Performance of a new integrated CADe/CADx system for detection and characterization of colorectal neoplasia

Background Artificial Intelligence (AI) may reduce miss rate of colorectal neoplasia at colonoscopy by improving lesion recognition (CADe), and cost of pathology by improving optical diagnosis (CADx). Methods To train a combined CADe and CADx (CAD-EYE,Fujifilm,Japan) based on deep learning, a multicenter library of >200,000 images from 1,572 polyps was used, while testing was performed on two independent image sets (CADe: 446 with polyps and 234 without; CADx: 267) from 234 polyps that was also evaluated by 6 endoscopists (3 experts, 3 non-experts). Results CADe showed a sensitivity, specificity and accuracy of 92.9%, 90.6% and 91.7%, respectively. Experts showed slightly higher accuracy and specificity and a similar sensitivity, while non-experts+CADe showed comparable sensitivity, but lower specificity and accuracy. CADx system showed a sensitivity, specificity and accuracy of 85%, 79.4% and 83.6% for polyp characterization, respectively. Experts comparable performances, while non-experts using CADx showed comparable accuracy, but lower specificity. Conclusions The high accuracy shown by CADe and CADx systems is similar to expert endoscopists, prompting its implementation in clinical practice. When using CAD, non-expert endoscopists achieve similar performances to those of expert endoscopists, with suboptimal specificity.