Detection And Segmentation Of Pneumonia And COVID-19 Based On Deep Convolutional Network From X-Ray Images

Pneumonia is a respiratory tract infection caused by a variety of pathogenic bacteria,often manifesting as cough,fever,chest pain and other symptoms,and can lead to death in severe cases.Since 2019,the COVID-19 has ravaged the world and is still in the stage of a pandemic,which has caused great harm to the global economic situation and ecological environment.Doctors can diagnose whether a patient has pneumonia and COVID-19 or not by examining the X-ray images.However,in some areas,there is a lack of professional testing doctors and lack of necessary medical treatments,leading to problems such as high pressure on doctors to diagnose and long time consuming for diagnosis.Therefore,this study uses deep convolutional networks to assist doctors in classification and segmentation through X-ray images.The first part of the paper is the problem of common pneumonia binary classification(pneumonia/normal)based on X-ray images.For the problem that conventional deep learning models have a large number of parameters,a custom lightweight deep learning network model is designed to diagnose pneumonia.Secondly,in order to enhance the amount of medical image data and the richness of the data,and improve the generalization ability of the model for different data,methods such as dynamic histogram contrast enhancement are used to preprocess the X-ray images.Then,the cross-validation method is used to systematically compare the influence of different input image sizes and loss functions on the detection results,and select the optimal parameter combination.Through experimental comparison,although the model proposed in this paper has few parameters,the classification performance can reach or even exceed conventional mainstream deep learning models,and the final classification accuracy rate exceeds 95%.The second part of the paper is the three-category(COVID-19/pneumonia/normal)problem of COVID-19 and common pneumonia based on X-ray images.For the problem that X-ray images have similar characteristics to various lung diseases,the selective kernel attention mechanism is first studied and improved which is added to each hidden layer of SENet,and the performance of each position is compared to construct the most optimal attention mechanism network model.The experimental process adopts the method of transfer learning,pre-training on the contrast-enhanced data,and then fine-tuning on the original data without contrast-enhancing.The classification accuracy of the models proposed in this section generally exceed 93%.In addition,in order to have a deeper understanding of the detection process of the proposed model,the visualization results of the internal feature maps and convolution kernels of the network during the training process and the class activation maps are shown.The segmentation of lung parenchyma also plays an important role in assisting doctors to detect lung diseases.The third part of the paper is the segmentation problem of lung parenchyma.Combining the improved attention mechanism of the second part with the UNet medical segmentation network,a U-Net segmentation network framework combined with the improved attention mechanism is proposed,and finally a Dice coefficient of more than 0.98 is obtained.At the same time,the visualization results of each model segmentation are shown.It can be seen from the experimental results that the application of deep learning models to X-ray imaging disease detection and segmentation has high feasibility and effectiveness,which can reduce the burden on doctors and overcome potential problems in detection.