Deep learning-based diabetic retinopathy recognition and grading: Challenges, gaps, and an improved approach — A survey

Diabetic Retinopathy (DR) is a leading cause of vision impairment and blindness worldwide. Early diagnosis is crucial for preventing irreversible vision loss, but manual screening methods are time-consuming and often inconsistent. Deep learning (DL) techniques have shown promise in automating DR detection; however, many existing models still struggle to capture subtle lesions and distinguish fine-grained severity stages. In this survey, we comprehensively review recent DL-based approaches for DR classification, emphasizing attention mechanisms, feature fusion strategies, and stage-wise grading. To address current gaps, we propose a hybrid taxonomy that identifies effective combinations such as texture-based attention, CNN-Transformer fusion, and multi-modal integration. Additionally, we validate our previously published model, STMFNet, a spatial texture-aware attention network based on EfficientNet, across four benchmark datasets. On EyePACS and Messidor, STMFNet achieves up to 98.10% accuracy, outperforming several state-of-the-art (SOTA) models under similar settings. This study provides both a consolidated overview of DR detection advancements and a practical benchmark framework to guide future research in AI-assisted DR classification.