A Lightweight Deep Learning Framework for Multi-Class Severity Grading of Diabetic Retinopathy from Fundus Images

Despite a limited number of ophthalmologists available compared to the steadily increasing population of diabetics in developing countries, DR continues to be one of the most important causes of preventable adult diabetic blindness, with early detection being even more important. This work proposes a lightweight deep learning framework that grades diabetic retinopathy severity directly from retinal fundus images, using the publicly available APTOS 2019 dataset (3,662 labeled images spanning five severity levels: No DR, Mild, Moderate, Severe, and Proliferative DR). Prior Iraqi studies on this dataset stopped at a binary decision (DR present or absent), whereas the framework presented here performs the full five stage grading task, which is clinically far more useful but noticeably harder because adjacent stages overlap visually. EfficientNet-B0 is used as a frozen feature extractor paired with a small trainable classification head, so that fewer than 165,000 of the roughly 4.2 million total parameters are ever updated during training; class weighted loss was used throughout to offset the dataset's imbalance. The resulting model reached 78.04% accuracy on the full five class task (weighted F1 = 0.777), and when its output was collapsed into a simple binary decision it reached 96.18% accuracy. In combination these numbers indicate that a fairly simple, transfer learning model can be used to produce a meaningful multistage grading without requiring the training of deeper or ensemble models, something that is relevant in a resource limited environment like Iraq.