Multi-Cluster Self-Paced Diverse Learning and Feature Fusion for Autism Spectrum Disorder Identification With Resting-State fMRI

Brain functional connectivity (FC) derived from resting-state functional magnetic resonance imaging (rs-fMRI) has been widely applied to study autism spectrum disorder (ASD). Existing studies usually suffer from 1) significant data heterogeneity, 2) high dimensional rs-fMRI data, 3) inevitable noise. To this end, we propose a multi-cluster self-paced diverse learning and feature fusion (MCSDLFF) method for rs-fMRI-based ASD identification. Specifically, we first divide multi-site training subjects into ASD and healthy control (HC) groups. To model data heterogeneity within each category, we employ a similarity-driven multi-kernel learning method to cluster the FC networks within each group into different subpopulations. To alleviate the impact of noise lies in the data and improve the robustness of the model, we design a self-paced diverse learning method to extract FC features within each subpopulation, which can take both easiness and diversity of the subjects into consideration during the training process. Then we enhance the cluster-consistent features within each category via intra-class feature weight fusion and category-specific features across ASD and HC groups via inter-class feature weight fusion, respectively. Finally, a linear support vector machine (SVM) is used for ASD identification. Experimental results on 1035 subjects from the ABIDE dataset on 17 imaging sites suggest that the proposed MCSDLFF outperforms several state-of-the-art methods in ASD identification. The most discriminative FCs identified by the MCSDLFF are mainly located in default mode network, somatomotor network and cerebellum region.