Adoption value of deep learning and serological indicators in the screening of atrophic gastritis based on artificial intelligence

Abstract

This work aimed to improve the early clinical diagnosis rate of atrophic gastritis (AG) and reduce the risk of disease deterioration or cancerization. Three hundred and eight patients with gastric disease were taken as the research object, who were divided into two groups: AG (n = 159) and non-AG (n = 149), according to the diagnosis results. The gastric antrum images of patients were collected, and the DenseNet model for gastric antrum image lesion screening was improved. Then, the differences in serum pepsinogen (PG I and PG II) of patients were detected, and the efficiency of different methods to screen AG was compared. The results revealed that the levels of PG I and PG II in AG patients were substantially reduced, and the sensitivity (70.44%), specificity (66.44%), and accuracy (68.51%) of AG diagnosis by indicator PG I were higher than that of PG II and joint diagnosis. The diagnosis accuracy rate of AG based on the improved DenseNet model was 98.63%. The accuracy of model recognition combined with serological indicators for disease diagnosis was as high as 99.25%, with a sensitivity of 96.17% and a specificity of 94.33%. In summary, the combination of deep learning-based image recognition methods and serological specific indicators could improve the clinical diagnosis rate of AG, which could provide a reference for the subsequent clinical adoption of artificial intelligence recognition technology.