Cross-Domain Segmentation of Fundus Vessels Based on Feature Space Alignment

The accurate segmentation of fundus vessels plays a very important role in the detection and treatment of fundus diseases. With the rapid development of Convolutional Neural Networks (CNN), some CNN-based methods have been proposed for the segmentation of fundus vessels which show a good segmentation performance, but they rely on much well-annotated data sets. Aimed at this issue, based on a small number of annotated images, a new segmentation network is proposed in this paper to realize the segmentation of fundus vessels in the cross-domain. Two different high-level feature space are aligned and the Wasserstein distance is used to train the antagonistic networks. Experiments show that the proposed method could acquire a good segmentation performance on the public data sets of the DRIVE and STARE data sets.