Microscope-Assisted Hypertensive Retinopathy Diagnosis Using Deep Learning Models

Abstract

The retina is the most crucial part of the human eye, and it can be affected due to hypertension. However, retinal abnormalities due to hypertension are termed hypertensive retinopathy (HR). A severe stage of HR can lead to complete blindness if not diagnosed and treated on time. Manually analyzing retinal images for HR diagnosis is time-consuming and prone to errors. This research article provides a novel technique based on U-Net and Dense-Net for automatic HR detection and grading through retinal images. The presented method consists of preprocessing, vessel segmentation, artery or vein (A/V) classification, and vessel width calculation to compute the arteriovenous ratio (AVR). In the preprocessing phase, the Gabor filter is applied to the retinal image to enhance the vascular network of the image. The preprocessed image is fed into the U-Net architecture to segment the vascular network image. The segmented vascular network image is fed into the Dense-Net architecture for A/V classification. The A/V classified vascular network is divided into several artery and vein segments at the bifurcation and crossover points. The A/V segments are labeled for width calculation to compute the AVR. The AVR is a standard parameter for HR detection and grading. The evaluation results show an average accuracy of 99.40% in HR classification and 99.77% in HR grading on the AVRDB dataset. The evaluated results are beneficial for the automatic HR detection and grading for clinical purposes.