Multi-site diagnostic classification of Autism spectrum disorder using adversarial deep learning on resting-state fMRI

Abstract

Existing imaging studies in Autism spectrum disorder (ASD) mainly focused on single-site resting-state functional magnetic resonance imaging (rs-fMRI) data, which may contain limited samples and suffer from geographic bias. The lack of effective detection in information mining of rs-fMRI is a main reason for affecting the recognition rate of multi-site diagnostic. This study aims to propose a two-stage adversarial learning model with sliding window that integrates information from multi-site rs-fMRI data at the cost of minimal information loss. First, single rs-fMRI data is sampled with a sliding window to preserve both the spatial and temporal information of original data. And then, site-shared features of these samples are extracted through an adversarial learning model. Finally, the model is fine-tuned to learn discriminative disease-related features. Experimental results show that through adversarial learning, the heterogeneity problem among multi-site data is solved. Furthermore, the spatial–temporal information on rs-fMRI is effectively extracted and yields better classification performance (0.80 accuracy, 0.81 sensitivity, 0.80 specificity) than the state-of-the-art methods. Our results demonstrate the feasibility of the proposed method in the ASD classification task and the importance of fully exploiting the site-shared and spatial–temporal information in rs-fMRI data for multi-site ASD study.