Diabetic Retinopathy (DR) remains a major cause of preventable blindness worldwide among individuals with diabetes, emphasizing the critical need for early and accurate detection to enable timely intervention. Traditional manual screening by ophthalmologists is time-consuming, resource-intensive, and subject to variability, particularly in resource-limited settings. This project proposes an automated Diabetic Retinopathy detection system leveraging deep learning techniques to classify retinal fundus images according to DR severity levels (ranging from No DR to Proliferative DR). The proposed framework incorporates advanced preprocessing steps, including contrast enhancement via CLAHE and data augmentation to address challenges such as image quality variations, class imbalance, and limited dataset sizes. Transfer learning is employed using state-of-the-art convolutional neural network architectures, such as DenseNet-121 or EfficientNet, fine-tuned on publicly available datasets like EyePACS, APTOS 2019, and Messidor. The model architecture features a pre-trained backbone for robust feature extraction, followed by custom classification heads with dropout regularization and focal loss to improve performance on imbalanced classes. Optional enhancements include lesion segmentation using U-Net for explainability and ensemble methods for boosted accuracy. Experimental evaluations demonstrate promising results, achieving classification accuracies in the range of 92–97%, high sensitivity (essential for detecting referable DR cases), and AUC-ROC scores exceeding 0.95 on benchmark datasets. This system offers a scalable, cost-effective solution for automated DR screening, with potential deployment in mobile or web-based applications to support telemedicine and large-scale population screening, ultimately contributing to reduced vision loss in diabetic communities. Future work may explore uncertainty quantification via Bayesian approaches and real-world validation across diverse populations.