Abstract
Purpose :
To utilize machine learning (ML) techniques - automated eyelid detection and tarsal plate segmentation - to standardize the analysis of meibography images in dry eye (DE) research. We hypothesize that raw meibography images from real-world dataset (i.e., DREAM dataset) may contain extraneous artifacts, impeding accurate meibography image quantification. The key objective is to bridge the gap between curated and real-world meibography datasets.
Methods :
A total of 4232 meibography images were collected, with 1563 sourced from a curated dataset and 2669 from a real-world dataset (i.e., DREAM dataset). This study employed a two-stage approach to standardizing meibography image analysis: automated eyelid detection and tarsal plate segmentation. The first stage utilized a detection model trained on curated images to identify relevant eyelid areas in real-world images. The subsequent stage focused on refining the eyelid area in meibography images, particularly in segmenting the tarsal plate to enhance the visualization of Meibomian gland (MG) structure and meibography image features. This method also incorporates techniques for specular reflection removal and tarsal plate mask refinement. An unsupervised feature learning model was then employed to classify images from DE and NonDE subjects.
Results :
To standardize the meibography analysis, raw images were first processed through the automated eyelid detection and tarsal plate segmentation stages. 2968 images were then used for training an unsupervised feature learning model while the remaining 634 images were used for evaluation. The model achieved an 80.8% instance-wise accuracy in distinguishing meibography images from 399 DE and 235 NonDE subjects. Visualizations through Uniform Manifold Approximation and Projection (UMAP) unveiled distinct clusters for DE and NonDE phenotypes. Furthermore, it was observed that the image features of NonDE subjects formed distinct clusters associated with the upper and lower eyelids.
Conclusions :
The ML-driven methodology in this study has the potential to simplify the integration of diverse meibography images into large-scale databases and to enhance the comparability of research findings across different sources by standardizing the meibography image processing pipeline.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.