A fully automated dry eye and breakup pattern detection based on deep learning

Dry Eye Disease (DED) is a chronic ocular condition characterized by a high prevalence rate. It arises from an imbalance between tear secretion and the evaporation of the tear film’s aqueous component, leading to tear film instability. The Tear film Break-Up Time (TBUT) test is one of the commonly employed clinical methods for diagnosing dry eye, while the classification of Tear film Break-Up Patterns (TBUPs) provides valuable assessment support for personalized treatment strategies. Currently, both TBUT testing and TBUP identification rely primarily on manual assessment. The process is cumbersome, time-consuming, and labor-intensive, and is further susceptible to subjective influences. Consequently, the development of fully automated systems based on deep learning techniques for TBUT and TBUP detection is a significant research field. We constructed a dataset of 248 fluorescent videos from 148 DED patients and 100 non-DED subjects. We developed a fully automated, real-time system integrating three networks: a Multi-Attention U-shape Network (MAUN) for tear film segmentation, a Lightweight and Efficient Classification Network (LECN) for break-up frame detection and TBUT calculation, and a Multi-Attention Classification Network (MACN) for TBUP recognition, enhanced by a log-ratio difference method for spatiotemporal feature extraction. The system achieved an accuracy (ACC) of 0.96 and an Area Under the Curve (AUC) of 0.98 for tear film breakup detection. In measuring TBUT, compared with clinical ophthalmologists, this system yielded a root mean square error of 0.112 s and a mean absolute error of 0.102 s, with a maximum deviation of ±0.21 s. For TBUP detection, the overall ACC was 0.96, with precision ranging from 0.94 to 0.98 across types. This fully automated system enables efficient, real-time, large-scale TBUT measurement and TBUP assessment and provides robust support for DED’s Tear Film-Oriented Diagnosis (TFOD) and Film-Oriented Therapy (TFOT).