A generalizable deep learning regression model for automated glaucoma screening from fundus images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma detection models often fail to generalize due to data variations.
• Existing models struggle with shifts in prevalence, camera types, and ground truth definitions.

Purpose of the Study:

• To evaluate the generalizability of the G-RISK regression network for glaucoma detection.
• To assess model performance across diverse, multi-source fundus image datasets.

Main Methods:

• Utilized 13 diverse data sources, including population cohorts (BMES, GHS) and 11 public datasets.
• Applied a standardized image processing strategy for 30° disc-centered images.
• Tested the G-RISK model on 149,455 fundus images.

Main Results:

• Achieved high Area Under the Curve (AUC) values on population cohorts: 0.976 (BMES) and 0.984 (GHS).
• Demonstrated high sensitivity (87.3% and 90.3% at 95% specificity), exceeding clinical recommendations.
• Reported AUC values on public datasets ranging from 0.854 to 0.988.

Conclusions:

• The G-RISK model exhibits excellent generalizability, even when trained on homogeneous data.
• Validated the model's effectiveness across varied clinical and imaging conditions.
• Further validation in prospective studies is recommended.