Studies Of Fast Magnetic Resonance Imaging Based On Dictionary Learning And GPU Acceleration

Magnetic Resonance Imaging(MRI)technology has been applied widely in clinical medicine because of its excellent soft tissue resolution and no radiation damage to the human body.However,the MRI technology is limited in slow sample speed,which become a bottleneck in the field of clinical application especially in the field of Dynamic Cardiac Imaging(DCI).The emergence of the compressed sensing(CS)theory and the application of CS technology based on sparse representation of data,have partly resolved this problem.In this paper,we will mainly study the dictionary learning(DL)algorithm and its optimization algorithm-Dual DL algorithm,and applied them to three-dimensional dynamic cardiac imaging(3D-DCI)respectively.Because of the time consuming large amount of 3D data reconstruction,this paper will use the GPU(Graphic Processing Unit)parallel computing to shorten reconstruction time.This paper mainly includes the following work:(1)The theoretical framework of CS theory was studied,including three core parts:signal sparse representation method,observation matrix and image reconstruction.We focused on the application of compressed sensing theory in fast magnetic resonance imaging(FMRI),that is,the sparse representation of MR images,the design of the undersampling method and the design of the image reconstruction algorithm.Finally,the CS-MRI implementation method based on wavelet transform was studied and verified by simulation experiments.(2)The CS-MRI algorithm,with DL a sparse representation was studied.The research data is also extended from the 2D to 3D.In the method of 3D reconstruction,the 3D reconstruction is implemented by the traditional method,that is the data are directly involved in the reconstruction.The purpose of this is to use the prior experience and the correlations between slices.Furthermore,the algorithm was optimized by the the dual dictionary algorithm,which can improve the quality of data reconstruction.In order to improve the quality of data reconstruction,the dictionary of different accuracy is used to improve the quality of data reconstruction.Finally,we apply this algorithm to 3D-DCI.The simulation experiment showed that the quality of the reconstructed image is improved more effectively than the other algorithms.(3)We considering the long time consuming of the algorithm and the large amount of3 D data,especially the 3D matrix multiplication in 3D reconstruction.In the process of the reconstruction of the algorithm,the new dictionary is by producing new atoms.The sparse(4)representation of the MR image is obtained by the orthogonal matching pursuit(OMP)algorithm,and the image is reconstructed according to the sparsity of the image and the consistency of the image.OMP algorithm is very suitable for parallel computing,which provides the possibility of speed up the algorithm running time by the GPU hardware.In this study we used MATLAB plug-in JACKET based on GPU parallel optimization.The simulation experiments showed that GPU can effectively shorten the time of reconstruction.