# Automated Diagnosis of Hypertensive Retinopathy Using Res‐UNet and Graph Convolutional Networks

## TL;DR

This paper introduces an automated system using AI to diagnose hypertensive retinopathy by analyzing retinal images, improving accuracy and efficiency.

## Contribution

A novel framework combining Res-UNet and graph convolutional networks for accurate retinal vessel segmentation and hypertensive retinopathy diagnosis.

## Key findings

- The system achieves 96.45% accuracy in retinal blood vessel segmentation.
- Arteriovenous classification reaches 96.7% accuracy using the proposed method.

## Abstract

Hypertensive retinopathy (HR), a progressive retinal condition, is associated with both hypertension and diabetes mellitus. The development of HR is closely correlated with the severity and duration of hypertension. The results of the HR indicate that pathological eye issues include cotton‐like spots, macular oedema, constrained arterioles and retinal haemorrhage. An ophthalmologist would still often undertake a manual physical examination using an ophthalmoscope to detect HR in a patient's body. It is time‐consuming for a physician to identify HR in a patient based only on retina fundus imaging when done manually. An automated approach for identifying the retinal fundus image is required to solve this issue. One crucial component in the diagnosis of many eye diseases is the condition of the blood vessels in the retina. Researchers have found great interest in the blood vessel segmentation of fundus images for this reason. The knowledge of blood vessel changes associated with various disorders, such as cardiovascular diseases and retinopathy, depends on the accurate segmentation of arteries and veins (A/V) from fundus images. The arteriovenous ratio displays the proportion of vein to artery diameters. The precision with which vessels are divided into veins and arteries determines the significance of this measure. To increase the accuracy of classifying retinal blood vessels and HR phases, a novel technique combining deep residual UNET (Res‐UNet) and a graph convolutional network is suggested in this research. Pre‐processing (green channel, contrast‐restricted adaptive histogram equalisation) was done before identification. Graphs are used to depict the features of the vessel that are extracted from the spatial domain. The DRIVE‐AV image dataset is used to execute the proposed approach and the results show that the system achieves a blood vessel segmentation accuracy of 96.45% and an A/V classification of 96.7%.

The proposed framework integrates retinal vessel segmentation and artery/vein classification using deep residual UNET and GCN models for hypertensive retinopathy diagnosis. By computing the arteriovenous ratio, the method accurately detects HR stages from retinal fundus images. The system achieves reliable performance using the DRIVE dataset.