Unsupervised Domain Adaptation for Multi-Center Autism Spectrum Disorder Identification

Abstract

Improving autism spectrum disorder (ASD) diagnosis through effective utilization of multi-center data has attracted increasing attention recently. However, most previous studies do not take the distribution discrepancy among multi-center datasets into consideration, which may degrade the diagnosis performance based on multi-center data. To address this issue, we propose a multi-center domain adaptation (MCDA) method for ASD diagnosis. Most prior researches have explored two learning strategies independently for domain adaptation: feature matching and instance reweighting. In this paper, MCDA is designed to model these two strategies in a unified optimization problem. Specifically, we first choose one center as the target domain and other centers as source domains. Then, we reduce the cross-domain difference by jointly matching the features and reweighting the instances in a principled dimension reduction procedure. Through the procedure our method construct a new feature representation which is invariant to both the distribution difference and the irrelevant instances for both target and source domains. Based on the learned feature representation, we employ the k-nearest neighbor (KNN) algorithm to perform disease classification. Out method has been evaluated on the ABIDE database, and the superior classification results verify that MCDA can significantly outperform competitive methods for multi-center ASD diagnosis.