Title: An efficient hybrid model for localisation and grading of diabetic retinopathy using fundus images

Authors: Pammi Kumari; Priyank Saxena

Addresses: Department of Electronics and Communication Engineering, Birla Institute of Technology, Mesra, Ranchi, India ' Department of Electronics and Communication Engineering, Birla Institute of Technology, Mesra, Ranchi, India

Abstract: Diabetic retinopathy (DR) is the leading factor affecting the visions of many. This study aims to develop a computationally efficient deep learning (DL) framework for DR grading (0 to 4) to overcome the limitations of computationally inefficient existing DL models. This prompted us to use a small-scale architecture (MobileNetV2) integrated with a support vector machine (SVM) for DR grading on the APTOS dataset. A computationally light MobileNetV2 has considerably fewer trainable parameters, making it suitable for edge devices. The integration of SVM provides flexibility in tuning the essential characteristics of the dataset and enhances the grading performance efficaciously. The gradient-weighted heatmap technique is incorporated for disease localisation to visualise the affected regions adequately. The investigation's outcome substantiates the proposed architecture's efficiency over the existing DL methods, achieving a test set accuracy of 80% for multilevel and 96% for binary classification with a minimum testing loss.

Keywords: diabetic retinopathy; DR; support vector machine; SVM; Grad-CAM; deep learning; hybrid architecture; APTOS; MobileNetV2.

DOI: 10.1504/IJCVR.2026.154149

International Journal of Computational Vision and Robotics, 2026 Vol.16 No.4, pp.411 - 426

Received: 06 Mar 2023
Accepted: 08 Feb 2024
Published online: 15 Jun 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article