Development and multicenter validation of an AI driven model for quantitative meibomian gland evaluation

Abstract

Abstract This multicenter retrospective study developed and validated an AI driven model for automated segmentation and quantitative evaluation of meibomian glands using infrared meibography images acquired by the Keratograph 5M device. A total of 1350 infrared meibography images were collected and annotated for model training and validation. The model demonstrated high segmentation performance, with an Intersection over Union of 81.67% (95% Confidence Interval [CI]: 81.03–82.31) and accuracy of 97.49% (95% CI: 97.38–97.62), outperforming conventional algorithms. The agreement was observed between AI-based and manual gland grading (Kappa = 0.93) and gland counting (Spearman r = 0.9334). Repeatability analysis confirmed the model’s stability, and external validation across four independent centers yielded consistent results with AUCs exceeding 0.99. This AI tool offers a standardized, efficient, and objective method for meibography image analysis, which may improve diagnostic precision and assist in the clinical management of meibomian gland dysfunction across diverse populations.