Automatic Meibomian Gland Segmentation and Assessment Based on TransUnet with Data Augmentation

Meibomian glands (MG) are important for eyes and meibomian gland dysfunction (MGD) would lead to dry eye disease. With the development of eye imaging technology, non-contact infrared imaging named meibography has become mainstream. Physicians determine the meiboscore based on meibography images, which serves for the follow-up diagnosis. In this paper, a deep learning-based MG segmentation approach has been proposed to accurately segment meibography images. The method is based on TransUnet, which outperforms other deep learning algorithms in MG segmentation. Moreover, data augmentation is utilized to expand the limited dataset. Experiments show the dice coefficient, jaccard index, precision and recall score were achieved 0.93, 0.87, 0.94 and 0.92. Single image segmentation took 0.3 s on average. To our best knowledge, this is the first time that TransUnet and data augmentation are applied in the MG segmentation and assessment tasks.