Clustering-Based Analysis of Meibomian Gland Morphology for Automated Assessment of Meibomian Gland Dysfunction Using the MGD-1k Dataset

Purpose: To develop a clustering-based framework for automated categorization of Meibomian gland morphology, addressing subjectivity in Meibomian Gland Dysfunction (MGD) diagnosis. Methods: The MGD-1k dataset of 1,000 infrared meibography images was analyzed using k-means clustering to classify gland areas into three groups: low (Cluster 1), medium (Cluster 2), and high (Cluster 3). Pixel-based segmentation determined white pixel counts as a surrogate for gland area. Statistical validation, including ANOVA, was performed, and results were visualized with scatter plots, bar charts, and box plots. Results: Clustering identified 417, 358, and 225 images in Clusters 1, 2, and 3, comprising 41.7%, 35.8%, and 22.5% of the dataset, respectively. Mean pixel counts were 39,123.99 (SD = 4,213.89), 23,960.30 (SD = 5,930.49), and 54,413.95 (SD = 6,378.15). ANOVA confirmed significant inter-cluster differences (p < 0.0001). Conclusions: This framework objectively quantifies gland morphology, enabling severity stratification. The approach offers scalable diagnostic support and could be integrated with clinical tools or AI models. Future validation on diverse datasets is needed to confirm its broader applicability.