Unsupervised Learning Based on Meibography Enables Subtyping of Dry Eye Disease and Reveals Ocular Surface Features

Abstract

Purpose: This study aimed to establish an image-based classification that can reveal the clinical characteristics of patients with dry eye using unsupervised learning methods.

Methods: In this study, we analyzed 82,236 meibography images from 20,559 subjects. Using the SimCLR neural network, the images were categorized. Data for each patient were averaged and subjected to mini-batch k-means clustering, and validated through consensus clustering. Statistical metrics determined optimal category numbers. Using a UNet model, images were segmented to identify meibomian gland (MG) areas. Clinical features were assessed, including tear breakup time (BUT), tear meniscus height (TMH), and gland atrophy. A thorough ocular surface evaluation was conducted on 280 cooperative patients.

Results: SimCLR neural network achieved clustering patients with dry eye into six image-based subtypes. Patients in different subtypes harbored significantly different noninvasive BUT, significantly correlated with TMH. Subtypes 1 and 5 had the most severe MG atrophy. Subtype 2 had the highest corneal fluorescent staining (CFS). Subtype 4 had the lowest TMH, whereas subtype 5 had the highest. Subtypes 3 and 6 had the largest MG areas, and the upper MG areas of a person's bilateral eyes were highly correlated. Image-based subtypes are related to meibum quality, CFS, and morphological characteristics of MG.

Conclusions: In this study, we developed an unsupervised neural network model to cluster patients with dry eye into image-based subtypes using meibography images. We annotated these subtypes with functional and morphological clinical characteristics.

Figure 1.

Schematic of the analysis pipeline. Meibography images of patients with dry eye and related clinical information were collected from 2021 to 2022. Unsupervised SimCLR network was trained using meibography images, and the images were embedded into vectors using the network. Average vector of four images of one patient were calculated and considered as the image vector of the patient. Then, consensus clustering was applied to group patients into six major image subtypes, and the differences of the clinical features among image subtypes were analyzed.

Figure 2.

Consensus clustering of dry eye patients using meibography images. (A) Clustering metrics of different number of clusters (k = 2, 3, …, 16). Davies-Bouldin score (DB) and Bayesian Information Criterion (BIC) of different number of clusters were calculated, and k = 8 achieves lowest DB and less BIC. (B) Consensus matrix of k = 8. Six major clusters (> 50 patients) and two minor clusters (< 50 patients) were obtained. (C) Examples of clustering results of six major clusters (subtypes). For each cluster, 12 images of 3 patients are shown.

Figure 3.

Area of meibomian gland regions differ among image subtypes. (A) Illustration of deep learning models segment eyelid area (E model) and meibomian gland regions (G model). The ratio of outputs of two models is defined as the normalized gland area of the meibomian gland region. (B) Dice coefficient of E model and G model in training processes with an early-stop patience. (C) Normalized gland area of 6 subtypes was compared using 1-way ANOVA analysis, and F-test was applied to calculate the P values. (D) Normalized gland area of different eyelids in patients with dry eye. Spearman's rank correlation coefficient was calculated and shown.

Figure 4.

Image-based subtypes are corelated with clinical features. (A, B) Tear meniscus height of the left and right eyes of 6 subtypes were compared using 1-way ANOVA analysis, and F-test was applied to calculate the P values. (C) Noninvasive tear breakup time of 6 subtypes was compared using 1-way ANOVA analysis, and F-test was applied to calculate the P values. (D) Noninvasive tear breakup time is correlated with tear meniscus height. Patients are merged into 100 bins, and the related 2-sided P value was shown. Clinical features, including meibum expressibility (E), CFS (F), fluorescein tear break-up time (G), meibum quality score (H), and OSDI (I) of 6 image-based subtypes was compared using 1-way ANOVA analysis, and F-test was applied to calculate the P values.