UU-Net:基于U-Net的U形多路径网络的视网膜血管分割

To solve the problem wherein the complex shape of retinal blood vessels and the small contrast ratio of the image make it difficult to detect vascular endings, a multi-path U-shaped network model based on U-Net (UU-Net for short) is proposed in this paper. First, inside U-Net, the residual module is used to replace the ordinary convolution to prevent gradient vanishing caused by the excessive depth of the model. Second, as the core module, the U-Net is stacked with the U-shaped structure to obtain more detailed information. Finally, U-Net modules are interconnected by addition, forming multiple paths from input to output. Each path is equivalent to a variant of FCN, enabling the UU-Net model to capture more complex features with higher accuracy. On the DRIVE dataset, the UU-Net model has achieved excellent performance on multiple test indicators, with an average accuracy rate of 0.956 1. The area under the receiver operating characteristic curve is 0.985 1. The area under the accuracy-recall rate curve is 0.982 6. In addition, the UU-Net model provides an idea based on the U-Net to improve model performance, which can be used as the infrastructure of a dense module or a residual module.