Standalone Performance of a New Integrated CADE/CADX System For Detection and Characterization Of Colorectal Neoplasia

Aims Artificial Intelligence (AI) may reduce miss rate of colorectal neoplasia at colonoscopy by improving lesion recognition (CADe), and it may also reduce the cost of pathology by improving optical diagnosis (CADx).

Methods We trained and tested the first regulatory-approved AI system combining CADe and CADx in the same platform. A library of ≥200,000 images of 1 063 polyps from 4 European centers was created, and used to train two Convoluted Neural Network models for CADe (1 063 polyps white-light [WL] and 963 Linked Colour Imaging [LCI]) and CADx (662 WL and 1,202 Blue light Imaging [BLI]). The systems were subsequently tested with two image sets (CADe: 459 WL, 455 LCI; CADx: 133WL, 134 BLI) that were independent from the training set. The testing sets were also evaluated by 3 expert endoscopists, and 3 non-expert endoscopists using the AI for benchmarking.

Results Overall, 914 and 267 images were used to test CADe and CADx, respectively. CADe system showed a sensitivity, specificity and accuracy of 92.9 %, 90.6 % and 91.7 %, respectively, in detecting colorectal neoplasia. Experts showed higher accuracy (94.6 % vs 91.7 %; p<0.05) and specificity (94.2 % vs. 90.6 %: p<0.05), and a similar sensitivity, while non-experts+CADe showed comparable sensitivity (94.8 %, p = 0.27), but lower specificity (76.2 %, p = <0.0001) and accuracy (85.4 %, p = <0.0001) compared to both experts and CADe alone. CADx system showed a sensitivity, specificity and accuracy of 85 %, 79.4 % and 83.6 %, respectively, in the characterization of colorectal neoplasia. Experts alone showed comparable performances, while non-experts using CADx showed comparable accuracy (80.1 %, p = 0.4), but lower specificity (62.7 %, p = <0.0001) compared to experts or CADx alone. No difference in CADe and CADx accuracy was noticed when comparing white-light with advanced imaging.

Conclusions The high accuracy shown by CADe and CADx systems is similar to that of expert endoscopists, prompting its implementation in clinical practice. When using CAD, inexpert endoscopists achieve similar performances to those of expert endoscopists.

Citation: Antonelli G, Repici A, Weigt J et al. OP118 STANDALONE PERFORMANCE OF A NEW INTEGRATED CADE/CADX SYSTEM&NBSP;FOR DETECTION AND CHARACTERIZATION OF COLORECTAL NEOPLASIA. Endoscopy 2021; 53: S50.

Publication History

Article published online:
19 March 2021