Diabetic retinopathy is one of the leading causes vision loss, particularly among adults living with diabetes. Early detection is vital but often relies on manual examination of retinal images, which can be slow and subject to human error. This research focuses on automating the classification of diabetic retinopathy severity levels by applying convolutional neural networks (CNNs). APTOS 2019 dataset served as the foundation for this work, and several deep learning models were examined to assess their effectiveness. Preprocessing steps like image normalization, cropping, contrast adjustment (CLAHE), and Gaussian blur were applied to enhance image quality and reduce noise. To address class imbalance and expand the dataset, various augmentation techniques were used. DenseNetl21, enhanced through transfer learning, stood out among the models tested, reaching an accuracy of 97.29% across five DR categories. Grad-CAM was used to demonstrate the model's decision-making process based on regions of interest (ROI), enhancing interpretability. Overall, the results suggest that deep learning models, particularly DenseNetl21, can offer reliable support for early and accurate diagnosis of diabetic retinopathy.