A Generalized and Interpretable Multi-Label Multi-Disease Screening System for Ocular Anterior Segment Disease...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Ocular anterior segment diseases require timely and accurate diagnosis.
• Current screening methods may lack efficiency and broad applicability.
• Developing an automated system can aid in early detection and management.

Purpose of the Study:

• To develop and validate a multi-label, multi-disease artificial intelligence (AI) screening system.
• To apply the system to the detection of common ocular anterior segment diseases using slit-lamp images.
• To ensure the system is well-generalized and interpretable for clinical use.

Main Methods:

• A multicenter, retrospective study involving 1990 patients and 5132 slit-lamp images.
• Training and validation on an internal dataset, with testing on internal and external datasets including less-trained phenotypes.
• Comparison of the AI system's performance against ophthalmologists.

Main Results:

• The AI system achieved high accuracy across various detection levels (image, region, lesion), with average accuracies ranging from 0.827 to 0.972.
• Performance was comparable to ophthalmologists, even on less-trained or novel phenotypes (average accuracy 0.852-0.950).
• The system demonstrated excellent multi-label, multi-disease detection and generalization capabilities.

Conclusions:

• The developed AI screening system shows strong performance in detecting multiple ocular anterior segment diseases.
• The system exhibits robust generalization, even with limited training data phenotypes.
• It is poised to provide accessible primary medical information and assist ophthalmologists in clinical practice.