Retinal blood vessel segmentation model based on improved U-Net

Morphological changes in retinal blood vessels, such as bifurcation angle and degree of expansion, are of great significance to the diagnosis of fundus diseases. In recent years, the use of deep learning technology to assess the degree of retinopathy has become a research hotspot. This paper proposes a retinal blood vessel segmentation model based on improved U-Net. This model introduces a channel-enhanced residual network in the encoding and decoding part to optimize the U-Net architecture, thereby improving the network’s ability to identify retinal microvessels. In addition, a multi-level residual atrous convolution layer is added at the bottom of the model, which not only alleviates the problem of network gradient disappearance, but also enhances the model’s utilization efficiency of blood vessel information. Comparative experimental results on the DRIVE public data set show that the network proposed in this article performs well in segmentation performance.