A Comprehensive Deep Learning Framework for Diabetic Retinopathy Severity Grading: Integration of LDA-Based Segmentation and DenseNet121

Abstract

Diabetic Retinopathy (DR) is a progressive retinal disorder caused by diabetes that can lead to irreversible vision loss if not detected early. This study aims to develop a reproducible deep learning workflow for the early detection and severity grading of DR. The proposed framework integrates image enhancement, lesion segmentation using Linear Discriminant Analysis (LDA), and a transfer learning classifier based on DenseNet121. Evaluation was conducted on a synthesized multi-source retinal fundus image dataset comprising EyePACS, APTOS, Messidor, and IDRiD. Experimental results demonstrate that segmentation-guided pre-processing significantly improves lesion visibility and enhances classification performance across DR severity levels. The model achieved sensitivity ranging from 71% to 90%, specificity ranging from 95% to 96%, and accuracy ranging from 92% to 95%. These findings indicate that the proposed approach offers reliable automated DR screening, with strong potential for clinical application and translation into real-world healthcare settings.