NeuroPathNet: Dynamic Path Trajectory Learning for Brain Functional Connectivity Analysis

TL;DR

The paper addresses the challenge of capturing temporal evolution in brain functional networks by introducing NeuroPathNet, a path-level trajectory learning framework. It constructs dynamic connectivity paths between static functional partitions using sliding-window fMRI correlations and models each path with a temporal Transformer, followed by a global cross-path fusion for classification. NeuroPathNet achieves state-of-the-art results across three public datasets (ADNI, ABIDE, HCP) on multiple tasks, demonstrating stronger discriminative power and balanced performance. This approach enhances interpretability of dynamic brain interactions and holds promise for clinical applications in neurological disease diagnosis.

Abstract

Understanding the evolution of brain functional networks over time is of great significance for the analysis of cognitive mechanisms and the diagnosis of neurological diseases. Existing methods often have difficulty in capturing the temporal evolution characteristics of connections between specific functional communities. To this end, this paper proposes a new path-level trajectory modeling framework (NeuroPathNet) to characterize the dynamic behavior of connection pathways between brain functional partitions. Based on medically supported static partitioning schemes (such as Yeo and Smith ICA), we extract the time series of connection strengths between each pair of functional partitions and model them using a temporal neural network. We validate the model performance on three public functional Magnetic Resonance Imaging (fMRI) datasets, and the results show that it outperforms existing mainstream methods in multiple indicators. This study can promote the development of dynamic graph learning methods for brain network analysis, and provide possible clinical applications for the diagnosis of neurological diseases.

Figures (2)

• Figure 1: The model takes fMRI data as input, partitions the brain into functional modules, and computes inter-module correlations over time to construct dynamic path sequences. Each path, encoded as a time-series vector with positional information, is fed into a Path Transformer to model temporal interactions. Cross-path attention then integrates all path representations, followed by attention pooling and classifier to produce the final prediction.
• Figure 2: The similarity matrices of sample representations learned by TokenGT, Graphormer, and NeuroPathNet