Research On Detection Method Of COVID-19 CT Image Based On Deep Learning

The Corona Virus Disease 2019（COVID-19）is a new type of respiratory infectious disease.Its emergence poses a serious threat to the survival of mankind.Due to the shortage of nucleic acid testing reagents in many areas during the outbreak of COVID-19,and nucleic acid testing errors also occur from time to time.The lung CT images of patients with COVID-19 are specific.If they can be correctly interpreted,it can be an important supplement to nucleic acid detection and improve the accuracy of the diagnosis of COVID-19.However,due to the complexity and large quantity of such images,the competent doctors are relatively insufficient,and the current diagnosis efficiency and accuracy cannot meet the needs.Therefore,people urgently need new technologies and methods for their auxiliary diagnosis.As an important part in the field of artificial intelligence,computer vision technology is expected to be able to recognize and understand images like human eyes and brains,so it has been widely and successfully applied in the field of medical image processing.In the current critical situation of COVID-19,research on COVID-19 detection,diagnosis methods and technologies based on deep learning not only has a good foundation and feasibility,but is also very necessary and urgent.Based on the above situation,this thesis has conducted a systematic and in-depth study on the segmentation and classification methods of COVID-19 CT images based on deep learning.The specific work mainly includes:Firstly,we constructed a segmentation data set consisting of 5000 lung CT images of COVID-19 patients and corresponding four-label masks and a classification data set consisting of 10,000 normal lung CT images,10,000 COVID-19 lung CT images,and 10,000 other pneumonia lung CT images.Secondly,we constructed a model called Lung Seg-Net to pre-segment the lung CT image,remove the background area except the lung area in the CT image,and lay the foundation for the subsequent work.After that,in the segmentation data set,the U-Net,U-Net++,U-Net+Res Net101 and Deep Lab V3+ models were used to perform segmentation experiments respectively.Then we adjusted the Deep Lab V3+ model by adding convolutional block attention module and depthwise separable convolution to improve the performance of the model.Finally,three supervised classification models,Mobile Net V1,VGG16 and Res Net50,were used in the pneumonia classification task.The SSGAN model based on GAN was introduced into the pneumonia classification task to achieve the purpose of accurately classifying images with a small amount of label data.In this thesis,the improved methods proposed were all verified,and compared with a variety of similar methods to prove their feasibility and advancement.