Artificial Intelligence in the Characterization of Colorectal Polyps: A Prospective Study In a Clinical Setting Using Cadeye

Aims Current guidelines recommend resection and histopathological analyses of all colorectal polyps. Real-time optical diagnosis can obviate non-neoplastic polyp resection (“diagnose-and-leave-behind”) and histopathological analyses of diminutive polyps (“predict-resect-and-discard”) reducing healthcare and cost burden. We aimed to evaluate the diagnostic accuracy of computer-aided diagnosis using CADEYE (Fujifilm, Germany) in real-time optical characterization of colorectal polyps compared to endoscopic diagnosis with histopathology as the gold-standard.

Methods Single-centre prospective study of diminutive/small colorectal polyps, between September-November/2020. Thirteen participating endoscopists were previously submitted to a brief online course of Blue Laser Imaging (BLI) chromoendoscopy. First, two independent endoscopists performed a blind optical characterization, with high-definition colonoscopy without amplification using BLI. Second, CADEYE on BLI mode was applied for optical characterization. Third, all polyps were resected and submitted to a blind histopathological evaluation by two independent gastrointestinal pathologists.

Results A total of 159 polyps (mean size 5.0±2.4mm;2-9mm) were included, being 115 (72.3 %) adenomas and 44 (27.7 %) non-adenomas by histopathology. Regarding neoplastic/hyperplastic polyp differentiation, CADEYE had a diagnostic accuracy of 81.1 % versus 82.1 % (p=0.132) compared to experienced endoscopists with sensitivity, specificity, positive and negative predictive values of 78.4 %, 83.7 %, 92.9 % and 59.0 % for CADEYE and 75.9 %, 88.4 %, 94.6 % and 57.6 % for experienced endoscopists. After excluding polyps with low-confidence characterization, diagnostic accuracy was 83.9 % versus 87.3 % (p=0.014) for CADEYE and experienced endoscopists, respectively.In the trainee setting (148 polyps), diagnostic accuracy was 77.6 % (versus 80.4 % for CADEYE; p=0.201); considering high-confidence characterization, the diagnostic accuracy was 84.0 % (versus 83.3 % for CADEYE; p=0.705).

Conclusions CADEYE is an user-friendly tool with high accuracy in optical characterization of colorectal polyps, that may not be superior to high-confidence characterization by the endoscopist. However, both endoscopic and CADEYE optical characterization are still under the minimum desirable target (90 %PIVI). Ongoing optimization of artificial intelligence technology may allow future clinical implementation.

Citation: Correia C, Gravito-Soares E, Gravito-Soares M et al. eP184 ARTIFICIAL INTELLIGENCE IN THE CHARACTERIZATION OF COLORECTAL POLYPS: A PROSPECTIVE STUDY IN A CLINICAL SETTING USING CADEYE. Endoscopy 2021; 53: S156.

Publication History

Article published online:
19 March 2021