Study On CT Sparse Reconstruction Based On Deep Learning

Computed Tomography(CT)has greatly promoted the development of medical imaging technology,and has a very important significance for the diagnosis and treatment of various diseases.Due to the potential risk of disease caused by X-ray,low-dose CT has become the focus of current research.Sparse reconstruction,that is,the projected reconstruction image acquired from a sparse angle,is an effective method to achieve low-dose CT.However,the sparsely reconstructed image of the classic analytical method often contains severe streak artifacts,which makes it impossible to correctly interpret the disease.Since 2006,deep learning technology has demonstrated its powerful performance in industry and academia due to the combination of big data,big network and big computing power.The image processing methods based on deep learning have achieved better processing effects than traditional methods in the fields of image recognition,target detection,image segmentation,image denoising,and super-resolution.Since 2016,deep learning image reconstruction has become a hot research direction in the field of reconstruction.In order to effectively suppress the streak artifacts introduced by analytic sparse reconstruction,a deep learning-based sparse reconstruction method is systematically studied in this paper.The specific work contents are as follows:(1)Based on the in-depth analysis of the performance of the classical Residual UNet network,Full Convolutional Network,and Autoencode-decoding Network,the three networks were introduced into the sparse reconstruction task,and the artifact suppression ability of the three networks was compared.The experimental results show that the performance of the Residual UNet network is better than the other two network structures in suppressing streak artifacts.Therefore,this paper will optimize the network structure on the basis of the Residual UNet network to further improve the performance.(2)A high precision sparse reconstruction method based on Multiply residual UNet(Mr-UNet)network is proposed.The classical residual UNet is composed of several convolution units,and each convolution unit is a straightthrough structure.Mr-UNet constructs each convolution unit as a residual structure to improve the training performance of the network.Firstly,the ability of this network to suppress streak artifacts was systematically compared with traditional TV(Total Variation)method and residual UNet network.The experimental results show that the proposed reconstruction method more effectively suppresses the streak artifacts,and retains the detailed information and texture features of the reconstructed image better.Secondly,the evolution law of the streak artifact suppression performance of the network structure is explored under different sparsity.The experimental results show that with the increase of the number of projections,the performance of Mr-UNet network to suppress the streak artifacts is better and better.Finally,a sparse reconstruction system based on Mr-UNet is designed.(3)A high precision sparse reconstruction method based on adversarial MrUNet is proposed.Based on the principle of GAN network,a discrimination network based on adversarial training mechanism is introduced on the basis of Mr-UNet to form adversarial Mr-UNet.Experimental results show that this network model can further improve the performance of suppressing artifacts.Experiments also show that the reconstruction performance of the network is better and better with the increase of the number of projections.In this paper,the performance of CT image sparse reconstruction based on convolutional neural network is further improved by introducing multiple residuals and confrontation mechanism,which has certain theoretical significance and practical value.