Abstract
Purpose :
Meibomian gland morphology is a sensitive early indicator of meibomian gland dysfunction, yet there exist few methods that automatically provide standard quantification of morphological features. We developed an automated deep learning approach to segmenting meibomian glands, analyzing their morphology and predicting ghost types based on morphology.
Methods :
A total of 1432 meibography images were collected and annotated for each image with detailed gland regions. The dataset was then divided into the development and evaluation sets. The development set was used to train and tune the deep learning model, while the evaluation set was used to evaluate the performance of the model. The gland segmentations were further used to analyze individual gland features including gland length, width, curvature and average intensity. The features were further fed to a machine learning model to predict the ghost type of each individual gland.
Results :
1028 meibography images (including 486 upper lid and 542 lower lid) were used for training and tuning the deep learning model while the remaining 404 images (including 203 upper lid and 201 lower lid) were used for evaluations. The algorithm achieves 82% F1 score in detecting meibomian glands and 67% mean IoU in segmenting gland regions. Morphological features of each gland were analyzed and further fed to a SVM model for predicting ghost gland, resulting 69% prediction accuracy. Analysis of the model coefficient indicates that average gland intensity is the primary indicator for predicting ghost glands.
Conclusions :
The proposed approach can automatically segment individual meibomian gland, quantitatively analyze gland morphological features and efficiently predict ghost glands from meibography images. Our approach as a standard quantification tool may be helpful for further quantitative studies regarding meibomian gland dysfunction and dry eye syndrome.
This is a 2020 ARVO Annual Meeting abstract.