VAULT-OCT: Vault Accuracy Using Deep Learning Technology – An AI Model for Predicting Implantable Collamer Lens Postoperative Vault with AS-OCT

Purpose: To develop an accurate deep learning model, VAULT-OCT, to predict postoperative vault of phakic Implantable Collamer Lenses (ICLs) based on pre-operative OCT. Setting: Parkhurst NuVision LASIK Eye Surgery, San Antonio, Texas, USA Design: Retrospective Machine Learning Study. Methods: A total of 324 eyes from 162 consecutive patients who underwent ICL implantation were included. VAULT-OCT, the neural network, was trained on pre-operative anterior segment (AS) optical coherence tomography (OCT) images paired with postoperative vault measurements for different ICL sizes. Incomplete data were excluded, and the images were consistently resized and normalized. A custom classifier was used in VAULT-OCT, and model performance was evaluated using Root Mean Squared Error (RMSE) on the test set, with Mean Absolute Error (MAE) reported as the primary performance metric. Results: A mean absolute error (MAE) of 22.3 µm, 21.7 µm, and 98.1 µm and a standard deviation of 13.5 µm, 17.8 µm, and 105.9 µm were achieved with 100%, 100%, and 89.1% of predictions within 200 µm, for the 12.1 mm, 12.6 mm, and 13.2 mm size respectively. Conclusions: This OCT-based deep learning model, VAULT-OCT, achieved a high level of accuracy in predicting postoperative ICL vault, with most predictions falling within a clinically acceptable margin of vault, suggesting the feasibility of basing ICL sizing on pre-operative AS-OCT.