Abstract
Background Use of artificial intelligence may increase detection of colorectal neoplasia at
colonoscopy by improving lesion recognition (CADe) and reduce pathology costs by improving
optical diagnosis (CADx).
Methods A multicenter library of ≥ 200 000 images from 1572 polyps was used to train a combined
CADe/CADx system. System testing was performed on two independent image sets (CADe:
446 with polyps, 234 without; CADx: 267) from 234 polyps, which were also evaluated
by six endoscopists (three experts, three non-experts).
Results CADe showed sensitivity, specificity, and accuracy of 92.9 %, 90.6 %, and 91.7 %,
respectively. Experts showed significantly higher accuracy and specificity, and similar
sensitivity, while non-experts + CADe showed comparable sensitivity but lower specificity
and accuracy than CADe and experts. CADx showed sensitivity, specificity, and accuracy
of 85.0 %, 79.4 %, and 83.6 %, respectively. Experts showed comparable performance,
whereas non-experts + CADx showed comparable accuracy but lower specificity than CADx
and experts.
Conclusions The high accuracy shown by CADe and CADx was similar to that of experts, supporting
further evaluation in a clinical setting. When using CAD, non-experts achieved a similar
performance to experts, with suboptimal specificity.
