This article is more than 5 years old. Information may no longer be current.
Computer-assisted polyp detection similar to expert colonoscopists
Click Here to Manage Email Alerts
WASHINGTON — A machine learning model successfully identified all polyps in colonoscopies previously reviewed by expert colonoscopists and even found a few more, according to research presented at Digestive Disease Week.
Priyam Vada Tripathi, MD, MPH, of the division of gastroenterology and hepatology at the University of California Irvine, said that colorectal cancer rates are directly linked to adenoma detection rate, with every 1% increase in ADR associated with a 3% reduction in interval cancer rate.
Tripathi said other studies have shown that colonoscopists in the top quintile of ADR are associated with an 82% reduction in interval cancers. “Thus, efforts are being made to improve polyp detection and close this gap between ADR and true adenoma prevalence,” she said in her presentation.
Tripathi and colleagues built their machine learning model by training a convolutional neural network on 8,641 images of polyps and normal colon. They achieved 96% accuracy, as well as a processing speed that allowed for its use on live video.
Investigators performed two video validation studies that assessed the effect of machine learning on expert polyp detection and determined sensitivity and specificity for the machine learning model.
Four expert colonoscopists with ADR of at least 50% were asked to review nine colonoscopy videos and identify all the polyps by consensus. The videos were then reviewed by a senior colonoscopy expert with assistance from the machine learning model. After the model detected a polyp, the senior expert assigned uniqueness, as well as a confidence level that a true polyp had been detected.
The four-expert panel identified 36 polyps in their review, while the senior expert identified 45 polyps with help from the machine learning model. The model did not miss any unique polyps identified by the expert reviewers, according to Tripathi.
Investigators determined that the sensitivity of the machine learning model was 0.98 and the specificity was 0.93 (P < .00001).
Tripathi said future randomized trials will help them determine if the model will encourage more careful examination and detection of additional polyps.
“Our artificial intelligence machine learning model is well-positioned for validation in prospective clinical trials during live colonoscopies,” she said. – by Alex Young
Reference:
Tripathi PV, et al. Poster 133. Presented at: Digestive Disease Week; June 2-5, 2018; Washington, D.C.
Disclosures: Tripathi reports no relevant financial disclosures. Please see the DDW faculty disclosure index for a list of all other authors’ relevant financial disclosures.
Perspective
Back to Top
Carol Burke, MD
No less than 10 presentations on artificial intelligence (AI) were shared at DDW this year. The utility of AI was demonstrated in a plethora of gastrointestinal disease states including accurately diagnosing non-alcoholic steatohepatitis and malignancies in the esophagus, stomach and within pancreatic cysts. Many reports of AI for the detection of colorectal polyps were also presented. AI is an exciting and quickly evolving technology that will be part of our armamentarium during endoscopy. The accuracy of AI to detect colorectal polyps is higher than detection by expert colonoscopists. This study by Tripathi and colleagues on machine learning to detect polyps on colonoscopy videos demonstrated high sensitivity and specificity and machine learning discovered additional polyps not seen by expert endoscopists.
A presentation by Byrne, et al, at DDW presented an AI model which incorporates a continuous detection mode triggering either white light or narrow band imaging without the need to switch the signal from the endoscopy tower or the endoscopist. The tool runs in real-time during the colonoscopy and has shown to discriminate the pathology of polyps including type 1 and type 2 polyps on narrow band imaging. Importantly, polyps were reported to be detected on average 403 milliseconds after their first appearance. The promise of AI is a substantial decrease in missed colorectal lesions, an increase in ADR and enhanced effectiveness of colonoscopy in the prevention of colorectal cancer. Many more uses of AI in endoscopy will be coming.
Click Here to Manage Email Alerts