A Retinal Vessel Segmentation Network With Dual-Stage Network and Vessel Pixel Emendation

An automatic retinal vessel detection method is vital for the computer-aided diagnosis of eye diseases. Nevertheless, high-precision fundus image segmentation still faces some challenges, such as the intraclass variation of vessels, the class imbalance issue between background pixels and vessel pixels, and different pixel ratios between thin vessels and thick vessels. In addition, the quality issue of fundus image will also bring a certain challenge for accurate retinal vessel segmentation. In recent years, combined with strong feature representation ability, deep learning has got fast development and application on fundus image segmentation, particularly U-Net and its variant networks. Nevertheless, single segmentation network with specific loss function could not well process these two different retinal vessels simultaneously, which will greatly affect the whole segmentation precision on retinal vessels. Therefore, the precision segmentation task of retinal vessels still remains to be overcome. In order to solve these problems, a novel deep segmentation network with dual-stage network and vessel pixel emendation, named DS-VPE, is proposed for automatic retinal vessel segmentation, which includes a dual-branch segmentation network and a vessel pixel emendation network. Faced with different pixel ratios and signal-to-noise ratios between thin vessels and thick vessels, a dual-branch segmentation network is proposed for thin vessels segmentation and thick vessels segmentation, which consists of two independent segmentation branches: the thin vessel segmentation branch and the thick vessel segmentation branch. In addition, face the high feature similarity between certain vessel pixels and background pixels, a vessel pixel emendation and fusion network (VPFE-Net) is designed for further refinement and fusion to improve the issue of undersegmentation in retinal vessels. Quantitative and qualitative experiments are conducted on three public datasets: DRIVE, CHASE_DB1, and STARE. Experiment results show that the proposed method could achieve a better performance compared with advanced methods.