A PROSPECTIVE, MULTI-CENTER VALIDATION STUDY FOR AUTOMATED POLYP DETECTION AS A SECOND OBSERVER

Aims Last year, we presented a deep learning framework for automated polyp detection. Contrary to classical CNNs, we use ‘memory cells’ enabling more accurate predictions. Little evidence is available on the performance of AI for polyp detection in clinical practice. The aim of this study is to prospectively validate our system in a multi-center clinical setting to obtain an estimate power calculation for future trial design and to assess preliminary performance compared to experienced endoscopists.

Methods Our system was trained with 131.619 frames from 825 polyps from 206 patients and was implemented in a bedside module for real-time analysis. In this study, an experienced endoscopist (ADR>35%) does not see the system output while a second observer looks at the AI-enhanced screen. We define four different situations: (1) Obvious false positive - the system gives an obviously false detection (stool, air bubbles, …). (2) Other positive - after its location disappears from the image, the endoscopist is asked to return. If there is a polyp, this is an additional detection. (3) False negative - the endoscopist found a polyp, but the system didn’t. (4) True positive - the system and endoscopist found the polyp.

Results Currently, 99/300 patients are included from three European centers. In total 199 polyps were found of which 181 were detected by the system and endoscopist. There were 13 false negatives (all diminutive) and 5 additional detections by the system. Combined, this corresponds to a 3% increase in polyps-per-colonoscopy. A low average of 1 false positive per minute was recorded.

Conclusions The interim analysis shows promising results for the clinical validation of a novel AI system. These exploratory studies are important to power future trials and to estimate the optimal trial design (non-inferiority versus superiority). We plan to include 300 patients by the ESGE days 2020 and present the full results.