Eye Tracking Analysis to Determine the Endoscopist's Recognition Rate for Artificial Intelligence-Detected Sites in Colonoscopy.

PURPOSE: Computer-aided detection systems (CADe) are used in screening colonoscopy; however, no study has quantitatively analyzed how endoscopists recognize candidate sites identified by CADe. The aim of this study was to determine the extent to which endoscopists could recognize CADe-detected sites.

METHODS

This prospective study used a dedicated eye tracking technology-based system to record the endoscopist's eye position and CADe detection sites in a single video during colonoscope withdrawal. All videos were reviewed to calculate the endoscopist's recognition rate, which was defined as the percentage of sites identified by CADe at which the endoscopist stared for > 300 ms. A logistic regression model was used to determine the predictors for quicker lesion detection by endoscopists than by CADe.

RESULTS

In 55 colonoscopies, 447 sites (306 artifacts, 98 adenomas, 28 hyperplastic polyps, 12 sessile serrated lesions [SSLs], 3 intramucosal cancers) were detected by CADe. The overall endoscopist recognition rate was 82.1%; the rate was 95.9% for adenomas, 92.8% for hyperplastic polyps, 100% for SSLs, 100% for intramucosal cancers, and 75.8% for artifacts. However, tracking speed analysis showed that only 33.3% CADe-detected SSLs were tracked by endoscopists within 300 ms. Logistic regression analysis revealed SSLs as an independent predictor for faster detection by endoscopists than by CADe (odds ratio: 6.1, 95% confidence interval: 4.6-807.3, P = 0.002).

CONCLUSION

Endoscopists ignored a substantial proportion of artifacts and successfully recognized most lesions detected by CADe. These findings provide insights into the CADe-endoscopist interaction. Nevertheless, SSL detection should be conducted with caution, even with CADe support.

TRIAL REGISTRATION: UMIN-CTR (UMIN000052891).