The American Gastroenterological Association has published a Clinical Practice Update (CPU) on artificial intelligence (AI) for diagnosing and managing colorectal polyps.
The CPU, authored by Jason Samarasena, MD, of UCI Health, Orange, Calif., and colleagues, draws on recent studies and clinical experience to discuss ways that AI is already reshaping colonoscopy, and what opportunities may lie ahead.
Dr. Jason Samarasena
“As with any emerging technology, there are important questions and challenges that need to be addressed to ensure that AI tools are introduced safely and effectively into clinical endoscopic practice, ”they wrote in Gastroenterology.
With advances in processing speed and deep-learning technology, AI “computer vision” can now analyze live video of a colonoscopy in progress, enabling computer-aided detection (CADe) and computer-aided diagnosis (CADx), which the panelists described as the two most important developments in the area.
CADe
“In the last several years, numerous prospective, multicenter studies have found that real-time use of AI CADe tools during colonoscopy leads to improvements in adenoma detection and other related performance metrics,” Dr. Samarasena and colleagues wrote.
CADe has yielded mixed success in real-world practice, however, with some studies reporting worse detection metrics after implementing the new technology. Dr. Samarasena and colleagues offered a variety of possible explanations for these findings, including a “ceiling effect” among highly adept endoscopists, reduced operator vigilance caused by false confidence in the technology, and potential confounding inherent to unblinded trials.
CADe may also increase health care costs and burden, they suggested, as the technology tends to catch small benign polyps, prompting unnecessary resections and shortened colonoscopy surveillance intervals.
CADx
The above, unintended consequences of CADe may be counteracted by CADx, which uses computer vision to predict which lesions have benign histology, enabling “resect-and discard” or “diagnose-and-leave” strategies.
Such approaches could significantly reduce rates of polypectomy and/or histopathology, saving an estimated $33 million–150 million per year, according to the update.
Results of real-time CADx clinical trials have been “encouraging,” Dr. Samarasena and colleagues wrote, noting that emerging technology–compatible white-light endoscopy can achieve a negative predictive value of almost 98% for lesions less than 5 mm in diameter, potentially reducing polypectomy rate by almost half.
“Increasing endoscopist confidence in optical diagnosis may be an important step toward broader implementation of leave in situ and resect-and-discard strategies, but successful implementation will also require CADx tools that seamlessly integrate the endoscopic work flow, without the need for image enhancement or magnification,” the panelists wrote.
Reimbursement models may also need to be reworked, they suggested, as many GI practices depend on a steady stream of revenue from pathology services.
Computer-aided quality assessment systems
Beyond optical detection and diagnosis, AI tools are also being developed to improve colonoscopy technique.
Investigators are studying quality assessment systems that use AI offer feedback on a range of endoscopist skills, including colonic-fold evaluation, level of mucosal exposure, and withdrawal time, the latter of which is visualized by a “speedometer” that “paints” the mucosa with “a graphical representation of the colon.”
“In the future, these types of AI-based systems may support trainees and lower-performing endoscopists to reduce exposure errors and, more broadly, may empower physician practices and hospital systems with more nuanced and actionable data on an array of factors that contribute to colonoscopy quality,” the panelists wrote.
