BrainMAP: Learning Multiple Activation Pathways in Brain Networks

TL;DR

BrainMAP addresses the challenge of learning multiple long-range activation pathways in brain networks by transforming FC graphs into informative sequences and applying a hierarchical MoE framework. It introduces Adaptive Graph Sequentialization to reveal information flow and a two-stage Pathway Integration (within-order MoE and cross-order aggregation) to capture diverse pathways. Empirical results on HCP and NeuroGraph datasets show BrainMAP achieving state-of-the-art predictive performance and providing meaningful explanations of task-related brain regions, with ablations highlighting the importance of both sequentialization and MoE. The approach offers a principled means to analyze brain activity with improved accuracy and interpretability, with code available for reproducibility.

Abstract

Functional Magnetic Resonance Image (fMRI) is commonly employed to study human brain activity, since it offers insight into the relationship between functional fluctuations and human behavior. To enhance analysis and comprehension of brain activity, Graph Neural Networks (GNNs) have been widely applied to the analysis of functional connectivities (FC) derived from fMRI data, due to their ability to capture the synergistic interactions among brain regions. However, in the human brain, performing complex tasks typically involves the activation of certain pathways, which could be represented as paths across graphs. As such, conventional GNNs struggle to learn from these pathways due to the long-range dependencies of multiple pathways. To address these challenges, we introduce a novel framework BrainMAP to learn Multiple Activation Pathways in Brain networks. BrainMAP leverages sequential models to identify long-range correlations among sequentialized brain regions and incorporates an aggregation module based on Mixture of Experts (MoE) to learn from multiple pathways. Our comprehensive experiments highlight BrainMAP's superior performance. Furthermore, our framework enables explanatory analyses of crucial brain regions involved in tasks. Our code is provided at https://github.com/LzyFischer/Graph-Mamba.

Figures (5)

• Figure 1: An illustration of the transition from fMRI data to FC graphs, along with two challenges for learning from pathways in FC graphs: (1) The sequential dependency, a fundamental feature of human brain activity, is not naturally presented in FC graphs; (2) Multiple pathways exist in FC graphs, making the extraction of them more difficult.
• Figure 2: The overall process of BrainMAP. We first adaptively learn $M$ ($M=3$ in the figure) orders with three order-learners (GNNs). Then these orders are input into the gating function to select $K$ ($K=2$ in the figure) experts from a total number of $P$ ($P=4$ in the figure) experts. Each expert is implemented as a sequential model. The output of these experts will be aggregated into a representation for each order. Finally, the representations from all orders are aggregated into the output.
• Figure 3: Interpretation results of BrainMAP for the task MOTOR in HCP-Task. The average salient regions from random samples. The color bar ranges from 0.4 to 1. The bright-yellow color indicates a high score, while dark-red color indicates a relatively lower score. The ground-truth brain regions given by domain experts are circled in blue.
• Figure 4: The results of varying the number of experts in the MoE on four HCP benchmark datasets.
• Figure 5: The activation distribution of the experts across different model layers. BrainMAP consistently maintains high activation rates on various HCP datasets.

Theorems & Definitions (4)

• Theorem 4.1
• Theorem 4.2
• proof
• proof