DetectDiab: Observation of Diabetic Retinopathy Progression Based on Deep Learning Object Detection Using a Web Application with Client-Side Processing

Abstract

Diabetic retinopathy is a serious complication of diabetes that affects the retinal tissue and can cause oxygen deficiency, abnormal blood vessel growth in the eye, and severe visual impairment or even blindness. Limited access to healthcare services in rural and remote areas in Indonesia affects this condition, as it is difficult for people to get regular eye examinations.
On the other hand, the increasing use of biometric data in digital services requires serious attention to security and privacy aspects, given that people's biometric data is very important to protect. Therefore, this study aims to develop an AI-based application that can recognize the severity of diabetic retinopathy from retinal images uploaded by users online or offline
(without a network). Since this study uses Client-Side Processing, fundus photos will not be uploaded to any server to protect users' personal data. The application was created by collecting retinal data from the APTOS 2019 Blindness Detection public dataset, followed by AI training using Edge Impulse with the Bring Your Own Mode (BYOM) method. The AI
training results were launched on the Detectdiab website application. It was then tested and validated by ophthalmologists, medical personnel, and other volunteers. This research has successfully developed DetectDiab, a prototype website application that functions for early detection and observation of the progression of diabetic retinopathy. This application was built using the MobileNetV2 deep learning model with the highest accuracy of 78% compared to other models. However, optimization is still needed in the mild class to achieve higher accuracy. Client-Side Processing technology using TensorFlow WebAssembly (WASM)
enables retinal image analysis to be performed entirely on the user's device, ensuring data privacy and security.