AGA Clinical Practice Update on the Role of Artificial Intelligence in Colon Polyp Diagnosis and Management: Commentary

Abstract

DescriptionThe purpose of this American Gastroenterological Association (AGA) Institute Clinical Practice Update (CPU) is to review the available evidence and provide expert commentary on the current landscape of artificial intelligence in the evaluation and management of colorectal polyps.MethodsThis CPU was commissioned and approved by the AGA Institute Clinical Practice Updates Committee (CPUC) and the AGA Governing Board to provide timely guidance on a topic of high clinical importance to the AGA membership and underwent internal peer review by the CPUC and external peer review through standard procedures of Gastroenterology. This Expert Commentary incorporates important as well as recently published studies in this field, and it reflects the experiences of the authors who are experienced endoscopists with expertise in the field of artificial intelligence and colorectal polyps. The purpose of this American Gastroenterological Association (AGA) Institute Clinical Practice Update (CPU) is to review the available evidence and provide expert commentary on the current landscape of artificial intelligence in the evaluation and management of colorectal polyps. This CPU was commissioned and approved by the AGA Institute Clinical Practice Updates Committee (CPUC) and the AGA Governing Board to provide timely guidance on a topic of high clinical importance to the AGA membership and underwent internal peer review by the CPUC and external peer review through standard procedures of Gastroenterology. This Expert Commentary incorporates important as well as recently published studies in this field, and it reflects the experiences of the authors who are experienced endoscopists with expertise in the field of artificial intelligence and colorectal polyps. Colorectal cancer (CRC) is the second most common cause of cancer deaths worldwide.1Bray F. Ferlay J. Soerjomataram I. et al.Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries.CA Cancer J Clin. 2018; 68: 394-424Google Scholar Screening colonoscopy reduces the risk of CRC through the removal of precancerous polyps.2Brenner H. Stock C. Hoffmeister M. Effect of screening sigmoidoscopy and screening colonoscopy on colorectal cancer incidence and mortality: systematic review and meta-analysis of randomised controlled trials and observational studies.BMJ. 2014; 348: g2467Google Scholar Polyp detection is operator-dependent, with adenoma detection rates (ADRs) ranging widely from 7% to 53% among colonoscopists.3Corley D.A. Jensen C.D. Marks A.R. et al.Adenoma detection rate and risk of colorectal cancer and death.N Engl J Med. 2014; 370: 1298-1306Google Scholar Failure to detect and remove neoplastic polyps is associated with the development of post-colonoscopy interval CRC, which accounts for nearly 8% of all diagnosed CRC.4Cooper G.S. Xu F. Barnholtz Sloan J.S. et al.Prevalence and predictors of interval colorectal cancers in medicare beneficiaries.Cancer. 2012; 118: 3044-3052Google Scholar Conversely, a 1% increase in a colonoscopist’s ADR has been associated with a 3% decrease in future CRC risk.3Corley D.A. Jensen C.D. Marks A.R. et al.Adenoma detection rate and risk of colorectal cancer and death.N Engl J Med. 2014; 370: 1298-1306Google Scholar However, the majority of polyps detected during colonoscopy are diminutive and non-neoplastic.5Ponugoti P.L. Cummings O.W. Rex D.K. Risk of cancer in small and diminutive colorectal polyps.Dig Liver Dis. 2017; 49: 34-37Google Scholar Unnecessary resection and pathologic evaluation of these non-neoplastic lesions can be associated with increased costs and risk for adverse events. These critical issues around colon polyp detection and diagnosis have been a central focus for a range of artificial intelligence (AI) tools that have recently been introduced to the field of gastrointestinal endoscopy. 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. This commentary incorporates important and recently published studies in the field and elaborates on the future directions of AI in colonoscopy. The term artificial intelligence refers to computer systems performing complex tasks that would normally require the use of the human brain, such as visual perception (“computer vision”), speech recognition, and decision making.6Colom R. Karama S. Jung R.E. et al.Human intelligence and brain networks.Dialogues Clin Neurosci. 2010; 12: 489-501Google Scholar Early attempts at polyp detection required explicit programming of software to recognize certain polyp features (eg, textures and shapes).7Iakovidis D.K. Maroulis D.E. Karkanis S.A. An intelligent system for automatic detection of gastrointestinal adenomas in video endoscopy.Comput Biol Med. 2006; 36: 1084-1103Google Scholar These early efforts were focused on recognizing still images because computer-processing speed at that time could not support real-time, live video image analysis. Since then, major advances in deep-learning algorithms using convolutional neural networks have dramatically expanded the capabilities of computer vision for endoscopy. These contemporary algorithms are trained on large data sets and can adapt and “learn” to recognize complex objects in live video.8Tajbakhsh N. Gurudu S.R. Liang J. Automated polyp detection in colonoscopy videos using shape and context information.IEEE Trans Med Imaging. 2016; 35: 630-644Google Scholar The most important applications of AI computer vision in colonoscopy today include computer-aided detection (CADe) and computer-aided diagnosis (CADx). CADe is designed to help the endoscopist detect polyps during colonoscopy and CADx is intended to accurately predict polyp histology without the need for a tissue biopsy. AI detection of colorectal polyps was the first target for AI technology in gastroenterology and now a myriad of studies has reported the successful application of AI for the recognition of colon polyps using CADe. These algorithms are the equivalent of a highly trained set of eyes relentlessly scanning the monitor alongside the endoscopist, while simultaneously “flagging” lesions that potentially represent precancerous polyps (Figure 1). Urban and colleagues9Urban G. Tripathi P. Alkayali T. et al.Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy.Gastroenterology. 2018; 155: 1069-1078.e8Google Scholar reported one of the earliest applications of convolutional neural network–based CADe on video clips. Their algorithm showed 97% sensitivity, 95% specificity, and 96% accuracy for detection of colorectal polyps, which was superior to the performance of the endoscopists. Importantly, 92% of polyps missed by the endoscopists were detected by the CADe.9Urban G. Tripathi P. Alkayali T. et al.Deep learning localizes and identifies polyps in real time with 96% accuracy in screening colonoscopy.Gastroenterology. 2018; 155: 1069-1078.e8Google Scholar 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. A meta-analysis by Huang et al10Huang D. Shen J. Hong J. et al.Effect of artificial intelligence-aided colonoscopy for adenoma and polyp detection: a meta-analysis of randomized clinical trials.Int J Colorectal Dis. 2022; 37: 495-506Google Scholar of 10 randomized controlled trials with 6629 patients found that both ADR (relative risk [RR], 1.43; P < .001) and polyp detection rate (RR, 1.44; P < .001) were significantly greater with AI-aided colonoscopy compared with routine colonoscopy. The adenomas detected per colonoscopy (APC) and polyps detected per colonoscopy were also significantly higher in the AI-aided group compared with the routine colonoscopy group.10Huang D. Shen J. Hong J. et al.Effect of artificial intelligence-aided colonoscopy for adenoma and polyp detection: a meta-analysis of randomized clinical trials.Int J Colorectal Dis. 2022; 37: 495-506Google Scholar Despite these positive early results, recent studies have suggested that CADe may not improve adenoma detection in every clinical setting. In a retrospective single-center pragmatic trial, Ladabaum et al11Ladabaum U. Shepard J. Weng Y. et al.Computer-aided detection of polyps does not improve colonoscopist performance in a pragmatic implementation trial.Gastroenterology. 2023; 164: 481-483.e6Google Scholar reported that CADe did not improve ADR, APC, or other detection metrics when compared with historic and concurrent controls.11Ladabaum U. Shepard J. Weng Y. et al.Computer-aided detection of polyps does not improve colonoscopist performance in a pragmatic implementation trial.Gastroenterology. 2023; 164: 481-483.e6Google Scholar In a separate, large, retrospective, observational study, Levy et al12Levy I. Bruckmayer L. Klang E. et al.Artificial intelligence-aided colonoscopy does not increase adenoma detection rate in routine clinical practice.Am J Gastroenterol. 2022; 117: 1871-1873Google Scholar reported a lower ADR in the CADe group compared with a pre-CADe retrospective control (30.3% vs 35.2%; P = .001), as well as a lower polyp detection rate and lower APC.12Levy I. Bruckmayer L. Klang E. et al.Artificial intelligence-aided colonoscopy does not increase adenoma detection rate in routine clinical practice.Am J Gastroenterol. 2022; 117: 1871-1873Google Scholar In addition to these observational studies, a recent trial from the United Kingdom with patients randomized to CADe vs standard colonoscopy also failed to demonstrate a difference in ADR. There are several plausible explanations for the lack of benefit of CADe across these studies. For one, the possibility of a “ceiling” effect for polyp detection among high-performing endoscopists may have accounted for the lack of incremental benefit with CADe. In addition, there may be an unconscious behavioral change, for instance, degradation in the quality of mucosal exposure, possibly due to a false sense of security that CADe ensures a high-quality examination. Alternatively, we should also acknowledge that in the majority of published randomized controlled trials, the endoscopists are unblinded and it is possible that this could introduce performance bias, favoring CADe performance. Cost-effectiveness of AI-assisted colonoscopy also needs to be examined carefully, as there are several ways this technology may increase health expenditure. The adenoma detection improvements noted using CADe are attributed mainly to the increased detection of small, nonadvanced adenomas. Indeed, one could argue that the increased detection of small benign polyps by AI could inadvertently lead to more unnecessary resections, thereby increasing cost and procedural risks. Shaukat and colleagues' recent US multicenter, randomized, parallel study13Shaukat A. Lichtenstein D.R. Somers S.C. et al.Computer-aided detection improves adenomas per colonoscopy for screening and surveillance colonoscopy: a randomized trial.Gastroenterology. 2022; 163: 732-741Google Scholar at 5 academic and community centers by 22 US board-certified gastroenterologists helped address this concern. The primary end points were APC, total number of adenomas resected divided by the total number of colonoscopies; and true histology rate, defined as the proportion of resections with clinically significant histology divided by the total number of polyp resections. There were 677 patients in the standard colonoscopy group and 682 patients in the CADe group. APC was significantly higher in the CADe group (standard vs CADe, 0.83 vs 1.05) with no decrease in true histology with use of the CADe device. Overall, the results suggest that CADe may help improve APC without a concomitant (and potentially costly) increase in the resection of non-neoplastic lesions. Another mechanism by which CADe may lead to increased health expenditure is higher polyp detection leading to shorter colonoscopy surveillance intervals. One recent study by Mori and colleagues14Mori Y. Wang P. Loberg M. et al.Impact of artificial intelligence on colonoscopy surveillance after polyp removal: a pooled analysis of randomized trials.Clin Gastroenterol Hepatol. 2023; 21: 949-959.e2Google Scholar estimated that the use of AI during colonoscopy increased the proportion of patients requiring intensive colonoscopy surveillance by approximately 35% in the United States and 20% in Europe, with absolute increases of 2.9% and 1.3%, respectively. However, increases in ADR can increase early detection of colon cancer and save costs due to cancer management. Areia and colleagues15Areia M. Mori Y. Correale L. et al.Cost-effectiveness of artificial intelligence for screening colonoscopy: a modelling study.Lancet Digit Health. 2022; 4: e436-e444Google Scholar conducted a study using a Markov model microsimulation to investigate the effect of implementing AI-assisted colonoscopy on colon cancer incidence, mortality, and cost-effectiveness. Compared with no screening, the relative reduction of CRC incidence with screening colonoscopy without AI tools was 44.2% and for screening colonoscopy with AI tools was 48.9% (4.8% incremental gain). AI detection tools decreased the discounted costs per screened individual from $3400 to $3343 (a saving of $57 per individual). At the US population level, the implementation of AI detection during screening colonoscopy resulted in yearly additional prevention of 7194 CRC cases and 2089 related deaths, and a yearly saving of $290 million.15Areia M. Mori Y. Correale L. et al.Cost-effectiveness of artificial intelligence for screening colonoscopy: a modelling study.Lancet Digit Health. 2022; 4: e436-e444Google Scholar Naturally, however, the cost-effectiveness of this technology relies on the assumptions that ADR will increase from baseline in real-world clinical practice, an outcome that still needs to be corroborated in further studies. We are just beginning to understand more about AI–human interaction and physician attitudes toward AI in gastroenterology. A study from Nehme et al16Nehme F. Coronel E. Barringer D.A. et al.Performance and attitudes toward real-time computer-aided polyp detection during colonoscopy in a large tertiary referral center in the United States.Gastrointest Endosc. 2023; 98: 100-109.e6Google Scholar evaluated the real-world performance of AI-assisted colonoscopy using a commercially available CADe platform in a fully democratized fashion, in which the entire faculty of a large US academic institution was given the opportunity to perform AI-assisted colonoscopy. The decision to activate the CADe system was at the discretion of the endoscopist. An anonymous survey was circulated among endoscopy physicians and staff at the beginning and at the conclusion of the study period regarding their attitudes toward AI-assisted colonoscopy. CADe was activated in 52.1% of cases. Survey results demonstrated mixed attitudes toward AI-assisted colonoscopy, with some physicians expressing concerns regarding the number of false-positive signals, the possibility of distraction (eg, due to false positives or audio alerts), and the impression that it could prolong procedure time. In our own experience, the concern for false-positive alerts remains one of the largest concerns and possible deterrents to the incorporation of AI-assisted colonoscopy. This is an area where human–AI interaction comes to the fore. Although there have been assumptions that CADe could be a “plug and play” technology, we propose there may be a role for more formalized training to teach endoscopists how to use CADe most effectively, including best approaches to efficiently recognize and dismiss false positives. As with many new clinical technologies, there may be a physician learning curve for CADe. Optimizing this and other aspects of the physician–AI collaboration will be an important area of ongoing investigation. Finally, it is imperative for CADe developers to continue to optimize CADe system performance to balance the high sensitivity needed for avoidance of missed lesions, while limiting the noise from false-positive signals. For many years, expectations have grown that optical diagnosis of colorectal polyps would enable “resect-and-discard” or “diagnose-and-leave” strategies, reducing the need for polypectomy and/or histopathology processing for diminutive polyps. The American Society for Gastrointestinal Endoscopy created an in and of that optical of colorectal polyps in as a area for new endoscopic In to safely a for polyps in the by endoscopic technologies, should provide a for In to a for colorectal polyps the endoscopic technology to histology of these polyps should provide in of surveillance compared with on of all D.K. C. M. et American Society for Gastrointestinal Endoscopy and of on real-time endoscopic of the histology of diminutive colorectal Endosc. Scholar Overall, adenomas in the colon and polyps in the for of all polyps detected during screening and surveillance N. A. D. et al.Prevalence of features in diminutive and small colon Endosc. 2012; Scholar or were more in clinical practice, the could be with yearly estimated and per Y. et in colonoscopy with artificial intelligence-aided polyp an analysis of a clinical trial Endosc. C. Rex D.K. A and would improve cost-effectiveness of colorectal cancer Gastroenterol Hepatol. 2010; Scholar CADx tools may an important role in to perform optical diagnosis during colonoscopy by real-time for colon polyps (Figure Y. Y. et performance of real-time of colorectal lesions using an artificial a 2022; 163: Scholar Although current of optical polyp diagnosis on such as there is that CADx tools may be to perform this with CADe polyp detection with CADx for more approaches toward and polyps during colonoscopy. CADx clinical trials are to and results are and C. G. R. et al.Artificial intelligence colorectal Gastroenterol Hepatol. 2022; Scholar recently conducted a study using CADe and CADx algorithms during screening colonoscopy. The CADx with endoscopy and a convolutional neural that in real time on several images of the to a of adenoma or For CADx and endoscopist diagnosis with were the histology Overall, polyps were in of which were lesions CADx diagnosis was in of polyps, and the for lesions was There were of lesions that were for a in on CADx of would be to a in a CADx and surveillance to US This study important evidence that a real-time CADx system was to the for histology required for the in in standard endoscopy. In the study a would in a reduction of all of the and related with the addition of a for certain small polyps requiring this would have further the need for to of the lesions detected in the study, with a with An important of AI CADx algorithms is the possibility of endoscopists to perform optical diagnosis with a higher baseline of and A recent study by et I. P. S. et artificial optical diagnosis of neoplastic polyps during 2022; Scholar demonstrated that a this requiring a did not increase sensitivity, the of endoscopists in a diagnosis diagnosis for during standard visual vs when by endoscopist in optical diagnosis may be an important toward implementation of in and strategies, successful implementation will also require CADx tools that the endoscopic without the need for image or In addition, CADx will on in are an important for many gastroenterology and around the concerns by polyp in or resected polyps after adenomas during colonoscopy can from when the endoscopist to recognize adenomas that are on the or due to and mucosal that can be related to the speed during colonoscopy, endoscopist of and other CADe systems not address during colonoscopy. that colonoscopy is associated with a lower rate for adenomas and to higher quality to mucosal and of time D.K. is associated with adenoma Endosc. Scholar in computer-aided are to focus on to These tools can be defined as computer-aided systems and represent a group of computer-aided in colonoscopy with CADe and is a critical of or not polyps in the field of D. Y. L. et in Trans Med Imaging. Scholar Although may a critical role in AI systems to support quality control of are also now and Y. et al.Artificial of a new for and the quality of 2022; Scholar and an system for evaluation quality One and by were for the of colonoscopy, after which the were evaluated by the AI The system was by evaluation expert ADR, and time of The AI of the of endoscopist were significantly with expert ADR, and time. For by with AI significantly the evaluated by both the AI system and computer-aided quality include technology to the colon on a of the and systems that the of mucosal during S. et of artificial from live to optimize the quality of the colonoscopy in real of Scholar In the these of systems may support and endoscopists to more may physician and systems with more and data on an of that to colonoscopy CADe systems have commercially available in the United States and across the in the last several years, the early data a role for CADe in adenoma clinical has been Rex and D.K. Mori Y. intelligence improves detection at colonoscopy: we all using 2022; 163: Scholar have several there is a of polyp detection that have from the US and failed to include endoscopy and or that can on The of this entire that a to the of increasing detection by there is no for CADe or other and physicians are not on the that from a higher number of and future increased procedure to shorter surveillance the cost of these A second that may to CADe is that the technology is not into or video This is a from and all of which have been as of as these have The current of CADe as a separate, an additional of and cost that may for some endoscopy A in the CADe as is the distraction and that some endoscopists have reported with CADe false positives. This performance is by of CADe early may physicians from AI in the Overall, CADe and CADx diagnosis are just the first of a of AI tools that will support gastrointestinal endoscopic practice. The of AI for colonoscopy will toward computer-aided quality and AI of procedures These and other AI tools are of the and we predict an AI of tools for colonoscopy will as a to support and clinical practice. AI tools that improve colonoscopy quality may more and by and possibly by patients who to ensure the of their the applications of AI for colonoscopy, AI has also as a through which to into how perform with all our and As we can that the future for AI in endoscopy will be a where the capabilities of physicians and our AI tools will be to optimize

Topics & Concepts