Retinal vessel detection by machine learning

Abstract

Retinal vessels are one of a kind as they are the only blood vessels in the body that can be seen in real time. Retinal diseases are recognized as one of the most significant public health problems in the working and aged population over the world[1]. Any changes in the retinal vessel would induce serious vision impairment such as Diabetic Retinopathy (DR), Glaucoma, arteriosclerosis as well as aged-related macular degeneration (AMD). Retinal vessel image segmentation plays a vital part in performing accurate assessment by detecting the abnormal signs of retinal vessels. However, the segmentation of retinal vessel images has become a difficult task due to the low contrast and complicated features of the vessels. The aim for this project is to study the various neural networks applied in vessel extraction and segmentation as well as developing a deep learning retinal vessel segmentation method with U-Net approach. The model training was conducted on the public dataset DRIVE. To get a better training result, the number of sample images was increased using the data augmentation techniques such as horizontal and vertical flipping, elastic transforming, grid and optical distortion. Comparisons between the used method and other models are included in the study. The result shows that the U-Net could achieve an average accuracy score at 0.94 over 50 epochs training, at the learning rate of 0.04 and batch size at 2.