PLPP: A Pseudo Labeling Post-Processing Strategy for Unsupervised Domain Adaptation
Tomer Bar Natan, Hayit Greenspan, Jacob H. Goldberger · 2023
A well known problem in medical imaging is the ability to use an existing model learned on source data, in a new site. This is known as the domain shift problem. We propose a pseudo labels procedure, which was originally introduced for semi-supervised learning, that is suitable for unsupervised domain adaptation (UDA). We iteratively improve the pseudo labels of the target domain data only using the current pseudo labels without involving the labeled source domain data. We applied our method to several medical MRI image segmentation tasks. We show that, by combining our approach as a post-processing step in standard UDA algorithms, we consistently and significantly improve the segmentation results on test images from the target site.