P125 LCI/BLI chromoendoscopy plus CAD-EYE artificial intelligence for the detection and characterization of endoscopic visible lesions in ulcerative colitis

Abstract

Abstract Background LCI/BLI chromoendoscopy and CAD-EYE artificial intelligence have been developed for the detection and characterization of colorectal neoplasia in the general population. In this study, we analyzed their diagnostic accuracy in the detection and characterization of endoscopic visible lesions (EVL) during endoscopic surveillance of ulcerative colitis (UC). Methods A prospective tandem study on consecutive UC patients was performed, using Eluxeo white light endoscopy (WLE), LCI and LCI+CAD-EYE during a back-to-back withdrawal phase. Lesion characterization of all EVL was performed according to three classifications (Kudo, NICE and Kudo-IBD) and, finally, by BLI+CAD-EYE. The lesion/neoplasia miss rates were analyzed during the detection phase. Sensitivity (SE), specificity (SP), positive- and negative-predictive values (PPV/NPV) and accuracy (ACC) for lesion characterization by Kudo, NICE and Kudo-IBD were calculated when applied to BLI and compared with CAD-EYE, using histology as reference test. Results 62 patients (mean age 54 years, mean disease duration 18 years) were included; 133 lesions (mean size 6 mm, 18 neoplastic) were found in 35 patients. The lesion miss rate was significantly higher (p<0.05) with WLE (9.8%) and LCI (5.3%) than CAD-EYE (0%), with no differences in neoplasia miss rates. For lesion characterization, the SE, SP, PPV, NPV and ACC of BLI+CAD-EYE (83%, 65%, 27%, 96%, 68%, respectively) were not inferior to the conventional Kudo (67%, 71%, 27%, 93%, 71%) and NICE (72%, 68%, 26%, 94%, 68%) classifications, while Kudo-IBD had higher SE (89%) and significantly (p<0.05) higher SP (83%) and ACC (84%) than Kudo, NICE and CAD-EYE. Conclusion LCI+CAD-EYE and the Kudo-IBD criteria applied to BLI can improve, respectively, the detection and characterization of EVL in UC.