REAL-TIME COMPUTER-AIDED DIAGNOSIS SYSTEM FOR OPTICAL DIAGNOSIS OF DIMINUTIVE COLORECTAL POLYPS INCLUDING SESSILE SERRATED LESIONS: A PROSPECTIVE, MULTICENTER STUDY WITH BENCHMARKING AGAINST SCREENING ENDOSCOPISTS

Aims We aimed to develop and validate a robust computer-aided diagnosis (CAD) system, designed to use in real-time colonoscopy, to improve the accuracy of endoscopic characterization of diminutive polyps, including sessile serrated lesions (SSLs).

Methods We developed a CAD system (POLyp Artificial Recognition [POLAR]) to characterize diminutive colorectal polyps during live endoscopy, using a maximum of three non-magnified narrow-band imaging images. For pre-training the Microsoft-COCO dataset with a variety of object images (>300k) was used. For training, the prospectively collected data from 8 hospitals were used (2.637 images from 1,339 polyps). For clinical validation, POLAR was tested during colonoscopies in a fecal immunochemical test (FIT)-screening setting, and compared with the performance of 20 endoscopists from 8 hospitals. Primary outcome was the accuracy of differentiating neoplastic (i.e. adenomas, SSLs) from non-neoplastic (i.e. hyperplastic polyps) diminutive polyps by POLAR, compared with the accuracy of endoscopists. Histopathology served as reference standard.

Results During clinical validation, a total of 429 diminutive polyps detected in 195 FIT-positive patients were included for analysis. POLAR differentiated neoplastic from non-neoplastic lesions with 79% accuracy, 89% sensitivity and 37% specificity, while endoscopists achieved 83% accuracy, 93% sensitivity, and 44% specificity. No significant difference was observed in optical diagnosis accuracy between POLAR and endoscopists (P=.07). Success rate for acquiring a histological prediction by POLAR was 98%.

Conclusions We developed, validated, and benchmarked a trustworthy CAD system for optical diagnosis of diminutive polyps during real-time colonoscopy. The system differentiated neoplastic from non-neoplastic diminutive polyps with an accuracy comparable to screening endoscopists, with near-perfect technical efficacy.

Publication History

Article published online:
14 April 2022