Artificial Intelligence Model for the Characterization of Biliary Strictures During Real-Time Digital Cholangioscopy: A Pilot Study

Aims We aimed to develop an artificial intelligence model for a real-time evaluation during digital cholangioscopy.

Methods A single-center, pilot study. We collected 23 digital cholangioscopy videos for the training of the AI models using automated machine learning (AI Works, MD Consulting group, Ecuador). Three parameters were trained by two expert endoscopists. The AI classifies cholangioscopy findings as normal aspect, inflammatory aspect, and suggestive of malignancy. Biliary strictures’ final diagnosis was based on cholangioscopy visual impression, intraductal biopsy, and 6-months follow-up outcomes.

Results A total of 1903 samples (1714 training and 189 testings) were used to train the AI. The automated learning process took 75 hours (2000 badges per parameter). The developed model reached a mean average precision (mAP) of 94.64 %. The developed model had a total loss of 0.1988. The F1-score (harmonic mean of sensitivity and precision) of 92 %. The average IoU (overall between model prediction and expert marking) was 81.63 %. For real-time detection, the processing speed of the model using an IntelCoreI7 processor with 16 Gb of RAM was 5-7 FPS (frames per second). Two expert endoscopists tested the AI with twenty videos and twenty-live cholangioscopy procedures from histology-confirmed cholangiocarcinoma patients. The AI model accurately detected the malignancy pattern in all cases.

Conclusions The proposed AI model accurately recognizes and classifies biliary strictures during recorded videos and on real-time digital cholangioscopy procedures. Future evaluations of AI cholangioscopy are necessary to confirm these results.

Citation: Robles-Medranda C, Oleas R, Alcivar-Vasquez J et al. OP53 ARTIFICIAL INTELLIGENCE MODEL FOR THE CHARACTERIZATION OF BILIARY STRICTURES DURING REAL-TIME DIGITAL CHOLANGIOSCOPY: A PILOT STUDY. Endoscopy 2021; 53: S24.

Publication History

Article published online:
19 March 2021