| Lens-coupled detector based Micro-CT has been widely used in the field of biomedicine,material science and many other fields due to its nature of non-destructive and high resolution.The spatial resolution of such kind CT system is ultimately limited by some physical factors including detector element size,image degradation due to focal spot size and lens system,as well as some model problems such as signal to noise ratio of imaging system and low computational efficiency due to mass data.In this paper,on the existing lens-coupled high-resolution micro-CT in our laboratory,a sub-pixel information based super-resolution CT reconstruction algorithm is proposed,which not only makes use of sub-pixel information,but also constructs a forward model that includes blur and noise correlation associated with finite focal spot size and scintillator thickness.Moreover,the proposed method also takes advantage of deep learning and dual GPU techniques to accelerate the reconstruction procedure of super-resolution CT.The main contents of this paper are organized as follows:Firstly,this paper proposes a super-resolution FDK(SR-FDK)algorithm,which utilizes the complementary information of multiple sub-pixel shifted projection images.A new CT scanning scheme is put forward,in which the 3-Axis Nanometer-Precision Linear stage is served as shifting the scanned object in nanoscale for sub-pixel projection images capturing.Based on the ITK framework,a high-precision displacement matrix correction algorithm is designed,which can be directly calculated by the coordinates of the stage moving.And then,the FDK algorithm is carried out with sub-pixel projections and up-sampling datagrid to obtain the super-resolution images.The numerical simulations and real experiments results indicate that,the spatial resolution of reconstructed images are improved by SR-FDK,though those still suffer from some degradation due to focal spot size and lens system.In addition,since sub-pixel motion of the scanned object destroys the formation mechanism of ring artifact in CT,SR-FDK demonstrates a good mitigation on ring artifacts.To suppress the system blur metioned above,a super-resolution MBIR(SR-MBIR)algorithm is presented based on sub-pixel information.The scanning scheme in SR-MBIR is same as the one in SR-FDK.By optimizing the forward model and designing the system blur measurement method based on blind deconvolution,the source and lens system blurs as well as noise correlation can be introduced into the iterative reconstrction.Further experiment shows that the SR-MBIR can improve the spatial resolution of CT reconstruction images significantly,suppress the blurs introduced by focal spot size and lens system,and improve the signal-to-noise ratio in reconstructed images.Although SR-MBIR algorithm has a better performance on enhancing the spatial resolution of CT reconstruction images,larger amount of projection data and slow convergence of high frequency information obstruct its extensive use in practical applications.In this paper,deep learning and dual GPU programming are applied respectively to accelerate the convergence of iterative algorithm and improve the computational efficiency.A super-resolution convolutional neural network(SRCNN)is contrstructed,which takes the Gauss blurred FDK results as the input dataset and the iteration results as the label dataset for supervised learning procedure.SRCNN is trained to learn high-frequency features of iteration results,and after training,its output image can be taken as the initial image of iterative reconstruction to improve the convergence speed.The dual-GPU programing technique is implemented to improve the degree of parallelism and the computational efficiency.The experiment results demonstrate that the reconstruction time of the proposed method takes only 25% that of conventional reconstruction. |
PDF Full Text Request