Panoramic brain network analyzer: A residual graph network with attention mechanism for autism spectrum disorder diagnosis

Abstract

Autism Spectrum Disorder (ASD) is a prevalent neurodevelopmental disorder nowadays, which is featured by the deficits in reciprocal social communication and the presence of restricted and repetitive patterns of behaviors. It is generally acknowledged that the resting-state functional magnetic resonance imaging (fMRI) for the brain functional connectivity (FC) detection is one of the most effective ways in predicting ASD. However, many challenges still exist, e.g., the gradient vanishing in deep GCN networks, the difficulties in localizing potential biomarkers for diagnosis. To address these issues, in this paper we propose a new ASD diagnostic model, called Panoramic Brain Network Analyzer (PBNA). The main advantage of our new model is to introduce the residual techniques and various attention mechanisms to deepen GCN architecture, which enables to learn more advanced information. Additionally, an innovation of the current graph pooling methods is also given, in which we incorporate the softmax and straight-through to alleviate dimensionality explosion. The empirical results on the ABIDE CC200, CC400 and AAL datasets demonstrate the superiority of PBNA, these evidences support PBNA to be a more accurate and efficient clinical diagnosis. More precisely, by utilizing a five-fold cross-validation strategy, the ACC indicators of PBNA on the three datasets could reach 75.77%, 74.11%, 74.65%, respectively, surpassing most of the state-of-the-art diagnostic methods.