Deep Learning for Automated Percutaneous Cholangioscopy Video Review

Wednesday, April 2, 2025

10:57 AM – 11:06 AM CT

Location: Music City Center, 202 A

Presenting Author(s)

Christine Yoon (she/her/hers)

Medical Student
Albert Einstein College of Medicine, United States

Author/Co-author(s)

AR

Anna Romanowska-Pawliczek, PhD, MS

Deputy Director of Research and Industry Relations
Stanford Medical Center, Department of Rad/Interventional Radiology, United States
AV

Alexander Vezeridis, MD, PhD

Assistant Professor, Interventional Radiology
Stanford University Medical Center, United States

Purpose: Percutaneous cholangioscopy (PCS) allows direct visualization of the biliary tree and the recording of videos for review. These videos can be several hours in duration, making manual review both labor-intensive and prone to error. We introduce a novel AI-based tool designed to automate PCS video review, focusing on improving documentation, education, and clinical efficiency.

Materials and Methods: Fourteen video recordings of Spyglass Discover system-assisted cholangioscopies were used in this study. Videos were converted into arrays of keyframes, with initial filtering to remove black, white, and blurry frames. A convolutional autoencoder neural network was employed for feature extraction, allowing for the grouping of similar images into continuous scenes. From each scene, a representative keyframe was selected. Additionally, a convolutional neural network (CNN) model was trained to identify pathologies such as gallstones, blood clots, suspicious bile duct mucosa, and various tools or debris. These elements were integrated into an interactive software tool designed for clinical application.

Results: The video summarization tool achieved a mean squared error (MSE) of 0.0004 and a peak signal-to-noise ratio (PSNR) of 35.3874 for feature extraction. The blurriness classification model demonstrated an accuracy of 77.1%, outperforming traditional methods such as the variance of Laplacian (59.4%) and the Canny Edge Detector (62.2%). Keyframe extraction achieved a precision of 1.0 across all tested videos, with compression ratios ranging from 0.02 to 0.09. The lesion detection model, based on YOLOv8, demonstrated high precision and recall, with a mean average precision (mAP) of 98%. The use of this tool reduced the time required for video review by an average of 46%.

Conclusion: This AI tool for PCS video review presents substantial potential for improving clinical documentation, enhancing educational resources, and optimizing workflow efficiency. The tool demonstrated high accuracy in keyframe extraction, significant time savings in video review, and effective lesion detection, suggesting its utility in clinical practice. Future steps include further model refinement, validation in broader clinical settings, and the development of a web-based platform for easier accessibility.