Segmentation for Retinal Blood Vessels Using Channel Attention U-Net and Pix2Pix

In medicine, segmentation of medical images is an extremely important process that allows doctors to locate and extract different elements faster. In order to speed up and automate the analysis of patient data, artificial intelligence models are very often used. To date, U-NET models in the application of binary or multi-class segmentation, with many modifications, are the most popular. The researchers also take on the use of generative neural networks to support model training, using different types of GANs. For this purpose, in this article we have researched the use of the generative network called Pix2Pix, which has been little analyzed so far in the problem of blood vessel segmentation. Our overall results presented that the proposed Channel Attention U-NET model in the Generator in the Pix2Pix architecture outperformed previous research on this type of model. The test data set confirmed the superiority of our architecture by obtaining results for Accuracy: 95.72% and for Dice: 82.23%.