External validation of a deep learning detection system for glaucomatous optic neuropathy: a real-world multicentre...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucomatous optic neuropathy (GON) is a leading cause of irreversible blindness worldwide.
• Accurate and timely diagnosis of GON is crucial for preventing vision loss.
• Automated diagnostic systems using artificial intelligence (AI) show promise in improving diagnostic efficiency.

Purpose of the Study:

• To externally validate an automated AI diagnostic system for detecting GON using fundus photographs.
• To assess the AI system's performance in a real-world, multicentre cohort.
• To evaluate the AI system's generalizability across different clinical scenarios and patient populations.

Main Methods:

• External validation of an AI diagnostic system using three multicentre datasets (totaling 11,360 fundus images).
• Datasets included images from different hospitals and a high myopia (HM) population.
• Performance metrics including sensitivity, specificity, accuracy, and area under the receiver operating curve (AUC) were calculated.

Main Results:

• The AI system achieved high accuracy (91.40%-93.18%) and AUC (95.17%-96.64%) in general validation datasets, outperforming manual graders in sensitivity.
• Performance remained robust on subsets with retinal comorbidities (e.g., diabetic retinopathy, age-related macular degeneration), with AUCs up to 97.46%.
• In the high myopia population, the AI system demonstrated comparable accuracy (81.98%) and AUC (87.49%) for GON detection.

Conclusions:

• The automated AI diagnostic system exhibits acceptable generalization capabilities across varying image quality, clinical centers, and retinal comorbidities.
• The AI system has the potential to provide expert-level glaucoma detection.
• This validates the AI system's utility in real-world clinical practice for identifying glaucomatous optic neuropathy.