Development and Preliminary Validation of an AI-Derived Hemorrhage Metric for Surgical Skill Assessment in Laparoscopic Cholecystectomy
Related Experiment Video
Updated: Nov 1, 2026
Application of Hemostatic Devices in Laparoscopic Hepatectomy
Published on: April 19, 2022
Zhixiang Chen1, Xiaojun Zeng1, Haoyuan Qiu2
1Department of Hepatobiliary Surgery, Zhujiang Hospital, Southern Medical University, Guangzhou, China; Institute of Digital Intelligent Minimally Invasive Surgery, Zhujiang Hospital, Southern Medical University, Guangzhou, China; Guangdong Provincial Clinical and Engineering Center of Digital Medicine, Guangzhou, China; South China Institute of National Engineering Research Center of Innovation and Application of Minimally Invasive Instruments, Guangzhou, China.
Introduction:
The objective assessment of technical skill remains a significant challenge in surgical training and quality control. Intraoperative hemorrhage, a direct consequence of tissue handling, may serve as a quantifiable proxy for surgical proficiency. This study aims to develop and validate an artificial intelligence-derived metric based on intraoperative hemorrhage for objective skill assessment in laparoscopic cholecystectomy.
Methods:
Frames were sampled at 1 fps (every 25th frame) from videos recorded at 25 fps, and the 70 curated cases were split at the video (case) level into training, validation, and test sets (8:1:1; 56/7/7 videos). An external test set included 30 cases from Zhujiang Hospital. Hemorrhagic regions were manually annotated on 84,680 frames from 56 training videos. A PoolNet-based model quantified hemorrhagic pixels per frame during critical view of safety. Model outputs were validated against surgeon assessments (Fleiss' κ = 0.72) and operative metrics, with the modified Global Operative Assessment of Laparoscopic Skills (mGOALS) scale as the skill reference. Mean hemorrhagic pixels were correlated with skill scores, ratings, and maneuver frequencies.
Results:
The hemorrhage identification model achieved dice scores of 0.819 ± 0.054 (internal) and 0.685 ± 0.031 (external). The proficiency metric negatively correlated with expert skill scores (Spearman's r = -0.93, P < 0.001), and strongly positively with subjective bleeding assessments (ρ = 0.89, P < 0.001) and hemostasis frequency (ρ = 0.94, P < 0.001). It effectively stratified surgeons by skill, with high-skill operators having fewer hemorrhagic pixels (Kruskal-Wallis H = 21.0-31.1, P < 0.001).
Conclusions:
The developed model provides an objective, automated hemorrhage metric that, in this preliminary validation, aligned with expert skill ratings and operative behavior during laparoscopic cholecystectomy. Pending difficulty-adjusted, multicentre validation, it may serve as a potential adjunct for postoperative, competency-based feedback rather than as a stand-alone measure of surgical skill.