Research On The Detection Of COVID-19 Based On CT Images

Computed tomography is currently one of the most effective auxiliary diagnostic methods for chest imaging in clinical practice.However,manual analysis of CT images often requires high professional knowledge,and the analysis of image features is timeconsuming and labor-intensive.Using deep learning to identify and diagnose CT images can not only improve the efficiency of diagnosis,but also avoid misdiagnosis and missed diagnosis to a certain extent.This thesis hopes to obtain a model with good classification effect on CT images of COVID-19 through deep learning methods,so as to help clinicians identify and diagnose new coronary pneumonia as soon as possible.The main content of this thesis is to improve the classification accuracy of COVID-19 CT images by improving the convolutional neural network structure and transfer learning.In the way of improving the convolutional neural network,this thesis adopts the way of improving the classification layer structure of the convolutional neural network and the way of embedding the SE module and the CBAM module in the convolutional neural network.Among them,the three models of Inception V3,Res Net50 and Dense Net169 in the experiments of improving the classification layer structure of the convolutional neural network all showed good performance.In the classification of the SARS-Co V-2 CT dataset,the accuracy rate has reached about 95%,which is improved compared to the original model.For the improved Res Net50 model,SE module and CBAM module are embedded,and its classification accuracy has been further improved.When the amount of data is not sufficient,transfer learning can often reflect better learning performance.This thesis studies the effect of the transfer learning method based on convolutional neural network in the classification of Corona Virus Disease2019 CT images.Among them,the fine-tuned transfer learning training accuracy of Inception V3,Res Net50,and Dense Net169 model has reached more than 97%,and the classification accuracy of fine-tuned transfer learning training of VGG19 model is also91.54%.Through ensemble learning of Inception V3,Res Net50,and Dense Net169 models,the ensemble model obtains the best classification effect in this thesis,and the classification accuracy rate is 98.59%.For the COVID-CT image dataset with a smaller amount of data,this thesis proposes a transfer learning method based on a combination of fine-tuning and feature extraction.The SARS-Co V-2 CT dataset is used as the source domain for transfer learning,and the transfer learning method is compared directly with Image Net data.The experimental results show the effectiveness of this method.Finally,this thesis uses the Flask framework to build a COVID-19 detection system to detect the uploaded COVID-19 CT images,providing a simple and fast COVID-19 detection method.