Study Of Deep Learning Based Super Resolution Image Reconstruction In Spectral CT

Multi spectral CT(Computer Tomograph,CT)uses photon counting detector to directly convert optical signal into digital signal,which can obtain images of different energy bins.It can use K-edge imaging to reduce radiation or contrast agent dose,and can also use multi spectral characteristics to improve soft tissue contrast.However,it is difficult to distinguish the substance from the background when the substance concentration is low.When two kinds of substances with very close atomic numbers are very close,they will be mixed together and difficult to distinguish in the image.The super-resolution image reconstruction aims to improve the image resolution and solve the problem that it is difficult to distinguish the material from the background and the material with similar atomic number.It is a research hotspot of multi spectral CT imaging in recent years.Based on the research of super-resolution reconstruction technology in the field of natural image,this paper proposes a dual network model of global feature fusion Unet(Global Feature Fusion Unet,GFFUnet)and efficient sub pixel convolutional network(Efficient Sub-Pixel Convolutional Network).Compared with other network models,such as residual network,it is simpler and more effective.In order to make the multispectral CT image more clear in the part with filling material,according to the characteristics of multispectral CT image,this paper proposes two methods,that is,based on the pixel binary classification results and the material decomposition results to improve the preliminary reconstruction results.In this paper,the multi spectral CT image data collected by Philips machine in Creatis Laboratory is used for super-resolution image reconstruction research,and the F-norm of error matrix is used as the error to measure various methods.The main research contents and contributions of this paper are as follows:(1)According to the characteristics of multispectral CT image data,this paper proposes a deep learning structure based on global feature fusion Unet(GFFUnet)and ESPCN for super-resolution reconstruction.The GFFUNet structure is used for feature extraction,the first three convolutional neural networks in ESPCN for nonlinear mapping,and the sub-pixel convolution layer for reconstruction.The experimental results show that compared with SRCNN structure,Unet-ESPC structure(original Unet as feature extraction part,ESPCN as the part of nonlinear mapping and reconstruction),Resnet-ESPC structure(Resnet as the feature extraction part,ESPCN as the nonlinear mapping and reconstruction part)and GFFUnet-transpose structure(GFFUnet as the feature extraction part,three layer convolution as nonlinear mapping part,and the transpose convolution as the reconstruction part),the results of the network model(GFFUnet-ESPC-Cascaded)are improved.(2)In order to improve the preliminary super-resolution reconstruction results,this paper proposes a method based on pixel binary classification results.Firstly,the binary classification model is trained with the labeled pixel data,and then the preliminary super-resolution reconstruction results are put into the binary classification model to get the binary image.Finally,the binary image is used to enhance the boundary and filling range of high concentration area.The experimental results show that the boundary is clearer,the range of high concentration filling area is closer to the original image and the error is reduced.(3)In order to improve the preliminary super-resolution reconstruction results,this paper also proposes a method based on material decomposition results.Firstly,we decompose the high-resolution image to get the material decomposition images,and then the boundary and high concentration region are obtained according to the decomposition images,so as to improve the preliminary super-resolution reconstruction results.The experimental results show that the material decomposition images from the high-resolution image reconstructed by iterative back projection,can be used to improve the image quality obviously,the range of high concentration filling area after enhancement is closer to the original image and the error is reduced.