Autism Classification Based on Multi-atlas Graph Attention Networks and Metric Learning

Autism spectrum disorder (ASD) is a complex neurodevelopmental disorder characterized by significant impairments in brain function. However, current research primarily focuses on extracting brain network features in Euclidean space, often overlooking the critical topological properties of brain networks. Moreover, the reliance on a single brain atlas for segmenting brain regions restricts the utilization of complementary insights provided by multiple atlases. To overcome these limitations, we introduce a novel approach called multi-atlas graph attention networks and metric learning (MAGM). Our method utilizes two brain atlases to define brain regions more comprehensively. We then implement a metric learning-based graph attention network to extract discriminative features for ASD classification. Additionally, we propose a dynamic attention fusion mechanism to effectively integrate features obtained from both atlases. The effectiveness of our MAGM model is demonstrated through classification results on the ABIDE dataset. Our method not only improves the classification performance but also identifies crucial brain regions associated with ASD.