人工知能搭載シミュレーターを用いた乳房再建術の結果の可視化

Visual aids that simulate personalized postoperative outcomes following breast reconstruction may facilitate shared decision-making and help manage expectations. This study aimed to develop and validate an AI-based surgical simulator trained on real patient photographs to predict breast reconstruction outcomes.

A cohort of women who underwent bilateral two-stage implant reconstruction or abdominal-based autologous breast reconstruction between January 2010 and June 2024 at Memorial Sloan Kettering Cancer Center was retrospectively identified. Standardized frontal preoperative and postoperative photographs were collected and preprocessed. An AI-based simulator was developed using a generative adversarial network (GAN) and trained on photographs over 250 epochs using a 75:25 training/testing split. AI-generated postoperative predictions were evaluated against real postoperative results both qualitatively and quantitatively using standard metrics including structural similarity index (SSIM) and Fréchet Inception Distance (FID).

A total of 1,405 patients were included (implant: n=1006 [72%]; autologous: n=399 [28%]). The cohort was predominantly White (83.5%), with smaller representations of Asian (5.5%), Black (5.0%), and Other (6.0%) patients. Quantitative assessment of image fidelity supported accurate model predictions (implant: SSIM 0.61, FID 23.77; autologous: SSIM 0.50, FID 38.21). AI-generated predictions demonstrated visual concordance with real postoperative photographs in terms of breast shape, volume, symmetry, and nipple reconstruction.

This study demonstrates the feasibility of an AI-based model trained on real-world clinical photographs to simulate breast reconstruction outcomes. These findings support further exploration of AI-based models for personalized surgical simulation in breast reconstruction and their integration into clinical care.