Retinal Vessel Image Segmentation Based on Guided Filter and Neural Networks

Abstract

Accurate segmentation of retinal blood vessels is essential for the early diagnosis and monitoring of various ophthalmic and systemic diseases. This work shows how to use a modified U-Net architecture to improve how retinal blood vessels are segmented. The method uses rolling guidance filters with two different sets of parameters to make the edges of the vessels sharper and make it easier to tell them apart. Separate U-Net models process each filtered image independently, and then the outputs are combined to make the final segmentation result. Both networks are trained at the same time using a shared loss function. Tests on many benchmarks retinal image datasets show that the suggested method makes segmentation far more accurate than the usual U-Net model.