Enhanced Spatial and Temporal Feature Integration for fMRI-Based ASD Classification

Autism Spectrum Disorder (ASD) poses diagnostic challenges due to its varied manifestations and the absence of the specific biological test. To improve the auxiliary diagnosis of ASD, this work proposes a novel deep learning framework called BWA-GCN, which capitalizes on the strengths of Graph Convolutional Networks (GCN), Bidirectional Long Short-Term Memory (BiLSTM) networks, and attention mechanisms. The advantage of this model is by incorporating BiLSTM for the significant temporal feature extraction and GCN for the spatial information aggregation. The attention mechanisms are adopted in this model to enhance the precision of ASD classification. Extensive experiments based on the ABIDE dataset with CC200 and AAL brain atlas prove the good performance of this model, with the accuracy of 70.92% on the CC200 atlas and 68.72% on the AAL atlas. This research contributes a new method to the research of ASD classification, shedding light on the spatial-temporal dynamics of fMRI time-series data in ASD diagnosis.