Study Of CT-Based Auxiliary Diagnosis Method For Pneumonia

In recent years,the outbreak of COVID-19 has attracted wide attention,and CT is an effective tool for rapid screening of pneumonia.However,a large amount of CT data brings great diagnostic pressure to imaging doctors.Computer-aided diagnosis method can effectively improve the diagnostic efficiency of doctors.The auxiliary diagnosis of pneumonia is based on the classification,segmentation and quantitative analysis of pneumonia.For classification problems,existing methods feed the entire CT scan into the network for training.But slices not associated with pneumonia increased the amount of time spent online training.For the pneumonia segmentation task,most of the existing models can only process the data from the same machine,and the model lacks generalization ability.In addition,during the follow-up of COVID-19 patients,it was found that the lungs of COVID-19 patients were affected by fibrosis.There are some subvisual lesions that are not visible below the lung window,which are ignored in most studies.In this paper,deep learning algorithm is used to study the classification of pneumonia,segmentation of pneumonia lesions and segmentation of subvisual lesions.The research content is as follows:(1)Aiming at the task of pneumonia classification,this paper proposes an improved Resnet based pneumonia classification method.First,the CT selection algorithm was used to remove the closed lung sections in the CT scan so that the model could focus on extracting effective pneumonia features.Next,based on Resnet50 classification network,RReLU activation function is used to avoid the "neuron death" problem.In addition,due to the large differences in the shape and size of pneumonia,a characteristic pyramid mechanism was added to the network.By integrating the rich semantic features of the high level with the spatial location features of the low level,the multi-scale feature extraction is improved.In the data set of novel coronavirus pneumonia and immune pneumonia,the classification performance is effectively improved.(2)Aiming at the task of pneumonia segmentation,a novel pneumonia segmentation method based on 2.5D U-net model was proposed.Firstly,spatial normalization and signal normalization are used to embed CT into standard space so that the model can process CT data from different machines.2.5D U-net segmentation model carries out 2D U-net segmentation from x-y plane,x-z plane and y-z plane respectively.Three 2D segmentation results were fused to make the model learn the features inside slices and the spatial features between slices at the same time.On COVID-19 data sets,a balance between segmentation performance and accuracy was achieved.(3)A subvisual lesion segmentation method based on lung parenchyma enhancement is proposed for the subvisual lesion segmentation task.First,the lung tracheal vessels are extracted using a segmentation method based on the 2.5D U-net model with a two-stage protocol,which solves the problem of inaccurate segmentation of small targets in large scenes.Next,extraneous tissues such as tracheal vessels and visible lesions are removed from the lung parenchyma,while the optimal window width window position for observing subvisual lesions is calculated based on CT values,at which time the window significantly enhances the lung parenchymal abnormalities compared with the lung window.Subvisual lesions were identified in the enhanced CT images and quantified for analysis.