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

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Manual diagnosis of retinal fundus images (RFIs) for glaucoma detection (GD) is inefficient.
• Existing automated GD methods often require manual hyperparameter tuning.

Purpose of the Study:

• To develop an improved automated glaucoma detection system.
• To enhance diagnostic accuracy and model generalization using ensemble learning.
• To integrate deep learning models with automated hyperparameter optimization.

Main Methods:

• Utilized 1,355 RFIs from ACRIMA and ORIGA datasets.
• Employed an ensemble of a custom CNN, MobileNet, and DenseNet201.
• Applied Bayesian Optimization for automated hyperparameter tuning.
• Combined model predictions using a Dirichlet-based Weighted Average Ensemble (Dirichlet-WAE).

Main Results:

• Achieved state-of-the-art performance with 95.09% accuracy, 95.51% precision, 94.55% sensitivity, 94.94% F1-score, and 0.9854 AUC.
• The Dirichlet-WAE significantly reduced the false positive rate.
• The ensemble model outperformed individual models across all metrics.

Conclusions:

• Ensemble learning and automated optimization significantly improve glaucoma detection accuracy.
• The Dirichlet-WAE is crucial for balanced and accurate diagnostic performance.
• Ensemble methods are vital for robust medical diagnosis in ophthalmology.