Unpaired Image-to-Image Translation Based Domain Adaptation for Polyp Segmentation

Deep polyp segmentation methods have made tremendous progress recently. However, due to the domain shift among different imaging modalities, existing methods learned on white-light imaging (WLI) achieve inferior results on other modalities such as narrow-band imaging (NBI), which limits their clinical usage. To tackle this problem, we propose a Polyp Style Translation Network (PST-Net). Specifically, test images from the NBI domain are translated by PST-Net to have the style and features of WLI images. In this way, the already deployed segmentation model can be easily generalized to images from the unseen NBI domain, without the need for tedious re-training and re-labeling. Besides, three additional designs, content consistency, attention map consistency, and adversarial segmentation loss, are proposed to achieve better translation as well as domain adaptation. Extensive experiments demonstrate that PST-Net achieves state-of-the-art performance.