Aiming at the problems of loss of detailed feature information and blurred contours of blood vessels in the segmentation process of retinal vessels, an improved cycle segmentation confrontation network algorithm is proposed. The algorithm improves the network model of the segmenter, adds dense connection structure in the up and down sampling processes of the U-Net network, fully retains the image feature information, which improves the generalization ability and robustness of the model, and alleviates the over segmentation phenomenon. In order to prevent network degradation, the loss function is replaced by the least square function, which improves the quality of image segmentation and the stability of network model training. The experimental results show that the segmentation accuracy and sensitivity of the network are 96.93%, 84.30% and 96.94%, 79.92% in the DRIVE and CHASE datasets respectively. The algorithm has good network generalization ability and segmentation accuracy, which can provide an important basis for disease diagnosis.