Clinical Study Of Gastroscopy Image Recognition Model Based On Artificial Intelligence In The Diagnosis Of Chronic Atrophic Gastritis

ObjectiveChronic atrophic gastritis(CAG)is a common chronic inflammatory disease of digestive system,which is recognized by the World Health Organization as a precancerous disease.At present,the diagnosis of CAG mainly depends on endoscopic biopsy.However,under white light endoscopy,the coincidence rate of CAG endoscopy and pathological diagnosis is low,and the diagnosis of CAG under endoscopy depends on the level of endoscopists.Artificial intelligence(AI)technology has a strong ability of visual recognition and data processing.It can be applied in the field of digestive endoscopy to assist doctors in endoscopic diagnosis of diseases and reduce missed diagnosis of diseases.This study is based on deep learning to build a gastroscope image recognition system,evaluate its diagnostic performance,and explore its clinical application value in CAG endoscopic diagnosis.MethodsWe selected patients who underwent gastroscopy and pathological examination in the Endoscopy Center of Qingdao Municipal Hospital from April 2018 to August2020,and collected their endoscopic images.In addition,we collected the basic information of patients and pathological results for statistical analysis.A total of 3813 endoscopic images of 1563 patients were included,including 1927 CAG images and 1886 images of Chronic non-atrophic Gastritis(CNAG).According to the 8:1:1 ratio,the images were randomly divided into training set,adjustment set and test set.3055 images(CAG1541,CNAG 1514)were selected as the training set to train the model and automatically learn the features.379 images(CAG 193,CNAG 186)were used as adjustment sets to verify the model performance and adjust the model parameters according to the results.The remaining 379 images are test sets,used to test the diagnostic performance of the final model.Finally,we drew the receiver operating characteristic curve(ROC)and P-R curve of the model,and calculated the AUC under ROC curve,AP under P-R curve,sensitivity,specificity,accuracy,Positive Predictive Value(PPV)and Negative Predictive Value(NPV)of the deeplearning model.At the same time,3 endoscopic physicians with low seniority and 3 endoscopic physicians with high seniority were invited to diagnose the images of the test set,and the diagnostic results of the physicians and the model were compared to evaluate the diagnostic performance of the model.Results1.A total of 802 CAG patients were included,including 407 males and 395 females,aged from 26 to 89 years old,with an average of(59.72±9.99)years old;There were 761 CNAG patients,including 376 males and 385 females,aged from 27 to 88 years old,with an average of(58.32±10.20)years old.Among the CAG patients,there were 519 patients with mild mucosal atrophy,227 patients with moderate atrophy,56 patients with severe atrophy,793 patients with intestinal metaplasia,and 9 patients without intestinal metaplasia.Among the patients with CNAG,236 patients were associated with intestinal metaplasia,while 525 patients were not.2.Among the 1927 endoscopic images of CAG,there were 1250 cases of mild atrophic gastritis,540 cases of moderate atrophic gastritis and 137 cases of severe atrophy of gastric mucosa;There were 1903 cases with intestinal metaplasia and 24 cases without intestinal metaplasia.Among the 1886 CNAG images,there were 579 with intestinal metaplasia,and 1289 without intestinal metaplasia.3.The final test results show that the AUC of the CAG image recognition model is0.9168,AP value is 0.9316,sensitivity is 89.1%,specificity is 74.2%,accuracy is 81.8%,PPV is 78.18%(172 / 220),NPV is 86.79%(138 / 159),recall is 89.1% and F1-score is83.3%.The average diagnostic time of one image was(30 ± 3)ms.4.In 379 test images,48 cnag images were mistakenly identified as CAG,the false positive rate was 25.81%,21 CAG images were mistakenly identified as cnag,the false negative rate was 10.89%.5.The sensitivity,specificity and accuracy of three junior endoscopists were 73.4%,59.3% and 66.5%,respectively;The sensitivity,specificity and accuracy of three senior endoscopists were 86.4%,68.8% and 77.7%,respectively.Compared with the diagnostic results of the model,the sensitivity,specificity and accuracy of the model were higher than those of the junior endoscopists(P< 0.05),reaching the level of experienced endoscopists.Conclusion1.The diagnosis model of chronic atrophic gastritis based on deep learning has higher area under ROC curve and P-R curve,which means the model can improve the diagnosis rate of chronic atrophic gastritis under endoscopy.2.The diagnosis model of chronic atrophic gastritis based on deep learning technology has high sensitivity,specificity and accuracy,which can effectively improve the endoscopic diagnosis and treatment level of young endoscopists and grassroots doctors.