An Interpretable Deep Graph Neural Network Based On Attentional Multi-Scale Feature Fusion for FMRI Analysis

Abstract

Translate Article

Understanding which brain regions are associated with specific neurological disorders has been an important area of neuroimaging research, which has important implications for biomarker and diagnostic studies. In this paper, we propose an interpretable deep graph neural network (GNN) framework to analyze functional magnetic resonance images (fMRI) and discover neurological biomarkers, namely IDGNN. Specifically, we design a novel deep graph convolutional layer to better utilize the spatial and functional information of fMRI, and aggregate feature information of multi-hop neighbor regions of interest (ROIs) more efficiently. Considering the need for interpretability in brain image analysis, we improve the Gradient Class Activation Mapping (Grad-CAM) technique for fMRI brain graphs as a way to discover the most significant ROIs identified by IDGNN from brain connectivity patterns. Furthermore, we employ an attentional feature fusion mechanism to better perform fusion of multi-scale features. We apply our IDGNN on ABIDE fMRI dataset. Results show that our method outperforms several competing methods, and successfully identifies biomarkers of autism spectrum disorders (ASD).