Research On Deep Learning Noise Reduction And Filter Design Of X-ray CT Imaging

CT image reconstruction is currently divided into two categories: analysis reconstruction algorithm and iterative reconstruction algorithm.The analysis reconstruction algorithm has the advantages of simple mathematical form,easy implementation,fast reconstruction speed,and high reconstruction image quality.Its disadvantage is that it is susceptible to noise,requires complete projection data,and the filter function in the algorithm needs to be windowed to derive.Due to the relationship of the window function,it will cause Gibbs phenomenon,resulting in noise and artifacts in the reconstructed image.When using the analysis reconstruction algorithm,these problems will reduce the quality of the reconstruction results,affect the application of CT in real life,and in particular cause doctors to make misjudgment to patients in medical,and the consequences are unimaginable.Therefore,the direction of this paper is the suppress Gibbs phenomenon and the CT noise reduction.First of all,for the Gibbs phenomenon,the FDK algorithm is the research object,because the cone beam CT is the most widely used type of CT in addition to the fan beam CT,and in recent years,the cone beam CT is one of the hot areas of CT theory research,plus the FDK The algorithm is one of the most widely used algorithms in many cone-beam CT algorithms.Therefore,it is necessary to improve the filter function of the FDK algorithm.A new filter function NEW-MS-L is proposed according to the existing filter function and an improved method.The MS-L filter function is obtained by the weighted average SL filter function and the NEW filter function and the MS-L filter are processed by the hybrid filter function.The functions are superimposed to suppress the Gibbs phenomenon,and the effect is better than the existing filter function.For the noise problem,the filter function in the analysis reconstruction algorithm can only have a good noise reduction effect for a single noise and the noise intensity is small,when the noise intensity is strong or has multiple noise,the noise reduction effect become weakly.Therefore,based on this problem and considered the random nature of noise,this paper uses deep learning to solve the noise problem in CT imaging,and proposes a new neural network structure DSResNet based on deep residual network to replace conventionalconvolutions with depthwise convolution is used to reduce the amount of calculation,add channel shuffle methods to enhance the connection between the channels and remove the pooling layer to keep the details to the utmost.Noise(e.g.gaussian noise)with different intensities or different types of noise is added to the projection data,and then using fan beam reconstruction algorithm to reconsturct the improved Shepp-Logan model for DSResNet training.In the simulation experiments and error analysis,the DSResNet can achieve a good effect of noise reduction,and it can suppress Gibbs phenomenon in some degree.