Estimation Of Missing CT Projection Data Based On Multiple Deep Networks

As an effective way of medical treatment,X-ray computed tomography(CT)technology provides clear anatomical images of human body,which brings convenience to doctors in diagnosis and treatment.The complete CT scan has a high degree of ionizing radiation.Although low-dose CT scan can effectively reduce the radiation injury of patients,however,because the scanning range is not complete,the missing projection data is obtained,so the reconstructed CT image is not clear enough to affect the clinical treatment.Therefore,it is of great significance to reconstruct high quality CT images without CT projection data.In this paper,a new method based on the depth multiple analytic network is proposed to estimate the missing CT projection data.The method flow includes training phase and testing phase.In the training stage,the generated network uses the encoder based on convolutional neural network(CNN)to encode the input missing CT projection data,and then the decoder outputs the estimation results of the missing part or the truncated missing part of the limited angle.The estimation result of the missing part and its corresponding real projection data are used as the input of the local discriminator.The estimation results of the missing part and the real projection data of the non missing part are spliced to form the composite projection data.The composite projection data and the complete CT projection data are used as the input of the global discriminator.The discriminator determines whether the missing part of the projection data estimated by the birth network is the true probability.The trained generation network can effectively predict the missing CT projection data.Finally,the filtered back projection(FBP)algorithm is used to reconstruct the CT image from the synthetic projection data.In addition,for sparse CT projection data,this paper also proposes a method of missing CT projection data estimation based on multi discriminator multi analytic network.It is also divided into training stage and testing stage.Compared with the previous network model,a number of region discriminators similar to the generated network encoder structure are added to deal with the corresponding sparse missing region.The experimental results show that compared with the existing methods,the CT projection data generated by the method proposed in this paper is closer to the real value of CT projection data,the pixel value has a high consistency on the boundary of the missing area,and the reconstructed CT image is clearer,which can meet the clinical needs.