Deep learning segmentation and quantification of Meibomian glands

Meibomian gland dysfunction is the most common cause of the dry-eye syndrome, and it refers to deterioration of the Meibomian glands that are present in the eyelids. This paper presents a strategy for segmentation of Meibomian glands using Convolutional Neural Networks. We also present a set of clinically-relevant metrics to quantify the health of the glands. In order to model the possible variations in data using a limited representative training image database, our work proposes several custom augmentation strategies to use the available data efficiently. We have collected Meibography images from two sources: (i) Oculus Keratograph-5M, which is a high-end table-top equipment, and (ii) Prototype Hand-held camera. We have found that the images captured from the hand-held imager are of sufficient quality and comparable to the Oculus Keratograph. We present the analysis of the results of gland segmentation on test-sets from both these imagers and compare against the results from ground-truth markings by clinical experts. Our deep learning-based segmentation model is tested on an equal number of diseased images as well as healthy images. We conclude that the metrics from our segmentation results are close to those derived from ground-truth, and also that the metrics are useful for differentiating between healthy versus diseased eyes. The p-values between the ground-truth and the proposed method is p > 0.005 consistently for all the metrics, and therefore, the segmentation approach is quite accurate. There is no overlap in the intervals between the healthy and diseased cases for majority of the metrics. Average metrics for diseased cases using the proposed algorithm on the images captured by the prototype are: number of glands 13.11, tortuosity 1.29, width 4.3, length 40.31 and gland-drop 0.70 and for healthy cases: number of glands 14.40, tortuosity 1.31, width 4.20, length 47.88 and gland-drop 0.56.