<![CDATA[Diabetic retinopathy (DR) is a leading cause of preventable blindness worldwide and is becoming a significant public health concern in Indonesia due to the rising prevalence of diabetes. Early detection is critical, yet access to ophthalmologists and conventional fundus cameras remains limited in many primary healthcare facilities. To address these challenges, this study proposes a cost-effective, web-based intelligent system for early detection of DR using smartphone-based fundus adapters and deep learning.A hybrid dataset was employed, combining publicly available fundus image repositories with locally collected retinal images from Indonesian healthcare facilities, annotated by ophthalmologists. Images were preprocessed through normalization, cropping, artifact removal, and augmentation to address variability, particularly from smartphone acquisitions. A DenseNet-121 convolutional neural network was fine-tuned on this hybrid dataset to classify DR into five severity levels according to the International Clinical Diabetic Retinopathy Disease Severity Scale. Model performance was evaluated using accuracy as the primary metric, with results compared against ophthalmologist annotations.The proposed system demonstrated promising performance in classifying DR severity levels, showing that combining public and local datasets improves contextual relevance and model robustness. Furthermore, integration into a web-based platform enables healthcare workers in primary care to upload fundus images, obtain real-time classification results, and facilitate referral decisions for severe cases.This study contributes to the development of an accessible and scalable screening tool for DR in Indonesia by integrating affordable imaging hardware, locally relevant datasets, and an AI-powered classification system. The approach has the potential to reduce reliance on expensive equipment and specialists, supporting national efforts to prevent diabetes-related blindness.]]>
