AI-guided colonoscopy results in more small adenomas detected

Artificial intelligence technologies significantly improved detection of precancerous polyps during colonoscopy, according to results from a large meta-analysis of randomized controlled trials.

But while adenoma detection rates increased by nearly 25%, compared with conventional care, the AI-guided procedures were also associated with an increase in unnecessary removal of non-neoplastic polyps. Little effect was seen on the detection of larger, advanced lesions.

The findings, published in Annals of Internal Medicine, are likely to change clinical guidelines in favor of AI-assisted procedures, said Cesare Hassan, MD, PhD, of Humanitas Research Hospital and University in Milan, Italy.

For their research, Dr. Hassan and his colleagues looked at results from 21 trials randomizing more than 18,000 patients to colonoscopy with computer-aided detection (CADe) or standard colonoscopy. Colonoscopy could be carried out for screening, surveillance, or diagnostic purposes. The included trials, which were all published in 2019 or later, took place in Asia, North America, and Europe. In most of the studies endoscopists were not blinded to treatment allocation.

The adenoma detection rate, or the proportion of individuals undergoing colonoscopy who had at least one adenoma detected and removed, was 44% in the CADe arms, compared with 35.9% assigned to standard care (relative risk [RR], 1.24; 95% CI, 1.16-1.33). The CADe patients also saw more non-neoplastic polyps removed: 0.52 per colonoscopy, compared with 0.34 for standard care. The increased adenoma detection rate appeared to be driven by a 55% decrease in the error or miss rate of adenomas at per-polyp analysis, the investigators wrote in their paper.

CADe did not increase the number of advanced adenomas (defined as greater than 10 mm, with high-grade dysplasia and villous histology) detected per colonoscopy. Rather, the benefit was “mainly limited to increased detection of diminutive (≤5 mm) adenomas,” the investigators wrote. Dr. Hassan commented that the lack of benefit for detecting larger adenomas was expected, as these are easier for endoscopists to identify visually.

The key limitation of the meta-analysis, Dr. Hassan said, was the fact that endoscopists in the studies could not be blinded. “We can assume there is a risk of bias in our estimates — that is why we describe the quality of evidence as low or moderate, never high. Through randomization we can control other aspects, especially the prevalence of disease, which avoids a scenario in which the endoscopist opts to treat riskier patients with CADe.” But the possibility of a change in the endoscopist’s performance when using these systems cannot be excluded.

Dr. Hassan commented that quantifying the risks, and costs, of overtreatment linked to CADe would require more investigation. “Any time I remove a polyp there’s a risk of perforation and bleeding,” he noted, though most of the unnecessary resections seen in the meta-analysis were of small hyperplastic polyps considered to be low risk for complications. Use of CADe was associated with only slight increases in procedure time, the investigators found.

In a multicenter Spanish study also published in Annals of Internal Medicine, Carolina Mangas-Sanjuan, MD, PhD, of the Hospital General Universitario Dr. Balmis, Alicante, Spain, and her colleagues looked at computer-aided detection of advanced colorectal neoplasias in a higher-risk cohort and saw little advantage over standard colonoscopy.

This study, which randomized 3,213 subjects, is the largest to date aimed at learning whether AI can improve the detection of advanced lesions. As in Dr. Hassan’s meta-analysis, the researchers did not see significant differences in the rates of detection for these larger lesions. Nor, in this study, did CADe did improve the global adenoma detection rate among the FIT positive individuals undergoing screening.

The detection rate of advanced colorectal neoplasias (advanced adenoma or advanced serrated polyp) was 34.8% with CADe (95% CI, 32.5%-37.2%) and 34.6% for standard colonoscopy (95% CI, 32.2%-36.9%); adjusted risk ratio, 1.01 [95% CI, 0.92-1.10]. The mean number of advanced colorectal neoplasias detected per colonoscopy was 0.54 for the intervention group, compared with 0.52 for standard care. For all adenomas, the detection rate was 64.2% with CADe vs 62% for controls.

Dr. Rodrigo Jover of the Hospital General Universitario Dr. Balmis, the study’s corresponding author, commented to this news organization that “while CADe systems are able to improve detection of small low-risk lesions, these devices are not yet able to detect more significant high-risk lesions. Therefore, there is still room for improvement if these systems are adequately trained with datasets of large, difficult-to-detect lesions.”

In an editorial comment on the Spanish and Italian studies Dennis Shung, MD, PhD of Yale University in New Haven, Connecticut, concluded that “this recent evidence suggests that CADe systems do not meaningfully improve the detection of larger (≥10 mm) clinically significant polyps. This tempers enthusiasm for CADe but does not negate the clear performance benefit for detecting adenomas of all sizes.”

How to integrate the AI systems into real-world practice is the real challenge ahead, Dr. Shung argued, noting that, in contrast to randomized trials, “several recent real-world studies have found no improvement in [adenoma detection rate] when CADe is deployed.” Lower trust in the systems can result in their underutilization, Dr. Shung argued, while higher trust can lead to overreliance. “How lgorithmic systems partner with clinicians and how these should be designed and refined across heterogeneous systems and contexts are necessary questions that must be explored to minimize disruption and lead to real-world effectiveness.”

Dr. Mangas-Sanjuan’s study was funded by a grant from Medtronic; Part of Dr. Hassan’s meta-analysis was supported by a European Commission grant to one co-author. Drs. Shung, Hassan, Manguas-Sanjuan, and Jover declared no financial conflicts of interest.

Artificial intelligence technologies significantly improved detection of precancerous polyps during colonoscopy, according to results from a large meta-analysis of randomized controlled trials.

But while adenoma detection rates increased by nearly 25%, compared with conventional care, the AI-guided procedures were also associated with an increase in unnecessary removal of non-neoplastic polyps. Little effect was seen on the detection of larger, advanced lesions.

The findings, published in Annals of Internal Medicine, are likely to change clinical guidelines in favor of AI-assisted procedures, said Cesare Hassan, MD, PhD, of Humanitas Research Hospital and University in Milan, Italy.

For their research, Dr. Hassan and his colleagues looked at results from 21 trials randomizing more than 18,000 patients to colonoscopy with computer-aided detection (CADe) or standard colonoscopy. Colonoscopy could be carried out for screening, surveillance, or diagnostic purposes. The included trials, which were all published in 2019 or later, took place in Asia, North America, and Europe. In most of the studies endoscopists were not blinded to treatment allocation.

The adenoma detection rate, or the proportion of individuals undergoing colonoscopy who had at least one adenoma detected and removed, was 44% in the CADe arms, compared with 35.9% assigned to standard care (relative risk [RR], 1.24; 95% CI, 1.16-1.33). The CADe patients also saw more non-neoplastic polyps removed: 0.52 per colonoscopy, compared with 0.34 for standard care. The increased adenoma detection rate appeared to be driven by a 55% decrease in the error or miss rate of adenomas at per-polyp analysis, the investigators wrote in their paper.

CADe did not increase the number of advanced adenomas (defined as greater than 10 mm, with high-grade dysplasia and villous histology) detected per colonoscopy. Rather, the benefit was “mainly limited to increased detection of diminutive (≤5 mm) adenomas,” the investigators wrote. Dr. Hassan commented that the lack of benefit for detecting larger adenomas was expected, as these are easier for endoscopists to identify visually.

The key limitation of the meta-analysis, Dr. Hassan said, was the fact that endoscopists in the studies could not be blinded. “We can assume there is a risk of bias in our estimates — that is why we describe the quality of evidence as low or moderate, never high. Through randomization we can control other aspects, especially the prevalence of disease, which avoids a scenario in which the endoscopist opts to treat riskier patients with CADe.” But the possibility of a change in the endoscopist’s performance when using these systems cannot be excluded.

Dr. Hassan commented that quantifying the risks, and costs, of overtreatment linked to CADe would require more investigation. “Any time I remove a polyp there’s a risk of perforation and bleeding,” he noted, though most of the unnecessary resections seen in the meta-analysis were of small hyperplastic polyps considered to be low risk for complications. Use of CADe was associated with only slight increases in procedure time, the investigators found.

In a multicenter Spanish study also published in Annals of Internal Medicine, Carolina Mangas-Sanjuan, MD, PhD, of the Hospital General Universitario Dr. Balmis, Alicante, Spain, and her colleagues looked at computer-aided detection of advanced colorectal neoplasias in a higher-risk cohort and saw little advantage over standard colonoscopy.

This study, which randomized 3,213 subjects, is the largest to date aimed at learning whether AI can improve the detection of advanced lesions. As in Dr. Hassan’s meta-analysis, the researchers did not see significant differences in the rates of detection for these larger lesions. Nor, in this study, did CADe did improve the global adenoma detection rate among the FIT positive individuals undergoing screening.

The detection rate of advanced colorectal neoplasias (advanced adenoma or advanced serrated polyp) was 34.8% with CADe (95% CI, 32.5%-37.2%) and 34.6% for standard colonoscopy (95% CI, 32.2%-36.9%); adjusted risk ratio, 1.01 [95% CI, 0.92-1.10]. The mean number of advanced colorectal neoplasias detected per colonoscopy was 0.54 for the intervention group, compared with 0.52 for standard care. For all adenomas, the detection rate was 64.2% with CADe vs 62% for controls.

Dr. Rodrigo Jover of the Hospital General Universitario Dr. Balmis, the study’s corresponding author, commented to this news organization that “while CADe systems are able to improve detection of small low-risk lesions, these devices are not yet able to detect more significant high-risk lesions. Therefore, there is still room for improvement if these systems are adequately trained with datasets of large, difficult-to-detect lesions.”

In an editorial comment on the Spanish and Italian studies Dennis Shung, MD, PhD of Yale University in New Haven, Connecticut, concluded that “this recent evidence suggests that CADe systems do not meaningfully improve the detection of larger (≥10 mm) clinically significant polyps. This tempers enthusiasm for CADe but does not negate the clear performance benefit for detecting adenomas of all sizes.”

How to integrate the AI systems into real-world practice is the real challenge ahead, Dr. Shung argued, noting that, in contrast to randomized trials, “several recent real-world studies have found no improvement in [adenoma detection rate] when CADe is deployed.” Lower trust in the systems can result in their underutilization, Dr. Shung argued, while higher trust can lead to overreliance. “How lgorithmic systems partner with clinicians and how these should be designed and refined across heterogeneous systems and contexts are necessary questions that must be explored to minimize disruption and lead to real-world effectiveness.”

Dr. Mangas-Sanjuan’s study was funded by a grant from Medtronic; Part of Dr. Hassan’s meta-analysis was supported by a European Commission grant to one co-author. Drs. Shung, Hassan, Manguas-Sanjuan, and Jover declared no financial conflicts of interest.

Artificial intelligence technologies significantly improved detection of precancerous polyps during colonoscopy, according to results from a large meta-analysis of randomized controlled trials.

But while adenoma detection rates increased by nearly 25%, compared with conventional care, the AI-guided procedures were also associated with an increase in unnecessary removal of non-neoplastic polyps. Little effect was seen on the detection of larger, advanced lesions.

The findings, published in Annals of Internal Medicine, are likely to change clinical guidelines in favor of AI-assisted procedures, said Cesare Hassan, MD, PhD, of Humanitas Research Hospital and University in Milan, Italy.

For their research, Dr. Hassan and his colleagues looked at results from 21 trials randomizing more than 18,000 patients to colonoscopy with computer-aided detection (CADe) or standard colonoscopy. Colonoscopy could be carried out for screening, surveillance, or diagnostic purposes. The included trials, which were all published in 2019 or later, took place in Asia, North America, and Europe. In most of the studies endoscopists were not blinded to treatment allocation.

The adenoma detection rate, or the proportion of individuals undergoing colonoscopy who had at least one adenoma detected and removed, was 44% in the CADe arms, compared with 35.9% assigned to standard care (relative risk [RR], 1.24; 95% CI, 1.16-1.33). The CADe patients also saw more non-neoplastic polyps removed: 0.52 per colonoscopy, compared with 0.34 for standard care. The increased adenoma detection rate appeared to be driven by a 55% decrease in the error or miss rate of adenomas at per-polyp analysis, the investigators wrote in their paper.

CADe did not increase the number of advanced adenomas (defined as greater than 10 mm, with high-grade dysplasia and villous histology) detected per colonoscopy. Rather, the benefit was “mainly limited to increased detection of diminutive (≤5 mm) adenomas,” the investigators wrote. Dr. Hassan commented that the lack of benefit for detecting larger adenomas was expected, as these are easier for endoscopists to identify visually.

The key limitation of the meta-analysis, Dr. Hassan said, was the fact that endoscopists in the studies could not be blinded. “We can assume there is a risk of bias in our estimates — that is why we describe the quality of evidence as low or moderate, never high. Through randomization we can control other aspects, especially the prevalence of disease, which avoids a scenario in which the endoscopist opts to treat riskier patients with CADe.” But the possibility of a change in the endoscopist’s performance when using these systems cannot be excluded.

Dr. Hassan commented that quantifying the risks, and costs, of overtreatment linked to CADe would require more investigation. “Any time I remove a polyp there’s a risk of perforation and bleeding,” he noted, though most of the unnecessary resections seen in the meta-analysis were of small hyperplastic polyps considered to be low risk for complications. Use of CADe was associated with only slight increases in procedure time, the investigators found.

In a multicenter Spanish study also published in Annals of Internal Medicine, Carolina Mangas-Sanjuan, MD, PhD, of the Hospital General Universitario Dr. Balmis, Alicante, Spain, and her colleagues looked at computer-aided detection of advanced colorectal neoplasias in a higher-risk cohort and saw little advantage over standard colonoscopy.

This study, which randomized 3,213 subjects, is the largest to date aimed at learning whether AI can improve the detection of advanced lesions. As in Dr. Hassan’s meta-analysis, the researchers did not see significant differences in the rates of detection for these larger lesions. Nor, in this study, did CADe did improve the global adenoma detection rate among the FIT positive individuals undergoing screening.

The detection rate of advanced colorectal neoplasias (advanced adenoma or advanced serrated polyp) was 34.8% with CADe (95% CI, 32.5%-37.2%) and 34.6% for standard colonoscopy (95% CI, 32.2%-36.9%); adjusted risk ratio, 1.01 [95% CI, 0.92-1.10]. The mean number of advanced colorectal neoplasias detected per colonoscopy was 0.54 for the intervention group, compared with 0.52 for standard care. For all adenomas, the detection rate was 64.2% with CADe vs 62% for controls.

Dr. Rodrigo Jover of the Hospital General Universitario Dr. Balmis, the study’s corresponding author, commented to this news organization that “while CADe systems are able to improve detection of small low-risk lesions, these devices are not yet able to detect more significant high-risk lesions. Therefore, there is still room for improvement if these systems are adequately trained with datasets of large, difficult-to-detect lesions.”

In an editorial comment on the Spanish and Italian studies Dennis Shung, MD, PhD of Yale University in New Haven, Connecticut, concluded that “this recent evidence suggests that CADe systems do not meaningfully improve the detection of larger (≥10 mm) clinically significant polyps. This tempers enthusiasm for CADe but does not negate the clear performance benefit for detecting adenomas of all sizes.”

How to integrate the AI systems into real-world practice is the real challenge ahead, Dr. Shung argued, noting that, in contrast to randomized trials, “several recent real-world studies have found no improvement in [adenoma detection rate] when CADe is deployed.” Lower trust in the systems can result in their underutilization, Dr. Shung argued, while higher trust can lead to overreliance. “How lgorithmic systems partner with clinicians and how these should be designed and refined across heterogeneous systems and contexts are necessary questions that must be explored to minimize disruption and lead to real-world effectiveness.”

Dr. Mangas-Sanjuan’s study was funded by a grant from Medtronic; Part of Dr. Hassan’s meta-analysis was supported by a European Commission grant to one co-author. Drs. Shung, Hassan, Manguas-Sanjuan, and Jover declared no financial conflicts of interest.