Automatic identification of meibomian gland dysfunction with meibography images using deep learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Dry eye disease is a prevalent condition often linked to meibomian gland dysfunction (MGD).
• Noninvasive infrared meibography is a key diagnostic tool for MGD, enabling objective gland assessment.
• Current AI applications show limited progress in addressing MGD diagnostics.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) method for measuring and assessing meibomian glands using meibography images.
• To enhance the accuracy and efficiency of MGD diagnosis through automated image analysis.

Main Methods:

• Utilized the Mask R-CNN deep learning framework.
• Trained the DL model on 1878 manually annotated meibography images (conjunctiva and meibomian glands).
• Validated the model's performance against manual annotations from two licensed eyelid specialists using an independent test set of 58 images.

Main Results:

• The DL model achieved high accuracy in identifying conjunctiva (mAP > 0.976) and meibomian glands (mAP > 0.92).
• The model demonstrated precise calculation of meibomian gland loss ratio, with minimal differences compared to specialist evaluations.
• Automated evaluation per image took 480 ms, approximately 21 times faster than human specialists.

Conclusions:

• The developed DL model significantly improves the accuracy of meibography image evaluation for MGD.
• This AI tool assists specialists in grading meibomian gland status more effectively.
• The model offers substantial time savings for clinical specialists in MGD assessment.