Will computer-aided detection of polyps decrease colorectal cancer incidence and mortality?

Artificial intelligence (AI) is arriving with great fanfare across all human endeavors. Some applications will fulfill their promises, and some will disappoint. It remains to be seen whether computer-aided detection (CADe) of polyps will deliver decreases in colorectal cancer (CRC) incidence and mortality.

The case for CADe seems straightforward: removal of CRC precursors (best reflected by the adenoma detection rate [ADR]) decreases CRC incidence and mortality; humans miss things, and despite years of attention to colonoscopy quality, substantial interoperator ADR variability persists; computers can overcome human limitations; therefore, computers should solve our nagging “low detector” problem – with eventual benefits in CRC control.

Merke

“Beyond a certain ADR level (is it 35%–40%?), the impact on post-colonoscopy CRC may be minimal. The urgent practical question is whether “lower detectors” (ADR <30%–35%) can be helped by CADe.”

Initial randomized controlled trials (RCTs) reported impressive improvements in ADR and adenomas per colonoscopy with CADe, as referenced by Maas et al. [11], with improvements attributable to detection of nonadvanced lesions. These studies were followed by some RCTs with less impressive results, and by pragmatic trials that found little or no benefit from CADe [22]. In this issue of Endoscopy, Maas et al. report results from an international, multicenter RCT of CADe vs. conventional colonoscopy [11] in this rapidly moving context. I agree with the authors that, even though this is the first reported study using the DISCOVERY CADe system (PENTAX Medical, Tokyo, Japan), the results are probably reflective of current CADe systems in general.

In this RCT involving 14 experienced endoscopists from 7 centers in 6 countries, and including screening, surveillance, and diagnostic colonoscopies, there were no differences between CADe and conventional colonoscopy in ADR (38.4% vs. 37.7%; P = 0.43) or adenomas per colonoscopy (0.66 vs. 0.66; P = 0.97), but detection of sessile serrated lesions [SSL] per colonoscopy was better with CADe (0.30 vs. 0.19; P = 0.049) [11].

We should remind ourselves of the primary problem we are trying to solve: the persistence of suboptimal quality in some colonoscopies, which is largely reflected by a colonoscopist’s ADR. Maas et al. suggest that their study may have lacked power because their sample size calculations assumed a lower ADR with conventional colonoscopy and a much larger CADe effect than observed – but, in my view, the issue is not power (a very large study that detects a trivial clinical difference would not make the case for CADe), but rather a possible ceiling effect. What should we expect from CADe if colonoscopists are already performing well? Even if “high detectors” marginally improve their ADRs or adenomas per colonoscopy by detection of diminutive adenomas, will that decrease CRC incidence and mortality? Beyond a certain ADR level (is it 35%–40%? [33]), the impact on post-colonoscopy CRC may be minimal. The urgent practical question is whether “lower detectors” (ADR <30%–35%) can be helped by CADe.

Maas et al. analyze post hoc their results by ADR tertile. The ADR tertiles were based on ADRs in the conventional colonoscopy arm data (247 colonoscopies by 14 endoscopists, with a range of 1–43/colonoscopist), and not a baseline audit with larger sample sizes. Thus, substantial random variability in individual ADR is expected, and the assignment to low (ADR <25%), intermediate, and high tertiles was probably not reliable. In fact, Fig. 1s is consistent with classic regression to the mean. It is not clear, therefore, whether the study included any colonoscopists with a truly suboptimal ADR who stood to improve meaningfully with CADe.

In our pragmatic trial, we were disappointed that colonoscopists in our low baseline ADR tertile (i.e. those with the need to improve), and even those in our middle ADR tertile, did not improve in any detection metric with CADe [44]. Very similar results were reported by Nehme et al., whose low ADR tertile had higher ADRs than our low ADR tertile [55]. Although some pragmatic studies do report improvements even in “high detectors” [22], the accumulated evidence suggests that CADe is not a simple fix to the “low detector problem.”

Just as we must keep focus on the big picture goal (decreases in CRC incidence and mortality), we should also remember the basics. Colonoscopy quality rests on a foundation of meticulous mucosal exposure and inspection leading to detection. CADe does not directly improve exposure, though it can indirectly motivate it (perhaps with a range of Hawthorne effects across studies). Studies of mucosal exposure devices and CADe are beginning to shed light on the relative contributions of better exposure vs. inspection/detection. In one RCT, a mucosal exposure device (Endocuff; Olympus Corp., Tokyo, Japan) and CADe both improved ADR substantially, and the combination also improved advanced adenoma detection; ADR was slightly higher numerically with Endocuff than with CADe, and CADe added relatively little on top of Endocuff [66]. In another study, CADe did not improve ADR among “high detectors” in the English national bowel cancer screening program who used Endocuff in approximately 70% of cases [77]. In another RCT, ADR was higher with Endocuff added to CADe than with CADe alone (49.6% vs. 44.0%; P = 0.04) [88]. Given the choice of only one, would we pick a mucosal exposure device or CADe? All uncertainties aside, and even in light of possibly marginal differences of uncertain long-term benefit, patients probably want us to use both.

Beyond adenomas, the improvement in number of SSLs per colonoscopy with CADe in the study by Maas et al. is notable. The authors cite unpublished data that 11% of lesions used to train the study’s CADe system were SSLs. I am not sure whether the “deep features” detected by CADe algorithms are fundamentally different for SSLs or adenomas, or whether systems trained primarily on adenomas are demonstrating a “halo effect” extending to SSLs. Regardless, if CADe improves SSL detection, which is emerging as a quality metric, this could improve CRC outcomes.

Most of us believe, or want to believe, that AI can truly revolutionize endoscopy – for the benefit of patients and practitioners. The case of CADe of polyps, which seemed straightforward, is proving to be more nuanced than expected. More challenging extensions of AI and CADe, including real-time assessment of mucosal exposure that allows intraprocedural corrective action, or detection of subtle CRCs and advanced precancerous lesions that even “high detectors” may be missing, could help deliver on a touted promise of AI in endoscopy – applications that help decrease cancer morbidity and mortality.

Publication History

Article published online:
07 August 2024

• References

• 1 Maas MHJ, Rath T, Spada C. et al. A computer-aided detection system in the everyday setting of diagnostic, screening, and surveillance colonoscopy: an international, randomized trial. Endoscopy 2024; 56 DOI: 10.1055/a-2328-2844. (PMID: 38749482)
• 2 Wei MT, Fay S, Yung D. et al. Artificial intelligence-assisted colonoscopy in real-world clinical practice: a systematic review and meta-Analysis. Clin Transl Gastroenterol 2024; 15: e00671 DOI: 10.14309/ctg.0000000000000671. (PMID: 38146871)
• 3 Schottinger JE, Jensen CD, Ghai NR. et al. Association of physician adenoma detection rates with postcolonoscopy colorectal cancer. JAMA 2022; 327: 2114-2122
• 4 Ladabaum 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.e486
• 5 Nehme F, Coronel E, Barringer DA. 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.e106
• 6 Aniwan S, Mekritthikrai K, Kerr SJ. et al. Computer-aided detection, mucosal exposure device, their combination, and standard colonoscopy for adenoma detection: a randomized controlled trial. Gastrointest Endosc 2023; 97: 507-516
• 7 Ahmad A, Wilson A, Haycock A. et al. Evaluation of a real-time computer-aided polyp detection system during screening colonoscopy: AI-DETECT study. Endoscopy 2023; 55: 313-319
• 8 Spadaccini M, Hassan C, Rondonotti E. et al. Combination of mucosa-exposure device and computer-aided detection for adenoma detection during colonoscopy: a randomized trial. Gastroenterology 2023; 165: 244-251.e243