Abstract
PURPOSE: To evaluate the performance of a new deep learning-based implantable collamer lens (ICL) sizing model that uses raw swept-source optical coherence tomography (SS-OCT) images as an input. SETTING: Multicenter European study. DESIGN: Retrospective external validation study. METHODS: Patients implanted with EVO ICL V4 at 2 European clinics between October 2019 and April 2024 were analyzed. Preoperative OCT images and implanted ICL data were processed by the model under study, which predicted vault, confidence levels, and the probability of achieving a postoperative vault within 250 to 750 μm (P250-750). Predictions were compared with actual postoperative vaults, and performance metrics such as mean absolute error (MAE), P250-750 accuracy, and postoperative vault distribution were assessed. Preoperative images were evaluated blinded from the postoperative results. RESULTS: 848 eyes from 429 patients were included. The mean postoperative vault was 476 ± 235 μm, with a model MAE of 146 ± 113 μm, significantly outperforming the STAAR nomogram (186 ± 149 μm; P < .001). The model correctly predicted vaults within ±250 μm in 81.7% of cases and ±300 μm in 90.7%. Among cases outside the 250 to 750 μm range (28.9%), the model recommended more appropriate sizes in 70.6%. P250-750 was comparable with the actual proportion of eyes achieving a satisfactory vault for P250-750 >60%. In cases requiring lens exchange, the model's suggested size aligned with the final implanted size in 81.8% of cases. CONCLUSIONS: This deep learning-based model, using raw OCT images, provided accurate ICL sizing predictions and valuable metrics such as P250-750 to assist clinical decision-making. This approach may reduce sizing errors and improve patient outcomes. The model is available for ANTERION SS-OCT users at safevaulticl.com .