Mapping malignancy: Multicenter study addressing topographic challenges in biliary stricture artificial intelligence analysis.

BACKGROUND AND STUDY AIMS: Cholangiocarcinoma (CCa) is a complex malignancy of the biliary tract, classified as intrahepatic, perihilar, or distal. Digital single-operator cholangioscopy (D-SOC) enhances evaluation of biliary strictures, although it remains limited by suboptimal biopsy yield and technical constraints. Artificial intelligence (AI), particularly convolutional neural networks (CNNs), has emerged as a promising adjunct. However, performance across anatomical subtypes is not well defined. This study evaluated diagnostic performance of an AI-based model in detecting CCa lesions by location.

PATIENTS AND

METHODS

A YOLOv8-based CNN was trained and validated using 315,993 D-SOC images from 183 patients across six international high-volume centers. Images were labeled as benign or malignant based on expert consensus. Frame-based analysis assessed diagnostic performance using macro-average F1-score, precision, and recall. Subgroup analysis explored anatomical site-specific performance.

RESULTS

Among the included patients (mean age 66.1±12.1 years; 64.5% male), 43.2% had perihilar, 37.2% intrahepatic, and 19.7% distal biliary strictures. The model demonstrated high overall performance: F1 score 95.3%, precision 95.5%, and recall 95.1%. Site-specific analysis revealed an F1 score of 89.6% for distal strictures and 91.1% for perihilar strictures. Receiver operating characteristic curves showed areas under the curve of 0.980 for intrahepatic strictures and 0.990 for perihilar and distal strictures.

CONCLUSIONS

This is the first study to demonstrate AI-based diagnostic performance across CCa topography using a large, multicenter D-SOC dataset. Although anatomical complexity affects detection, the model's high precision and generalizability suggest potential for clinical utility. These results support application of an AI-enhanced algorithm decision algorithm for cholangioscopy.