Deep learning models for predicting opaque bubble layer morphology of keratorefractive lenticule extraction before laser scanning.

Purposes: To develop deep learning (DL) models for predicting opaque bubble layer (OBL) morphology and area before femtosecond laser scanning in keratorefractive lenticule extraction (KLEx) procedures. Methods: A total of 10276 frames from 5138 KLEx surgical videos, involving 2698 patients, were used to construct and validate the DL models. Suction-initiated frames captured before laser scanning were used as input to construct an OBL perceptual attention network (OBLPA-Net) and a DL-based OBL prediction regression model, which were developed to predict the OBL morphology and area in posterior lenticular cut frames during laser scanning. The performance of models was primarily evaluated by Dice coefficient, intersection over union (IoU), and mean absolute error (MAE). Results: The OBLPA-Net demonstrated strong predictive performance for OBL morphology, achieving a Dice coefficient of 0.918 (95% confidence interval (CI): 0.915-0.921) and an IoU of 0.849 (95% CI: 0.844-0.854) on the validation set. Good generalizability was observed across external test sets. Moreover, to predict the global quantitative OBL measurement (OBL area in cornea), the regression model yielded an MAE of 0.221% (95% CI: 0.201-0.242) in the validation set, with similar results in external test sets. Conclusions: Developing accurate prediction models for intraoperative complications in KLEx across surgical steps is feasible and may aid in refractive surgery decision-making, improve surgical techniques, and enhance surgical education.