ORLANDO — The use of artificial intelligence during colonoscopy appears to boost the accuracy and sensitivity of adenoma detection, even for lesions as small as 5 mm, according to recent studies.
The goal of integrating this technology into routine colonoscopy is to get the adenoma detection rate closer to the 50% reported in studies of screening-age populations, said William Karnes, MD, from the University of California, Irvine.
"Adenoma detection rates among different colonoscopists vary between 7% and 53%. The adenoma detection rate should equal prevalence," he told Medscape Medical News.
Convolutional neural network technology was able to distinguish between images with and without polyps with 96% accuracy in a study of 9000 screening colonoscopy images, he reported at the World Congress of Gastroenterology 2017.
The machine-learning system can read up to 170 images per second, making it "easily applied to live video," said Dr Karnes when he presented the research, which won the 2017 Presidential Poster Award for Colorectal Cancer Prevention.
"Artificial intelligence for polyp detection has the potential to help all colonoscopists achieve detection rates closer to true prevalence, and to further reduce the risk of interval colorectal cancers," he added.
Another computer-aided system, known as EndoBRAIN artificial intelligence, is easy to use and provides an automatic diagnosis at the push of a button, reported Yuichi Mori, MD, from Showa University in Shinagawa-ku, Japan.
"Recognizing a lesion as neoplastic is very useful information during a colonoscopy," Dr Mori emphasized at the recent United European Gastroenterology (UEG) Week 2017. With EndoBRAIN, lesions identified as neoplastic can be considered as such "with very high confidence."
In their prospective evaluation, Dr Mori and his colleagues found that the sensitivity of EndoBRAIN computer-aided endocytoscopy for the detection of neoplastic polyps exceeded 90%. And in a subanalysis of rectosigmoid colon polyps smaller than 5 mm, the negative predictive value was 99%, which meets the criteria for optical biopsies of diminutive polyps.
In the intent-to-treat analysis of 726 patients undergoing computer-assisted colonoscopy, 306 small polyps were detected and evaluated with EndoBRAIN. The primary outcome was the accuracy of EndoBRAIN when lesions were enhanced with methylene blue staining. A secondary outcome was the accuracy when lesions were enhanced with narrow-band imaging.
Table. Primary and Secondary Outcomes With the EndoBRAIN System
|Accuracy||Methylene Blue Staining, %||Narrow-Band Imaging, %|
Images were produced faster with the narrow-band imaging mode than without (24 vs 88 seconds), so that "might be the optimal method for the management of diminutive polyps," Dr Mori said.
Potential limitations of the study include some bias in the selection of target lesions and the single-center setting. "We are now doing a multicenter clinical trial in cooperation with the National Cancer Center in Japan," he reported.
Dr Mori was awarded €10,000 for the study by the UEG leadership, which he is required to put toward future gastroenterology research.
If EndoBRAIN continues to improve the ability of gastroenterologists to distinguish neoplastic from non-neoplastic lesions during colonoscopy, it could ultimately save money by reducing the need for pathology evaluation, said Dr Mori.
Gastroenterologists in Pathology
In previous studies, he and colleagues showed that human eyes can detect neoplasticity on endocytoscopic imaging, and that performance is "excellent" in expert hands, which means detection is limited by available expertise. The EndoBRAIN system could expand the use of endocytoscopy, and could even be used to shorten the learning curve for trainees, he pointed out.
EndoBRAIN currently costs about $50,000, but if and when the system becomes commercially available outside of Japan, the price will likely drop, he told Medscape Medical News.
Another recent study looked at the computer-assisted classification of colorectal polyps 5 mm and smaller (Gastroenterology. Published online October 14, 2017).
The team of researchers from Taiwan found that the classification of polyps as neoplastic or hyperplastic was 90.1% accurate using computer-aided diagnosis with a deep neural network.
Accuracies of 90.5% and 87.0% were achieved by two endoscopists with at least 5 years of experience. Accuracies ranging from 88.0% to 80.3% were achieved by four novice endoscopists who each had only 1 year of experience.
Readings by two of the less-experienced endoscopists were significantly less accurate than the computer-aided analysis.
Dr Mori's study was supported by a research grant from the Japan Society for the Promotion of Science. Dr Mori and Dr Karnes have disclosed no relevant financial relationships.
World Congress of Gastroenterology at ACG 2017: Abstract P1032, presented October 16, 2017; and United European Gastroenterology (UEG) Week 2017: Abstract OP001, presented October 30, 2017.
Follow Medscape on Twitter @Medscape and Damian McNamara @MedReporter
Medscape Medical News © 2017 WebMD, LLC
Send comments and news tips to firstname.lastname@example.org.
Cite this: Artificial Intelligence Advances Colonoscopy - Medscape - Dec 04, 2017.