Application Of Deep Neural Network In Diagnosis And Treatment Of Early Gastric Cancer And Pre-Cancerous Lesions

Objective:In view of the difficulty of diagnosis of pre-cancerous lesions and early cancer screening under endoscopy,and the uneven limitation of diagnosis level by doctor’s experience and regional development level,this study uses deep neural network as the main tool,supplemented by traditional computer vision and pattern recognition algorithm.Method:1,Based on the convolutional neural network training model implemented by the caffe framework;2,Develop a high-quality endoscopic image acquisition standard;3,Carry out gastroscopy image labeling and preprocessing;4,Establish an early cancer gastric cancer and precancerous lesion endoscopic image database;5,Build a deep neural network based gastric early cancer detection artificial intelligence model and atrophic gastritis detection and typing artificial intelligence model,and evaluate the model;6,The results of the early detection model and human gastric cancer doctors to evaluate the model.Results 1,Efficacy assessment of the AI model for gastric early cancer detection : The internal validation set sensitivity of the convolutional neural network system(CNN)to diagnose gastric early cancer is 95.5 %,The specificity is 81.7 %,The AUC is0.940;Three external validation sets whose sensitivity is 85.9 %-92.1 %,Specific to84.4 %-90.3 %,AUC 为0.887-0.925.CNN diagnosed early gastric cancer,In intra epithelial neoplasia,Sensitivity:95.8 %,Specific to 80.3 %,AUC 0.938;’s In intra mucosal carcinoma,Sensitivity : 95.6 %,Specific to 82.1 %,AUC 0.946;’s In sub mucosal carcinoma,Sensitive properties : 95.0 %,Specific to 83.7 %,AUC0.937;’s The experiments showed higher sensitivity of CNN systems(93.0 %)than expert level(82.7 %)and beginners(50.2 %).The CNN systems had better sensitivity and specificity than the two groups of endo do scopes.Although the experts and the trained group achieved considerable specificity,But the panel is relatively highly sensitive.The sensitivity of the experts has increased by 14.7 % after using the CNN,Beg increased sensitivity by 0.6 %.2,Detection of CNN-based atrophic gastritis and efficacy evaluation of classified AI model :(1)White-light endoscopic diagnosis of atrophic gastritis,The CNN’s sensitivity to the diagnosis of atrophic gastritis was 85.7 %,Specific is 72.5 %,by pixel to distinguish the boundaries of atrophic gastritis using the image segmentation model,The CNN sensitivity to atrophic gastritis reached 92.1 %,The specificity is 85.7 %.(2)The anatomical recognition rate of each part in the classification of Mucun bamboo in atrophic gastritis is higher,They were: cardia 95.7,82.3% in the upper part of the stomach,83.5% in the lower part of the stomach,Antrum 94.2%,99.8% of the stomach horn,92.3% of stomach floor,pylorus 94.4%,87.5% in the upper middle of the inverted gastroscope,The lower part of the inverted gastroscope was 89.4%,Reverse mirror cardia 92.3.And the combination of single frame image atrophy and automatic recognition of anatomical location,The complete model of wood village bamboo classification is realized.3,The early gastric cancer detection model is deployed to the auxiliary software of endoscopic hardware(hardware equipment with artificial intelligence auxiliary diagnosis model).Conclusions The convolution neural network system(CNN)can effectively improve the diagnostic efficiency of early gastric cancer,with the help of CNN,can effectively improve the diagnostic specificity of experts and beginners to early gastric cancer,and can CNN improve the diagnostic efficiency of atrophic gastritis.CNN can carry artificial intelligence platform connected with endoscopic equipment.