BASEMENT TO BEDSIDE: REAL-TIME CHOLANGIOSCOPY ARTIFICIAL INTELLIGENCE FOR BILIARY STRICTURES

Background: Accurate classification of biliary strictures continues to be a challenge, in part due to inadequate sampling techniques. Cholangioscopy provides direct visualization of biliary pathology, but it has been reported that blinded experts are inaccurate and have poor agreement for stricture classification. Thus, we developed a cholangioscopy artificial intelligence (AI) that can autonomously classify strictures as benign or malignant. In this video, we demonstrate how we designed and deployed a real-time, point-of-care cholangioscopy AI system at our hospital.

Case presentations:
Case 1 – A 51-year-old female with a history of severe COVID-19 pneumonia presented with fatigue and pruritus. An MRCP demonstrated filling defects as well as irregular areas of strictures and dilation of the intrahepatic biliary tree. During ERCP multiple large casts were removed, so a stent was placed. Cholangioscopy was performed and the AI correctly predicted this patient had benign disease as she was later diagnosed with COVID cholangiopathy.

Case 2 – A 78-year-old male presented to the ED with 1 month of painless jaundice. A CT of the abdomen demonstrated concerning intrahepatic biliary dilation. ERCP and cholangioscopy were performed and the AI correctly predicted this patient had malignant disease as he was later diagnosed with cholangiocarcinoma.

Case 3 –An 81-year-old male presented with fatigue and weight loss. CT scan demonstrated hyperenhancement of the common bile duct with upstream biliary dilation. Cholangioscopy was performed and the AI predicted benign disease in a patient later diagnosed with IgG4-sclerosing cholangitis.

Case 4 – A 79-year-old female with Stage IV ovarian cancer presented with jaundice. Imaging demonstrated the presence of a hilar stricture. Cholangioscopy was later performed and confirmed the presence of a metastatic nodule at the hilum that the AI correctly identified as cancer.

Endoscopic methods
Using commercially available materials we built a point-of-care computer system that could be used in the operating room during live cholangioscopy cases. The computer system directly receives the cholangioscopy video stream, processes the images, and outputs predictions as to whether malignancy is present over the course of the video. Over the course of a pilot and feasibility study, the AI has been more sensitive and more accurate for stricture classification than standard sampling modalities.

Conclusions
In this video we demonstrate that it is feasible to build a real time cholangioscopy AI system using commercially available materials. In this prospective deployment the AI continues to be more effective for stricture classification than sampling techniques. In the future, cholangioscopy AI may be included as part of the diagnostic assessment of biliary strictures.