Mobile-based deep learning system for early detection of diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is a leading cause of vision loss globally, especially in regions with limited access to eye care. Early detection is essential to prevent irreversible damage and improve patient outcomes. In this study, a portable, real-time Assisted Mobile Diagnostic (AMD) system for DR detection, which integrates an optimized deep learning model into a mobile platform, is presented. Unlike conventional AI-based approaches that require high-performance computing and stationary fundus cameras, our system combines a non-mydriatic retinal camera with a mobile device, enabling point-of-care diagnostics. Captured retinal images are preprocessed using techniques such as blurring and contrast enhancement before being analyzed by a fine-tuned DenseNet-121 model. The model is trained using a private dataset along with two large public datasets: APTOS (Asia Pacific Tele-Ophthalmology Society) and EyePACS (Eye Picture Archive Communication System). The proposed approach achieved a high accuracy: 97.38% on APTOS, 90.90% on EyePACS, and 98.61% on the private dataset. The system delivers real-time performance on mobile devices, with an average processing time of 162.5 ms, making it well-suited for rapid screening. This Deep learning-based mobile application is part of a multi-platform tele-ophthalmology framework that includes both tablet and desktop integrations, facilitating accessible and remote DR diagnosis.