Research On Computer Aided Diagnosis Methods Of COVID-19 Based On CT Images

SARS-associated Coronavirus 2(SARS-Co V-2)is characterized by high infectiousness,severe pathogenicity,high concealment and long incubation period.The COVID-19 caused by being infected with SARS-Co V-2 has given enormous impact on the public healthcare system and economic activities worldwide,it also poses severe tests on how to protect human health and life safety at the same time.Therefore,rapid diagnosis of patients with COVID-19 will act an essential part in tracking or controlling the progression of the disease and precise treatment of patients.The clinical outcome time of nucleic acid RT-PCR test is usually long and the positive rate of detection is low.As a routine diagnostic tool for pneumonia,CT images can provide the assistance for screening patients and evaluating the progress of disease.While the different types of pneumonia have multiple pulmonary infections and similar radiological features,and it is a labor-intensive and challenging task to use CT images for manual screening lesion areas.In this thesis,based on the relevant researching development about auxiliary diagnostic methods of COVID-19,with the purpose of ensuring and improving the efficiency and reliability of COVID-19 diagnosis and lesion location,the author investigates the computer aided diagnosis of COVID-19 based on deep learning by algorithm analysis,model design and experimental verification.A clinically-oriented preprocessing method for lung CT images based on the multi-instance learning strategy is proposed.The proposed preprocessing method can represent a complete CT image through a bag composed of a set of CT axial planes,which can ease the problem with uneven data format and quality caused by different data sources.A multiple instance diagnostic model based on the similarities adaptive pooling(SAP)layer and a lesion location model based on the attention mechanism are established respectively.The SAP layer in the diagnostic model can strengthen the learning of the features of instances with higher distinguishing abilities,and weaken the influence of instances with lower distinguishing abilities.The lesion attention module in the lesion location model can establish connection among image global information to focus on task-related areas.On the basis of two single-task models,a multi-task model of diagnosis and lesion location with adaptive task weight is established to perform the task of classification and segmentation at the same time,and it can adaptively adjust the contribution ratio of tasks to optimize the parameters of model.A quantitative assessment method of lung lesions is proposed based on the results of lesions location,which calculates the proportion of lesion areas to lung areas by quantitative analysis,provides the infection ratio in CT images and the progress of the patients’ condition.In addition,in view of the black box of neural networks,a posterior method of diagnosis results based on similarities is proposed.It can select the instances which play a decisive role in determining the diagnosis result according to the attribute similarities,and intuitively explain the basis for the model to make predictions.The above proposed methods have performed ablation and comparison study on a real COVID-19 CT images dataset CC-CCII,and the results suggest the effectiveness and advancement of proposed auxiliary diagnostic methods of COVID-19.