Research On Reconstruction Network And Suppression Of Sparse-view Artifact Based On Deep Learning

Reconstruction algorithm is the key part of micro-CT imaging.Conventional analytic reconstruction algorithm is high-efficiency but performs poorly in noise suppression.The iterative reconstruction algorithm has a good effect on noise suppression,while it is computationally demanding and time-consuming.Therefore,it is difficult to take advantages of both to achieve the outcome of high-efficiency and high-quality.Deep-learning based methods present a new perspective to solve the imaging problem of micro-CT.Currently,the deep-learning based methods to solve these problems only focus on image domain or projection domain,which breaks up the inherent relevance between two domains.As a novel reconstruction model,the deeplearning based reconstruction network can not only take advantages of two conventional reconstruction algorithms,but also integrate image domain and projection domain into the same architecture.Owing to the end-to-end characteristics and the powerful learning ability of neural network,high-efficiency and high-quality reconstruction can be achieved at the same time.Furthermore,the noise and artifacts can be taken as a whole to deal with,such that the better performance can be achieved than the preceding conventional methods.Currently,the existing reconstruction networks are based on fully connected layers and consist of huge amounts of parameters,which is not capable for 3D reconstruction of micro-CT.Therefore,it is a great issue currently for reducing the number of network parameters and realizing high-quality reconstruction.This paper focuses on building highly-sparse reconstruction network with good reconstruction quality.Meanwhile,some deep-learning based methods are developed to suppress sparse-view artifacts in the image domain and projection domain respectively.The main contents are as follows:Firstly,according to the principle of sinusoidal-connection and rotation based reconstruction network,the formula of forward propagation and gradient backward propagation of these two networks are derived.The corresponding networks are realized by using C++ and Tensorflow respectively.The experiment results indicate that the sinusoidal-connection based network with highly sparse parameters is easy to be trained and to achieve high precision.As for the rotation based network,one group parameters are shared over all projection views,which reduces the number of parameters but is easy to introduce ring artifact into the reconstructed image.Furthermore,by combining the characteristics of the two networks mentioned before,this paper proposes a sparse-rotation based CT projection neural network(SRPNet)and CT reconstruction neural network(SRRNet).The proposed network is highly-sparse and its parameters are shared over all projection views,which reduces the number of network parameters by two to three orders of magnitude.In terms of 2D reconstruction,the SRRNet is realized and the experiment results show that the reconstruction results of SRRNet are inferior to that of conventional reconstruction algorithm.By increasing the number of network parameters,four group parameters are shared over all projection views,and the results show great consistency.Then the SRRNet is extended to 3D reconstruction subject to the same principle of 2D reconstruction.The data of 50 layers up and down the midplane are reconstructed by SRRNet.Due to the difference between 2D and 3D reconstruction model,the experiment results demonstrate that ten group parameters shared over all views are enough for SRRNet to achieve high-quality reconstruction.Meanwhile,the reconstruction result of the 50 th layer above the midplane has a quite agreement with the original image than the FDK algorithm.Finally,by using deep-learning techniques,the suppression approaches of sparseview artifact are researched in the image domain and projection domain respectively.In the image domain,the Inception module is employed to extract the multiscale features of sparse-view artifacts and the residual mechanism is employed to make the neural network focus on learning the artifact features.From the results,it can be seen that the sparse-view artifacts are suppressed obviously.In the projection domain,the features of projection image in different direction are extracted by anisotropic convolution kernels to complete projection image.The experiment results indicate that the processing in the projection domain can preserve more details and achieve good effect on sparse-view artifacts suppression,but it’s easy to introduce other artifacts in the process of interpolating and network training.