Real-Time Characterization of Colonic Polyps: A Multicenter Prospective Study Evaluating the CAD-EYE System in...

Polypectomy-related costs could potentially be reduced through optical diagnosis strategies, such as "diagnose-and-leave" and "resect-and-discard." Artificial intelligence, using computer-aided diagnosis (CAD), may provide a reproducible optical diagnosis of colorectal lesions. This study aimed to assess the performance of the CAD-EYE system in the real-time characterization of colonic polyps.

We conducted a cross-sectional, multicenter study evaluating the CAD-EYE system in patients undergoing screening colonoscopies at 5 French centers. CAD-EYE predictions and assessments by endoscopists (hyperplastic vs neoplastic) were compared with histopathology results. The primary outcome was the sensitivity of CAD-EYE for predicting neoplastic polyps, compared with the predefined threshold of 85%. The secondary outcomes were the specificity, positive predictive value, negative predictive value, endoscopists' performance, and polyp detection rates.

Of 398 polyps analyzed, 343 were included in the primary analysis. CAD-EYE characterization was feasible in 96% of cases. The sensitivity was 0.80 (95% confidence interval, 0.74-0.85), which failed to achieve the predefined threshold of 85% ( P = 0.064). The specificity, negative predictive value, and positive predictive value were 0.79, 0.64, and 0.90, respectively. Performance was higher for diminutive rectosigmoid polyps (DRSPs). Endoscopists showed higher sensitivity than CAD-EYE (0.90 vs 0.80, P = 0.001). CAD-EYE-assisted colonoscopies detected more polyps per procedure (3.3 vs 2.3, P < 0.001) than endoscopists alone.

The performance of CAD-EYE was insufficient for the characterization of neoplastic colonic polyps. CAD-EYE performed better for DRSPs. AI seems to be beneficial for polyp detection.