Promoting smartphone-based keratitis screening using meta-learning: A multicenter study.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Keratitis is a leading cause of corneal blindness globally.
• Early detection and referral are crucial for improving patient outcomes.
• Smartphones offer potential for keratitis screening in underserved areas, but data scarcity challenges traditional deep learning.

Purpose of the Study:

• To propose a meta-learning framework, cosine nearest centroid-based metric learning (CNCML), for smartphone-based keratitis screening.
• To develop a robust model despite limited smartphone data by utilizing prior knowledge from slit-lamp images.

Main Methods:

• Developed and assessed CNCML using 13,009 slit-lamp and 4,075 smartphone photographs from 3 clinical centers.
• Trained CNCML with varying small datasets (0-20 images/class) from HUAWEI smartphones to simulate real-world scarcity.
• Evaluated performance on internal and external smartphone datasets (VIVO, XIAOMI) and compared with traditional deep learning models.

Main Results:

• CNCML achieved high accuracies (83.15%-89.99%) and macro-AUCs (0.95-0.98) with only 15 smartphone images per class.
• CNCML outperformed traditional deep learning models by 0.56% to 9.65% in accuracy on smartphone datasets.
• Demonstrated fast learning capabilities and remarkable performance with minimal training samples.

Conclusions:

• CNCML offers a viable solution for keratitis screening using smartphones, even with limited data.
• This meta-learning approach facilitates the transition of intelligent keratitis detection from professional equipment to ubiquitous devices.
• Enhances convenience and effectiveness of keratitis screening, particularly in remote or underserved regions.