Establishment and preliminary clinical verification of an artificial intelligence-assisted colorectal polyps classification system
10.3760/cma.j.cn311367-20200612-00387
- VernacularTitle:人工智能辅助结直肠息肉性质鉴别系统的建立与临床初步验证
- Author:
Peng PAN
1
;
Shengbing ZHAO
;
Rundong WANG
;
Zhaoshen LI
;
Yu BAI
Author Information
1. 海军军医大学长海医院消化科 国家消化系疾病临床医学研究中心，上海 200433
- From:
Chinese Journal of Digestion
2020;40(11):758-762
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To establish an artificial intelligence (AI)-assisted colorectal polyps classification system (AI polyps system) by using clinical big data, and to conduct the clinical verification.Methods:From June 2018 to June 2019, the colonoscopy images of polyps from 16 participating research centers were prospectively collected. The basic information of the polyps (location, size, shape and pathological biopsy results) in colonoscopy images of colorectal polyps were marked by senior colonoscopist, and the outline of the polyp was circled for the development of the AI polyps system. Taking pathological biopsy results of polyps as the gold standard, the sensitivity, specificity, and positivity predictive value (PPV), negative predictive value (NPV) and accuracy of white light model, narrow band imaging (NBI) model, the combination of white light and NBI model and colonoscopists′ identification of polyps were calculated respectively. Paired McNemar test and Kappa test were used for statistical analysis. Results:A total of 15 441 qualified colonoscopy images were collected, including 9 109 images in white light model and 6 332 images in NBI model. At laboratory level, the sensitivity, specificity, PPV, NPV and accuracy of white light model and NBI model in the identification of the polyps were 90.3%, 98.3%, 89.8%, 98.4%, 97.2%, and 90.5%, 92.5%, 92.3%, 90.6%, 91.5%, respectively. In clinical verification phase, a total of 78 polyps of 56 patients with colorectal polyps were enrolled. The sensitivity, specificity, PPV, NPV and accuracy of the white light model and NBI model in the identification of polyps were 70.3%, 82.1%, 78.8%, 74.4%, 76.3%, and 78.4%, 87.2%, 85.3%, 81.0%, 82.9%, respectively. There were no statistically significant differences between the diagnostic results of colonoscopists, the white light model, the NBI model and the results of pathological results (all McNemar test, all P>0.05), but the consistency were general and the Kappa values were 0.632, 0.525 and 0.657, respectively (all P<0.01). The Kappa value of combination of the white light and NBI model and the pathological results was 0.575, however the consistency was general, but the difterence was statistically significant (McNemar test, P=0.004). Conclusions:The established AI polyps system has a certain role in assisting diagnosis, but the accuracy still needs to be improved.
