Abstract:Abstract Autism spectrum disorder (ASD) is a neurodevelopmental disorder that causesrepetitive and stereotyped behavior and social dysfunction. Early diagnosis andintervention can improve the effectiveness of treatment. However, the currentdiagnostic methods of ASD still use the subjective symptom criteria of clinicalobservation, which is time-consuming and labor-intensive. In recent years, theemergence of functional magnetic resonance imaging (fMRI) neuroimaging techniques has facilitated the identification of potential biomarkers for the diagnosisof ASD. In this study, we developed a deep learning framework called Spatio-Temporal Autism Spectrum Disorder Network (ST-ASDNet) to distinguish ASDsfrom normal controls (NCs) as well as to distinguish ASD subtypes based onthe fMRI data. Specifically, two modules: 1) bidirectional long and short-termmemory Transformer (BLSTM-Transformer) and 2) full convolutional networkTransformer (FCN-Transformer) are proposed to obtain the spatial and the temporal characteristics of the fMRI data, respectively. Additionally, we explored theinfluence of different brain atlas for preprocessing the fMRI data to recognize theASDs. The study verifies that a more detailed division of brain regions is helpfulfor the recognition of ASDs. The proposed model is evaluated on the Craddock200 (CC200) atlas and the Automated Anatomical Labeling (AAL) atlas with the robust accuracy of 70.26% and 67.80%, respectively. The effect of the proposed model is competitive in diagnosing ASD compared with other methods,which could provide a useful and effective way to assist the clinical diagnosis.

ST-ASDNET: A BLSTM-FCN-Transformer basedASD Classification Model for Time-series fMRI