In order to solve the defect that the classical segmentation algorithm is not accurate enough for retinal vascular segmentation, U-net3+ (full-scale connected U-shaped network) is applied to retinal microvascular segmentation and improved to improve the segmentation accuracy. Firstly, the full-scale jump connection in U-net3 + is used to extract more scale retinal microvascular features. Aiming at the problem that it is difficult to capture small blood vessels, the ordinary convolution in the network is replaced by deformable convolution, which can change the size of receptive field according to the shape and size of blood vessels and improve the segmentation accuracy of the algorithm. Then SFAM module is used to optimize the feature fusion part in U-net3+ network to retain more useful information. The algorithm is tested on the retinal image (drive) database. The results show that the average accuracy (ACC) of segmentation is 97.63%, which is 2.35% and 0.99% higher than the traditional u-net network and u-net3 + network respectively. It can be seen that the improved algorithm effectively improves the segmentation accuracy.