Diabetic Retinopathy Early Analysis and Detection System

This preprint presents an AI-driven system for the early analysis and detection of Diabetic Retinopathy (DR) using ensemble deep learning techniques. Diabetic Retinopathy is a leading cause of preventable blindness worldwide, and timely diagnosis plays a critical role in reducing irreversible vision loss. The proposed framework integrates multiple deep learning architectures, including convolutional neural networks and transformer-based models, to improve diagnostic accuracy, robustness, and generalizability across diverse retinal fundus image datasets. Both publicly available datasets (such as Kaggle EyePACS and Messidor) and real-world clinical samples were utilized to enhance model reliability and real-world applicability. The system performs multi-class classification of DR severity levels based on the International Clinical Diabetic Retinopathy (ICDR) scale. Extensive preprocessing, data augmentation, and ensemble learning strategies were applied to address challenges such as class imbalance, image quality variation, and limited dataset diversity. This work aims to contribute to AI-assisted medical imaging research by providing a scalable, low-resource, and clinically relevant diagnostic framework suitable for deployment in under-resourced healthcare environments. The preprint is shared for academic dissemination, citation, and feedback purposes and has not yet undergone formal peer review. The authors retain copyright of this work. This research is made publicly available under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.