Adaptive Embedding for Long-Range High-Order Dependencies via Time-Varying Transformer on fMRI

Rundong Xue, Xiangmin Han^*, Hao Hu, Zeyu Zhang, Shaoyi Du^*, Yue Gao

Accepted by MICCAI 2025

[Paper]

Overview

Figure 1. The framework of the proposed LHDFormer.

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.

Dependencies

• python=3.9
• cudatoolkit=11.3
• torchvision=0.13.1
• pytorch=1.12.1
• torchaudio=0.12.1
• wandb=0.13.1
• scikit-learn=1.1.1
• pandas=1.4.3
• hydra-core=1.2.0

Get Started

1. Data Preparation

Download the ABIDE dataset from here.

2. Usage

Run the following command to train the model.

sh main.sh

Cite our work

@inproceedings{xue2025adaptive,
  title={Adaptive Embedding for Long-Range High-Order Dependencies via Time-Varying Transformer on fMRI},
  author={Xue, Rundong and Han, Xiangmin and Hu, Hao and Zhang, Zeyu and Du, Shaoyi and Gao, Yue},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={46--55},
  year={2025},
  organization={Springer}
}

License

The source code is free for research and educational use only. Any commercial use should get formal permission first.

This repo benefits from BNT and ALTER. Thanks for their wonderful works.