CFCNet: A coarse-to-fine cascade network for retinal vessel segmentation

In this paper, a new retinal vessel segmentation method is proposed to study the early symptoms and prevention of certain eye diseases. There are still no effective algorithm for existing automatic vessel segmentation methods to obtain precise results by reason of the complexity of retinal vessels and therefore a coarse-to-fine cascade network named CFCNet is formulated and aiming at improving the accuracy of retinal vessel segmentation. Here, two U-shaped sub-networks are contained in the proposed network. Vessels are coarsely segmented by the first sub-network (Net1) and then refined by the second sub-network (Net2). To enrich features, Cross-Level Connections are introduced in the CFCNet. Res-ASPP and CBAM are further implemented for the sake of extracting the multi-scale features and enhancement, respectively. Two databases of DRIVE and STARE are then employed in several experiments for the purpose of evaluating te proposed network. The results show that the proposed CFCNet achieves more competitive effect for retinal vessel segmentation than the existing networks and is of great clinical significance.