Manual annotation of 3D medical images for segmentation tasks is tedious and time-consuming. Moreover, data privacy limits the applicability of crowdsourcing to perform data annotation in medical domains. As a result, training deep neural networks for medical image segmentation can be challenging. We introduce a new source-free Unsupervised Domain Adaptation method to address this problem. Our idea is based on estimating the internally learned distribution of a relevant source domain by a base model and then sampling synthetic latent features that guide distribution alignment through a sliced Wasserstein loss. We demonstrate that our approach achieves competitive performance on real-world 3D medical datasets, including cross-modality cardiac segmentation (MMWHS) and cross-sequence brain tumor segmentation (BraTS), matching or approaching state-of-the-art results without requiring access to source data, which is a critical advantage for privacy-sensitive medical imaging applications.
Sun et al. (Fri,) studied this question.