Prediction of Early Visual Outcome of Small-Incision Lenticule Extraction (SMILE) Based on Deep Learning.

INTRODUCTION

Deep learning (DL) has been widely used to estimate clinical images. The objective of this project was to create DL models to predict the early postoperative visual acuity after small-incision lenticule extraction (SMILE) surgery.

METHODS

We enrolled three independent patient cohorts (a retrospective cohort and two prospective SMILE cohorts) who underwent the SMILE refractive correction procedure at two different refractive surgery centers from July to September 2022. The medical records and surgical videos were collected for further analysis. Based on the uncorrected visual acuity (UCVA) at 24 h postsurgery, the eyes were divided into two groups: those showing good recovery and those showing poor recovery. We then trained a DL model (Resnet50) to predict eyes with early postoperative visual acuity of patients in the retrospective cohort who had undergone SMILE surgery from surgical videos and subsequently validated the model's performance in the two prospective cohorts. Finally, Gradient-weighted Class Activation Mapping (Grad-CAM) was performed for interpretation of the model.

RESULTS

Among the 318 eyes (159 patients) enrolled in the study, 10,176 good quality femtosecond laser scanning images were obtained from the surgical videos. We observed that the developed DL model achieved a high accuracy of 96% for image prediction. The area under the curve (AUC) value of the DL model in the retrospective cohort was 0.962 and 0.998 in the training and validation datasets, respectively. The AUC values in two prospective cohorts were 0.959 and 0.936. At the video level, the trained machine learning (ML) model (XGBoost) also accurately distinguished patients with good or poor recovery. The AUC value of the ML model was 0.998 and 0.889 in the retrospective cohort (training and test datasets, respectively) and 1.000 and 0.984 in the two prospective cohorts. We also trained a DL model which can accurately distinguish suction loss (100%), black spots (85%), and opaque bubble layer (96%). The Grad-CAM heatmap indicated that our models can recognize the area of scanning and precisely identify intraoperative complications.

CONCLUSIONS

Our findings suggest that artificial intelligence (DL and ML model) can accurately predict the early postoperative visual acuity and intraoperative complications after SMILE surgery just using surgical videos or images, which may display a great importance for artificial intelligence in application of refractive surgeries.