12 - U-net autoencoder architectures for retinal blood vessels segmentation

Abstract

India, home to one-third of the world's blind population finds its major causes to be diabetic retinopathy and glaucoma. An effective diagnosis of diabetic retinopathy, glaucoma, and other eye-related diseases will require an effective segmentation of retinal blood vessels from retinal images. The need for an effective segmentation technique is to aid ophthalmologists in the diagnosis of glaucoma, diabetic retinopathy, and to overcome problems in retinal blood vessels segmentation. One such effective segmentation technique is the training of a deep learning model for segmentation. In this work, the segmentation of retinal blood vessels from the retinal image using a combination of deep learning models based on U-net, including ResConU-net, dense residual path U-net, inception block residual U-net, serial residual U-net, and parallel residual U-net has been proposed. The major contribution of this work is the introduction of concatenated filter-based residual paths in the U-net so that the disparity between the encoder and decoder is reduced. To leverage the availability of Graphics Processing Unit (GPU), models have been trained and tested using Google Colaboratory on the Digital Retinal Images for Vessel Extraction (DRIVE) dataset. On the DRIVE dataset, these five U-net variants offered accuracy values of 0.9561, 0.9562, 0.9560, 0.9623, and 0.9559, respectively, and exhibited improved performance metrics compared to most state-of-the-art alternatives.