Hierarchical feature extraction on functional brain networks for autism spectrum disorder identification with resting-state fMRI data

Autism spectrum disorder (ASD) is a pervasive developmental disorder of the central nervous system, which occurs most frequently in childhood and is characterized by unusual and repetitive ritualistic behaviors. Currently, diagnostic methods primarily rely on questionnaire surveys and behavioral observation, which may lead to misdiagnoses due to the subjective evaluation and measurement used in these traditional methods. With the advancement in medical imaging, MR imaging-based diagnosis has become an alternative and more objective approach. In this paper, we propose a Hybrid neural Network model for ASD identification, termded ASD-HNet, to hierarchically extract features on the functional brain networks based on resting-state functional magnetic resonance imaging data. This hierarchical method can better extract brain representations, improve the diagnostic accuracy, and help us better locate brain regions related to ASD. Specifically, features are extracted from three scales: local regions of interest (ROIs) scale, community-clustering scale, and the whole-communities scale. For the ROI scale, graph convolution is used to transfer features between ROIs. At the community cluster scale, functional gradients are introduced, the clustering algorithm K-Means is used to automatically cluster ROIs with similar functional gradients into several communities, and features of ROIs belonging to the same community are extracted to characterize these communities. At global information integration scale, we extract global features from community-scale brain networks to characterize the whole brain networks. We validate the effectiveness of our method using the public dataset of Autism Brain Imaging Data Exchange I (ABIDE I), and elucidate the interpretability of the method. Experimental results demonstrate that the proposed ASD-HNet can yield superior performance than compared methods.