Optimizing glaucoma detection using hybrid deep-evolutionary framework from retinal images

Glaucoma represents a heterogeneous group of progressive optic neuropathies characterized by the irreversible degeneration of retinal ganglion cells and their axons, culminating in structural optic nerve head damage and permanent visual field loss. Early and accurate detection is critical for preventing permanent vision loss, yet existing automated systems suffer from inadequate generalization across diverse datasets, poor handling of class imbalance, and sub-optimal hyperparameter selection in deep learning pipelines.To address these critical clinical challenges, this research proposes a novel hybrid evolutionary-optimized deep learning framework that systematically integrates Grasshopper Optimization Algorithm (GOA) with a custom CNN-DNN architecture specifically designed for glaucoma detection from fundus images.Unlike existing approaches that rely on manual or grid-based hyperparameter tuning, our GOA-based optimization systematically navigates a 7-dimensional hyperparameter space comprising learning rate, batch size, dropout rate, number of convolutional filters, dense layer units, regularization coefficient, and optimizer types. The proposed framework is rigorously evaluated across five publicly available benchmark datasets (RIM-ONE, ACRIMA, REFUGE, ORIGA, and DRISHTI-GS) representing diverse imaging protocols, patient demographics, and class distributions. Comprehensive statistical validation using Friedman test demonstrates that the GOA-CNN-DNN model achieves statistically superior performance over contemporary state-of-the-art approaches, with mean accuracy of 97.8%, sensitivity of 96.4%, specificity of 98.2%, and AUC-ROC of 0.982 averaged across all five datasets.