<p>Diabetic retinopathy (DR) is considered a common and serious complication of diabetes, leading to preventable blindness worldwide. Early detection is thus essential, but manual grading by ophthalmologists is time-consuming, expensive, and inherently variable. Deep learning for automatic DR screening has therefore attracted considerable attention as a promising alternative. In this paper, we aim to clarify how each stage of the learning pipeline contributes to the overall performance and generalization of diabetic retinopathy classifiers.Thus, we bring in a three-stage workflow, which is our solution to the issues raised by the scarcity of medical imaging data. We first improve the quality of retinal images by implementing specially designed preprocessing, which is a cropping strategy that not only helps in noise reduction but also focuses on the regions that are clinically relevant. In the next step, we overcome the problem of limited data by the use of MixUp augmentation, which not only extends the training distribution but also makes the learning more stable. At last, a transfer-learning approach is put into effect to make use of the power of the pretrained models in terms of their representational ability and to adapt them for the task of multi-class DR diagnosis.Our approach is evaluated on a publicly available DR dataset. The proposed model shows substantial improvement in robustness and accuracy, achieving 94.36% overall accuracy and 99% precision, recall, and F1-score across severity classes. These results reveal the potential of the proposed workflow for reliable, scalable, early DR detection in clinical settings.</p>