Dictionary Learning Based Quantitative Susceptibility Mapping Reconstruction

Quantitative Susceptibility Mapping(Quantitative Susceptibility Mapping)technology is a novel magnetic resonance imaging technology that can reconstruct tissue susceptibility mapping from phase signal which containing rich susceptibility information,and can quantitatively analyze iron content,calcification,blood oxygen saturation and other related information in the body,can provide diagnostic information for many diseases,has important clinical application value.The reconstruction process of the susceptibility mapping includes four steps: field map fitting,phase unwrapping,background field removal and dipole kernel inversion.Since the position of the dipole kernel in the K space at an angle of 54.7 ° with the main magnetic field is 0,the dipole kernel inversion is an ill-posed inverse problem.Currently,there are many methods to solve this inverse problem,such as The calculation of susceptibility through multiple orientation sampling method,the thresholded K-space Division method in the Fourier domain,the total variation method in the spatial domain,the L2-regularized reconstruction algorithms with closed-form method and the morphology enabled dipole inversion method.However,these methods have certain limitations.In the reconstruction of susceptibility mapping,how to effectively suppress artifacts and improve the reconstruction quality is still an important challenge.In recent years,compressed sensing technology has achieved good results in solving the inverse problem in image reconstruction.The reconstruction method based on dictionary learning can well suppress noise and artifacts in the reconstruction of low-dose CT and undersampled MRI data.Based on the dictionary learning method,this paper proposes to use the feature dictionary to constrain the reconstruction of the susceptibility mapping and solve the inverse problem in the dipole kernel inversion.The main work is as follows:(1)A quantitative susceptibility mapping reconstruction method based on Bregman iterative dictionary learning(BI-DL)is proposed.In the reconstruction process,the dictionary is updated by the Bregman iterative method and used to constrain the reconstruction of the susceptibility mapping.The effectiveness of the algorithm is verified by 3T and 7T human brain data.The results show that the method has an improvement in artifact suppression and quantitative metrics results compared to mainstream reconstruction methods at low field MRI.At high field MRI,there is still a higher accuracy of reconstruction of the region of interest.(2)A quantitative susceptibility mapping reconstruction method based on the magnitude edge prior feature dictionary(EP-DL)is proposed.Train the feature dictionary with magnitude images for the edges in the susceptibility mapping closely match the structures in the magnitude images of gradient recalled echo(GRE)obtained in the same acquisition,and then used feature dictionary to constrain the susceptibility mapping Reconstruction.The effectiveness of the algorithm is verified by 3T and 7T human brain data.Compared with the BI-DL method,the reconstruction quality is further improved.(3)Design a quantitative magnetic susceptibility imaging software based on the MATLAB GUI tool,which can interactively complete the entire process from the original data to the susceptibility mapping.The software only supports the reconstruction of the susceptibility mapping of the brain data but also adds additional water-lipid separation method can better support the reconstruction of the susceptibility mapping of the abdomen data,expanding the scope of application of the software.At the same time,the software provides a variety of reconstruction methods for users to choose flexibly.Also,the software provides data post-processing analysis function,which can analyze and process the reconstruction results,which provides convenience for subsequent experimental research and clinical application.