Colonoscopies with artificial intelligence (AI) demonstrate significantly better adenoma detection rates (ADRs) than most other endoscopic interventions, according to a new report.
AI-based tools appear to outperform other methods intended to increase ADRs, including distal attachment devices, dye-based/virtual chromoendoscopy, water-based techniques, and balloon-assisted devices, researchers found in a systematic review and meta-analysis.
Dr Muhammad Aziz
“ADR is a very important quality metric. The higher the ADR, the less likely the chance of interval cancer,” first author Muhammad Aziz, MD, co-chief gastroenterology fellow at the University of Toledo, told Medscape Medical News. Interval cancer refers to colorectal cancer that is diagnosed within 5 years of a patient’s undergoing a negative colonoscopy.
“Numerous interventions have been attempted and researched to see the impact on ADR,” he said. “The new kid on the block — AI-assisted colonoscopy — is a game-changer. I knew that AI was impactful in improving ADR, but I didn’t know it would be the best.”
The study was published online in the Journal of Clinical Gastroenterology.
Analyzing Detection Rates
Current guidelines set an ADR benchmark of 25% overall, with 30% for men and 20% for women undergoing screening colonoscopy. Every 1% increase in ADR results in a 3% reduction in colorectal cancer, Aziz and his co-authors write.
Several methods can improve ADR over standard colonoscopy. Computer-aided detection and AI methods, whch have emerged in recent years, alert the endoscopist of potential lesions in real time with visual signals.
No direct comparative studies had been conducted, so to make an indirect comparison, Aziz and colleagues undertook a systematic review and network meta-analysis of 94 randomized controlled trials that included 61,172 patients and 20 different study interventions.
The research team assessed the impact of AI in comparison with other endoscopic methods using relative risk for proportional outcomes and mean difference for continuous outcomes. About 63% of the colonoscopies were for screening and surveillance, and 37% were diagnostic. The effectiveness was ranked by P-score (the probability of being the best treatment).
Overall, AI had the highest P-score (0.96), signifying the best modality of all interventions for improving ADR, the study authors write. A sensitivity analysis using the fixed effects model did not significantly alter the effect measure.
The network meta-analysis showed significantly higher ADR for AI compared with autofluorescence imaging (relative risk [RR], 1.33), dye-based chromoendoscopy (RR, 1.22), Endocap (RR, 1.32), Endocuff (RR, 1.19), Endocuff Vision (RR, 1.26), EndoRings (RR, 1.30), flexible spectral imaging color enhancement (RR,1.26), full-spectrum endoscopy (RR, 1.40), high-definition (HD) colonoscopy (RR, 1.41), linked color imaging ( 1.21), narrow-band imaging (RR, 1.33), water exchange (RR, 1.22), and water immersion (RR, 1.47).
Among 34 studies of colonoscpies for screening or surveillance only, the ADR was significantly improved for linked color imaging (RR, 1.18), I-Scan with contrast and surface enhancement (RR, 1.25), Endocuff (RR, 1.20), Endocuff Vision (RR, 1.13), and water exchange (RR, 1.24), compared with HD colonoscopy. Only one AI study was included in this analysis, because the others had significantly more patients who underwent colonoscopy for diagnostic indications. In this case, AI did not improve ADR in comparison with HD colonoscopy (RR, 1.44).
In addition, a significantly improved polyp detection rate (PDR) was noted for AI compared with autofluorescence imaging (RR, 1.28), Endocap (RR, 1.18), Endocuff Vision (RR, 1.21), EndoRings (RR, 1.30), flexible spectral imaging color enhancement (RR, 1.21), full-spectrum endoscopy (RR, 1.39), HD colonoscopy (RR, 1.34), linked color imaging (RR, 1.19), and narrow-band imaging (RR, 1.21). Again, AI had the highest P-score (RR, 0.93).
Among 17 studies of colonoscopy for screening and surveillance, only one AI study was included for PDR. A significantly higher PDR was noted for AI as compared with HD colonoscopy (RR, 1.33). None of the other interventions improved PDR over HD colonoscopy.
No AI Advantage for Serrated Polyps
Twenty-three studies evaluated detection for serrated polyps, including three AI studies. AI did not improve the serrated polyp detection rate (SPDR) compared with other interventions. However, several modalities did improve SPDR: G-EYE compared with full-spectrum endoscopy (RR, 3.93), linked color imaging compared with full-spectrum endoscopy (RR, 1.88) and HD colonoscopy (RR, 1.71), and Endocuff Vision compared with HD colonoscopy (RR, 1.36). G-EYE had the highest P-score (0.93).
AI significantly improved adenomas per colonoscopy compared with full-spectrum endoscopy (mean difference [MD], 0.38), HD colonoscopy (MD, 0.18), and narrow-band imaging (MD, 0.13), the authors note. However, the number of adenomas detected per colonoscopy was significantly lower for AI compared with Endocap (-0.13). Endocap had the highest P-score (0.92).
Dr Jeremy Glissen Brown
“The strengths of this study include the wide range of endoscopic add-ons included, the number of trials included, and the granularity of some of the reporting data,” Jeremy Glissen Brown, MD, a gastroenterologist and an assistant professor of medicine at Duke University, told Medscape Medical News.
Glissen Brown, who wasn’t involved with this study, researches AI tools for polyp detection. He and colleagues have found that AI decreases adenoma miss rates and increases the number of first-pass adenomas detected per colonoscopy.
“The limitations include significant heterogeneity among many of the comparisons, as well as a high risk of bias, as it is technically difficult to achieve blinding of provider participants in the device-based RCTs [randomized controlled trials] that this analysis was based on,” he said.
Additional Considerations
Aziz and colleagues note the need for additional studies of AI-based detection, particularly for screening and surveillance. For widespread adoption into clinical practice, new systems must have higher specificity, sensitivity, accuracy, and efficiency, they write.
“AI technology needs further optimization, as there is still the aspect of having a lot of false positives — lesions detected but not necessarily adenomas that can turn into cancer,” Aziz said. “This decreases the efficiency of the colonoscopy and increases the anesthesia and sedation time. In addition, different AI systems have different diagnostic yield, as it all depends on the images that were fed to the system or algorithm.”
Glissen Brown also pointed to the low number of AI-based studies involving serrated polyp lesion detection. Future research could investigate whether computer-aided detection systems (CADe) decrease miss rates and increase detection rates for sessile serrated lesions, he said.
For practical clinical purposes, Glissen Brown highlighted the potential complementary nature of the various colonoscopy tools. When used together, for instance, AI and Endocuff may increase ADRs even further and decrease the number of missed polyps through different mechanisms, he said.
“It is also important in device research to interrogate the cost vs benefit of any intervention or combination of interventions,” he said. “I think with CADe this is still something that we are figuring out. We will need to find novel ways of making these technologies affordable, especially as the debate of which clinically meaningful outcomes we examine when it comes to AI continues to evolve.”
No funding source for the study was reported. Two authors have received grant support from or have consulted for several pharmaceutical and medical device companies. Glissen Brown has disclosed no relevant financial relationships.
J Clin Gastroenterol. Published online November 28, 2022. Abstract
Carolyn Crist is a health and medical journalist who reports on the latest studies for Medscape, MDedge, and WebMD.
For more news, follow Medscape on Facebook, Twitter, Instagram, and YouTube.
Source: Read Full Article