A Robust Approach to Early Glaucoma Identification from Retinal Fundus Images using Dirichlet-based Weighted Average Ensemble and Bayesian Optimization.

Abstract

Objective: Glaucoma is a leading cause of irreversible visual impairment and blindness worldwide, primarily linked to increased intraocular pressure (IOP). Early detection is essential to prevent further visual impairment, yet the manual diagnosis of retinal fundus images (RFIs) is both time-consuming and inefficient. Although automated methods for glaucoma detection (GD) exist, they often rely on individual models with manually optimized hyperparameters. This study aims to address these limitations by proposing an ensemble-based approach that integrates multiple deep learning (DL) models with automated hyperparameter optimization, with the goal of improving diagnostic accuracy and enhancing model generalization for practical clinical applications.

Materials and methods: The RFIs used in this study were sourced from two publicly available datasets (ACRIMA and ORIGA), consisting of a total of 1,355 images for GD. Our method combines a custom Multi-branch convolutional neural network (CNN), pretrained MobileNet, and DenseNet201 to extract complementary features from RFIs. Moreover, to optimize model performance, we apply Bayesian Optimization (BO) for automated hyperparameter tuning, eliminating the need for manual adjustments. The predictions from these models are then combined using a Dirichlet-based Weighted Average Ensemble (Dirichlet-WAE), which adaptively adjusts the weight of each model based on its performance.

Results: The proposed ensemble model demonstrated state-of-the-art performance, achieving an accuracy (ACC) of 95.09%, precision (PREC) of 95.51%, sensitivity (SEN) of 94.55%, an F1-score (F1) of 94.94%, and an area under the curve (AUC) of 0.9854. The innovative Dirichlet-based WAE substantially reduced the false positive rate, enhancing diagnostic reliability for GD. When compared to individual models, the ensemble method consistently outperformed across all evaluation metrics, underscoring its robustness and potential scalability for clinical applications.

Conclusion: The integration of ensemble learning (EL) and advanced optimization techniques significantly improved the ACC of GD in RFIs. The enhanced WAE method proved to be a critical factor in achieving well-balanced and highly accurate diagnostic performance, underscoring the importance of EL in medical diagnosis.