Deep Learning Architectures for Diabetic Retinopathy Detection: A Comparative Study of CNNs, Transformers, and Explainable Hybrid Models

This review paper comprehensively examines how deep learning has transformed traditional screening methods for diabetic retinopathy (DR) diagnosis, providing highly accurate and scalable solutions for medical image analysis. We analyze the evolution of diverse model architectures, including classic convolutional neural networks (CNNs), advanced vision transformers, and hybrid attention-based networks, in the detection and classification of DR from fundus images. The paper also discusses the critical role of explainable artificial intelligence (XAI) tools such as SHAP and Grad-CAM in enhancing clinician trust and fulfilling regulatory requirements, while recognizing their associated computational costs. Furthermore, we explore the integration of XAI, lightweight model deployment strategies, and multi-modal data as defining trends in next-generation DR screening systems. Through these advances, we highlight the potential to improve diagnostic accuracy, increase affordability, and expand DR screening accessibility-particularly in underserved regions.