Diabetic Retinopathy Detection and Grading: A Comparative Survey of Existing Techniques

Diabetic retinopathy (DR), a primary contributor to vision impairment in individuals with diabetes, requires automated systems for prompt detection and classification. This paper offers a thorough comparative analysis of traditional machine learning and deep learning methodologies for DR assessment, emphasising architectures, datasets, evaluation metrics, and clinical relevance. Despite the superior performance of convolutional neural networks (CNNs) and attention-based models, issues such as class imbalance, dataset bias, and model interpretability remain unresolved. We assess the shortcomings of accuracy-focused metrics in imbalanced datasets and recommend supplementary measures such as F1-score and AUC-ROC. The survey emphasises the inadequate use of cross-dataset validation and external testing, which are essential for real-world generalisation. Strategies to alleviate these challenges—such as lightweight CNNs, federated learning, and explainable AI (XAI)—are presented as viable solutions. Additionally, we examine hybrid frameworks (e.g., ensemble models, domain adaptation) and their function in addressing computational efficiency with diagnostic precision. The paper concludes with practical recommendations for future endeavours, highlighting the necessity for ethical AI implementation, integration of multimodal imaging, and collaborative frameworks to mitigate global diabetic blindness. This review connects technological innovations with clinical applicability, providing researchers a systematic framework for creating resilient, interpretable, and implementable diabetic retinopathy screening systems.