U-Net Fundus Retinal Vessel Segmentation Method Based on Multi-scale Feature Fusion

Abstract

In the task of retinal blood vessel segmentation, the existing algorithm has the problem of low segmentation accuracy due to the complex morphology of the retinal small blood vessel structure and the uneven illumination of the image. To solve this problem, this paper improves on the traditional U-Net and proposes a Res-HSPP U-Net. Firstly, the original convolution of the encoding and decoding parts is changed to the Res-HSPP structure, which can balance the width and depth of the network. At the same time, the dilated convolution group is used to obtain a larger receptive field and the fusion of multi-scale feature vector information to enhance the ability to extract the features of small blood vessels and reduce the problem of loss of detailed information. Secondly, the ECA attention module is introduced to learn the feature channels after convolution, increase the weight of the effective feature channels, and strengthen the features of small blood vessels. Compared with other channel attention modules, the ECA module has the characteristics of light weight and higher computational efficiency. The algorithm in this paper is verified on the public dataset DRIVE (Digital Retinal Images For Vessel Extraction). The final performance indicators are: accuracy rate of 96.87%, specific rate of 98.54%, sensitivity of 82.87%, F1-score of 82.51%. Experiment results show that the performance of this algorithm is better than that of advanced algorithms in recent years.