AI assistance enables surgical residents to accurately annotate anatomical structures in laparoscopic surgery

Abstract

Background The development of artificial intelligence in surgery is limited by the scarcity of annotated datasets. Recent advances in medical image segmentation, such as the Medical Segment Anything Model (MedSAM), may facilitate efficient annotation of surgical videos by residents. Methods Eight surgical residents annotated intra-abdominal structures from two laparoscopic datasets using MedSAM. The anatomical structures included the abdominal wall, colon, gallbladder, liver, small bowel, spleen, and stomach. Annotation accuracy was evaluated by comparing resident annotations to reference annotations from expert surgeons using the Dice Similarity Coefficient (DSC). A sensitivity analysis compared MedSAM annotation time with traditional watershed segmentation methods. Results A total of 25,516 annotations were created across 8,390 frames. Most annotations (87%) were performed by first- to third-year residents with no prior annotation experience. Median DSC ranged from 86.5% for gallbladder to 94.4% for liver annotations. False positive rates ranged from 7.7% to 14.6%, while false negative rates ranged from 4.7% to 13.6%. MedSAM reduced annotation time by 59.7% compared to watershed segmentation (25.4 vs 63.0 seconds per image). Conclusion Surgical residents can achieve high annotation accuracy using AI-assisted tools, with median DSC exceeding 86% across all structures. This demonstrates the potential for AI-assisted annotation to accelerate the development of annotated surgical datasets while maintaining annotation quality, even when performed by less experienced annotators.