Deep learning in population-based screening for glaucoma.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) shows potential for glaucoma detection via retinal imaging.
• The Northern Finland Birth Cohort Eye Study evaluated glaucoma screening in middle-aged individuals.
• This study assessed the G-RISK DL model in a low-prevalence (1.1%) cohort aged 45-49 years.

Purpose of the Study:

• To evaluate the performance and generalisability of the G-RISK deep learning model for glaucoma screening.
• To assess G-RISK performance across different screening scenarios and photographic modalities.
• To understand the impact of low glaucoma prevalence and age on DL model performance.

Main Methods:

• Four screening scenarios were tested: glaucoma eyes, glaucoma patients, and combined glaucoma/suspects.
• Performance was evaluated using Area Under the Receiver Operating Characteristic Curve (AUC), Precision-Recall AUC (PR-AUC), and sensitivity at 95% specificity.
• The study analyzed results across four photographic modalities.

Main Results:

• G-RISK achieved the highest performance for detecting glaucoma eyes using color fundus photographs (AUC 0.83).
• Consistent discrimination was observed across other modalities (AUC 0.75-0.78).
• Performance was higher for confirmed glaucoma cases compared to combined glaucoma and suspect cases (AUC 0.83 vs. 0.67).

Conclusions:

• The G-RISK model performed best on color fundus images but maintained discrimination across other modalities.
• Screening metrics were limited by the cohort's young age and low glaucoma prevalence.
• Inclusion of glaucoma suspects reduced performance, potentially due to definition variability.