Using DeepGCN to identify the autism spectrum disorder from multi-site resting-state data

Abstract It is challenging to discriminate Autism spectrum disorder (ASD) from a highly heterogeneous database, because there is a great deal of uncontrollable variability in the data from different sites. The enormous success of graph convolutional neural networks (GCNs) in disease prediction based on multi-site data has sparked recent interest in applying GCNs in diagnosis of ASD. However, the current research results are all based on shallow GCNs. The main objective of this research was to improve the classification results by using DeepGCN. We constructed a deep ASD diagnosing framework based on 16-layer GCN. And ResNet units and DropEdge strategy were integrated into the DeepGCN model to avoid the vanishing gradient, over-fitting and over-smoothing. We combined the scale information with neuroimaging to form a graph structure based on the ABIDE dataset I, which contains a total of 871 subjects from 17 sites. We compared the DeepGCN results with well-established models based on the same subjects. The mean accuracy of our classification algorithm is 73.7% for classifying ASD versus normal controls (GCN: 70.4%, SVM-l2: 66.8%, Metric Learning: 62.9%). We introduce a new perspective for studying the biological markers of early diagnosis of ASD based on multi-site and multi-modality data. Meanwhile, it can be easily applied to various mental illnesses.