CHARACTERIZATION COMPARISON BETWEEN TWO CAD SYSTEMS (COMBO CAD STUDY) IN REAL-LIFE ENDOSCOPY: AN INTERIM ANALYSIS

Aims Implementation of clinical strategies based on optical diagnosis of < 5 mm rectosigmoid polyps may lead to a substantial saving of economic and financial resources. Artificial intelligence(AI)has the potential to help in making the characterization process more reliable and objective. We compared the performances in rectosigmoid polyp characterization of two recently approved CADx systems(CAD-EYE,CAD A,and GI-Genius,CAD B)working in blue- and white-light.

Methods We performed an interim analysis of a prospective observational study(NCT05141409)enrolling all average-risk patients who underwent screening or surveillance colonoscopy from November 2021. All recto-sigmoid diminutive lesions(≤5 mm)were included and categorize as adenoma or non-adenoma by the endoscopist using both white-light and chromoendoscopy(BLI)(Endoscopist-alone performance).Then, the output automatically provided by the two AI appeared on two different screens(AI-alone performance)in white(B)and blue light(A). The final diagnosis provided by endoscopist combining the results of the first two steps were reported as AI-assisted diagnosis. Concordance and comparison between the AI-alone performance of the two CADx were assessed as primary outcome.

Results One-hundred-twenty-five diminutive rectosigmoid polyps(51adenomas,40.8%,and 74hyperplastic,59.2%)were included in the final analysis. The diagnostic sensitivity/specificity/Negative Predictive value of the CAD A was 92.2% / 88.7% / 97.2%, and it was 96.1% / 86.0% / 97.1% for CAD B, respectively. No difference in overall accuracy and NPV between the two systems was found(p value:0.9).The concordance rate between the systems were 92%. The diagnostic accuracy of endoscopist before and after AI-assistance was 91.2%and92.8%.

Conclusions The NPV for rectosigmoid diminutive lesions was higher than the PIVI for both of the machines, irrespectively of the need of advanced imaging.

Publication History

Article published online:
14 April 2022