Related Experiment Video
Updated: Jan 10, 2026
Endoscopic Ultrasound-Guided Biliary Drainage: Endoscopic Ultrasound-Guided Hepaticogastrostomy in Malignant Biliary Obstruction
Published on: March 25, 2022
Artificial intelligence-assisted biliary stent length selection for common bile duct strictures in endoscopic
Wen-Lin Zhang1, Xue-Jun Shao2, Xuan-Yuan Dong3
1Department of Gastroenterology, Qilu Hospital, Cheeloo College of Medicine, Shandong University, Jinan 250012, China; Laboratory of Translational Gastroenterology, Qilu Hospital, Cheeloo College of Medicine, Shandong University, Jinan 250012, China; Robot Engineering Laboratory for Precise Diagnosis and Therapy of GI Tumor, Qilu Hospital, Cheeloo College of Medicine, Shandong University, Jinan 250012, China.
Background:
Biliary stent placement during endoscopic retrograde cholangiopancreatography (ERCP) is important for drainage in common bile duct (CBD) strictures, while the stent length is associated with many stent-related complications. We aimed to develop an artificial intelligence (AI) model for stent length selection during ERCP.
Methods:
Images of the patients who underwent ERCP and were diagnosed with CBD strictures were collected. Training involved identifying and delineating the duodenoscope, CBD and guidewire, calculating the pixel distance of the target guidewire and determining the required biliary stent length based on the diameter of the duodenoscope. The performance of the model, accuracy for length calculation and the assistance for endoscopists were validated using the testing set.
Results:
A total of 794 images from 431 patients were included and data augmentation was conducted. The mean intersection over union (mIoU) for duodenoscope, CBD and guidewire were 90.46%, 84.79% and 84.64%, respectively. The accuracy in identifying the strictures was 97.58% (121/124). The accuracy for stent length calculation achieved 85.95% (104/121) with an error margin of ± 1 cm. The mean absolute error (MAE) and mean relative error (MRE) of the AI model was 0.81 cm and 0.13, respectively. The AI model could reduce approximately 202 mGycm2 of the radiation exposure for each patient. It significantly improved both MAE and MRE for less experienced endoscopists (P = 0.01 and P = 0.02, respectively).
Conclusions:
The AI model could accurately identify duodenoscope, CBD and guidewire, enabling accurate strictures identification and stent length selection.
