Prospective Video-Based Study Assessing Effect of Computer-Assisted Optical Diagnosis on Distinguishing Serrated,...

The current computer-aided diagnosis (CADx) systems are not trained to recognize sessile-serrated lesions (SSL) and currently classify SSL incorrectly either as hyperplastic or adenomatous. We aimed to assess the effect of current untrained CADx systems on diagnostic accuracy of SSL recognition among a diverse group of endoscopists.

17 endoscopists reviewed the same video and image series of 115 polyps [33% SL, 33% Hyperplastic Polyps (HP), and 34% tubular adenoma (TA)]. Presentation of polyps (SSL, TA, HP) was randomized with 50% of the polyps being displayed with a CADx diagnosis, and other 50% without.

CADx identified 52.6% of SSL as HP and 47.4% as TA. CADx identified 68.4% of HP correctly and 82.1% of TA correctly. The accuracy of the endoscopists for SSL did not differ significantly when the CADx diagnosis was displayed compared to no CADx being displayed (64.1% vs 64.9%, p > 0.05). The diagnostic accuracy for HP and TA improved among endoscopists when a correct CADx diagnosis was displayed to the endoscopists compared to when no CADx information was displayed but worsened when the displayed CADx diagnosis was incorrect.

Current CADx systems when displaying correct diagnostic outputs for HP or TA improve diagnostic accuracy of the endoscopists. CADx systems not trained for SSL recognition do not significantly affect SSL recognition among endoscopists even if an incorrect diagnostic output (TA or HP) is displayed.