Research And Implementation Of COVID-19 Lesion Detection Model Based On Deep Learning

During the global epidemic of COVID-19,how to improve the resolution and the localization accuracy between COVID-19 and common pneumonia remains a research hotspot in the scientific research field.In the traditional imaging diagnosis of pneumonia,X-ray,CT,MRI and other forms of images are used as the detection objects,among which the X-ray generated images are cheap and the radiation amount is low,so X-ray imaging detection is widely used in remote areas and relatively backward third world countries.With the rapid development of deep learning research in recent years,it has become the mainstream research direction to classify medical images of pneumonia and detect the lesion area by using deep learning.Since most of the existing deep learning models are trained through non-medical public data sets,it is difficult to generalize their models to the COVID-19 data sets.Even if the pneumonia public data sets are trained,the accuracy of the deep learning model will decline due to the particularity of the data and the small scale of the data set.If too many convolution layers are added to the model,the nodet will produce overfitting problem.This thesis proposed CPC-VGG19 pneumonia lesions detection.The model improved the VGG19 model by adding baatch standardization layer and L2 regularization of the model,and then conducted comparative learning pre-training,saving the weight,improving the deep learning dependence on data sets and overfitting problems,improving the recognition rate of pneumonia image classification,and then The classified pneumonia images were input with the Xception modelof capsule network to further detect the pneumonia lesion area and output the location information.The main work of this thesis is as follows:(1)Classification of COVID-19 images based on CPC-VGG19 modelA CPC-VGG19 model based on contrast learning was proposed to classify the X-ray images of COVID-19.In the first step,the mage is preprocessed and the data set is expanded by rotation.The second step is to reduce the noise in the data set by filtering algorithm and so on,and to enhance the data by median filtering.The third step is to improve the VGG19 model by adding a loatch normalization layer and applying L2 regularization to the entire model.In the fourth step,the improved model was pre-trained by comparison learning and self-supervised,and then the weight of the trained model was saved.Finally,the COVID-19 data set was used for classification experiment,and the accuracy of the experiment reached 94%.(2)Localization of pneumonia lesions based on CAPs-Xception modelIn order to further locate and detect the classified images,a CAPs-Xception model based on capsule network was proposed to locate the COVID-19 lesions.In the first step,the pneumonia images separated from the classification model were preprocessed.In the second step,Xception model is constructed,capsule network is added as feature extraction layer,and the Dropout layer is added to the capsule network to shield half of the vector neurons to prevent overfitting and further improve the localization and recognition rate of the localization model.The experimental results show that the improved model with capsule network can locate the lesion area with 92% accuracy and the Io U index reached 0.564.Through relevant experimental,the pneumonia detection model can improve the accuracy of classification and location,reduce the misdiagnosis rate,and improve the accuracy of auxiliary diagnosis.