Author(s)

Xiao Yang, Yirong Guo, Lihuang She and Shi Zhang

Corresponding Author

Xiao Yang

ABSTRACT

Automated segmentation of the retinal vasculature is a critical task for precise geometric analysis and automated diagnostics. In recent years, convolutional neural networks have shown excellent performance compared to traditional segmentation methods. In this paper, we choose the Generative Adversarial Network with game ideas in the field of computer vision to segment the retinal images, and improve the network framework on the basis of the original version. This paper mainly improves the network of generators by combining the Generative Adversarial Network with residual learning, dense connection and U-Net, and increases the utilization of image features and solves the problem of less training data. Through experimental evaluation, the performance gain of the Generative Adversarial Network when applied to the segmentation task is generated. The results show that the network can finely segment small blood vessels and achieve statistically significant improvements in the analysis of accuracy, sensitivity, specificity, and ROC curves on two common data sets, DRIVE and STARE.

KEYWORDS

Retinal vasculature, segmentation, CNN, GAN