Retinal Blood Vessel Segmentation Using Pix2Pix GAN

In the past few years, scientists have developed new models that are based on the U-Net structure and require a small number of ground truth images for training. Regarding retinal blood vessel segmentation, these models have proven higher accuracy in small vessels and microvascularization detection. After investigating GAN-type neural networks, the authors chose the Pix2Pix structure to implement the neural network for segmenting blood vessels in retinal images. For exemplification, we trained and tested the GAN model using retinal images from three public databases: CHASE, DRIVE and STARE. Training and testing of Pix2Pix GAN, as well as image preprocessing, were performed in Matlab R2020a. Globally, the proposed neural network had an accuracy of 92.36%, considering a total of 188 images for training and testing.