Study On Sparse CT Reconstruction Method Based On Intelligent Interpolation Of Sinogram

Computed Tomography(CT)is the most widely used medical imaging technique.However,the radiation of radiation to human body during the scanning process brings the potential risk of disease to patients,so low-dose CT has become the current research hotspot.There are two strategies to realize low-dose scanning: one is to reduce the current intensity of the tube when the projection is collected at each Angle;the other is to reduce the number of projections by sparse acquisition.The reconstruction corresponding to strategy 2 is CT sparse reconstruction.However,using traditional analytical methods,such as filtering backprojection algorithm,the sparse reconstructed images contain serious stripe artifacts,which affect the doctors' interpretation of the disease.The iterative method based on compressed sensing can effectively suppress strip artifacts because the prior information of the image can be modeled into the optimization model,but its reconstruction speed is too slow,which restricts its wide application.Deep learning sparse reconstruction method is expected to achieve high-speed and high-precision sparse reconstruction,which includes two methods based on image processing strategy.One is to use deep learning to remove the artifact from the image reconstructed by the analytical method.The other is to use deep learning to carry out intelligent interpolation processing on the projection sinusographs formed by sparse collection,and then use analytical method to reconstruct them.This paper studies the second method.The main work is as follows:(1)inspired by classic DnCNN network,this paper proposes a receptive field of diminishing the intensive DnCNN network structure(Decresing receptive field density,DFD-DnCNN),interpolation,learning the sinogram is implemented and realized the high precision sparse reconstruction.(2)In order to solve the problems of large network and slow training in the intensive DNCNN network structure with decreasing receiving field,a lightweight,three-channel parallel convolutional network is further proposed.Experimental results show that the sparse reconstruction method based on this network is not only more accurate than the traditional interpolation method,but also better than the intensive DNCNN method with decreasing receptive field.(3)On the basis of algorithm research,a CT sparse reconstruction system based on sinusoidal image intelligent interpolation is designed.The system is divided into five modules: sparse projection construction module,traditional interpolation module,intelligent interpolation module,image reconstruction module and image quality assessment module.Based on the above specific research on CT sparse reconstruction,this paper found a new method to reconstruct high-precision images under a small number of projections,thus reducing CT radiation and fundamentally solving the harm of CT diagnosis to patients.