Deep learning based segmentation of Meibomian glands for quantitative image analysis

Abstract

Meibomian glands (MG) are glands along the rims of the eyelid to produce oily substance to prevent the evaporation of eye’s tear film. Recently, MG dysfunction is one of the main reason for dry eye disease. The major challenge in this field is to examine the MG through gland expression with infrared images to segment MG dropout, changes to the lid morphology and determine the area of each gland segment. Despite advances in deep learning, its application in MG segmentation is limited by a lack of abundant images and expert labels of those infrared images. We introduced a novel semi-supervised system to generate labeled training images using local adaptive threshold techniques. A U-Net based convolutional neural network(CNN) was trained using augmented images from those labeled images combined with MG Infrared images taken from expert ophthalmologist using LipiView II, an ocular surface interferometer. The CNN has shown promising results segmenting MG with much greater accuracy compared to all traditional methods. Besides, we resembled a fully supervised method to train the CNN with hand-labeled infrared images of MG images and validated our model. While analyzing MG manually by expert physicians is time-consuming and traditional techniques examine that in a fraction of time by sacrificing quality with low accuracy result, our approach provides significant higher accuracy compared to traditional techniques as well outperform human-level accuracy by 7%. Also, we optimized the model to detect region of interest (ROI, i.e. eyelid within the image) & remove glare from MG Images. Finally, analysis of fraction of MG within the ROI with our method to detect MG dysfunction is presented to make our technique highly reliable and effective Tobe utilized in the clinical workflow