Retinal Vessel Segmentation via A Coarse-to-fine Convolutional Neural Network

Retinal vessel segmentation has drawn great attention in various medical applications, such as the registration for fundus images and the early treatment of fundus diseases. Accurate segmentation of small retinal vessels under the noise background is still difficult. In this paper, we propose a coarse-to-fine convolutional neural network (CTF-Net) to address above problem. The proposed network has a cascaded architecture that consisted of several basic networks and each basic network is a simple encoder-decoder network based on the modification of U-Net. To improve the feature propagation of network, we introduce an ensemble strategy by concatenating the input image with outputs of later basic networks sequentially, which helps to process the image step by step. Experiments on the DRIVE dataset show our proposed CTF-Net achieves the state-of-the-art segmentation performance with 79.79% sensitivity, 98.57% specificity and 96.85% accuracy respectively.