The ocular surface · 2026
A hybrid system integrating deep learning and computer vision for automated blink monitoring and tear film break-up pattern classification.
СтатьяПоддержка исследования: вне правительства СШАНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Li Yike, Pan An-Peng, Sun Yiting, Jin Xiuming, Yu A-Yong
Аннотация
PURPOSE: To construct and validate a hybrid system integrating deep learning and computer vision for real-time blink monitoring, tear film Break-Up Patterns (BUPs) classification, and Dry Eye (DE) subtype diagnosis and treatment recommendation.
METHODS: This prospective study included 95 subjects. Three YOLOv8 models (YOLOv8-n/s/m) were trained and compared to detect ocular structures and Tear Film Break Up (TFBU) regions. The best-performing model was integrated with OpenCV image processing algorithms based on geometric morphology to classify five specific BUPs (Area, Line, Spot, Dimple, and Random) and perform real-time blink monitoring. The overall performance of the automated system was validated on an independent test set.
RESULTS: Significant differences in Ocular Surface Disease Index (OSDI) and Tear-film Break Up Time (TBUT) were observed between the DE (n = 65) and non-DE (n = 30) groups (P < 0.001). The YOLOv8-m model was selected for its optimal performance (Precision: 74.6% for TFBU detection) to construct the automated blink monitoring and BUPs classification system. In the test set, this system achieved a 100% success rate in blink monitoring and an overall BUPs classification accuracy of 75.47%. System-measured TBUT (STBUT) and Manually-measured TBUT (MTBUT) demonstrated excellent agreement. Notably, STBUT (1.99 ± 1.36 s) was significantly shorter than MTBUT (2.21 ± 1.47 s, P < 0.01), reflecting the system's higher sensitivity in detecting early TFBU, alongside inherent methodological differences. Bland-Altman analysis confirmed that differences were predominantly distributed within the 95% limits of agreement.
CONCLUSION: The hybrid system enables real-time blink monitoring and precise BUPs classification, supporting Tear Film-Oriented Diagnosis (TFOD) and personalized DE management.
METHODS: This prospective study included 95 subjects. Three YOLOv8 models (YOLOv8-n/s/m) were trained and compared to detect ocular structures and Tear Film Break Up (TFBU) regions. The best-performing model was integrated with OpenCV image processing algorithms based on geometric morphology to classify five specific BUPs (Area, Line, Spot, Dimple, and Random) and perform real-time blink monitoring. The overall performance of the automated system was validated on an independent test set.
RESULTS: Significant differences in Ocular Surface Disease Index (OSDI) and Tear-film Break Up Time (TBUT) were observed between the DE (n = 65) and non-DE (n = 30) groups (P < 0.001). The YOLOv8-m model was selected for its optimal performance (Precision: 74.6% for TFBU detection) to construct the automated blink monitoring and BUPs classification system. In the test set, this system achieved a 100% success rate in blink monitoring and an overall BUPs classification accuracy of 75.47%. System-measured TBUT (STBUT) and Manually-measured TBUT (MTBUT) demonstrated excellent agreement. Notably, STBUT (1.99 ± 1.36 s) was significantly shorter than MTBUT (2.21 ± 1.47 s, P < 0.01), reflecting the system's higher sensitivity in detecting early TFBU, alongside inherent methodological differences. Bland-Altman analysis confirmed that differences were predominantly distributed within the 95% limits of agreement.
CONCLUSION: The hybrid system enables real-time blink monitoring and precise BUPs classification, supporting Tear Film-Oriented Diagnosis (TFOD) and personalized DE management.