Artificial Intelligence Can Segment Meibomian Glands, Analyze Features

Young Woman Getting Eye Test at Optometrists
Portrait of young ophthalmologist testing sight of female patient using slit lamp exam n eye clinic
The proposed method, which would be the first of its kind, also identifies ghost glands.

A new artificial intelligence approach can automatically segment individual meibomian glands in infrared meibography images, identify ghost glands, and quantitatively analyze morphological features, according to findings published in Optometry and Vision Science.

Researchers assessed the effectiveness of an automated method that provides standard quantifications of morphological features for individual meibomian glands. The study explains that meibomian gland morphological abnormality is a common clinical sign of meibomian gland dysfunction; however, no such automated methods currently exist.

In the current study, researchers collected and annotated a total of 1443 meibography images and divided the dataset into 2 sets. First, a development set, which was used to train and tune deep learning models for segmenting glands and identifying ghost glands from images, was created. Subsequently, an evaluation set was created to evaluate the performance of the model. The researchers further used the gland segmentations to analyze individual gland features, including gland local contrast, length, width, and tortuosity.

The study used a total of 1039 meibography images, including 486 upper and 553 lower eyelids, for training and tuning the deep learning model; the remaining 404 images, including 203 upper and 201 lower eyelids, were used for evaluations. The mean intersection over union in segmenting glands achieved by the algorithm on average was 63%, as well as 84.4% sensitivity and 71.7% specificity in identifying ghost glands. To analyze their associations with ghost glands, morphological features of each gland were also fed to a support vector machine, and low gland local contrast was indicated via analysis of model coefficients to be the primary indicator for ghost glands.

The researchers note that the proposed approach only analyzed glands in the central eyelid region, which is a study limitation. Additionally, they explain that accurate identification of meibomian gland morphological features was challenging due to the possible occlusion of far peripheral glands caused by incomplete eyelid eversion.

Another limitation in terms of global morphological features is that gland density and percent atrophy were not analyzed for the lower eyelids due to the inability to accurately annotate boundaries for many participants.

Researchers say the deep learning model “can potentially be helpful in furthering our understanding of the interplay between meibomian gland features and clinical signs and symptoms.”


Wang J, Li S, Yeh T, et al. Quantifying meibomian gland morphology using artificial intelligence. Optom Vis Sci. 2021;98(9):1094-1103. doi:10.1097/OPX.0000000000001767