Artificial intelligence for characterization of colorectal polyps: Prospective multicenter study.

Abstract

Background and study aims Optical diagnosis poses challenges to implementation of "resect and discard" strategies. This study aimed to assess the feasibility and performance of a new commercially available system for colorectal polyps. Patients and methods Nine expert endoscopists in three centers performed colonoscopies using artificial intelligence-equipped colonoscopes (CAD EYE, Fujifilm). Histology and predictions were compared, with hyperplastic polyps and sessile serrated lesions grouped for analysis. Results Overall, 253 polyps in 119 patients were documented (n=152 adenomas, n=78 hyperplastic polyps, n=23 sessile serrated lesions). CAD EYE detected polyps before endoscopists in 81 of 253 cases (32%). The mean polyp size was 5.5 mm (SD 0.6 mm). Polyp morphology was Paris Ip (4 %), Is (28 %), IIa (60 %), and IIb (8 %). CAD EYE achieved a sensitivity of 80%, specificity of 83%, positive predictive value (PPV) of 96%, and negative predictive value (NPV) of 72%. Expert endoscopists had a sensitivity of 88%, specificity of 83%, PPV of 96%, and NPV of 72%. Diagnostic accuracy was similar between CAD EYE (81%) and endoscopists (86%). However, sensitivity was greater with endoscopists as compared with CAD EYE ( P <0.05). CAD EYE classified sessile serrated lesions as hyperplasia in 22 of 23 cases, and endoscopists correctly classified 16 of 23 cases. Conclusions The CAD EYE system shows promise for detecting and characterizing colorectal polyps. Larger studies are needed, however, to confirm these findings.