Diabetic Retinopathy is the most prevalent and common diabetes-linked ocular complication that remains one of the leading causes of vision impairment worldwide. The disease develops as a result of chronic hyperglycemia, which further promotes the development of retinal microvasculature, leading to eventual vision loss if not detected and addressed in a timely manner. In this case, it is essential to establish an effective and accurate diagnostic framework to facilitate the early detection and prevention of blindness. This paper outlines a hybrid framework for the classification of DR which integrates DenseNet121’s ability to extract features and the Machine learning group’s inherent classification with Support Vector Machines to diagnose DR into five phases: No DR, Mild, Moderate, Severe, and Proliferative DR. The presented model appropriately manages the most common issues associated with classification, including class imbalance, image noise, andimaging acquisition variance using machine learning pre-processing strategies and data augmentation. For the performance of the ten most popular pre-trained convolutional neural network architectures, a Stratified Whole Dataset was applied, conditioned on the source of the image: IDRiD, Kaggle, and AEH. DenseNet121 was identified as the best-performing pre-trained architecture, exhibiting a high sensitivity of 93.169 and an F1-score of 87.726, ahead of EfficientNetB0 and Xception. Taken together, the results provide strong evidence that strong pre-processing and model selection can significantly improve classification accuracy. The proposed methodology offers a valid, scalable, and patient-centered technique for self-regulating retinal image evaluation, possessing considerable potential for application in patient care. Subsequent experiments will be directed toincreasing the performance of the models with suboptimal performance and the Ensemble learning technique and fulfill the technique verification over bigger and more varied datasets for practical application.