Clinical Study on Gastroscopy Image Recognition Model Based on Artificial Intelligence in Diagnosis of Chronic Atrophic Gastritis
10.3969/j.issn.1008-7125.2020.10.003
- Author:
Wenqiao HUA
1
;
Xin GUAN
1
;
Xiangjun JIANG
1
Author Information
1. Department of Gastroenterology, Qingdao Municipal Hospital Affiliated to Qingdao University
- Publication Type:Journal Article
- Keywords:
Artificial Intelligence;
Deep Learning;
Diagnosis;
Endoscopy;
Gastritis, Atrophic
- From:
Chinese Journal of Gastroenterology
2020;25(10):588-593
- CountryChina
- Language:Chinese
-
Abstract:
Background: Artificial intelligence (AI) is a research hotspot in various fields of medicine at present. Its powerful image recognition and processing ability make it having a strong advantage in the field of digestive endoscopy. Aims: To construct a gastroscopy image recognition system based on AI, and to explore its value in the diagnosis of chronic atrophic gastritis (CAG). Methods: A total of 3 813 gastroscopy images were collected from patients who underwent gastroscopy and biopsy for pathological examination from April 2018 to August 2020 at Qingdao Municipal Hospital, including 1 927 images of CAG and 1 886 images of chronic non-atrophic gastritis (CNAG). Among them, 3 055 images were selected as training set (CAG/CNAG, 1 541/1 514) and 379 images (CAG/CNAG, 193/186) as the adjustment set, the remaining images as the test set. Deeping learning model was trained and verified. The receiver operating characteristic curve (ROC curve) and P-R curve were calculated. The sensitivity, specificity and accuracy of deep learning model and that of 3 less experienced endoscopists, 3 experienced endoscopists for diagnosis of CAG were compared. Results: The area under ROC curve of deep learning model for CAG was 0.916 8, the area under P-R curve was 0.931 6, and sensitivity was 89.1%, specificity was 74.2%, accuracy was 81.8%. The sensitivity, specificity and accuracy of deep learning model were superior to the three less experienced endoscopists, and even superior to some of the experienced endoscopists. Conclusions: The CAG diagnostic model based on deep learning technology has high sensitivity, specificity and accuracy, and can effectively identify CAG and assist the clinical endoscopists to diagnose CAG in gastroscopy.
