List of Papers Browse by Subject Areas Author List
Abstract
Dynamic functional brain network analysis using rs-fMRI has emerged as a powerful approach to understanding brain disorders. However, current methods predominantly focus on pairwise brain region interactions, neglecting critical high-order dependencies and time-varying communication mechanisms. To address these limitations, we propose the Long-Range High-Order Dependency Transformer (LHDFormer), a neurophysiologically-inspired framework that integrates multiscale long-range dependencies with time-varying connectivity patterns. Specifically, we present a biased random walk sampling strategy with NeuroWalk kernel-guided transfer probabilities that dynamically simulate multi-step information loss through a $k$-walk neuroadaptive factor, modeling brain neurobiological principles such as distance-dependent information loss and state-dependent pathway modulation. This enables the adaptive capture of the multi-scale short-range couplings and long-range high-order dependencies corresponding to different steps across evolving connectivity patterns. Complementing this, the time-varying transformer co-embeds local spatial configurations via topology-aware attention and global temporal dynamics through cross-window token guidance, overcoming the single-domain bias of conventional graph/transformer methods. Extensive experiments on ABIDE and ADNI datasets demonstrate that LHDFormer outperforms state-of-the-art methods in brain disease diagnosis. Crucially, the model identifies interpretable high-order connectivity signatures, revealing disrupted long-range integration patterns in patients that align with known neuropathological mechanisms.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0949_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{XueRun_Adaptive_MICCAI2025,
author = { Xue, Rundong and Han, Xiangmin and Hu, Hao and Zhang, Zeyu and Du, Shaoyi and Gao, Yue},
title = { { Adaptive Embedding for Long-Range High-Order Dependencies via Time-Varying Transformer on fMRI } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15971},
month = {September},
page = {45 -- 54}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes LHDFormer, a transformer-based framework for modeling dynamic functional connectivity in resting-state fMRI (rs-fMRI) data.
- 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 work targets known limitations in modeling long-range and temporal dependencies in dynamic brain networks, offering a compelling neurophysiologically-inspired solution.
- 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 NeuroWalk kernel is time-varying, the sliding-window mechanism itself is fixed. This may constrain temporal adaptivity in capturing event boundaries or abrupt changes in connectivity.
The evaluation omits comparisons with recent foundation models or multi-resolution temporal architectures that may offer alternative solutions to long-range modeling.
The model ignores the complex information hindered in the BOLD signal fluctuations and only utilizing the constructed FC to extract BOLD signal features.
- 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
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 paper addresses an important challenge in dynamic brain network modeling with a creative and neurobiologically plausible framework. Its empirical improvements and architectural novelty are commendable. However, further work is needed to enhance interpretability, evaluate generalization, and streamline complexity for broader adoption. With refinement, this direction has the potential to contribute meaningfully to neuroimaging-based diagnosis and representation learning.
- 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 answer of the author is not convincible
Review #2
- Please describe the contribution of the paper
Authors propose an advanced transformer architecture that integrates multiscale long-range time dependencies and time-varying connectivity patterns in rsfMRI data. They demonstrate superior performance in AD and ASD classification tasks and perform ablation studies to justify the architecture.
- 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.
- a random walk strategy guided by time-varying neuroadaptive factors to simulate biologically plausible information propagation is proposed to capture multiscale long range dependencies in temporal dimensions
- Another time-varying transformer is introduced, which encodes both spatial and temporal features of brain connectivity via topology-aware attention and cross-window temporal integration.
- Experiments and comparisons are performed on two well-known datasets, ADNI and ABIDE
- 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 details in the paper is insufficient to reproduce the work and authors do not provide access to the code
- the computational cost of the k-walk kernel and its scaling with larger brain parcellations is not discussed.
- Hyperparameter sensitivity analysis is insufficient, particularly for the number of walk steps, which is a key design choice in the proposed architecture.
- 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
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?
Pros:
- a random walk strategy guided by time-varying neuroadaptive factors to simulate biologically plausible information propagation is proposed to capture multiscale long range dependencies in temporal dimensions
- Another time-varying transformer is introduced, which encodes both spatial and temporal features of brain connectivity via topology-aware attention and cross-window temporal integration.
- Experiments and comparisons are performed on two well-known datasets, ADNI and ABIDE
Cons:
- The details in the paper is insufficient to reproduce the work and authors do not provide access to the code
- the computational cost of the k-walk kernel and its scaling with larger brain parcellations is not discussed.
- Hyperparameter sensitivity analysis is insufficient, particularly for the number of walk steps, which is a key design choice in the proposed architecture.
- 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 concern was the complexity and the reproducibility of the algorithm. Authors clarified some of the concerns and promised to release the code.
Review #3
- Please describe the contribution of the paper
This paper propose a framework for analyzing the rsfMRI. The key contribution focus on using the neurowalk kernel to formulate the long-range high-order dependencies and using the time-varying transformer to capture the spatiotemporal patterns of dynamic FC.
- 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 limitation of previous work, as well as the limitation of current work is clearly explained.
- The solutions for the corresponding limitations are well motivated.
- The ablation study is well organized so that the readers could see the contribution of each component.
- 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.
- It would be better to mention the computational cost to increase the steps for neurowalk, which may be important to apply this work in practice.
- In the last figure of this work, it is mentioned that the performance achieved the best at the step of 16. It would be very interesting if the authors could provide more discussion on how the brain formulate such long dependency, and if the order of brain regions in these long sequences mean anything.
- 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
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?
To be honest, there are lots of work concentrate on using rsfMRI for disease diagnosis. This work stands out for me because their formulation of using neurowalk to represent long-range dependency is interesting and may need further investigation.
- 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.
Thanks for the response from authors. I recommend to accept this paper for its novelty and clarity.
Author Feedback
We appreciate all reviews in depth. The main concerns are addressed below.
[R1] Reproducibility: We promise to release the source code after the paper is accepted to ensure reproducibility.
[R1, R2]Computational Cost: The computational complexity of the NeuroWalk kernel is principally governed by the walk length K and the number of brain regions N, scaling as O(KN2). Empirical evaluations on the ABIDE dataset demonstrate practical feasibility: with a walk length of 16, the model contains 7,642,586 parameters, achieves a training time of 11 ms per sample, an inference time of 2.12 ms, and a memory footprint of 2.56 GB. These metrics confirm the framework’s applicability in real-world clinical and research settings. Regarding scalability in large parcellations, the CC200 atlas used in this paper is already a large brain parcellation and is one of the commonly used atlases, balancing spatial resolution with computational cost. While larger parcellations could enhance spatial resolution, they exponentially increase computational demands without necessarily improving accuracy. Also, due to space limitations, we did not discuss scaling with larger parcellations in this paper, and we will improve this part in future work.
[R1, R2]Analysis of walk steps: The parameter sensitivity analysis reveals an optimal walk length of K=16, where shorter walks (K=1-4) primarily capture localized functional coupling within subnetworks (e.g., intra-default mode network synchronization), while medium-length walks (K=8-16) effectively encode long-range dependencies for disease characterization. Notably, our identified OLF-PHG-THA-PCG-MOG-MTG trajectory in ADNI analyses (Figure 2(a)) spatially aligns with Braak staging patterns of Alzheimer’s pathology progression (Petersen et al., 2016), confirming biological relevance. Performance degradation at K>16 suggests diminishing returns from redundant pathway information.
[R3] Q1 Sliding-Window mechanism: Regarding the adaptability of the sliding window, we think that while the window length remains fixed, our approach effectively captures gradual evolutionary patterns in FC states through learning across windows. This design aligns with established evidence demonstrating that FC transitions typically exhibit temporal cumulative effects rather than abrupt changes (Hutchison et al., 2013). The sliding-window technique has been extensively validated as a benchmark approach in dynamic FC analysis (Gadgil S et al., 2020; Kong Y et al., 2021), with its scientific validity consistently demonstrated across numerous studies and clinical applications. [R3] Q2 Compared Methods. This study addresses the challenge of modeling long-range dependencies in brain networks via a graph transformer. For foundation models, their requirement for pretraining compromises fair performance comparison, so we do not compare them. As requested by the reviewer, we compare the SOTA foundation model HGFM (Han X et al., 2025), where our method demonstrated superior accuracy (71.87% vs. 74.29 on the ABIDE dataset). For multi-resolution approaches, we compared MSSTAN (TMI) and STGCN (MICCAI) in Tab 1. However, these GNN-based methods focus on the aggregation of neighborhood information, i.e., short-range dependencies, which cannot offer alternative solutions to the long-range modeling. [R3] Q3 About FC. The process for BOLD extraction follows standard fMRI processing pipelines (slice timing correction, head motion correction, etc.) that effectively isolate neural signals from physiological noise (Power et al., 2014; Ciric et al., 2017). The practice of deriving FC as the feature representing BOLD signal is the most common approach (mainstream paradigm) in brain network analysis (Li X et al., 2021; Cui H et al., 2022; Yu S et al., 2024), as the FC features have demonstrated consistent biological interpretability across clinical and cognitive studies(Park H J et al., 2013), thereby providing robust theoretical support for our approach.
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.
Reject
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The paper presents a creative and neurobiologically inspired approach to modeling dynamic brain networks, addressing an important challenge in the field. However, key concerns around temporal adaptivity, missing comparisons with relevant baselines, and limited handling of BOLD signal complexity remain insufficiently addressed in the rebuttal. Given these unresolved issues, I recommend rejection at this stage.