Artificial Intelligence for Digital Cholangioscopy Neoplasia Diagnosis

Clinical Validation of an Artificial Intelligence Software for Digital Cholangioscopy Diagnosis: an Observational Trial

Digital single-operator cholangioscopy (DSOC) findings achieve high diagnostic accuracy for neoplastic bile duct lesions. To date, there is not a universally accepted DSOC classification. Endoscopists' Intra and interobserver agreements vary widely. Cholangiocarcinoma (CCA) assessment through artificial intelligence (AI) tools is almost exclusively for intrahepatic CCA (iCCA). Therefore, more AI tools are necessary for assessing extrahepatic neoplastic bile duct lesions.

In Ecuador, the investigators have recently proposed an AI model to classify bile duct lesions during real-time DSOC, which accurately detected malignancy patterns. This research pursues a clinical validation of our AI model for distinguishing between neoplastic and non-neoplastic bile duct lesions, compared with high DSOC experienced endoscopists.

Distinguishing neoplastic from non-neoplastic bile duct lesions is a challenge for clinicians. Magnetic resonance (MR) and biopsy guided by endoscopic retrograde cholangiopancreatography (ERCP) reached a negative predictive value (NPV) around 50%. On the other hand, Digital single-operator cholangioscopy (DSOC) findings achieve high diagnostic accuracy for neoplastic bile duct lesions. DSOC could be even better than DSOC-guided biopsy, which is inconclusive in some cases. However, to date, there is no universally accepted DSOC classification for that purpose. Also, endoscopists' Intra and interobserver agreements vary widely. Therefore, a more reproducible system is still needed.

With interesting results, Cholangiocarcinoma (CCA) assessment through artificial intelligence (AI) tools has been developed based on imaging radiomics. Nevertheless, CCA AI resources are almost exclusively for intrahepatic CCA (iCCA), with an endoscopic technique. Therefore, more AI tools are necessary for assessing extrahepatic neoplastic bile duct lesions. Perihilar (pCCA) and distal (dCCA) cholangiocarcinoma as the most typical neoplastic bile duct lesions. Both represent 50-60% and 20-30% of all CCA, including secondary malignancies by local extension (hepatocarcinoma, gallbladder, and pancreas cancer).

A recent Portuguese proof-of-concept study developed an AI tool based on convolutional neuronal networks (CNNs). It let to differentiate between malignant from benign bile duct lesions or normal tissue with very high accuracy. Still, it needs more external validation, including endoscopists' Intra and interobserver agreement comparison. In Ecuador, the investigators recently proposed an AI model to classify bile duct lesions during real-time DSOC, which has been able to detect malignancy pattern in all cases.

This research pursues a clinical validation of our AI model for distinguishing between neoplastic and non-neoplastic bile duct lesions, compared with six endoscopists with high DSOC experience.

Inclusion Criteria:

• Patients referred to our center with an indication of DSOC due to suspicion of CBD tumor or indeterminate CBD stenosis.
• Patients who authorized for recording DSOC procedure for this study.

Exclusion Criteria:

• Any clinical condition which makes DSOC inviable.
• Patients with more than one DSOC.
• Low quality of recorded DSOC videos, even for AI model as for the expert endoscopists.
• Lost on a one-year follow-up after DSOC.
• Disagreement between DSOC findings vs. one-year follow-up, even after re-assessment of respective DSOC videos.