Abstract

Diabetic retinopathy (DR) is a leading cause of preventable vision impairment, requiring scalable and reliable screening solutions. Artificial intelligence (AI), particularly machine learning and deep learning, has significantly advanced the automation of DR classification from retinal fundus images. This paper presents a systematic review of AI-based DR classification approaches, focusing on binary classification, multi-class severity grading, and lesion-based methods, and introduces a taxonomy based on their clinical objectives. The review covers state-of-the-art models, including convolutional neural networks, vision transformers, attention mechanisms, and ensemble techniques, as well as their preprocessing strategies and evaluation protocols. Common benchmark datasets are analysed in terms of class distribution and their impact on performance. While current models achieve high accuracy on standard datasets, challenges remain in generalization, class imbalance, and reliable detection of early-stage DR. Lesion-based approaches improve interpretability but are often not fully integrated into grading frameworks. Overall, this review highlights key limitations and outlines future directions toward developing robust, interpretable, and clinically deployable DR classification systems for real-world screening

Key words: Diabetic Retinopathy, Automated Disease Screening, Deep Learning, Artificial Intelligence, Fundus Imaging, Explainable AI, Vision Transformers, Domain Adaptation, Class Imbalance.