Evolution and Future Directions of Deep Learning in Automated Diabetic Retinopathy Detection: A Comprehensive Review

Abstract

Diabetic retinopathy (DR) remains one of the leading causes of preventable blindness globally, thus early and accurate detection is crucial. In the last decade, significant advances in computational imaging, deep learning and artificial intelligence have resulted in creation of several automated techniques to detect DR from retinal images. In this review, we summarize and compare thirty-one (31) representative studies between 2015 to 2025 from both classification-based and detection-based methods. The study proposes five main research questions to explore the transition from Convolutional Neural Networks (CNNs) towards hybrid transformer architectures, the impact of dataset diversity and the barriers towards clinical interpretability. The paper also studies the evolution of various model architectures, including convolutional neural networks (CNNs), transformer models and YOLO-based methodologies, by analyzing their performance across popular datasets like EyePACS, Messidor and IDRiD. Salient challenges, e.g., imbalanced datasets, varying image qualities, and the desired interpretability of the outputs are addressed in this review together with possible research paths moving forward. Comprehensively, this precise paper is aimed at offering a coherent, organised and extended literature resource to those researchers that are heading to automatic detection of diabetic retinopathy with better accuracies.