Graph neural network with modular attention for identifying brain disorders

Abstract

Listen

Translate article

Functional Magnetic Resonance Imaging (fMRI), by detecting the cerebral Blood Oxygen Level-Dependent (BOLD) signals, has developed into an effective technique to aid in the early identification of neurological or psychiatric disorders. Traditionally, Functional Brain Graphs (FBGs) are first estimated from BOLD fMRI data, followed by feature extraction from these graphs as the input of the downstream classifier for disorder identification. Unlike traditional step-by-step approaches, recently deep learning methods, such as transformers and attention-based graph neural networks, provide an end-to-end architecture for learning and classifying FBGs. The representative methods include Graph Attention Networks (GATs) and Brain Network Transformer (BNT). Although they work well for identifying some brain disorders, these deep learning methods involve complex network architecture and fail to take full advantage of the prior knowledge from brain function, resulting in the difficulty of understanding the relationship between input features and model predictions. Inspired by the fact that brain structure and function can be divided into multiple relatively independent and specialized modules as found in cognitive neuroscience. In this paper, we propose a Graph Neural Network with Modular Attention (GNNMA), which introduces a modular attention mechanism into GATs and BNTs through Singular Value Decomposition (SVD). This method reduces the dimensionality of high-dimensional graph data while retaining the most important information, thereby better capturing the modular features within the graph and improving both the accuracy of recognition and the interpretability of the results. Our model is tested on the Autism Brain Imaging Data Exchange I (ABIDE I) dataset and the Major Depressive Disorder (MDD) dataset, achieving an accuracy of up to 72.1%, exhibiting encouraging performance. Additionally, the proposed scheme not only enhance the interpretability of deep models, but also effectively captures the modular functional structure of the brain.