Highly Undersampled Magnetic Resonance Image Reconstruction With Dictionary Learning Method

The application of compressed sensing(CS)in magnetic resonance imaging(MRI)has long been the research focus of accelerating the scanning speed of MRI(CS-MRI).The CSMRI methods are able to achieve good results with highly undersampled k-space data by adding sparsity constraints to the reconstruction problem.MR image is sparse in some transform domains,and CS technique can recover the true image by nonlinear reconstruction as long as the samples are collected incoherently.In the CS-MRI framework,appropriate sparse transforms can be used to improve the sparsity of MR images and promote the reconstruction quality.Dictionary learning is an outstanding technique to achieve a high level of sparsity,and it has been applied in various fields of CS-MRI problems.In dictionary learning methods,MR images are usually divided into many small image patches,which help to learn the local image details and features more effectively,and help to suppress aliasing artifacts and noise.In addition,dictionary learning method can learn prior knowledge from the training set images or prior images to further improve the results of CS-MRI reconstruction.Prior images can improve the sparsity of the target image by eliminating redundant information.It can also provide knowledge about the structural features and anatomical details of the target image to improve the reconstruction accuracy,and make the reconstruction process converge quickly.This thesis mainly focuses on the CS-MRI reconstruction algorithms based on dictionary learning technique.The contributions of our works are as follows:(1)The proposed self-adaptive dual-dictionary learning(Dual-DLSAD)method simultaneously updates the resulting image together with the high and low resolution dictionaries within the process of iterations.The updated dictionaries become more adaptive,containing prior information of both the training set images and the target image.Traditional dual-dictionary learning method requires high-quality training set images and is highly sensitive to the initial dictionaries and reconstruction parameters.Dual-DLSAD method overcomes these defects,successfully enhances the flexibility and adaptability of dictionary learning algorithms,and reduces the parameters sensitivity.(2)This thesis proposes a novel concept of self-prior image,which is reconstructed from the undersampled k-space data of the target image itself.The constraint of self-prior image is successfully combined with dictionary learning method to improve the quality of static MRI reconstructions.The self-prior image can be reconstructed from the undersampled target image directly or from the training set images with dictionary learning method.Both ways are very simple and efficient.The proposed method solves the problem that to obtain a valid prior image is hard for most static MRI applications,and it expands the usage of prior image greatly.The target image can be decomposed based on the self-prior image,resulting in a residual error image that is much sparser.CS and dictionary learning methods are applied on the residual error image to improve the reconstruction performance.(3)This thesis also proposes a new dynamic MR imaging method with parallel imaging,self-prior image constraint and dictionary learning technique.The supervised random sampling method makes it possible to form a synthetic temporally full k-space from the measurements of all frames.In turn,the missing data in each frame is complemented by the synthetic full k-space so that the self-prior image of each frame can be reconstructed using the inverse Fourier transform directly.The self-prior images are almost the same as the target images in static zones,so the residual error images further sparsify the dynamic MRI data,and highlight the dynamic structures.The proposed method also integrates the parallel imaging technique with dictionary learning method to effectively exploit the spatio-temporal correlation and the joint-sparsity among multi-coil and multi-frame dynamic MRI data,improving the reconstruction accuracy at highly undersampled rates.In a word,an efficient and simple method is proposed for the acquisition of self-prior images.Meanwhile,a few novel CS-MRI reconstruction algorithms are presented,in which the dictionary learning method is effectively integrated with self-prior image constraint and parallel imaging techniques.The dictionaries are updated adaptively along with the target images,and the CS-MRI reconstruction results are obviously improved.