Glaucoma detection and staging from visual field images using machine learning techniques.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma diagnosis relies on visual field (VF) testing.
• Accurate staging of glaucoma is crucial for effective management.
• Current methods may not fully capture subtle changes in VF data.

Purpose of the Study:

• To evaluate deep learning (DL) models for differentiating normal and glaucomatous visual fields.
• To assess DL model performance in staging glaucoma severity using pattern deviation (PD) plots.
• To compare DL model results with traditional machine learning (ML) classifiers.

Main Methods:

• Trained DL models (ResNet18, VGG16) on 265 PD plots from normal and glaucomatous eyes.
• Classified VFs into normal, early, moderate, and advanced glaucoma stages.
• Employed five-fold cross-validation and data augmentation techniques.
• Compared DL performance against ML models using global indices (MD, PSD, VFI).

Main Results:

• ResNet18, trained on balanced, augmented PD images, achieved 96.8% F1-score.
• DL models accurately localized visual field loss.
• Machine learning (Random Forest) achieved a 96% F1-score.
• DL performance was comparable to conventional global indices.

Conclusions:

• DL models trained on PD plots show promise for glaucoma detection and staging.
• The DL model can stage glaucoma severity, similar to Mills criteria, using only PD plots.
• This automated approach can assist clinicians in glaucoma screening and progression management.