Computer-aided polyp detection (CADe) in real life: not the “CADe-llac” we were promised

Colorectal cancer (CRC) is the second leading cause of cancer-related deaths in the United States and is expected to cause 52,550 deaths in 2023.1Siegel R.L. Miller K.D. Goding Sauer A. et al.Colorectal cancer statistics, 2020.CA Cancer J Clin. 2020; 70: 145-164Crossref PubMed Scopus (2671) Google Scholar Colonoscopy allows early detection of CRC and prevention via resection of premalignant lesions. The adenoma detection rate (ADR) is an established colonoscopy quality measure, inversely correlating with interval CRC rates.2Corley 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: 1298-1306Crossref PubMed Scopus (1259) Google Scholar Given its importance, achieving sustainable improvements in ADR has been the topic of numerous studies. Efforts have ranged from innovative yet impractical (eg, the Third Eye Retroscope [Avantis Medical Systems, Sunnyvale, Calif, USA], which allows an additional retrograde view during colonoscopy),3DeMarco D.C. Odstrcil E. Lara L.F. et al.Impact of experience with a retrograde-viewing device on adenoma detection rates and withdrawal times during colonoscopy: the Third Eye Retroscope study group.Gastrointest Endosc. 2010; 71: 542-550Abstract Full Text Full Text PDF PubMed Scopus (104) Google Scholar to simple yet effective (minimum withdrawal times4Barclay R.L. Vicari J.J. Greenlaw R.L. Effect of a time-dependent colonoscopic withdrawal protocol on adenoma detection during screening colonoscopy.Clin Gastroenterol Hepatol. 2008; 6: 1091-1098Abstract Full Text Full Text PDF PubMed Scopus (217) Google Scholar and high-definition endoscopy). Computer-aided polyp detection (CADe) is the “newest kid on the block.” Those who have experienced CADe firsthand can attest to its novelty, with the sheer “wow” factor leading to considerable excitement about this technology. Outside of its novelty, several randomized clinical trials (RCTs) have now shown that CADe platforms can significantly decrease adenoma miss rates and increase both ADR and adenomas per colonoscopy (APC) rates. A study by Shaukat et al5Shaukat 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: 732-741Abstract Full Text Full Text PDF PubMed Scopus (16) Google Scholar at 5 academic and community centers with 1440 subjects randomized to CADe versus standard colonoscopy found a significantly higher APC rate with CADe (1.05 vs 0.83, P = .002). Another randomized study by Wang et al6Wang P. Liu X. Berzin T.M. et al.Effect of a deep-learning computer-aided detection system on adenoma detection during colonoscopy (CADe-DB trial): a double-blind randomised study.Lancet Gastroenterol Hepatol. 2020; 5: 343-351Abstract Full Text Full Text PDF PubMed Scopus (215) Google Scholar in China in 1010 patients found a significantly higher ADR in the CADe group than in those undergoing standard colonoscopy (34% vs 28%, P = .030). Finally, a randomized study of 685 patients by Repici et al7Repici A. Badalamenti M. Maselli R. et al.Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial.Gastroenterology. 2020; 159: 512-520.e7Abstract Full Text Full Text PDF PubMed Scopus (252) Google Scholar using the GI Genius (Medtronic) platform found a significantly higher ADR in the CADe group than in those undergoing standard colonoscopy (54.8% vs 40.4%) as well as a higher APC in the CADe group (1.07 vs 0.71; incidence rate ratio 1.46; 95% CI, 1.15-1.86). However, the increase in ADR and APC in several RCTs seems to be mostly attributable to increased detection of diminutive adenomas,7Repici A. Badalamenti M. Maselli R. et al.Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial.Gastroenterology. 2020; 159: 512-520.e7Abstract Full Text Full Text PDF PubMed Scopus (252) Google Scholar and it is unclear whether this translates to protection from postcolonoscopy colon cancer. In addition, there are practical concerns that CADe may lead to increased resection of nonneoplastic tissue (hyperplastic polyps), although this has not been clearly identified in RCTs.5Shaukat 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: 732-741Abstract Full Text Full Text PDF PubMed Scopus (16) Google Scholar,7Repici A. Badalamenti M. Maselli R. et al.Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial.Gastroenterology. 2020; 159: 512-520.e7Abstract Full Text Full Text PDF PubMed Scopus (252) Google Scholar In this issue of Gastrointestinal Endoscopy, Nehme et al8Nehme 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.e6Abstract Full Text Full Text PDF Scopus (3) Google Scholar report a large real-world experience with artificial intelligence (AI)-assisted colonoscopy in the United States. The authors also report the findings from a preintervention and postintervention survey they conducted for participating endoscopists. They retrospectively compared ADR and APC for the same set of endoscopists 6 months before and 6 months after the implementation of a commercially available CADe platform (GI Genius, Medtronic, Minneapolis, Minn, USA) at a tertiary referral center. Implementation in their center was “democratic”—activation of the CADe platform was at the discretion of the attending endoscopist. CADe was activated in only 52% of procedures in the intervention period, and their primary analysis comparing the historical with the intervention period included cases in which CADe was not activated. The authors found no improvement in APC from the historical control period (935 colonoscopies) compared with the intervention period (1041 colonoscopies) (1.08 vs 1.04, P = .65). Somewhat surprisingly, they found a small but statistically significant decrease in ADR with the use of CADe (48.2% vs 43.6%, P = .04). When the authors limited their analysis to screening and surveillance colonoscopies where CADe was activated, the results were similarly disappointing, with no improvement in either ADR or APC. Stratifying endoscopists by their baseline ADR (less than or higher than 45%) yielded similar results. Interestingly, when CADe was activated, endoscopists with <10 years of experience resected a higher number of hyperplastic polyps and normal colonic mucosa. Finally, endoscopists with baseline ADR >45% were also more likely to activate the CADe system as opposed to those with a lower baseline ADR. As for the survey portion of the study, before the use of CADe, most (87.5%) physicians had a positive attitude toward AI. After intervention, however, most physicians (82.4%) had concerns that the system resulted in too many false-positive signals and was too distracting (58.8%). However, many maintained their excitement, with 64.7% of physicians supporting a future role for AI-assisted colonoscopy. These results should be interpreted in the context of the study design. This was a retrospective, single-center, non-randomized study using a historical control group involving endoscopists with relatively high baseline ADR, in which the actual use of CADe during the intervention period was low. Therefore, the study may have been underpowered to detect differences in APC. Regardless, the results showing a lack of obvious benefit of CADe in terms of APC and ADR are somewhat sobering and may represent a more realistic picture of the utility of CADe in practice. Where do these findings fit in the broader literature? A recent real-life study by Levy et al9Levy 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: 1871-1873Crossref PubMed Scopus (11) Google Scholar at a large referral center in Israel included 4414 colonoscopies and had similar—perhaps even more striking—findings: a significant decrease in ADR after CADe implementation from 35.2% to 30.3%. One shared and concerning trend emerging from both real-world analyses is the small but significant decrease in ADR after the implementation of CADe. One hypothesis could be that endoscopists are spending additional time resecting smaller (even hyperplastic) lesions they are alerted to by CADe. Owing to concerns about the possibility of increased procedure time, they subsequently may perform a less careful withdrawal to stay on schedule. The findings presented by Nehme et al8Nehme 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.e6Abstract Full Text Full Text PDF Scopus (3) Google Scholar have clues to support this hypothesis: in their postintervention survey, 47% of endoscopists thought that the system prolonged procedure time; yet, there was no actual difference in total procedure time between the preintervention and postintervention periods, despite the decrease in ADR after intervention. Another hypothesis may be an over-reliance on CADe, which may in turn lead to less careful inspection behind folds and flexures during withdrawal. More broadly, a few points may explain the discrepancy between real-world and RCT findings. First, the “democratic” application of CADe introduces a selection bias and likely erodes into the ADR improvements CADe has yielded in RCTs. However, to the authors’ credit, a “democratic” approach is how many practices will likely choose to deploy CADe, at least initially. Second, the study endoscopists had a remarkably high ADR at baseline (48.2%), significantly higher than the minimum recommended ADR by the U.S. Multi-Society Task Force on CRC. Increases in ADR beyond this threshold are difficult, regardless of the modality used. To date, there is no evidence that driving up the ADR further by resecting small adenomas, most of which are unlikely to develop into advanced neoplasia, translates to a reduction in the risk of CRC. Third, human factors that are often dampened in carefully designed and monitored RCTs will arise in real life. For example, endoscopists can either miss or choose to ignore the alerts issued by the CADe system, unlike in a carefully monitored trial. Even when an alert is acknowledged, an endoscopist in a clinical trial is more likely to spend additional time carefully examining the area of interest and perhaps ultimately performing polypectomy. So where do we go from here? To answer the question of real-life effectiveness most accurately, a large, pragmatic, cluster-randomized trial in the community is needed. Community-based ambulatory surgical centers can be randomized to CADe-assisted or conventional colonoscopy. Such a study can provide the most valuable real-life insight into this technology. In an ideal setting, key outcomes such as CRC incidence and related mortality as well as postcolonoscopy CRC rates would be identified, rather than surrogate markers such as APC or ADR, although such a trial may be impractical and would take several years to accrue enough outcomes. In the interim, research and development in computer-aided diagnosis (CADx) and refinements in existing CADe platforms should continue. If one lesson is to be learned from the current study, it is that excitement about AI should not take our collective focus away from the primary goal of early detection and prevention of CRC. Simple, yet effective measures will continue to remain key for these efforts, including adherence to minimum withdrawal times and adequate bowel preparation regimens, as well as improving endoscopist quality overall, beyond the ability to resect small adenomas detected by CADe. Future AI tools such as those assessing the adequacy of visualization of the colon and more objectively assessing preparation quality during withdrawal are promising and may further bolster overall endoscopy quality. However, no matter where you stand on the topic of AI, a quick glance at the contents page of any recent leading GI journal sends a clear message: AI is here to stay. Both authors disclosed no financial relationships.