Hybrid Deep Learning and Morphological Processing for High-Accuracy Diabetic Retinopathy Detection and Grading

Diabetic retinopathy (DR) is a microvascular complication of diabetes and the leading cause of preventable blindness among working‑age adults. The rapid increase in global diabetes prevalence has motivated the development of automated screening systems that can support overburdened healthcare providers in resource‑constrained settings. In this paper we propose a comprehensive computer‑aided diagnosis (CAD) framework for DR that integrates deep learning with traditional image processing to deliver accurate and interpretable screening. The pipeline first uses a RetinalDenoiser network to suppress noise and enhance vasculature; the denoiser incorporates a residual attention mechanism to preserve fine lesions. A subsequent HyRetinalSegmentation module combines a U‑Net backbone with morphological opening, closing and top‑hat transforms to isolate vessels and lesions. Finally, a hybrid classification network (RetinalNet101) fuses high‑level features from ResNet‑101 and DenseNet‑201 branches to grade DR severity. The proposed framework is modular and can be trained end‑to‑end. We evaluate the system on publicly available EyePACS, APTOS 2019 and DDR datasets as well as a synthetic dataset built from the IDRiD segmentation ground truth. Comparisons against conventional preprocessing, segmentation and classification methods demonstrate significant improvements in peak signal‑to‑noise ratio (PSNR), Dice score and grading accuracy. The ablation study shows that each module contributes independently to performance; combining denoising, segmentation and hybrid classification yields a classification accuracy of 98. 5 % compared with 94. 0 % for a baseline pipeline. Our method outperforms state‑of‑the‑art approaches and offers interpretable intermediate outputs that could facilitate clinician trust. We conclude by discussing limitations and promising directions for future research.