Diabetic retinopathy (DR) is one of the leading causes of preventable blindness worldwide, with early detection being crucial for effective intervention and vision preservation. Recent advances in deep learning have transformed ophthalmic image analysis, enabling automated detection and grading of DR from retinal fundus images with unprecedented accuracy. This survey provides a comprehensive overview of recent developments (2023–2025) in deep learning-based approaches for early DR detection, covering publicly available datasets, preprocessing techniques, convolutional neural network (CNN) architectures, transfer learning strategies, attention mechanisms, and lesion localization methods. We systematically compare state-of-the-art models in terms of accuracy, sensitivity, specificity, and robustness, highlighting their strengths and limitations in real-world clinical deployment. Additionally, we critically examine the challenges of data imbalance, variability in image acquisition, and the need for explainability in AI-assisted diagnosis. Finally, we outline emerging research directions, including self-supervised learning, multimodal data integration, and explainable AI frameworks, to bridge existing gaps and accelerate the clinical adoption of deep learning-based DR screening systems.