Classification of Diabetic Retinopathy Severity Levels through Topological Feature Extraction Using Graph Neural Networks

Abstract

Diabetic retinopathy (DR) remains one of the leading causes of blindness worldwide and demands timely and precise diagnosis to enable early intervention. Conventional manual assessment of fundus images by medical professionals is both time-consuming and susceptible to human error. Consequently, computer-aided diagnostic techniques, particularly those based on Convolutional Neural Networks (CNNs), have emerged as promising solutions for automating DR detection. This research aims to enhance retinal image analysis by employing a Graph Convolutional Neural Network (GCNN) to classify diabetic retinopathy according to disease severity. By exploiting the topological relationships present within retinal images, GCNNs enable richer feature representation, leading to improved classification performance. The effectiveness of the proposed GCNN framework is evaluated using standard performance metrics, including accuracy, precision, recall, and F1-score. Experimental results demonstrate that the GCNN model outperforms conventional approaches, achieving an accuracy of 89% on the selected dataset. Furthermore, the study extends its scope by analysing advanced transfer learning architectures such as InceptionV3 and Xception, both of which attain classification accuracies exceeding 92%. The proposed methodology contributes to the early detection and effective management of diabetic retinopathy by providing clinicians with a fast, reliable, and automated diagnostic tool. As a future enhancement, the development of a user-friendly front-end interface using the Flask framework, along with secure authentication mechanisms for safe user interaction, is proposed.