[Paper Review] AI-based automated Meibomian gland segmentation, classification and reflection correction in infrared Meibography

Ripon K. Saha, A. M. Mahmud Chowdhury+7 more|arXiv (Cornell University)|May 31, 2022

Ocular Surface and Contact Lens4 citations

TL;DR

This study presents a deep learning-based framework for fully automated segmentation of meibomian glands (MG) and eyelids in non-contact infrared meibography, enabling accurate quantification of MG area, MG ratio, and meiboscore classification. It further introduces a GAN-based method to correct specular reflections without distorting gland morphology, achieving high consistency with expert annotations and improved diagnostic accuracy over manual grading.

ABSTRACT

Purpose: Develop a deep learning-based automated method to segment meibomian glands (MG) and eyelids, quantitatively analyze the MG area and MG ratio, estimate the meiboscore, and remove specular reflections from infrared images. Methods: A total of 1600 meibography images were captured in a clinical setting. 1000 images were precisely annotated with multiple revisions by investigators and graded 6 times by meibomian gland dysfunction (MGD) experts. Two deep learning (DL) models were trained separately to segment areas of the MG and eyelid. Those segmentation were used to estimate MG ratio and meiboscores using a classification-based DL model. A generative adversarial network was implemented to remove specular reflections from original images. Results: The mean ratio of MG calculated by investigator annotation and DL segmentation was consistent 26.23% vs 25.12% in the upper eyelids and 32.34% vs. 32.29% in the lower eyelids, respectively. Our DL model achieved 73.01% accuracy for meiboscore classification on validation set and 59.17% accuracy when tested on images from independent center, compared to 53.44% validation accuracy by MGD experts. The DL-based approach successfully removes reflection from the original MG images without affecting meiboscore grading. Conclusions: DL with infrared meibography provides a fully automated, fast quantitative evaluation of MG morphology (MG Segmentation, MG area, MG ratio, and meiboscore) which are sufficiently accurate for diagnosing dry eye disease. Also, the DL removes specular reflection from images to be used by ophthalmologists for distraction-free assessment.

Motivation & Objective

• To develop a deep learning model for fully automated segmentation of meibomian glands and eyelids in infrared meibography images.
• To enable quantitative assessment of MG area, MG ratio, and meiboscore with high accuracy.
• To correct specular reflections in meibography images using a generative adversarial network (GAN) to improve image clarity for clinical evaluation.
• To validate the model's performance against expert annotations across multiple metrics and independent datasets.
• To release the MGD-1K dataset, the first publicly available, high-quality dataset with precise MG and eyelid segmentation and meiboscore labels for open research.

Proposed method

• Trained two separate deep learning models to segment meibomian glands and eyelids from non-contact infrared meibography images.
• Used a classification-based deep learning model to estimate meiboscore from segmented MG and eyelid regions.
• Implemented a GAN-based architecture to remove specular reflections from original images while preserving underlying MG structures.
• Applied multiple validation strategies: linear regression, Bland-Altman analysis, and correlation with age and clinical grading.
• Utilized 1,600 clinical meibography images, with 1,000 annotated through multiple revisions by MGD experts across six grading sessions.
• Deployed the model for inference in 0.3 seconds per image, supporting both upper and lower eyelids without retraining.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve high-accuracy, automated segmentation of meibomian glands and eyelids in infrared meibography images?
• RQ2 How does the performance of the deep learning-based meiboscore classification compare to expert grading across different datasets?
• RQ3 To what extent does specular reflection correction improve image quality and diagnostic clarity without altering MG morphology or measurement accuracy?
• RQ4 How consistent are the deep learning-derived MG area and MG ratio measurements compared to manual annotations by MGD experts?
• RQ5 Can the proposed method generalize across independent clinical centers and image acquisition devices?

Key findings

• The deep learning model achieved a mean MG ratio of 25.12% in upper eyelids and 32.29% in lower eyelids, closely matching expert-annotated values of 26.23% and 32.34%, respectively.
• The model demonstrated 73.01% accuracy in meiboscore classification on the validation set and 59.17% accuracy on an independent test center dataset, outperforming MGD experts’ 53.44% validation accuracy.
• Specular reflection correction using the GAN-based model successfully restored reflection-affected regions without introducing artifacts, preserving MG structure and measurement integrity.
• Bland-Altman analysis and linear regression confirmed strong agreement between deep learning and expert-annotated MG area and MG ratio, with no significant differences after reflection correction.
• The model processed each image in 0.3 seconds, enabling real-time, fully automated analysis of both upper and lower eyelids without manual intervention.
• The study released the MGD-1K dataset, the first public, high-quality dataset with precise MG and eyelid segmentation and meiboscore annotations for open research.