COVID-19 CT Image Classification Based On Deep Learnin

Since its emergence in 2019,the COVID-19 has ravaged the word,causing a serious threat to the economic development,people’s lives and property safety,and social development of countries around the world.The effective solution to prevent and this highly infectious disease is early detection,early isolation and early treatment.Reverse transcription polymerase chain reaction(RT-PCR)assay has been considered the gold standard for the detection of COVID-19,but the method is time-consuming and has a high false-negative rate.The chest CT images of COVID-19 patients has specific lesion features and radiologists can quickly determine whether a person has COVID-19 through these features,but the huge workload may lead to misdiagnosis.Deep learning can effectively solve this problem,and its representative technology,convolutional neural network,can realize automatic classification of images,which is widely used in the classification and diagnosis of various disease images including breast cancer and lung nodules.Therefore,the development of COVID-19 diagnosis models using deep learning technology has become a hotspot in the field of medical image classification.However,the current deep learning-based CT image classification models for COVID-19 have low practicality.On the one hand,the scarcity of CT image datasets due to privacy issues makes the models inadequately trained,and many models are designed with deep convolutional neural networks as the backbone in order to pursue classification effects,which is inefficient.On the other hand,the existing COVID-19 classification models do not take into account the problem that the CT images of COVID-19 patients in the early stage are similar to the CT images of healthy people,which makes the models have low sensitivity in classifying chest CT images of such people and cannot makes effective diagnosis.In order to solve the above problems,this thesis launched a study on the classification of CT images of COVID-19 from two perspectives of model structure design and image feature processing based on deep learning techniques.The main research is as follows:(1)Aiming at the problems of scarce COVID-19 CT image data sets,high hardware requirements and low operation efficiency of models,a lightweight COVID-19 diagnosis model based on improved Mo Bile Net V2 is proposed.The Mobile Net V2 network is selected as the backbone of the model and improved by replacing the "inverted residual module" with a "multi-scale inverted residual module",which enables the module to extract multi-scale features and enhance the model’s ability to express global features,which improves the classification effect of the model.Adding L2 regularization to the loss function can prevent the model from overfitting,and then training the model by transfer learning to further accelerates the model training speed and alleviates the impact of insufficient dataset,finally selecting the optimal hyperparameters by Bayesian optimization algorithm.The lightweight model has a high operation efficiency while guaranteeing impressive classification results,with an accuracy of 92.08% on the SARS-COV-2 CT scan dataset provided by the Kaggle platform.(2)Aiming at the problem that the CT images of COVID-19 patients in early stage are similar to those of healthy people,a classification model of CT images of COVID-19 patients in the early stage based on COVID TCL is proposed.For the first time,triplet loss and center loss are applied in the field of medical image classification and combined with softmax loss to design a joint loss measure function COVID Triplet-Center loss,which increases the inter-class variability between different depth features and improve the intra-class compactness between the same features,making the extracted depth features more discriminating,and improving the model’s ability to classify the CT images of COVID-19 patients in early stage.At the same time,we use the XGBoost and Res Net50 to design a COVID-19 CT images classification model of CNN-XGBoost architecture,which not only reduces the overall parameter redundancy of the model,but also further improves the model classification accuracy.The classification accuracy on the SARS-COV-2 CT scan dataset provided by Kaggle platform is 97.41%,and the sensitivity is as high as 97.61%.