Abstract
Glaucoma remains a primary cause of irreversible visual impairment, necessitating precise diagnostic strategies. This study introduces an optimized classification framework integrating deep feature extraction and heuristic feature selection for enhanced detection of glaucomatous pathology in fundus imaging. The G1020 dataset, comprising 1020 retinal fundus images, was processed using a pre-trained ResNet50 to generate 2048-dimensional feature representations. Class imbalance was addressed using synthetic minority over-sampling (SMOTE). Feature selection was performed via particle swarm optimization (PSO), artificial bee colony (ABC), and binary cuckoo search (BCS) to remove redundant features and optimize dimensionality. Classification models, including random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN), and multi-layer perceptron (MLP), were assessed pre- and post-heuristic optimization. After optimization, KNN reduced false negatives by 73.5%, RF improved specificity by 80.8%, and MLP retained high sensitivity with minimal accuracy loss, while SVM exhibited an 8.5% increase in false positive classifications. The implemented heuristic feature selection method demonstrated potential in improving the accuracy of glaucoma detection and diagnosis through automated computational classification of fundus imaging data.