Hybrid AI Model For Diabetic Retinopathy Severity Detection Using Segmented Retinal Images

Abstract

Diabetic Retinopathy (DR) is a leading cause of preventable blindness worldwide and requires accurate severity grading for timely clinical intervention. Although deep convolutional neural networks have achieved strong performance in automated DR classification, most approaches rely solely on RGB fundus images without explicitly incorporating lesion-level pathological information. Since clinical DR grading depends heavily on the presence and distribution of lesions such as microaneurysms (MA) and hard exudates (HE), integrating lesion segmentation priors may improve grading robustness. In this work, we propose a two-stage lesion-aware deep learning framework for DR grading. First, U-Net-based models are trained to segment microaneurysms and hard exudates using patch-based training with lesion-aware oversampling to address extreme class imbalance. The best validation Dice scores achieved were $\mathbf{0. 5 3 8 9}$ for MA and $\mathbf{0. 7 4 8 3}$ for HE. The predicted lesion masks are concatenated with RGB channels to form a 5-channel representation, which is used as input to a modified EfficientNet-B3 classifier. The proposed model is compared with an RGB-only baseline. The RGB baseline achieved a macro F1-score of 0.7120, while the 5-channel lesion-aware model achieved a macro F1-score of $\mathbf{0. 6 9 4 5}$. Experimental results demonstrate improved minority class detection, particularly in severe DR stages, validating the effectiveness of lesion-aware multi-channel learning.