VESDNet: Deep Vessel Segmentation (U) Network for the Early Diagnosis of Diabetic Retinopathy

Abstract

Eye screening is the first level diagnosis method among the other medical diagnosis in finding the anomalies in human body. Diagnosing those diseases in the initial stage of the problem can save the people from losing visual and life. Deep analyzing of retinal anatomical components like as blood vessels, exudates, hemorrhages, fovea, microaneurysms and optical disk are the major tasks that contribute to the pre-screening of the anomaly detection using digital fundus images. Among them, accurate segmentation of blood vessels is the most widely accepted method of diabetic retinopathy diseases analysis as damage in the blood vessels is the sign for the diabetic retinopathy and other health issues. In this presented paper, we have addressed an approach based on deep convolutional neural network for the accurate blood vessel segmentation in the context of diabetic retinopathy analysis. In our approach, the retinal image is first pre-processed using computer vision algorithms and vessel segmentation is performed using multi-label deep learning segmentation model corresponding to thick and thin vessel pixels. We performed comparison of our approach with existing methods for both accuracy and speed using the open-source DRIVE and MESSIDOR fundus image databases.