Impact of biliary stricture topography on the performance of artificial intelligence algorithms for digital single-operator cholangioscopy: a multicenter, transatlantic, study

Aims Cholangiocarcinoma (CCa) is a complex neoplasm with poor prognosis, classified as intrahepatic (iCCa), perihilar (pCCa), or distal (dCCa) depending on its location. Its variable phenotypes pose diagnostic challenges. Digital single-operator cholangioscopy (DSOC) has emerged as an important diagnostic tool but still has limitations, including suboptimal biopsy sensitivity, lack of a standardized morphology classification, as well as technical issues. Emerging evidence highlights AI’s potential to enhance DSOC diagnostic capacity, but the differential performance according to tumor characteristics, namely its topography, remains unassessed. This study aimed to evaluate an AI algorithm’s effectiveness in detecting malignancy in different locations iCCa, pCCa, and dCCa.

Methods A deep learning model based on a convolutional neural network (CNN) was designed to automatically predict the malignancy status of biliary strictures. The model was built using DSOC images from three international high-volume centers: Centro Hospitalar Universitário de São João (CHUSJ, Porto, Portugal), Hospital Universitario Puerta de Hierro Majadahonda (HUPHM, Madrid, Spain), and New York University Langone Hospital (NYULH, New York, USA). Images showing strictures due to cholangiocarcinoma or benign conditions were included. The model was trained with a learning rate of 0.0001, a batch size of 32, and a number of epochs of 10, using PyTorch framework. The detection rate (DR) of the algorithm was compared for each location category.

Results A total of 18158 images from 31 patients were included (CHUSJ, n=21; HUPHM, n=6; NYULH, n=4). A total of 22 patients were male (71%) and the mean age was 63 years (SD: 11 years). Common bile duct (CBD) strictures represented 45% of the total (n=14) while intrahepatic and perihilar accounted for 23% (n=7) and 32% (n=10), respectively. The algorithm’s DR of strictures representing CCa was 91.6% for iCCa, 91.5% for pCCa and 79.6% for dCCa. Pairwise comparisons revealed a significantly difference in the DR between intrahepatic and CBD (p<0.001) and perihilar and CBD (p<0.001).

Conclusions Topographic characteristics of biliary strictures influence the diagnostic performance of DSOC, often posing difficulty to the evaluation of distal biliary strictures. While AI shows promise in improving the diagnostic capacity of DSOC, its performance is impacted by DSOC´s technical limitations. Developing models should address and control for these issues before widespread clinical integration. Additionally, this evidence stresses the importance of keeping a “human on the loop”, providing clinical reasoning to guide the application of AI algorithms.

 

Scroll to top

Conflicts of Interest

Authors do not have any conflict of interest to disclose.

Publication History

Article published online:
27 March 2025