Improved Segmentation of Retinal Vessel Pathways Using Enhanced U-Net Model

Abstract

Retinal imaging is a frequently employed method in medical practice to yield valuable data regarding modifications in the retinal blood vessel architecture, that are typical of ailments such as glaucoma, hypertension, or diabetes. Their timely treatment can prevent patients from suffering from blindness. Experienced physicians find the manual decomposition of the vascular structure from fundus images to be time-intensive and susceptible to errors due to the complexity involved in capturing the details of the blood vessel pathways. Lately, deep learning techniques have been implemented because of their superior speed and precision in comparison to manual segmentation. To address the problem of reduced accuracy caused by imbalanced categories, this paper proposes an enhanced U-NET structure with a combined loss function, referred to as Dice-BCE loss that specifically caters to imbalanced data to enable the model to concentrate on target characteristics. This proposed approach is analysed on the Digital Retinal Images for Vessel Extraction (DRIVE) dataset and it achieves better results than already proposed models with an accuracy of 96.75%, F1-Score of 81.09%, Jaccard-Index of 68.23%, a sensitivity of 80.28%, specificity of 98.35% and precision of 82.52%.