A Comprehensive Approach to Diabetic Retinopathy Detection: Preprocessing, Segmentation, and Ensemble Learning-Based Grading

Abstract

This paper presents a multistage computational pipeline for the automated detection and severity grading of diabetic retinopathy (DR) from retinal fundus images, integrating classical image preprocessing, lesion-oriented segmentation, texture-based feature extraction, and ensemble classification into a single reproducible workflow. Retinal images are first denoised by median filtering and reduced to their green channel for maximal vessel-to-background contrast, then enhanced by morphological operations before iterative thresholding separates anatomical structures (optic disc, vasculature) from candidate lesions (microaneurysms, haemorrhages, exudates). Multi-orientation, multi-scale Gabor filtering converts each segmented image into a 10,000-dimensional texture feature vector, which is passed to five classifiers such as Support Vector Machine (SVM), Support Vector Regression (SVR), k-Nearest Neighbours (KNN), Random Forest Classifier (RFC), and AdaBoost Classifier (ABC) evaluated individually and as an ensemble. Under stratified k-fold cross-validation on a combined DIARETDB1 / Messidor / DRIVE cohort, the SVM + RFC + ABC ensemble achieved the best overall performance (accuracy 97.1%, sensitivity 95.9%, specificity 98.2%, AUC 0.98), outperforming each constituent classifier individually and a conventional CNN baseline trained on the same data (93.5% accuracy). These results indicate that a carefully tuned classical ensemble, combined with domain-specific preprocessing and handcrafted texture features, remains a competitive and computationally lighter alternative to end-to-end deep learning for DR screening, with particular relevance to tele-ophthalmology deployments where GPU inference is not readily available.