Graph Convolutional Networks Based on Relational Attention Mechanism for Autism Spectrum Disorders Diagnosis

Nowadays, Autism spectrum disorder (ASD) is a neurodevelopmental disorder that severely affects social communication. The diagnostic criteria depend on clinicians' subjective judgment of the patient's behavioral criteria. Obviously, it is an urgent problem to establish an objective diagnosis method for patients with ASD. To address this problem, we propose a novel graph convolutional network(GCN) method based on relational attention mechanism. Firstly, we extract functional connectivity (FC) between brain regions from functional magnetic resonance (fMRI) effects that respond to blood oxygenation signals in the brain. Considering the different relationships between subjects, population relations are then modeled by graph structural models as a way to jointly learn population information. Finally, for individual-specific information, a relational attention mechanism is used to generate relationships between subjects and GCN is utilized to learn their unique representational information. Our proposed method is evaluated 871 subjects (including 403 ASD subjects and 468 typical control (TC) subjects) from the Autism Brain Imaging Data Exchange (ABIDE). The experimental results show that the mean accuracy and AUC values of our proposed method can obtained 90.57% and 90.51%, respectively. Our proposed method has achieved state-of-the-art performance in the diagnosis of ASD compared to some methods published in recent years. Overall, our method is effective and informative in guiding clinical practices.