Retinal vascular segmentation algorithm based on improved UNet

Aiming at the problems of complex retinal vascular structure, low image contrast and inaccurate detail region segmentation, a segmentation algorithm based on improved UNet is proposed. Firstly, combined with the idea of multi branch and multi-scale of divide branch block, a DBB-ConvNet module is added to the coding and decoding path. The module combines branches with different scales and complexity to enrich the diversity of feature space, so as to improve the ability of feature learning and expression of the network; Secondly, in order to strengthen feature reuse and avoid the influence of redundant features, dense connection is added at the bottom of the network; Finally, in order to further improve the segmentation effect, bconvlstm combined with nonlinear function is used to deal with the feature mapping between code and decoding paths in jump connection, replacing the original simple concatenation. Based on open datasets DRIVE and CHASE_DB1, The experimental results show that compared with other algorithms, the proposed algorithm has better segmentation effect in vision and various objective evaluation indexes. Compared with U-Net algorithm, the accuracy and F1 value of the proposed algorithm are improved by 1.97%, 2.04% and 2.05%, 5.86% respectively.