COVID-19 Image Classification Based On Convolution Neural Network And Transformer

With the rapid development of computer technology in the medical field,the research on the classification and recognition algorithm of novel coronavirus pneumonia image has become an important research direction in the computer-aided diagnosis medical image processing task.The image of novel coronavirus pneumonia has complex shape changes.It is not easy to observe hidden lesions in the early stage.Meanwhile,it is difficult to distinguish other viral pneumonia and bacterial pneumonia.In recent years,deep learning methods have been widely used in the field of computer vision and medical imaging,which have made good achievements.However,the current methods used for novel coronavirus pneumonia image classification have no obvious effect on feature extraction with complex shape changes.When processing high-resolution images,the classification effect is still low and the adaptability is not good.In order to further improve the classification effect and performance requirements,and accelerate clinical research and application,this thesis studies the image classification of novel coronavirus pneumonia based on the latest research achievements in the field of medical image processing.The main contents are as follows:(1)To solve the problem that the limited receptive field of convolutional neural network has poor ability to model the global information of medical images,in this thesis,based on the deep residual network ResNet50,a new network model(BoT-ViTNet)is designed to automatically classify Chest X-ray(CXR)images.This model introduces multi-headed self-attention in the last residual convolution of the first three stages of ResNet50 to enhance the ability to model global information.In addition,in the last stage of ResNet50,the introduction of TRT-ViT module can extract global and local feature information at the same time which enhance the expression ability of features and the correlation between the positions of features,and help to identify complex lesions in CXR images.Image classification is conducted on publicly accessed datasets and compared with other classification models.The experimental results show that this model is superior to its basic model ResNet50 and other classification models in many indicators.(2)In order to solve the problems of network load and feature information loss of high-resolution medical images,this thesis designs an efficient and unified network model of convolution and self-attention(DC-Transformer)to classify high-resolution pneumonia images into two and four categories.Firstly,the dilated depth separable convolution replaces the ordinary convolution to extract the local features of the image and perform the dual scale feature fusion;Secondly,the fused local features are input into the hybrid network of convolutional neural network fusion multi-headed self-attention mechanism to enhance the ability to extract the global features of the image.Finally,the obtained feature output is used for image classification.In the training process,residual connections and transfer learning methods are introduced to prevent network degradation and make the model converge quickly.Experiments show that the network model designed in this thesis has good classification performance.Figure[28] table[11] reference[90]...