Artificial intelligence for image recognition model construction: Using indocyanine green cholangiography to identify hepatocystic triangle during minimally invasive cholecystectomy

Jong-Uk Hou¹Tae Yoo^2*Seong Wook Park¹SeungLee Lee¹Won Tae Cho²Kyung Ho Park²Dong Woo Shin²Choon Hyuck David Kwon³

• ¹School of Software, Hallym University, Chuncheon-si, Republic of Korea
• ²Department of surgery, Hallym University Dongtan Sacred Heart Hospital, Hwaseong, Republic of Korea
• ³Department of Surgery, Cleveland Clinic, Cleveland, Ohio, United States

Minimally invasive cholecystectomy is one of the most frequently performed surgical procedures. However, iatrogenic injury related to hepatocystic triangle anatomy can occur even among experienced clinicians; therefore, an objective method can help in preventing this damage during cholecystectomy. The purpose of this study was to develop an image recognition model using the deep learning method to identify the hepatocystic triangle through indocyanine green (ICG) based near-infrared cholangiography (NIRC) during minimally invasive cholecystectomy Methods: Prediction of the object landmark (hepatocystic triangle) was evaluated using the YOLOv5 S model, an algorithm for real-time object detection in computer vision. We extracted 3,796 images from 200 cholecystectomy videos and divided them into 2979 training and 817 validation datasets. To annotate the process, we overlayed original and ICG images and estimated image prediction to identify the target, i.e., the hepatocystic triangle.When predicting landmarks, the YOLO model outlines the bounding boxes on each landmark against the image files.Results: Using the non-maximum suppression (NMS) algorithm, model performance changes according to the intersection over union (IoU) threshold; this high level of IoU threshold (0.7-0.9) results in duplicate predictions. The optimal IoU of NMS was 0.6 in multiple experiments, and the average precision score was 0.859.We successfully developed an image prediction model using intraoperative ICG NIRC to prevent bile duct damage during cholecystectomy. This experiment is a study based on the location of the actual hepatocystic triangle, and its findings are expected to be helpful in clinical practice because it can predict the bile duct before tissue dissection.