Deep learning on fundus images detects glaucoma beyond the optic disc

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma detection models using deep learning show high performance but lack transparency.
• Explainable AI is crucial for understanding model decisions in clinical applications.
• Vertical cup-disc ratio (VCDR) is a key indicator for glaucoma risk.

Purpose of the Study:

• To develop and evaluate an explainable deep learning methodology for glaucoma detection and VCDR estimation.
• To investigate the model's ability to detect glaucoma using image regions outside the optic nerve head (ONH).

Main Methods:

• Trained deep learning models on fundus images subjected to optic nerve head (ONH) cropping policies (10-60% radius) and periphery cropping.
• Evaluated model performance for glaucoma detection (Area Under the Curve - AUC) and VCDR estimation (R-squared).

Main Results:

• Models trained on original images achieved 0.94 AUC for glaucoma detection and 77% R-squared for VCDR estimation.
• Models trained without the ONH still demonstrated significant performance (0.88 AUC, 37% R-squared at 60% crop).

Conclusions:

• Deep learning models can effectively detect glaucoma using fundus image regions outside the ONH.
• This research provides evidence for explainable AI in glaucoma diagnosis, improving transparency in deep learning models.