Autonomous AI bests AI-assisted endoscopists in optical diagnosis of colorectal polyps
Click Here to Manage Email Alerts
Key takeaways:
- The overall accuracy of AI alone vs. AI-assisted human optical diagnosis of colorectal polyps was 77.2% vs. 72.1%.
- Autonomous AI demonstrated higher agreement with pathology-based surveillance intervals.
Autonomous AI exhibited “noninferior accuracy” for optical diagnosis of diminutive colorectal polyps vs. AI-assisted endoscopists, yet surpassed endoscopist accuracy in later pathology-based surveillance intervals, according to researchers.
“Optical diagnosis of colorectal polyps has been proposed as an alternative to histologic diagnosis of diminutive (1-5 mm) polyps,” Roupen Djinbachian, MD, of Montreal University Hospital Research Center, and colleagues wrote in Gastroenterology. “Optical diagnosis strategies have faced significant barriers to implementation, with most endoscopists citing fear of providing an incorrect diagnosis as the main barrier. AI-based systems (CADx) have been proposed as a solution to these barriers to implementation.”
They continued: “However, the efficacy and safety of autonomous AI-based diagnostic platforms have not yet been evaluated.”
To do this, researchers conducted a parallel-group, controlled noninferiority trial of 467 patients undergoing elective colonoscopy at Montreal University Hospital Center between September 2022 and June 2023. Patients were randomly assigned to one of two groups: autonomous AI, using CADx-only optical diagnosis (n = 238; mean age, 64.1 years; 48.3% women) or AI-assisted human diagnosis (AI-H), with optical diagnosis performed by endoscopists after review of CADx real-time diagnosis (n = 229; mean age, 64 years; 50.2% women).
Researchers analyzed 158 polyps from the AI group and 179 from the AI-H group, of which 63.8% and 65.1% were diminutive polyps, respectively.
According to results, accuracy of optical diagnosis in the autonomous AI group was 77.2% (95% CI, 69.7-84.7) vs. 72.1% (95% CI, 65.6-78.6) in the AI-H group, and 77.2% vs. 75.5%, respectively, for high-confidence diagnoses.
However, researchers reported statistically significantly higher agreement with pathology-based surveillance intervals with autonomous AI (91.5%; 95% CI, 86.9-96.1) vs. AI-H (82.1%; 95% CI, 76.5-87.7).
Sensitivity and specificity for adenoma diagnosis were slightly higher in the AI group vs. the AI-H group (84.8% vs. 83.6%; 64.4% vs. 63.8%), with similar findings for positive predictive values between groups (85.6% vs. 78.6%). Negative predictive values were 63% vs. 71%, respectively.
“Autonomous AI-based optical diagnosis had noninferior accuracy to endoscopist-based diagnosis,” Djinbachian and colleagues wrote. “Both autonomous AI and AI-H exhibited relatively low accuracy for optical diagnosis; however, autonomous AI achieved higher agreement with pathology-based surveillance intervals.”
They added, “The findings suggest that optical diagnosis of diminutive polyps can be accurately and safely performed using autonomous AI. This would allow the potential for eliminating the need for histologic assessment of diminutive polyps.”
Collapse
Read more about
Click Here to Manage Email Alerts