Introduction
In clinical practice, colonoscopy withdrawal time (CWT) is measured from the caecum to the anal verge without accounting for cleaning (washing or suctioning) and polypectomy events during withdrawal. The evidence for the correlation of CWT and adenoma detection rate (ADR) is somewhat conflicting, partially due to research relying on retrospective data, which does not account for cleaning and polypectomy events and is subject to inaccuracies in time measurement.
A randomized controlled trial recently demonstrated that ADR increases as colonoscopy inspection time (CIT) rise from 6 to 13 minutes 1. CIT was defined as withdrawal from the caecum to the anal verge with the timer manually stopped during cleaning and polypectomy events, but CIT measurement cannot be implemented in clinical practice without dedicated staff to pause a timer during these events.
Aims & Methods
Recent years have demonstrated the ability of artificial intelligence (AI) models to measure CWT by detecting cecal landmarks and the anal verge. Our aim was to develop convolutional neural networks (CNN) to detect phases of cleaning and polypectomy during withdrawal to facilitate automated measurement of CIT.
One hundred endoscopy videos were prospectively collected for training data. We reviewed each individual frame during withdrawal (defined as frames after a cecal landmark was first visualized). Each frame during ‘cleaning’ and ‘polypectomy’ events were annotated with respective image-level labels, and the remaining withdrawal frames labelled as ‘inspection’. Cleaning events was defined as frames during suctioning of intra-colonic content or washing. Polypectomy events was defined as the first frame with visualisation of the injection needle or forceps/snare if the polyp was not lifted until completion of inspection of post-polypectomy resection margins. These annotations were referenced as ground truth.
From the 100 videos, two ResNet-101 CNNs were developed to detect cleaning (213,936 training frames) and polypectomy (56,273 training frames). The CNNs were then evaluated with 80 colonoscopy videos. This included 40 colonoscopy videos from five expert (ADR >45%) and 40 videos from six non-expert (ADR <30%) endoscopists. Consecutive procedures were used to minimize selection bias. Videos were annotated following the above methodology.
Results
For the test set of 80 procedures, the median WT documented in the endoscopy reporting system (EPIC) was 15.0 minutes (IQR 10.0–23.0), annotated ground truth median CIT was 7.1 minutes (IQR 5.3–9.6), and AI predicted median CIT was 6.9 minutes (IQR 5.1–9.0) with a correlation coefficient (r) of 98.2%. For expert procedures (n=40), these were 16.5 minutes (IQR 11.0–25.3), 7.1 minutes (IQR 5.1–8.6) and 6.9 minutes (IQR 4.9–8.8), respectively. Amongst non-expert procedures (n=40), these were 14.0 minutes (IQR 10.0–22.0), 7.0 minutes (IQR 5.8–10.7) and 7.0 minutes (IQR 5.3–10.2), respectively. The AI system correctly categorized 96.3% (77/80) of procedures as less or more than 6 minutes CIT.
Conclusion
We have demonstrated the feasibility of AI to differentiate the phases of withdrawal to facilitate automated measurement of CIT and the considerable difference in time between CIT and the CWT documented in an endoscopy reporting system.
References
1. Desai M, Rex DK, Bohm ME, et al. Impact of withdrawal time on adenoma detection rate: results from a prospective, multi-center trial. Gastrointest Endosc. 2022 Oct 10:S0016-5107(22)02039-9
Disclosure
Professor Laurence B Lovat is a minority shareholder and consultant advisor for Odin Vision. Professor Danail Stoyanov is a shareholder for Odin Vision.