset.
  • 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. This project employs Hierarchical Graph Transformer (HGT) to model the Brain Arterial Network (BAN). However, the necessity of hierarchical modeling for capturing anatomical structures remains unclear. Please provide further explanation or evidence of what anatomical structures are captured by this hierarchical structure (for example, visualization, references, or more observations).
    2. The compared method are relatively out of date, thereby limiting the empirical validation of the proposed method’s superiority. It would help to add comparison with more up-to-date approaches.
  • 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

    I recommend changing the figures to vector format files such as .eps or .pdf for clearer presentation of details.

  • 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?

    This paper releases the first multimodal BAN dataset, presenting significant contributions to multiple cardiovascular prediction tasks. However, the methodology motivations needs further clarification as stated in major weaknesses.

  • 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

    The primary contribution of the paper is the development of a Hierarchical Graph Transformer (HGT) with edge-aware structural encoding for modeling the Brain Arterial Network (BAN), a complex vascular system critical for cardiovascular risk assessment. Unlike conventional Graph Neural Networks (GNNs) and standard graph transformers, HGT incorporates multi-scale structural information through topology-preserving reduction and heavy-edge matching, focusing on edge features (e.g., artery length, radius, tortuosity) rather than node features. This approach captures both local and global vascular patterns, addressing the limitations of prior methods in handling BAN’s hierarchical complexity. Additionally, the paper releases the first multimodal BAN dataset with clinical correlates, enabling evaluation across 11 predictive tasks (e.g., Framingham Risk Score classification, hypertension prediction). HGT achieves state-of-the-art performance across all tasks, demonstrating strong associations between BAN structure and cardiovascular risk factors, laying a foundation for personalized risk stratification.

  • 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 HGT introduces a modular extension to standard graph transformers [20] that incorporates edge-aware structural encoding through topology-preserving reduction and multi-level coarsening via heavy-edge matching (Section 2.1). Unlike prior graph transformers [12, 14, 20], which focus on node feature aggregation, HGT prioritizes edge features (artery length, radius, tortuosity) critical to BAN’s clinical significance. The multi-level edge encoding and structurally enriched attention mechanism (Equations 3-6) allow the model to capture hierarchical vascular patterns, modulated by a biologically meaningful weight function (Equation 2). The paper releases the first comprehensive multimodal BAN dataset (Section 2.2), comprising 402 participants from three East Asian cohorts [13, 4], with 3D MRA-derived vascular graphs and clinical correlates (e.g., blood pressure, lipid profiles, FRS). The dataset includes node features (3D coordinates) and edge features (radius, length, tortuosity), constructed using validated tracing methods [2, 3, 8, 24], and supports 11 predictive tasks (classification and regression).

  • 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 paper claims potential for personalized risk stratification (Section 4) but lacks direct clinical validation, such as integration into clinical workflows or evaluation by clinicians. While predictive performance is strong (Tables 1-2), there is no evidence of real-world testing or impact on clinical decision-making. The paper compares HGT to GCN [15], GAT [21], GT [5], and GPS [20] but does not benchmark against other hierarchical or edge-aware graph transformers, such as those in Ying et al. [26], “Hierarchical Graph Representation Learning with Differentiable Pooling,” NeurIPS 2018, or Wu et al. [14], which also incorporate hierarchical information. The choice of 𝐻 = 3 H=3 for hierarchical encoding (Section 3.1) and other hyperparameters (e.g., grid search via Neural Network Intelligence [16]) lacks justification or sensitivity analysis, reducing reproducibility.

  • 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

    nclude a pilot study or qualitative feedback from clinicians to demonstrate HGT’s utility in risk stratification. For example, evaluate how attention visualizations (Figure 3) guide identification of at-risk arterial segments, aligning with MICCAI’s translational focus. Discuss potential integration with existing tools (e.g., FRS [6]) to enhance clinical workflows.

  • 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 HGT’s edge-aware hierarchical encoding and focus on BAN’s edge features are innovative, addressing a gap in modeling complex vascular networks. The modular design (Section 2.1) and multi-scale coarsening (Equations 1-3) advance graph transformer research. The comprehensive evaluation across 11 tasks (Tables 1-2), supported by ablation studies (Table 3) and visualizations (Figure 3), demonstrates HGT’s superiority and interpretability. The new BAN dataset (Section 2.2) is a significant contribution to the field. The association between BAN structure and cardiovascular risk (Section 4) has strong potential for personalized risk assessment, though clinical validation is needed to fully realize this impact. The paper is well-written and organized, with minor issues in justifying hyperparameters and discussing limitations, which are addressable with revisions. Weaknesses: The lack of clinical validation, incomplete comparisons with hierarchical graph methods, dataset limitations, and missing computational cost analysis are notable but can be mitigated with minor revisions (e.g., adding discussions, benchmarks, or metrics).

  • 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 proposes a novel Hierarchical Graph Transformer (HGT) that incorporates edge-aware structural encoding for analyzing the Brain Arterial Network (BAN). This enables effective learning from long-range vascular dependencies. A new multimodal dataset of 3D MRA-derived vascular graphs with clinical cardiovascular risk profiles is introduced, addressing a critical gap in data availability for vascular modeling and risk prediction.

  • 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.

    Novel Architecture: The HGT integrates anatomical priors via multiscale hierarchical encodings, focusing on edge attributes like vessel length, radius, and tortuosity—critical for BAN analysis.

    Multimodal Dataset: A very good and important contribution—a BAN dataset with clinical correlates, public code, and graph annotations. This will be a valuable benchmark for future vascular studies.

    Multiscale and Multicentric Design: The hierarchical graph construction and dataset aggregation across three East Asian cohorts enhance robustness and generalizability.

    Interpretability via Attention Visualization: Edge and node importance are derived from learned attention maps, supporting explainability (though see comment on node importance below).

    Strong Performance Across Tasks: Close to SOTA results in both regression and classification tasks with clear ablation studies.

  • 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.

    • Number of Coarsening Levels: It’s unclear how the number of hierarchical coarsening levels H=3 was chosen. Is this empirically validated or guided by vascular topology?

    • Node Importance Visualization: While the edge-based attention makes intuitive sense, node importance is derived by summing edge values. But, it is unclear what does EdgeIMP(e_ij) does, is it the attention score? in that case if it is after the softmax? before?

    • Data Standardization: Details are missing on how features were standardized per subject. The length of the segments were normalized based on some subject-specific measurement?

    • Node Sampling: The augmentation strategy involves random node sampling, but it’s unclear if the underlying connectivity is modified or if the virtual node prevents disconnected components by being connected to all the nodes in such cases.

    • Significance of Results (Table 1): While the HGT outperforms others, the margin of improvement over GAT is modest in some tasks (e.g., FRS). Are these statistically significant? Indicating significance would strengthen the empirical claims.

    • Best-Performing Method Not Marked (Table 2): In some tasks (e.g., Smoking classification), GT performs better than HGT. The best values across methods should be marked for transparency, not only the authors’ results.

  • 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 is a well-executed and highly relevant contribution to graph-based biomedical modeling, with novelty in both method and data. The architectural design is domain-aware, the dataset is impactful and well-documented, and the experiments are thorough. Some methodological details (e.g., coarsening choice, node importance) could be clarified, and significance of results should be addressed, but overall, this is a paper that should be presented at MICCAI.

  • 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

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”.

    N/A



back to top