Establishment and evaluation of artificial intelligence image marking method for magnetically controlled capsule gastroscopy
10.3760/cma.j.cn311367-20210615-00337
- VernacularTitle:磁控胶囊胃镜人工智能标注方法的建立与评价
- Author:
Lijuan FENG
1
;
Lin TIAN
;
Qian ZOU
;
Zhongming DAI
;
Xiaojuan TIAN
;
Gongli YANG
;
Jingfeng DU
;
Mengqi XIANG
;
Yu MENG
;
Long XU
Author Information
1. 深圳大学总医院消化内科，深圳 518055
- Keywords:
Magnetically controlled capsule gastroscopy;
Artificial intelligence;
Image marking;
Method
- From:
Chinese Journal of Digestion
2022;42(1):14-18
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the marking method for magnetically controlled capsule gastroscopy (MCCG) pictures with artificial intelligence (AI), so as to improve the work efficiency of endoscopist and to reduce the blind area of AI image reading.Methods:According to the consensus of MCCG, 24 parts of stomach in 14 775 pictures of MCCG from 35 subjects in Shenzhen Zifu Medical Technology Co., Ltd received MCCG from March to August, 2020 were marked by ten gastroenterologists and one developer of MCCG with medical background, the marking shape included rectangles and polygons. Among the ten gastroenterologists, three were senior endoscopist (the total number of gastroenteroscopy operations over 80 000, chief physician or associate chief physician), four were medium seniority endoscopist (the total number of gastroenteroscopy operations between 10 000 and 80 000, associate chief physician), and three were junior endoscopist (the total number of gastroenteroscopy operations less than 10 000, attending physician). The pictures of the same subject were pre-marked by two selected senior endoscopists with blind method, and the standard of marking with most appropriate coincidence rate was determined. The qualified marked pictures were automatically learn with AI deep learning method, and the learning results were fed back. Chi square test was used for statistical analysis.Results:According to the pre-marked results, the standard of coincidence rate for rectangular marking area was set as 50.0% and that for polygon marking area was 70.0%. The first correction for qualified rate was 39.0% (5 762/14 775). A total of 9 013 pictures were corrected. After repeated training and correction for one to five times, all pictures were qualified marked. The marking qualified rate of senior endoscopist partners was higher than that of partners of different qualifications (48.7%, 1 200/2 466 vs. 19.0%, 825/4 337), and the difference was statistically significant ( χ2=659.20, P<0.001). There was no statistically significant difference in the marking qualified rate between the senior endoscopist partners and partners of senior endoscopist and capsule developer (48.7%, 1 200/2 466 vs. 49.6%, 1 496/3 019; P>0.05). Conclusions:Establishment of AI marking method for MCCG can provide technical support for AI non-blind area reading, and AI non-blind area monitoring during the operation of MCCG.
