Abstract
INTRODUCTION: Minimally-invasive cholecystectomy is one of the most commonly performed surgical procedures. However, iatrogenic injuries related to the hepatocystic triangle anatomy can occur even if the performing surgeon has extensive experience. Therefore, an objective method that could help prevent such damages during surgery is needed. AIM: This study aimed to develop an artificial intelligence (AI)-based image recognition model using indocyanine green (ICG)-based near-infrared cholangiography (NIRC) to identify the hepatocystic triangle during minimally-invasive cholecystectomy. MATERIALS AND METHODS: Anatomical landmark prediction of the hepatocystic triangle was evaluated using the YOLOv5s model, a real-time object detection algorithm in computer vision. From 200 cholecystectomy videos, 3796 images were extracted, of which 2979 were used for training and 817 for validation. Original and ICG-enhanced images were overlaid and annotated to identify the hepatocystic triangle, and the model generated bounding boxes for each predicted landmark. RESULTS: Using the nonmaximum suppression (NMS) algorithm, model performance changed according to the intersection over union (IoU) threshold. This high level of IoU threshold (0.7-0.9) resulted in duplicate predictions. The optimal IoU of NMS was 0.6 in multiple experiments, and the average precision score was 0.859. CONCLUSIONS: We successfully developed an AI-based image recognition model using intraoperative ICG-NIRC to predict the location of the hepatocystic triangle and help prevent bile duct injury during cholecystectomy. This model, based on real anatomical localization data, shows potential clinical utility by predicting the bile duct location before tissue dissection.