Genetic Fire Hawk Optimizer-Guided Entropy Segmentation with Deep Residual Autoencoder and Heterogeneous Ensemble Learning for Multi-Stage Diabetic Retinopathy Classification

Abstract

Diabetic Retinopathy (DR) is a leading cause of preventable blindness in working-age adults, yet automated five-stage severity grading remains difficult because lesion patterns overlap between adjacent stages and DR datasets are heavily class-imbalanced. This paper proposes a four-stage automated DR grading framework. First, retinal fundus images are denoised using Bilateral Filtering and Non-Local Means preprocessing. Second, a Genetic Fire Hawk Optimizer (GFHO) searches for multilevel segmentation thresholds that maximise Renyi entropy, isolating diagnostically relevant retinal structures. Third, a Deep Residual Autoencoder (DRA) extracts compact, hierarchical latent features from the segmented images. Fourth, a heterogeneous soft-voting ensemble of Extreme Gradient Boosting (XGBoost), Radial Basis Function Support Vector Machines (RBF-SVM), and Random Forest classifies each image into one of five DR severity stages. On the APTOS 2019 Blindness Detection benchmark (3,662 fundus images), the proposed framework attains 90.74% accuracy, a macro-averaged F1-score of 0.89, and an AUC of 0.95, exceeding single-classifier baselines by up to 3.83 percentage points.