Artificial Intelligence (AI) promises to identify pre-malignant and advanced malignant lesions during colonoscopy that might otherwise be overlooked. Does it deliver on this promise? It seems to depend on where, how, and by whom it is implemented.
Clinical Studies vs. Real World
The majority of globally conducted randomized clinical trials on AI use „clearly demonstrate an increase in adenoma detection rate (ADR) during colonoscopy,“ said Dr. Prateek Sharma, a gastroenterologist at the University of Kansas Cancer Center in Kansas City. „But the results in practice have been quite varied; some show improvement, others do not.“
Sharma is co-author of a recent pooled analysis of nine randomized controlled studies on the impact of AI on colonoscopy surveillance after polyp removal. It was found that the use of AI increased the proportion of patients requiring intensive surveillance by approximately 35% in the US and 20% in Europe (absolute increases of 2.9% and 1.3%, respectively). „While this may contribute to improved cancer prevention, it significantly increases patient burden and healthcare costs,“ the authors concluded.
A recently conducted retrospective analysis of the staged implementation of a computer-aided detection system (CADe) at a single academic center in Chicago found that endoscopists using CADe identified more adenomas and serrated polyps during combined screening and surveillance colonoscopy – but only if the endoscopists were regular CADe users („majority users“). A systematic review and meta-analysis of 21 randomized controlled trials comparing CADe with standard colonoscopy found increased detection of adenomas, but not advanced adenomas, as well as a higher rate of unnecessary removal of non-neoplastic polyps. Furthermore, a multicenter, randomized controlled trial in patients with a positive fecal immunochemical test found that the use of AI was not associated with better detection of advanced neoplasms. Lead author Carolina Mangas Sanjuán, MD, PhD, Hospital General Universitario Dr. Balmis, Alicante, Spain, told Medscape News, the results were „surprising,“ considering that previous studies had shown benefit.
Similarly, researchers in a pragmatic implementation study conducted in Stanford, California found no significant effect of CADe on ADR, adenomas per colonoscopy, or other detection metrics. Furthermore, CADe had no impact on procedure duration or non-neoplastic lesion detection rates. However, the authors cautioned against considering their study as an „outlier,“ pointing to an Israeli study comparing adenoma and polyp detection rates six months before and after the introduction of KI-assisted colonoscopy. These authors reported no performance improvement by the KI device and concluded that it was not useful in routine practice.
A Hodgepodge of Methods
„It’s not clear why some studies are positive and others negative,“ Sharma acknowledged. Study design is a factor, especially in studies under real-world conditions, he said. Some researchers use the pre-post approach, as in the Israeli study; others compare usage in different rooms – one with and one without a CADe device. As with the Chicago analysis, findings from such studies likely depend on whether the endoscopists who have the CADe device in the room actually use it.
Other real-world studies delve into temporal detection, Sharma said. For example, a study of 1780 colonoscopies in China found that KI systems exhibited higher supportability for colonoscopies performed later in the day, when adenoma detection rates typically declined, possibly due to fatigue. These authors suggest that KI may have the potential to maintain the high quality and homogeneity of colonoscopies and improve the performance of endoscopists in large screening programs and high-volume centers.
„A mix of different types of real-world studies is coming, and it’s very difficult to figure everything out,“ Sharma said. „We just have to look at these devices as innovations, embrace them, and work with them to figure out how they fit into our practice.“
Perceptions and Expectations
New evidence suggests that endoscopists‘ perceptions and expectations could influence their assessment of the potential benefit of AI in practice, Sharma noted. „Someone might say, ‚I’m a trained physician. Why do I need a machine to help me?‘ This can lead to a situation where the endoscopist is constantly questioning the device and trying to override it or not give it recognition.“ Others may believe that the AI device will definitely help, so they may not search as carefully for adenomas.
A study at MD Anderson Cancer Center at the University of Texas in Houston, where the activation of the AI system was at the discretion of the endoscopist, found that real-time CADe did not improve adenoma detection in endoscopists with high baseline detection rates. However, despite its availability, AI-assisted colonoscopy was only activated in half of the cases, and a post-procedural survey revealed numerous concerns from staff and endoscopists. In particular, endoscopists feared that the system would provide too many false-positive signals (82.4%), be too distracting (58.8%), and lengthen procedure time (47.1%). The authors of the Stanford study, which found no benefit of CADe in routine practice, noted: „Most concerning would be if the use of CADe were inadvertently associated with a simultaneous unconscious deterioration of mucosal exposure quality, possibly because of a false sense of well-being that CADe would ensure a high-quality examination.“ „We’re trying to evaluate some of these interactions between endoscopists and AI devices both pragmatically in practice and in clinical studies,“ Sharma said. „A lot depends on the context in which you approach and present the devices. We’re telling physicians, this is a tool, not something you’re competing with and not something that’s meant to replace you. This is something that can make your life easier, so try it out.“
Is AI More Helpful to Less Experienced Endoscopists?
It seems intuitive that less experienced endoscopists would be aided by AI, and indeed, some recent studies confirm this. A small randomized controlled study in Japan presented at the President’s Plenary Session of the American Society for Gastrointestinal Endoscopy (ASGE) annual meeting in May 2023 found that a CADe system was „especially useful“ for beginner endoscopists, where the adenoma miss rate was lower using the device compared to a white light control device. Another randomized controlled study in Japan found that the use of CADe was associated with an overall increased ADR in endoscopists in training.
However, experienced endoscopists are likely to benefit as well, noted ASGE President Jennifer Christie, MD, head of the Division of Gastroenterology and Hepatology at the Anschutz Medical Campus of the University of Colorado School of Medicine in Aurora. „We know that these AI devices can be useful in training our colleagues to recognize certain lesions in the colon,“ she said. „However, they are also helpful for many very experienced physicians, as they serve as a supportive diagnostic tool.“ Some studies support this dual benefit. The AID-2 study, specifically designed to investigate whether experience affects AI findings during colonoscopy, was conducted in non-expert endoscopists (lifetime volume of less than 2000 colonoscopies). The researchers, including Sharma, found that CADe increased ADR compared to the control group by 22%.
A previous study, AID-1, used a similar design but was conducted in experienced endoscopists. In AID-1, the ADR in the CADe group was also significantly higher (54.8%) compared to the control group (40.4%), and the adenomas detected per colonoscopy were significantly higher in the CADe group (mean 1.07) than in the control group (mean 0.71). A multivariate post hoc analysis summarizing the results of AID-1 and AID-2 showed that the use of CADe and the indication for colonoscopy, but not the degree of experience of the examiner, were associated with ADR differences. This led the researchers to conclude, „Experience seems to play a minor role as a determinant of ADR.“
Similarly, a 2023 study from China examined the average number of adenomas detected per colonoscopy according to the experience of the endoscopist. All rates were significantly higher in KI-assisted colonoscopies compared to conventional non-KI colonoscopy: overall ADR 39.9% vs. 32.4%; advanced ADR 6.6% vs. 4.9%; ADR of experienced endoscopists: 42.3% vs. 32.8%; ADR of non-expert endoscopists: 37.5% vs. 32.1%; and adenomas per colonoscopy, 0.59 vs. 0.45. The authors concluded that „KI-assisted colonoscopy improved overall ADR, advanced ADR, and ADR for both experienced and non-expert treating endoscopists.“
Improving the Algorithms
Experts agree that current and future research will improve the accuracy and quality of AI colonoscopy for all users, leading to new standards and more consistent results in both clinical studies and real-world applications. Ongoing…