Style Transfer Based Unsupervised Retinal Vessel Segmentation Adaptation

Abstract

Observing that the style variations mainly cause the performance drop of segmentor when encountering unseen retinal vessel images, we propose to employ Appearance Adaptation Networks (AAN) to perform style transfer before cross-domain testing. Specifically, AAN is adopted to synthesize source-alike target images by combining the content from target domain and the style from source domain. The synthetic images, instead of corresponding target domain images, are then utilized for testing. With this domain adaptation method, our unsupervised segmentation results are just slightly lower than those of supervised methods.