Meibography Phenotyping and Classification From Unsupervised Discriminative Feature Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Meibomian gland (MG) atrophy is a key indicator of Meibomian gland dysfunction (MGD).
• Accurate assessment of MG atrophy severity is crucial for diagnosis and management.
• Current methods often rely on subjective grading or require extensive manual annotation.

Purpose of the Study:

• To develop an unsupervised feature learning approach for automatic measurement of MG atrophy severity from meibography images.
• To discover visual similarities and relationships between meibography images using machine learning.

Main Methods:

• Utilized nonparametric instance discrimination (NPID), a convolutional neural network (CNN) model, to encode meibography images into feature vectors.
• Trained the network on 497 images and evaluated on 209 images from a dataset of 706 meibography images with meiboscores.
• Employed 3D feature visualization and agglomerative hierarchical clustering for relationship discovery.

Main Results:

• The NPID approach achieved an average meiboscore grading accuracy of 80.9%, surpassing the clinical team's performance by 25.9%.
• Successfully categorized MG characteristics and identified relationships between images through hierarchical clustering.
• Demonstrated the ability to provide quantitative analysis of MG atrophy severity based on phenotype.

Conclusions:

• The proposed unsupervised NPID approach enables automatic analysis of MG atrophy severity from meibography images without manual annotation.
• This method offers a quantitative, phenotype-based evaluation of MG atrophy.
• Potential to aid in the diagnosis and management of MGD, reducing reliance on time-consuming manual annotations.