An Enhanced Generative AI Model for Detecting Diabetic Retinopathy Anomalies Using Fundus Images: A Comprehensive Review

Abstract

Diabetic retinopathy (DR) is a leading cause of preventable blindness among working-age adults, driven by chronic hyperglycemia and the global rise in diabetes prevalence. Although early detection and treatment can prevent up to 90% of vision loss cases, limited access to ophthalmologists, poor screening adherence, and the manual burden of retinal image analysis remain major challenges [2], [9], [13]. To address these issues, artificial intelligence (AI), particularly deep learning (DL) and generative AI, has emerged as a scalable and cost-effective solution for DR screening. Advances in computer vision have enabled automated analysis of retinal fundus and optical coherence tomography (OCT) images with diagnostic performance approaching that of expert clinicians. Convolutional neural networks (CNNs) dominate DR detection and grading due to their ability to learn hierarchical features directly from image data, supported by public benchmarks such as the APTOS 2019 dataset [1]. However, challenges including class imbalance, image quality variability, and limited annotated data continue to affect model robustness. This review presents a systematic overview of DL and generative AI methods for DR detection, classification, and lesion analysis using both fundus imaging and OCT modalities [5], [17], [24]. It highlights the importance of preprocessing and augmentation techniques such as CLAHE, color normalization, and GAN-based image synthesis to enhance performance and mitigate dataset imbalance [6], [8], [14], [15]. The paper also discusses weakly supervised and semi-supervised learning strategies that reduce annotation costs while maintaining accuracy [3], [19], [20]. Further emphasis is placed on lightweight model design, knowledge distillation for deployment in resource-limited settings [4], and explainable AI techniques to improve clinical trust and interpretability [28], [29]. The review concludes that while AI has significantly advanced automated DR screening, future research must prioritize robustness, explainability, and ethical deployment to enable reliable large-scale clinical integration [17], [26].