Using AI and Fluorescence Guidance to Enhance Extrahepatic Bile Duct Identification Among Junior Surgeons During Laparoscopic Cholecystectomy

Surgical Education and Anatomical Identification in Laparoscopic Cholecystectomy

Laparoscopic Cholecystectomy, Artificial Intelligence, Common Bile Duct Identification, Indocyanine Green Fluorescence

The present study evaluates whether PGY trainees and surgical residents, with or without AI assistance, could accurately identify the presence and anatomical location of the CBD, as well as delineate intraoperative danger zones during LC.

This retrospective cohort study evaluated the impact of artificial intelligence (AI) assistance on anatomical recognition during laparoscopic cholecystectomy (LC). Between June 2022 and December 2024, indocyanine green (ICG) fluorescence-guided LC videos were prospectively collected at a tertiary referral center. After excluding duplicate cases, 177 videos were used for model training, 15 for validation, and 15 for testing. Frames were extracted at 1 frame per second, and key structures including the common bile duct (CBD), cystic duct, cystic artery, liver, gallbladder, and surgical instruments were annotated by board-certified hepatobiliary surgeons to generate the ground truth dataset. A YOLOv9 object detection model, incorporating Programmable Gradient Information (PGI) and Generalized Efficient Layer Aggregation Network (GELAN), was trained to recognize critical biliary anatomy. For the experimental phase, surgical trainees (postgraduate trainees, junior residents, and senior residents) reviewed condensed 2-3 minute surgical videos, segmented into 5-second clips. In the CBD recognition task, participants determined whether the CBD was visible in each clip. In the CBD annotation task, participants placed bounding boxes to indicate the CBD location on single frames, and additionally delineated a polygonal dangerous zone within Calot's triangle where further dissection is considered hazardous. Each participant first performed both tasks without AI assistance. After a one-week washout, the same tasks were repeated with AI support, which displayed YOLOv9-generated bounding boxes to guide decision-making. The task order was randomized to minimize learning bias. Performance metrics included recognition accuracy, precision, recall, F1-score, and intersection over union (IoU). This study was retrospectively registered after completion, as it used de-identified surgical videos and trainee assessments.

No related papers found yet.

None listed

Taiwan