Mapping malignancy: Multicenter study addressing topographic challenges in biliary stricture artificial intelligence
Miguel Mascarenhas1,2, Antonio Miguel Pinto da Costa3,4, Matheus Ferreira de Carvalho4
1Gastroenterology DepartmentCentro Hospitalar Universitário de São JoãoPortoPorto DistrictPortugal.
Endoscopy International Open
|June 25, 2026
Summary
An artificial intelligence (AI) model demonstrated high accuracy in detecting cholangiocarcinoma (CCa) using digital single-operator cholangioscopy (D-SOC) images. This AI tool shows promise for improving CCa diagnosis across different biliary tract locations.
Area of Science:
- Gastroenterology and Hepatology
- Medical Imaging
- Artificial Intelligence in Medicine
Background:
- Cholangiocarcinoma (CCa) is a complex biliary tract malignancy with intrahepatic, perihilar, and distal classifications.
- Digital single-operator cholangioscopy (D-SOC) aids biliary stricture evaluation but faces challenges with biopsy yield and technical limitations.
- Artificial intelligence (AI), particularly convolutional neural networks (CNNs), offers a potential solution for enhancing CCa detection.
Purpose of the Study:
- To evaluate the diagnostic performance of an AI-based model in detecting cholangiocarcinoma (CCa) lesions.
- To assess the AI model's performance across different anatomical subtypes of CCa (intrahepatic, perihilar, distal).
- To determine the generalizability of the AI model using a large, multicenter D-SOC image dataset.
Main Methods:
- A YOLOv8-based convolutional neural network (CNN) was trained and validated on 315,993 D-SOC images from 183 patients across six international centers.
- Images were classified as benign or malignant based on expert consensus.
- Diagnostic performance was assessed using frame-based analysis, including F1-score, precision, and recall, with subgroup analysis for anatomical sites.
Main Results:
- The AI model achieved high overall performance with an F1 score of 95.3%, precision of 95.5%, and recall of 95.1%.
- Site-specific analysis showed F1 scores of 89.6% for distal and 91.1% for perihilar strictures.
- Receiver operating characteristic curves indicated strong performance with areas under the curve of 0.980 for intrahepatic and 0.990 for perihilar and distal strictures.
Conclusions:
- This study is the first to report AI-based diagnostic performance across cholangiocarcinoma (CCa) topography using a large, multicenter D-SOC dataset.
- Despite anatomical complexities influencing detection, the AI model exhibits high precision and generalizability, suggesting significant clinical utility.
- The findings support the integration of AI-enhanced algorithms into clinical decision-making for cholangioscopy.