Evaluation of meibomian gland dysfunction with deep learning model considering different datasets and gland...

Meibomian gland dysfunction (MGD) is recognized as the primary cause of evaporative-type dry eye disease (DED). Diagnosis typically involves assessing meibomian gland (MG) morphology alongside symptom evaluation. Traditionally, experts manually grade meibography images to assign a meiboscore (0-4 or 0-3), where higher scores indicate more severe gland loss. However, manual grading is subjective, time-consuming, and labor-intensive. To address these limitations, recent studies have introduced automated methods for gland segmentation and classification, utilizing either raw meibography images or geometric features extracted from segmented glands. This study presents a deep learning (DL)-based framework for automated meiboscore prediction by integrating image embeddings with MG attributes such as gland area, length, thickness, and tortuosity (curvature measure). The model is trained and tested on two datasets: an in-house dataset (BCH) with 261 meibography images from 145 patients and an open-source dataset (MGD-1K) containing 1000 images from 320 patients. The proposed model achieves 77.86 % accuracy for 5-grade meiboscore classification on BCH and 78.03 % on MGD-1K, with 81.08 % accuracy for 4-grade classification. Despite differences in imaging devices (Sirius, LipiView), the model demonstrated robust mixed-dataset performance. Furthermore, clinical tests, including tear break-up time (TBUT), Schirmer test, and ocular surface disease index (OSDI), are analyzed for correlations with MG parameters. TBUT showed a positive correlation with MG area and length, while meiboscore correlated negatively with TBUT and Schirmer test. OSDI is negatively correlated with meiboscore and tortuosity. These findings highlight the potential for DL-based automated assessments in clinical applications, improving efficiency and consistency in MGD evaluation.