Diabetic Retinopathy (DR) is a major cause of avoidable blindness in the world, which is as a result of the high blood sugar levels that gradually destroy the blood vessels in the retina. Timely and prompt diagnosis of DR is essential to avoid permanent damage to vision, but the visual inspection of retinal fundus images by highly trained ophthalmologists is still time-consuming, costly, and unavailable in remote and resource-intensive health facilities. Although current automated DR detectors that use deep learning have shown expertise-level classification, and most are a black- box architecture, giving a prediction without explaining the reasoning behind it, a major impediment to clinical use and trust. This project outlines a deep learning framework of automated diagnosis of Diabetic Retinopathy and retinal fundus image classification of Diabetic Retinopathy severity. The proposed system categorizes the retinal fundus images into five globally accepted stages of DR severity, namely, No DR, Mild DR, Moderate DR, Severe DR and Proliferative DR, with the help of the Xception Convolutional Neural Network (CNN) architecture that is trained on ImageNet. Ben Graham preprocessing was used to improve retinal pathological characteristics (microaneurysms, hemorrhages, hard exudates, and neovascularization) prior to model training. Oversampling was used to balance all five classes in the dataset to solve the extreme class imbalance and data augmentation methods such as random flipping, rotation and brightness variation were used to enhance model generalization. The model was trained with the APTOS 2019 Blindness Detection dataset on a three-step progressive fine-tuning approach, initially training just the custom classification head, and then unfreezing and fine-tuning the Xception backbone layers, obtaining an overall validation accuracy of 77.9% and 97% accuracy on the clinically critical No DR class. In order to address the black-box nature of deep learning models, SHAP GradientExplainer and Gradient Saliency mapping were combined to create pixel-level visual heatmaps revealing the locations in the retina that had the greatest impact on any given prediction, and provided explanations that can be directly verified by ophthalmologists based on known retinal pathology characteristics. The entire system was implemented as a real-time web-based clinical decision support application on the Streamlit framework, where a clinician can input a retinal fundus photograph and in a few seconds obtain the predicted stage of DR severity, a confidence score, a visual explanation heatmap, and a clinical recommendation about proper patient management.