Unsupervised Site Adaptation by Intra-site Variability Alignment

A medical imaging network that was trained on a particular source domain usually suffers significant performance degradation when transferred to a different target domain. This is known as the domain-shift problem. In this study, we propose a general method for transfer knowledge from a source site with labeled data to a target site where only unlabeled data is available. We leverage the variability that is often present within each site, the intra-site variability, and propose an unsupervised site adaptation method that jointly aligns the intra-site data variability in the source and target sites while training the network on the labeled source site data. We applied our method to several medical MRI image segmentation tasks and show that it consistently outperforms state-of-the-art methods.