Objective:To investigate the diagnostic accuracy of an artificial intelligence(AI)based chronic atrophic gastritis identification model in a real-world clinical environment.Methods:Patients undergo-ing gastroscopy at Wuhan Jiangxia District Hospital of Traditional Chinese Medicine from August 2022 to August 2023 were selected as a real-world test set to evaluate the generalization ability of the model.Endoscopists marked the biopsy sites according to the pathological results as chronic atrophic gastritis(CAG)and chronic non-atrophic gastritis(CNAG)and divided them into six anatomical re-gions:middle-upper body in retroflex view,fundus,antrum,angulus,middle-upper body in ante-grade view,lower body in antegrade view.The evaluation indicators included the recognition sensitivity,specificity,and accuracy of all cases and regional sampling.At the same time,three en-doscopists with different experience levels(low,medium,and high)were selected for a human-machine comparison,and the AI model's CAG identification ability was compared with that of endoscopists.Results:A total of 2 248 clear gastric cavity samples(CAG/CNAG,1 165/1 083)from 200 patients were collected as a real-world test set.On this test set,the AI model's recognition sensitivity,specificity,and accuracy were 92.27%(1 075/1 165),88.73%(961/1 083),and 90.57%(2 036/2 248),respectively.Compared with the results of three endoscopists with different experience levels,the identification accuracy of the AI model was equivalent to that of the senior en-doscopist and significantly better than that of the middle and junior endoscopists,with statistically sig-nificant differences.Conclusion:The AI model has stable diagnostic ability in clinical application and is equivalent to the level of senior endoscopists,which is expected to become an effective auxiliary tool for endoscopists.
NETL
NSTL
万方数据
