A graph structure clustering-based spatial-temporal graph convolutional network model for autism spectrum disorder classification

In this paper, we propose a graph structure clustering-based spatio-temporal graph convolutional network (GSC-GCN) to classify Autism Spectrum Disorder. Firstly, a sliding window with fixed size is utilized to sample the preprocessed restingstate functional magnetic resonance imaging time series. Then functional connectivity matrices for all the subsequences are constructed and a series of dynamical temporal and spatial graphs are built accordingly. Next, to effectively reduce dimensionality and capture key functional connections, three different sizes of coarsening graphs are generated using clustering methods. Subsequently, a spatial-temporal graph convolution network is designed which focuses on the temporal dynamics of graph series as well as the spatial characteristics of functional brain network. In order to verify the effectiveness of our model, we test our GSC-GCN on the Autism Brain Imaging Data Exchange dataset and compare it with the state-of-the-art techniques. Results show that GSC-GCN outperforms other methods, indicating that the comprehensive use of both temporal and spatial features deserves further exploration.