Development of Human Conjunctival Goblet Cell Segmentation Datasets to Improve Quantitation

Abstract

Dry eye disease is an inflammatory disease of the ocular surface and one of the most common pathologies in medicine. This multifactorial disease can cause significant morbidity and visual disturbance. The ocular tear film consists of an outer lipid layer, an underlying aqueous layer and an innermost mucus layer produced by goblet cells interspersed in the conjunctiva. This inner mucus layer is vital for tear film stability, immunoregulation and ocular surface health. With dry eye disease and several other ocular surface pathologies the goblet cells can be affected, with decreased density and function. Much laboratory research is being performed on goblet cells necessitating manual counting and evaluation. This work presents the first comprehensive, publicly available dataset of semantically segmented goblet cells consisting of more than 65,000 instances. Moreover, we include versions of the dataset compatible with several state-of-the-art computer vision models and source code for training and testing. The current dataset can be used to train local models, as a basis for transfer learning on similar datasets, to streamline laboratory workflow, and thus save time and resources by reducing manual effort. Moreover, the dataset can play an important role in the development of improved computer vision algorithms in cellular detection.