Unsupervised domain adaptation for medical image segmentation via self-training of early features

U-Net models provide a state-of-the-art approach for medical image segmentation, but their accuracy is often reduced when training and test images come from different domains, such as different scanners. Recent work suggests that, when limited supervision is available for domain adaptation, early U-Net layers benefit the most from a refinement. This motivates our proposed approach for self-supervised refinement, which does not require any manual annotations, but instead refines early layers based on the richer, higher-level information that is derived in later layers of the U-Net. This is achieved by adding a segmentation head for early features, and using the final predictions of the network as pseudo-labels for refinement. This strategy reduces detrimental effects of imperfection in the pseudo-labels, which are unavoidable given the domain shift, by retaining their probabilistic nature and restricting the refinement to early layers. Experiments on two medical image segmentation tasks confirm the effectiveness of this approach, and compare favorably to a baseline method for unsupervised domain adaptation.