An artificial intelligence-based navigation system for enhancing anatomical recognition and safety during laparoscopic cholecystectomy: a pilot study

Background: Failure to incise ventral to the Rouvière’s sulcus and maintain the dissection plane on the gallbladder (GB) surface predisposes patients to bile duct injuries (BDIs), particularly during trainee-performed laparoscopic cholecystectomy (LC). No existing artificial intelligence (AI) tools offer Tokyo Guidelines 2018 (TG 2018)-anchored real-time guidance. We developed and externally validated an AI navigation system that highlights the alert zone (AZ) – the hepatoduodenal-ligament tissues lying below an imaginary line from the roof of the Rouvière’s sulcus to the base of segment 4 and the infundibulum–cystic duct junction – and GB surface specified in the TG 2018. Materials and Methods: Seventy-three LC videos (January 2022–March 2024) were used to train the DeepLab v3 + segmentation model. The AZ and GB surface were manually annotated. The performance was tested on 10 independent videos (100 frames) with an intersection-over-union (IoU) metric. A two-arm pilot usability study randomized 10 fifth- or sixth-year postgraduate surgeons to answer video-based safety questions with or without AI assistance (20 tasks each). Results: The AI achieved a mean IoU of 0.703 (AZ) and 0.735 (GB) compared to the developer ground truth and 0.706 (AZ) and 0.730 (GB) compared to the external ground truth. With AI navigation, the correct selection of a safe incision point increased from 58% to 90%, and contour recognition of the GB surface increased from 70% to 92% (both P < 0.05). Conclusion: The AI navigation system based on the TG 2018 reliably delineated critical landmarks and markedly improved intraoperative trainee decision-making. Prospective real-time trials should determine whether this technology reduces BDIs.