4th International Conference on Intelligent and Fuzzy Systems (INFUS), Bornova, Turkey, 19 - 21 July 2022, vol.505, pp.233-241
Recent deep learning methods have shown excellent performance in retinal vessel image segmentation, but the well-trained model will no longer be effective when applied to a dataset that has a large difference from the training dataset. Moreover, only a few annotated datasets can be used for supervised training, considering that clinical processes produce a large number of un-annotated images with diverse styles. As a result, it has become a big challenge to design effective deep learning segmentation models to be practically applicable. In this paper, an unsupervised cycle retinal generative adversarial network is proposed. It can realize the mutual transformation between annotated datasets and un-annotated datasets. The transformed images still retain the original vessel structure. Only the image style has been changed, so that the annotated vessel label can be shared in two domains. Furthermore, the synthetic images and the shared vessel labels can be used to train the deep learning segmentation model. We conducted experiments on annotated datasets and un-annotated datasets. The experiment results show that the cross-domain synthetic images have authentic appearance, vessel structure is well maintained. Both domains have good segmentation results.