Multi-Source Domain Adaptation via Optimal Transport for Brain Dementia Identification

Multi-site MRI data have been increasingly employed for automated identification of brain dementia, but are susceptible to large domain shift between different imaging sites/centers. Previous studies usually simply ignore the domain shift caused for instance by different scanners/protocols. Even though several studies proposed to reduce inter-domain discrepancy, they generally require a part of labeled target data and cannot well handle problems with multi-source domains. To this end, we propose a multi-source optimal transport (MSOT) framework for cross-domain Alzheimer’s disease (AD) diagnosis with multi-site MRI data. Specifically, we first project data from multi-source domains to target domain through optimal transport in an unsupervised manner. Based on projected representation, we calculate the similarity between each source and target domains, and use this similarity as the source domain weight. We then train a support vector machine (SVM) classifier based on projected samples from each source domain. Finally, an ensemble learning strategy via weighted voting is used to predict labels of target samples. The proposed MSOT does not require labeled target data and can be efficiently optimized. Experiments were performed on three benchmark neuroimaging datasets for AD identification, with results suggesting the superiority of MSOT over several state-of-the-art methods.