Retinal Vessel Segmentation Algorithm Based on U-NET Convolutional Neural Network

The segmentation of retinal vessels a crucial role in the accurate visualization, early intervention, and surgical planning for ophthalmic disorders. There are some problems in the process of retinal vascular imaging, such as noise, low contrast, imbalance of vascular background pixel ratio and distortion of capillary cutting. The retinal blood vessel images underwent a series of preprocessing steps to optimize the performance of image segmentation. These steps included converting the images to grayscale, normalizing the data, applying restricted contrast adaptive histogram equalization, performing gamma correction, and then normalizing the data again. The subsequent analysis utilized four segmentation algorithms based on the U-Net model, namely the U-Net segmentation algorithm, Res-UNet segmentation algorithm, DU-Net segmentation algorithm, and Sa-UNet segmentation algorithm, were selected to segment the retinal vessel images. The fundus images from the DRIVE public database were utilized to conduct simulation experiments in order to validate the efficacy of the adopted algorithms. The sensitivity, specificity, accuracy and AUC of Sa-UNet segmentation algorithm were 0.8573, 0.9835, 0.9905 and 0.9755, respectively.