S0259 Automated Polypectomy Time During Live Colonoscopy Using Convolutional Neural Networks - A Video Validation Study

Abstract

INTRODUCTION: Numerous factors have been shown to influence total colonoscopy time such as patient characteristics, bowel preparation, and the skills of the endoscopist. Awareness of polypectomy time (PT) can be a useful tool to improve efficiency and more accurately document inspection time (withdrawal time minus polypectomy time). Manually recording PT is time-consuming and impractical. Our objective was to accurately measure PT at the point of care utilizing established convolutional neural networks (CNNs). METHODS: We previously reported the development and validation of CNNs for polyp detection (>99% accuracy, AUC 0.995)1 and tool detection ( >99% accuracy, AUC 0.999)2. These algorithms are run live during colonoscopy while recording video and time-stamped CNN prediction CSV files. A total of 118 consecutive CNN-assisted colonoscopies (Jan 2019 to April 2019) containing 184 polypectomies were analyzed to compare CNN predictions (CNN-PT) with a blinded expert video review (Expert-PT), regarded as the gold standard. The CNN-PT was defined by the first simultaneous predictions of a polyp and a device (biopsy forceps or snare) and the last prediction of a device. To account for repeated biopsies, piecemeal resection, or temporarily lost views during a single polyp encounter, a 6-second gap in CNN predictions of the device was allowed before resetting to a new CNN-PT encounter. Expert-PT was defined by the first and last appearance of a device(s) utilized to complete each polypectomy. The expert reviewer was blinded to the CNN-PT. Each Expert-PT was linked to its paired CNN-PT by timestamp. RESULTS: Compared to Expert-PTs, CNN-PTs were very accurate for polypectomy time (PT) (R2 = 0.935). The mean difference in polypectomy time between the expert video review and CNN predictions was 0.16 seconds. In 117 of 184 cases, there was no absolute difference between PTs predicted by CNNs versus expert video review. Outliers included cases of lost and refound polyps, poor preps and spasm.Figure 1.: Scatter plot comparing CNN-predicted polypectomy times to the expert video-validated review.Figure 2.: Histogram displaying the difference in seconds between expert review video-validated review and CNN-predicted polypectomy times.CONCLUSION: We have demonstrated the feasibility of using CNNs live during colonoscopy to automatically and accurately record PT. Such data could be used to monitor colonoscopy trainees for polypectomy efficiency, assess relative efficiencies of various polypectomy techniques and devices, and provide standardized measures of inspection time during withdrawal when a device is not being used.