Multimodal deep learning for predicting postoperative vault and selecting implantable collamer lens sizes using AS-OCT and ultrasound biomicroscopy images

PURPOSE: To develop and validate a multimodal deep-learning model for predicting postoperative vault height and selecting implantable collamer lens (ICL) sizes using anterior segment optical coherence tomography (AS-OCT) and ultrasound biomicroscopy (UBM) images combined with clinical features. SETTING: West China Hospital, Sichuan University, Chengdu, Sichuan, China. DESIGN: Deep-learning study. METHODS: 626 AS-OCT and 1309 UBM images from 209 eyes of 105 participants with ICL V4c implantation were used. Features were extracted using a convolutional neural network (ResNet50) and combined with clinical data for model training. Machine learning algorithms including Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Random Forest (RF) were used to develop models for postoperative vault height prediction and ICL size selection. Models were validated using metrics such as mean absolute error (MAE), root mean squared error (RMSE), R2 , accuracy, sensitivity, specificity, and precision. RESULTS: The LightGBM, XGBoost, and RF models showed RMSE values below 150 μm, MAE values below 120 μm, and R2 values around 0.4 in predicting postoperative vault height. The LightGBM model achieved the best performance in ICL size selection, with an accuracy of 0.904, sensitivity of 0.935, specificity of 0.907, and precision of 0.873, outperforming traditional methods and nearing the performance of senior doctors. CONCLUSIONS: The multimodal deep-learning model significantly improved the accuracy of predicting postoperative vault height and selecting ICL sizes for ICL V4c implantation, overcoming the limitations of single-modal data analysis. Future studies should expand sample sizes and conduct multicenter validations to enhance model generalizability and clinical applicability.