Deep Learning Model Construction And Case-oriented Identification Of Benign And Malignant Gastric Ulcers

Objective:The deep learning models VGG19,ResNet50 and EfficientNet-B4 were compared in the single image-oriented classification of benign and malignant gastric ulcer.Research method:According to inclusion and exclusion criteria,a total of 3238 cases were included in the white light endoscope image dataset.Based on the patient’s final pathology at endoscopy,the data set was divided into malignant gastric ulcer(gastric cancer),benign gastric ulcer(gastric ulcer),and chronic superficial gastritis(normal stomach).In the data set of this study,1730 cases of normal gastric disease(chronic superficial gastritis)included 3242 normal images,761 cases of benign gastric ulcer included 2346 images,and 747 cases of malignant gastric ulcer(gastric cancer)included 3599 images.Then the collected data is preprocessed according to the requirements,and the data is randomly divided into "training set,verification set and test set" in the ratio of "8:1:1".Firstly,VGG 19,ResNet50 and EfficientNet-B4 deep learning models with pre-training weights are selected as learning models.The single image-oriented training set is used for fine-tuning training,the validation set image is used to select the optimal parameters of the model,and the optimized model is tested with test set.Finally,the optimal model is EfficientNet-B4.Secondly,the efficientnet-b4 model based on case oriented multi picture input is established,and the model data is practiced with the help of training set;Verify the deep learning model through the verification set,and optimize and adjust the super parameters of the deep learning model;324 cases in the test set were input into the trained artificial intelligence model to obtain the diagnosis results of the disease.At the same time,two experienced endoscopists were invited to interpret and diagnose 324 cases in the test set.The overall accuracy,sensitivity and positive predictive value of deep learning model and two doctors in disease diagnosis were obtained.:First,the results of EfficientNet-B4 single image recognition are compared by model VGG19,ResNet50,we find that EfficientNet-B4 was superior to ResNet50 and VGG 19 in overall accuracy,class accuracy and class recall rate.Secondly,based on the EfficientNet-B4 model,a deep learning model based on case multi-image input is established.The training model was obtained through training,verification and tuning,and was allowed to identify 324 cases in the test set and compared with two experienced endoscopists.The overall accuracy of deep learning artificial intelligence model in case-based identification of normal gastric mucosa,gastric ulcer and gastric cancer cases was 95.06%.The results were significantly better than those of the two endoscopists:92.09%and 91.36%.Research conclusions:EfficientNet-B4 deep learning model for case-oriented multi-image input,Through the training of gastroscopic images on a case-by-case basis,The patients with malignant gastric ulcer and benign lesions(benign gastric ulcer and chronic superficial gastritis)could be distinguished well,and the overall accuracy of identification was better than that of two experienced endoscopists.