Classification Of COVID-19 CT Images Based On Deep Learning

According to the World Health Organization,the novel coronavirus pandemic is putting healthcare systems around the world under unprecedented pressure.At present,the gold standard for diagnosing COVID-19 is RT-PCR testing,but due to the limited sampling method,high false negative rate,and shortage of resources,chest CT testing has become an effective method for clinically assisted diagnosis of COVID-19.However,with the increase in suspected cases,relying solely on radiologists to manually segment a large number of CT images is a serious challenge,and there is an urgent need to develop ways to automatically classify COVID-19 infections.With advances in computer algorithms,especially artificial intelligence,the detection of such viruses at an early stage will help with rapid recovery and help relieve stress on the healthcare system.As the core technology of emerging artificial intelligence in recent years,deep learning has been reported to have significant diagnostic accuracy in medical imaging.Based on COVID-19 CT images,this dissertation studies the classification of COVID-19 images based on deep learning,and the main research content is carried out from the following aspects:(1)Aiming at the problems of lack of samples in public datasets and the inefficiency of traditional machine learning algorithms,a COVID-19 CT image classification method based on improved U-Net network is proposed.In this model,the CT dataset of more samples is first enhanced by conditional generation adversarial network to obtain more samples,so as to reduce the risk of overfitting,and an improved U-Net network based on BIN residual block is proposed for image segmentation,which increases the convergence speed of the model and improves the generalization ability of the model,and then combines multilayer perceptrons for classification prediction.By comparing with Alex Net,Google Net and other network models,it is concluded that the BUF-Net network model proposed in this dissertation has the best performance and achieves an accuracy rate of 93%.Finally,the output of the system can be visualized using Grad-CAM technology,which can more intuitively illustrate the important role of CT imaging in diagnosing COVID-19.(2)Aiming at the fact that most of the current algorithms only focus on distinguishing between healthy patients and COVID-19 patients,ignoring the distinction between conventional pneumonia and COVID-19 patients,a weak label COVID-19 detection method based on the integrated network model is proposed.Using a deep learning integration model combined with Google Net,Alex Net and Res Net50,the output of all pre-trained models is combined,merged into integrated prediction,and the data imbalance problem is solved by three data processing methods,and finally the classification area is generated by Grad-CAM++,which can observe the classification results more clearly.Exploring the potential of deep learning-based methods to automatically detect COVID-19 in patient-level chest CT images using weak markers,the integrated network model achieved an F1 value of92.7%,based on experimental evaluation results,the weighted cost function method can be used to solve the problem of data imbalance well,and proved that the neural network integration method combining multiple deep architectures is more efficient than the structure based on a single model.(3)The method proposed in this dissertation is applied to the field of CT image classification of novel coronavirus pneumonia.First,the COVID-19 CT image dataset was preprocessed to remove the effects of some artifacts.Then the proposed method is applied to the COVID-19 CT image dataset,and through comparative experiments with several existing methods,it is proved that the proposed algorithm can correctly classify patients with COVID-19,which is of great significance in the field of practical medical applications.