Edge-preserving Domain Adaptation for semantic segmentation of Medical Images

Domain Adaptation is a technique to address the lack of massive amounts of\nlabeled data in unseen environments. Unsupervised domain adaptation is proposed\nto adapt a model to new modalities using solely labeled source data and\nunlabeled target domain data. Though many image-spaces domain adaptation\nmethods have been proposed to capture pixel-level domain-shift, such techniques\nmay fail to maintain high-level semantic information for the segmentation task.\nFor the case of biomedical images, fine details such as blood vessels can be\nlost during the image transformation operations between domains. In this work,\nwe propose a model that adapts between domains using cycle-consistent loss\nwhile maintaining edge details of the original images by enforcing an\nedge-based loss during the adaptation process. We demonstrate the effectiveness\nof our algorithm by comparing it to other approaches on two eye fundus vessels\nsegmentation datasets. We achieve 1.1 to 9.2 increment in DICE score compared\nto the SOTA and ~5.2 increments compared to a vanilla CycleGAN implementation.\n