Abstract

Accurate diagnosis of neurodegenerative disorders using functional magnetic resonance imaging (fMRI) remains challenging due to the intricate functional connectivity patterns and inherent imaging noise. Conventional Graph Neural Networks (GNNs) typically project the brain into a shared global space, overlooking that different brain regions reside on heterogeneous functional manifolds. While existing attention or heterogeneous graph methods merely adjust edge weights, they fail to capture the complex geometric misalignment among regional manifolds. To address this limitation, we propose Vector Bundle Graph Contrastive Learning (VB-GCL), a novel dual-stream framework grounded in vector bundle theory for brain disease classification. Unlike standard GNNs, vector bundles provide a natural framework for coordinating locally distinct geometries. VB-GCL lifts the brain connectivity graph to a vector bundle, where each region of interest (ROI) is associated with a learnable orthogonal subspace, enabling the model to explicitly capture region-specific functional geometry. Based on this structure, we introduce a Bundle Laplacian that jointly encodes topological connectivity and local geometric orientation. Our framework performs multi-hop geometric diffusion on the bundle manifold for structural feature aggregation and classification, while simultaneously employing Bundle-Laplacian-driven spectral perturbations to generate contrastive views that encourage noise-robust representations. Extensive experiments on the ADNI and an in-house clinical dataset demonstrate that VB-GCL achieves state-of-the-art performance across multiple diagnostic tasks. Ablation studies and interpretability analysis further confirm the effectiveness of each component and reveal clinically meaningful pathological biomarkers, providing a promising tool for precision brain disease diagnosis.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5462_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/CodeMedYY/VB-GCL

Link to the Dataset(s)

N/A

BibTex

@InProceedings{YinFei_Learning_MICCAI2026,
        author = { Yin, Feiyu AND Xu, Jie AND Zhang, Qi AND Zhang, Yipin AND Fu, Junhu AND Yu, Ziqi AND Zhang, Shengjie AND Yu, Jinhua},
        title = { { Learning Geometry-Aware Bundles with Spectral Heat Kernel Diffusion for fMRI-based Brain Disease Diagnosis } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16886},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper introduces VB-GCL, a graph learning framework for fMRI-based brain disease diagnosis that replaces standard Euclidean embeddings with vector bundle theory. The core idea is that each brain region (ROI) gets its own learnable orthogonal subspace, and a bundle Laplacian handles geometry-aware message passing between regions. On top of that, the paper proposes a spectral contrastive learning strategy that generates augmented views through heat kernel diffusion rather than random graph perturbations. Taken together, the goal is a framework that better captures heterogeneous functional geometry while being more robust to fMRI noise.

  • 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 formulation is genuinely interesting. Using vector bundle theory for brain networks is a novel move in this space. Rather than pushing all ROIs into a shared feature space, each region gets its own subspace with a learned orthogonal transformation — which is a principled and meaningful way to handle inter-regional heterogeneity. Standard GNN assumptions don’t really account for this, so the departure feels well-motivated.
    • Heat kernel augmentation is a smart choice. Swapping out random edge/node dropping for heat kernel-based perturbations is one of the more thoughtful design decisions in the paper. Brain graph topology carries physical meaning, so more structured, less destructive augmentations make real sense here.
    • Results are solid. The method consistently outperforms baselines on both datasets (Table 1), including recent GCL methods. The gains aren’t dramatic, but they’re stable across metrics.
    • The ablation and interpretability work is a real plus. Table 2 does a reasonable job supporting each component’s contribution, and the ROI-level interpretability analysis is a strong point — the identified regions align with known disease-related areas, which matters a lot for clinical relevance.
  • 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 theoretical side is thin. The geometric motivation is appealing, but the paper provides no formal analysis or evidence that the proposed formulation actually reduces geometric misalignment or improves representation quality over standard GNNs. The heat kernel approximation used for augmentation also feels somewhat heuristic without further justification.
    • Reproducibility is a concern. Several important implementation details are missing — diffusion time parameters, specifics of the orthogonality constraint, architectural choices. There is also no mention of code availability.
  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    VB-GCL is a novel and conceptually interesting paper. The vector bundle framing is non-trivial, the geometry-aware diffusion is well-motivated, and the empirical results hold up consistently across baselines. The ablation and interpretability work also add credibility.

    The main concern is reproducibility — several implementation details are missing (e.g., diffusion time parameters, orthogonality constraint specifics, architectural choices) and there is no mention of code availability. The theoretical justification could also be strengthened, as the geometric motivation is clear but not formally backed up.

    I would move toward a full Accept if the authors either release the code or provide sufficient implementation details to reproduce the results.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 main concern was reproducibility. Since the authors mentioned in the rebuttal that they will release the code upon acceptance, I will update my rate to “Accept”.



Review #2

  • Please describe the contribution of the paper

    The paper proposes VB-GCL, a framework for fMRI-based brain disease diagnosis that uses vector bundle theory to model the heterogeneous geometry of brain regions. By assigning each ROI to a learnable orthogonal subspace, the method aims to capture region-specific functional geometry and alleviate the geometric misalignment of conventional GNNs. The authors also introduce a dual-stream framework that combines bundle-based geometric diffusion with spectral contrastive learning. Instead of heuristic graph augmentations such as edge dropping, the method generates contrastive views through heat kernel-based spectral perturbations to better preserve topology and improve robustness to noise. Experiments on the ADNI and an in-house dataset show that the proposed method outperforms several baselines in brain disease 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.

    S1.The paper explores a geometry-aware formulation for brain connectivity modeling, aiming to account for heterogeneous functional structure across regions.

    S2.The integration of diffusion-based representations with contrastive learning is a reasonable design choice.

    S3.The method is evaluated on the ADNI dataset and an in-house dataset, and shows improved performance over several baselines, suggesting potential empirical effectiveness.

  • 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.

    W1.The paper claims to “introduce a Bundle Laplacian,” but this construction appears to have been already proposed in prior work, most notably in Bundle Neural Networks (Bamberger et al., ICLR 2025). The current manuscript does not cite this paper, despite the substantial overlap in the core formulation. This weakens the novelty claim and makes the positioning with respect to prior work unclear.

    W2.Several key motivations and claims in the paper are stated rather strongly, but are not sufficiently supported by evidence. In particular, the introduction argues that conventional GNNs fail because different brain regions reside on heterogeneous functional manifolds with distinct local geometric properties, yet the cited references [1,26] do not seem to directly substantiate this claim, and the paper does not provide its own empirical evidence for it. Likewise, the paper claims that the proposed multiscale spectral diffusion preserves the coherence of the vector bundle and ensures “strict physical and geometric consistency across views,” but this notion is neither clearly defined nor experimentally validated. As a result, the rationale behind the proposed method is not fully convincing beyond the observed performance gains.

    W3.The paper also lacks sufficient explanation and citation for heat kernel and heat diffusion, even though these are core components of the proposed method.

    W4.The paper mentions a first-order approximation of the heat kernel but does not present the corresponding equation or theoretical justification. In addition, Section 2.3 is difficult to interpret due to undefined variables (e.g., X_B, X_spec) and unclear module interactions (e.g., the shared backbone).

    W5.The implementation details of the spectral perturbation module are incomplete (\tau_1, \tau2).

    W6.The term “CI” is used without definition, making the interpretability analysis unclear.

  • 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 does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/en/REVIEWER-GUIDELINES.html

    1) In Eq. (2), are W_k shared across all layers, or are they layer-specific? 2) Is the in-house dataset used in this work publicly available, or are there plans to release it?

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper suffers from unclear novelty, insufficient justification of its core claims, and lack of technical clarity. Key components appear closely related to prior work but are not properly discussed, while several strong claims are not supported by evidence. In addition, the method is under-specified, with missing formulations and unclear implementation details, making it difficult to interpret and reproduce.

  • 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.

    My main concern regarding novelty and scholarly positioning remains unresolved. In particular, the rebuttal acknowledges that the core bundle Laplacian formulation and vector-bundle-based heat diffusion framework are closely related to prior work such as Bundle Neural Networks (Bamberger et al., ICLR 2025) and the earlier connection Laplacian literature. These components constitute a substantial portion of the paper’s conceptual and mathematical foundation, yet the original submission neither cited nor properly positioned itself with respect to this line of work. I do not view this as a minor citation issue, but rather as a scholarly positioning problem that materially affected the originality assessment during review.



Review #3

  • Please describe the contribution of the paper

    This paper proposes VB-GCL, a novel GNN framework based on vector bundle theory for brain fMRI disease classification. Unlike traditional methods that project brain functional features into a common latent space, VB-GCL utilizes a ROI-wise trainable orthogonal projection matrix to define a bundle Laplacian. Furthermore, to mitigate the effects of fMRI signal noise, the authors implement a heat diffusion-based contrastive learning scheme. The proposed VB-GCL achieves SOTA performance on binary classification tasks using the ADNI dataset and an in-house clinical dataset.

  • 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 application of vector bundle theory for ROI-level orthogonal feature projection and between-node feature alignment is innovative and well-motivated. The proposed contrastive learning strategy is both straightforward and highly effective. The method is rigorously validated across multiple datasets, demonstrating consistent state-of-the-art results.

  • 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.Ambiguity in Figure 1: Part 2 of Figure 1 is somewhat confusing. Because the bundle transformation should not introduce new topological connections that do not exist in the original functional graph, the graph connectivity shown in panel B (the bundle subspace) should align with panel A. However, they appear visibly different. The authors should clarify this discrepancy or revise the figure accordingly to avoid misleading the readers. 2.Missing Complexity Analysis: The bundle transformation increases the size of the Laplacian matrix by a factor of $d$. Consequently, the authors should provide an explicit discussion regarding the computational complexity and memory overhead of VB-GCL, particularly in comparison to standard GNN baselines.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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?

    Although there are minor presentational issues regarding the figures, the overall quality of this paper, particularly its technical novelty and robust experimental design, well exceeds the acceptance threshold for MICCAI. I recommend this paper for acceptance, and believe it is a strong candidate for a spotlight paper.

  • 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 have satisfactorily addressed my concerns, and I appreciate their comprehensive response.



Author Feedback

We thank all reviewers for their constructive feedback. All clarifications will be incorporated into the final version.

[R1 & R2: Relation to BuNN (Bamberger et al., ICLR 2025)] We will cite this work. The bundle Laplacian L_B = O^T (L ⊗ I_d) O traces back to Singer & Wu (2012, Vector Diffusion Maps and the Connection Laplacian) , and both BuNN and VB-GCL build on it. However, VB-GCL differs from BuNN in three aspects. First, BuNN is a general-purpose GNN for heterophilic and long-range graph tasks targeting over-smoothing, over-squashing, and limited expressivity. VB-GCL is the first to apply vector bundle theory to fMRI-based brain disease diagnosis, to model ROI manifold heterogeneity. Second, BuNN relies on the full heat kernel, often requiring costly spectral methods, to enable global node communication. By contrast, VB-GCL employs a localized K-order polynomial filter with residual connections for efficient structural feature aggregation, which is well suited to brain networks. Third, our spectral contrastive learning module is an independent contribution. BuNN does not include a contrastive mechanism, whereas VB-GCL uses a first-order heat-kernel expansion specifically to generate weak and strong contrastive views. This design avoids computationally expensive matrix exponentials while providing noise regularization for small clinical fMRI cohorts. In addition, ROI-level saliency analysis for disease biomarkers, as shown in Fig. 3, is also unique to our work.

[R2-W2: Insufficiently supported claims] We will replace [1,26] with Bijsterbosch et al. (eLife, 2018) on region-specific spatial configurations and Vidaurre et al. (PNAS, 2017) on region-specific hierarchical temporal dynamics, supporting manifold heterogeneity. We will formalize these as a mathematical property. Two views are generated by (I - τ_t L ̂_B) driven by the same bundle Laplacian, sharing the eigenbasis of L ̂_B and remaining on the manifold induced by (O, L ̂_B), unlike edge dropping which destroys topology.

[R2-W5: Perturbation parameters] τ_1, τ_2 are not fixed but sampled per batch from two non-overlapping ranges corresponding to weak and strong diffusion scales. This stochastic sampling acts as built-in augmentation, exposing the contrastive objective to multi-scale perturbations.

[R2-Q1: Are W_k shared across layers?] W_k are not shared. Within a layer, each polynomial order k ∈ {0, …, K} has its own W_k. Across layers, each layer maintains an independent set {W_0, …, W_k} to capture distinct geometric scales.

[R2-Q2: In-house dataset availability] Due to IRB constraints, public release is not permitted. Qualified researchers may contact the corresponding author and obtain authorized access for non-commercial research after a data-use agreement.

[R4-2: Complexity] VB-GCL has 13K parameters, 397 MB peak memory, and 5 ms inference per subject on a V100 GPU with batch size 32, compared with 3K parameters, 69 MB, and 3 ms for the GCN backbone. Despite the 3.8-fold parameter and 5.8-fold memory increase from bundle Laplacian operations, the absolute cost remains lightweight and clinically practical.

[R1: Reproducibility] To clarify the missing reproducibility details, we use d=8, K=3 polynomial filter, QR-based orthogonality enforcement, and Focal Loss with adaptive class weighting. Due to the double-blind policy, code is currently private; the full implementation will be released upon acceptance.

[Additional clarifications: R2-W3, R2-W4, R2-W6, and R4-1] We will define H(τ) = exp(−τ L̂_B) ≈ I − τ L̂_B. X_B denotes the bundle-transformed feature in Sec. 2.1, and X_spec^(i) denotes the contrastive views in Eq. 3.The shared backbone is the bundle diffusion encoder applied across streams. CI denotes Cognitive Impairment, consistent with Sec. 3.Fig. 1 will be revised to ensure that the bundle subspace preserves the original graph connectivity.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    This paper proposes a vector bundle-based graph contrastive learning framework for fMRI analysis, and reviewers acknowledge the conceptual novelty and promising empirical performance; The overall assessment is highly mixed (scores: 3 / 2 / 5). While the geometric formulation is interesting, several reviewers raise concerns about unclear novelty relative to recent work, insufficient theoretical justification, and lack of proper citation and positioning. In addition, important methodological details are missing, key claims are not fully supported by evidence, and reproducibility and clarity issues make the approach difficult to assess. The authors are encouraged to address these concerns in the rebuttal, particularly by clarifying novelty, strengthening theoretical and empirical justification, and improving methodological transparency.

  • 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.

    Based on the reviewer votes, acceptance is recommended. However, the authors should properly acknowledge that the core bundle Laplacian formulation and vector-bundle-based heat diffusion framework are closely related to prior works regarding bundle neural networks and other Laplacian literature.



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.

    After carefully reviewing the authors’ rebuttal, I have revised my initial assessment to Accept. The authors directly addressed key concerns including clarifying novelty relative to BuNN (ICLR 2025) with three distinct differences, replacing insufficient citations with evidence supporting ROI manifold heterogeneity, providing computational complexity and hyperparameter details, and committing to code release upon acceptance.



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.

    The paper presents a conceptually interesting geometry-aware graph contrastive learning framework for fMRI-based disease diagnosis, with promising empirical results and useful ablation and interpretability analyses. The reviewers were mixed, mainly due to concerns about novelty, missing methodological details, reproducibility, and insufficient positioning relative to prior vector-bundle and heat-kernel diffusion work. After rebuttal, most concerns were reasonably addressed, including code release, implementation details, complexity analysis, and clarification of the relation to prior work. I recommend acceptance, but the final version must include proper citations and a clear scholarly positioning with respect to Bundle Neural Networks, connection Laplacian, vector diffusion maps, and heat-kernel diffusion literature.



back to top