Similar Mask Retrieval with Contrastive Learning for Single Domain Generalization in Medical Imaging

Single-domain generalization for medical image segmentation is practical to use deep learning based medical image segmentation in actual clinical centers or hospitals. In medical image segmentation, consistency of anatomical structure can be assumed across domains, and that assumption is useful in single-domain generalization. However, segmentation masks of source domains that best represent anatomical structures have only been considered for use as supervised data in model training. Therefore, we have conducted a preliminary study of using source domain masks as shape templates for single-domain generalization in medical image segmentation. In this study, we propose a method for retrieving source domain masks that are similar to the ground truth masks of images in the unseen domain using contrastive learning. Our method outperformed the score of the baseline method, indicating the potential that our method can be effective in single-domain generalization for medical image segmentation.