Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Automated Web-Based AI Screening of Meibomian Gland Dysfunction for Tele-Ophthalmology Applications

Authors
Karpagam Damodaran, Gomathi Suresh, Maheshwari Srinivasan, Bemesha Smith, Saara Arsheen, Prathiksha Dency, Sofiya Grace, Anton Gil Christ, Nikitha R

Presenting
Karpagam Damodaran

PURPOSE:
To develop and validate a mobile-based AI system for remote and automated grading of meibomian gland dysfunction severity.

METHODS:
A mobile-compatible AI pipeline was developed using high-resolution images of upper and lower eyelid meibomian gland orifices. Images were captured under standardized illumination and fixed working distance, with eyelid eversion in primary gaze to ensure clear visualization. Six smartphones across four models were used to ensure device variability. An EfficientNet-based deep learning model was used for feature extraction and classification, optimized for small medical image datasets. The pipeline included eyelid and gland segmentation, automated extraction of morphological features (gland count, area, tortuosity, and gland area ratio), and direct grade prediction. Grades 0–1 were classified as normal and grades 2–3 as MGD for sensitivity and specificity analysis.

METHODS:

A mobile-compatible AI pipeline was developed using high-resolution images of upper and lower eyelid meibomian gland orifices. Images were captured under standardized illumination and fixed working distance, with eyelid eversion in primary gaze to ensure clear visualization. Six smartphones across four models were used to ensure device variability. An EfficientNet-based deep learning model was used for feature extraction and classification, optimized for small medical image datasets. The pipeline included eyelid and gland segmentation, automated extraction of morphological features (gland count, area, tortuosity, and gland area ratio), and direct grade prediction. Grades 0–1 were classified as normal and grades 2–3 as MGD for sensitivity and specificity analysis.

RESULTS:
A total of 1,746 images were graded by expert optometrists using a four-point scale (0–3); 1,197 were used for training and 549 for validation. The AI achieved a sensitivity of 95.5% (95% CI: 93.2–97.8%), specificity of 90.0% (95% CI: 86.2–93.8%), PPV of 92.5% (95% CI: 89.6–95.4%), NPV of 93.9% (95% CI: 90.8–97.0%), and accuracy of 93.2% (95% CI: 91.1–95.3%). Inter-rater agreement was high (Cohen's Kappa with quadratic weights κ = 0.921; 95% CI: 0.89 -- 0.94).

RESULTS:

A total of 1,746 images were graded by expert optometrists using a four-point scale (0–3); 1,197 were used for training and 549 for validation. The AI achieved a sensitivity of 95.5% (95% CI: 93.2–97.8%), specificity of 90.0% (95% CI: 86.2–93.8%), PPV of 92.5% (95% CI: 89.6–95.4%), NPV of 93.9% (95% CI: 90.8–97.0%), and accuracy of 93.2% (95% CI: 91.1–95.3%). Inter-rater agreement was high (Cohen's Kappa with quadratic weights κ = 0.921; 95% CI: 0.89 -- 0.94).

CONCLUSIONS:
A smartphone-compatible AI-based system for MGD grading has been successfully developed and validated