Artificial intelligence-based anatomical recognition improves surgeon decision-making during robotic gastrectomy.

BACKGROUND

Artificial intelligence (AI)-based anatomical recognition has emerged to support intraoperative cognition; however, its clinical utility beyond education remains limited. This study aimed to develop an AI model for suprapancreatic lymph node dissection during robotic distal gastrectomy (RDG) and evaluate its utility for intraoperative decision-making by surgeons.

METHODS

We developed a deep learning model using 67 RDG videos (54 for training and 13 for testing) to recognize the pancreas, common hepatic artery (CHA), left gastric artery (LGA), and left gastric vein (LGV). Model performance was evaluated using Intersection over Union (IoU). Twenty surgeons participated in two experiments: experiment 1 assessed peritoneal incision line selection, rated by three experts on a 5-point scale, and experiment 2 assessed the time to CHA identification in a crossover design.

RESULTS

The IoU values for the pancreas, CHA, LGA, and LGV were 0.66, 0.28, 0.216, and 0.232, respectively. In mixed-effects models, experiment 1 showed that AI assistance reduced the proportion of unsafe peritoneal incision lines (scores 1-2) compared with no assistance (odds ratio, 0.25; 95% confidence interval [CI], 0.15 to 0.41; P < 0.001) and improved the mean expert score by 0.64 points (95% CI, 0.39 to 0.89; P < 0.001). In experiment 2, AI assistance reduced the time to CHA identification by 9.5 s (95% CI, 2.20 to 16.81; P = 0.0124).

CONCLUSION

Our AI system successfully supported surgeons' intraoperative decision-making by improving anatomical recognition and thus can potentially enhance surgical safety.