AI in Colonoscopy: Disconnect Grows Between Academic, Community Studies

CHICAGO — Companies may have hit a patch of rough road as they race to put artificial intelligence (AI) to work in the colonoscopy suite, as results with systems deployed in community settings aren’t matching the numbers put up in university-based registration trials.

A series of studies presented here at the annual Digestive Disease Week (DDW) conference highlighted the disconnect. On the one hand, attendees heard several presentations of academic trials, touting great improvements in adenoma detection rates (ADRs) and other quality parameters when endoscopists used computer-aided detection systems (known as CADe) to flag polyps and other lesions, as opposed to conventional human visualization.

Yet other studies conducted in community hospitals and outpatient centers, including some randomized trials, for the most part failed to find significant improvements. Moreover, expectations that endoscopists whose past performance appeared suboptimal would see more improvement in ADR with CADe support weren’t borne out.

Explanations for the disconnect are only speculative at this point, with the inherent lack of operator blinding in these studies cited as one possible reason. But more important, perhaps, CADe only goes so far. It can’t pick up polyps hidden behind folds, and issues such as quality of bowel preparation may confound. And ultimately, the clinician-operators decide what to do with polyps flagged by these systems, and thus the systems’ success or failure rests ultimately with them.

One of the most prominent academic studies reported at DDW came from a group at the University of Kansas in Kansas City, evaluating Fujifilm’s EW10-EC02 add-on to its CAD EYE system for colonic lesion detection, approved in Europe in 2020 but still awaiting U.S. authorization.

Like most such systems, it runs in real time with the colonoscopy’s ordinary video feed, putting a green box around features likely to be a polyp — a form of augmented reality. (It bears little relation to the “generative AI” epitomized by ChatGPT, which can write essays and create images in response to users’ prompts. It, too, is finding uses in medicine, but diagnosing individual patients is not one of them, yet.) Also like other such systems, it was developed by first “training” an algorithm to recognize polyps and other lesions based on thousands of clinical images, then validated by applying it to an independent image set from patients with known outcomes.

Competing systems are also available in Europe, but thus far only one has been authorized in the U.S., the GI Genius module sold by Medtronic. And colonic polyp identification is not the only potential application for CADe in gastroenterology, as companies are also hard at work on systems to detect gastric cancers and esophageal dysplasias. These are farther back in the pipeline.

Underpinning the GI Genius authorization was a trial reported in 2020, in which rates of adenomas found per colonoscopy increased by 50% (0.71 to 1.07) when the device was used. That has been the standard others have hoped to achieve, and studies intended to support regulatory approvals have generally succeeded.

Madhav Desai, MD, formerly a fellow at the University of Kansas who is now at the University of Minnesota in Minneapolis, reported the multicenter, university-based trial of the Fujifilm system. (DDW organizers apparently thought it was important enough to have Desai present it in three separate sessions, including a presidential plenary.) The bottom line: with about 500 patients each assigned to be scoped with or without the Fujifilm system’s aid, 17% more adenomas (95% CI 3-33) were detected per patient when the operator used the system.

Similar findings were reported for a variety of other systems. Michiel Maas, MD, of Radboud University Medical Center in the Netherlands, reported on a randomized trial involving more than 900 patients scheduled for routine colonoscopy, testing a system under development by Israel-based Magentiq Eye. Improvements in adenoma detection were greater than in the Fujifilm study: 0.7 adenomas found per patient with the system running versus 0.51 without. The ADR (the number of patients with any adenomas found divided by the total number of patients) improved to 37% with the system, compared with 30% without.

Uri Ladabaum, MD, of Stanford University in California, followed Maas to the podium to deliver a very different set of results. With cooperation from Medtronic, Ladabaum and colleagues arranged to have GI Genius modules placed at several non-academic endoscopy clinics in the local community, and compared outcomes in those clinics to several others doing conventional colonoscopies over a 3-month period.

Crucially, Ladabaum said, the study used a “minimalist deployment strategy” in which clinicians were not specifically encouraged to use the Medtronic system nor were any expectations set with operators. This was an attempt to minimize “enthusiasm” for the novel system that, he said, may be inflating the apparent advantages seen in registration trials.

Absolutely no advantage for CADe was apparent in the study, not when detection prior to implementation was compared to post-implementation, nor between clinics using the devices and those not. Detection rates actually trended downward in most analyses.

Moreover, when outcomes were stratified by endoscopists’ pre-study detection rates (a parameter routinely captured at most clinics for quality-monitoring purposes), no improvement was seen with the systems.

Another community-deployment study yielded similar results. Mike Wei, MD, also at Stanford, reported on a study in which an investigational system developed by EndoVigilant was placed in community clinics in California (not Stanford-affiliated), Connecticut, New Jersey, and Maryland. More than 750 patients were randomly scoped with the system in place or not.

Again, there was no important difference in detection rates whether the system was used or not. With it running, 0.73 adenomas were found per colonoscopy, versus 0.67 without (P=0.496), Wei said. Detection of concerning sessile serrated polyps was also the same. One difference did emerge: rates of detection for non-adenomatous, non-serrated polyps was significantly higher with CADe (0.90 vs 0.51 per exam) but these lesions are relatively unimportant.

Ladabaum said he suspected a number of factors may be at work to diminish these systems’ utility in ordinary practice. Clinicians who see what might be a polyp on their screen that the system doesn’t flag might be induced to go against their own judgment and dismiss it; they might also ignore “boxed” polyps that don’t match what they think a polyp should look like.

The basic quality of the colonoscopy procedure — such as adequacy of bowel prep and the extent of “mucosal exposure,” that is, how much of the colon is actually visualized — may also be critical. In the Fujifilm study, Desai noted that patients with poor bowel prep were excluded.

Another study here suggested that mucosal exposure is indeed a determining factor. Isabella Bergagnini, DO, of Lenox Hill Hospital in New York City, presented a study in which GI Genius was used alongside an Olympus device called an Endocuff that mechanically expands the colon for better visualization. Her group found no particular improvement in detection rates with the CADe system, but the expansion device made a big difference in adenoma detection (adenomas per polyp excised: 61% with Endocuff vs 31% without).

Wei also cited a possible “frustration” factor with the CADe devices, which some endoscopists may find to be distracting. In addition to the highlighting box on the video screen, these systems typically emit a warning tone as well.

And not every community study found no benefit from CADe. Nikhil Thiruvengadam, MD, of Loma Linda University in California, led a trial targeting a historically underserved Hispanic population. Some 550 patients were randomized in blocks to undergo first-time colonoscopies either with or without help from GI Genius. Importantly, the study had no exclusions for bowel prep or indication for the procedure (and, in fact, bowel prep was judged poor in about 10% of participants).

ADRs in this study did show a benefit for CADe, at about 45% versus 32% for conventional decision-making. Findings were similar when analyzed for adenomas per exam and for polyps <5 mm. (Just about every study, irrespective of design, has found no important increase with CADe in detection of large polyps.)

So should every gastroenterologist think about installing a CADe system? It may be inevitable. As one speaker commented, surgeons who have shunned use of robotic aids are now seen as being “left behind,” and the same is likely to happen to endoscopists who don’t keep up with technology. For many, for now, it may be a question of dollars. GI Genius is reputed to cost about $100,000 per module, which must somehow be passed on to payers. Physicians here noted that there is currently no reimbursement code for CADe use, which may be needed before adoption becomes truly widespread.