Unsupervised Domain Adaptation With Boundary-Aware GAN for Medical Image Segmentation

Domain Adaptation is a technique to address the lack of massive amounts of labeled data in different application domains. Unsupervised domain adaptation is the process of adapting a model to an unseen target dataset using solely labeled source data and unlabeled target domain data. While many image-space domain adaptation methods have been proposed to capture pixel-level domain shift, such techniques may fail to maintain detailed information for the segmentation task. In the case of biomedical images, fine details such as blood vessels or boundaries between regions can be blurred during the image transformation operations between domains. In this work, we propose a model that adapts between domains using cycle-consistent loss while maintaining edge information of the original images by enforcing an edge-based loss during the adaptation process. We demonstrate the effectiveness of our algorithm by comparing it to other approaches on two biomedical image datasets. For the multi-modality eye fundus vessels segmentation datasets, we achieve a 3.1% increment in the Dice score when compared to the state-of-the-art and ∼7.02% increment compared to a vanilla CycleGAN implementation. For the CT/MRI multi-modality whole heart segmentation dataset, we achieve a 13% increment in Dice score when compared to the state-of-the-art and ∼6% increment compared to a vanilla CycleGAN implementation.