Study Of Dynamic MRI Image Sequence Reconstruction Methods Based On Compressed Sensing Theory

Magnetic resonance imaging(MRI)has become one of the most important medical imaging technology with many advantages such as no ionizing radiation and superior soft tissue contrast resolution.Compared with static MRI,dynamic magnetic resonance imaging(DMRI)which needs continuous MRI scanning over a period of time can display additional dynamic information of human body.However,the slow scanning speed of MRI greatly limits the wide application of DMRI,since it will lead to problems such as limited spatial-temporal resolution and motion interference in DMRI.This will further lead to the reduction of accuracy and reliability in some precision medical applications such as the evaluation of cardiac function after myocardial infarction surgery and the grading diagnosis of brain tumors.Among the existing DMRI acceleration methods,compressed sensing(CS)theory based reconstruction is the fastest developing and most important one at present,which mainly includes some classical CS reconstruction models such as low rank constraint-based model and dictionary learning(Di L)-based model.These classical CS reconstruction methods have achieved good reconstruction results in many DMRI applications such as cardiac cine and brain tumor enhancement,but they usually perform poorly in some complex imaging situations such as high acceleration factor,serious motion disturbance and high quality reconstruction of temporal dynamic information.Aiming at the shortcomings of the existing CS models and according to the different characteristics of different DMRI data,we have proposed several new CS reconstruction methods in this paper.The main research content and innovative points of this paper can be summarized as the following four aspects.(1)Low rank matrix constraint is one of the most commonly used sparsity constraints in the existing CS models,and the high-dimensional DMRI image sequence is usually rearranged into a two-dimensional matrix in these models.The destruction of the original high-dimensional data structure will degrade the final reconstruction performance.In order to directly exploit the internal correlation of DMRI data in highdimensional structure,we introduced the concept of low rank tensor constraint which is based on high-order tensor decomposition,then combined it with local method and a group-wise registration based motion compensation scheme to propose a new CS model called MALLRT,as well as its solving algorithm.Through a lot of reconstruction experiments based on two cardiac cine DMRI data,we have fully verified the superiority of MALLRT over some existing relevant CS reconstruction methods.(2)For DMRI reconstruction disturbed by large motion,the existing CS models often use a motion compensation scheme based on motion signal extraction to ensure the image quality of reconstruction results,but at the cost of time resolution.To solve this problem,this paper presents a new CS reconstruction model called MSC-Lp S,which is based on motion state classification.The main innovation of the model is that large motion is into the main motion and the remaining small motion,and deal with them separately.The main innovation of the model is to decompose the whole motion into the main motion and the remaining part,and compensate them separately.Reconstruction experiments were carried out on two free breathing liver DMRI data,results show that the proposed MSC-Lp S model can achieve high image quality as well as high temporal resolution.(3)When processing DMRI data with more dynamic information,the data-driven adaptive CS reconstruction models usually perform better.The Di L based CS model is the most famous one,but the corresponding optimization algorithm is usually computationally expensive.Therefore,a simpler scheme based on adaptive sparse transform is used as an alternative to Di L,and we combined it with local method and block matching to propose a new data-driven CS reconstruction model called LASTBM and corresponding solving algorithm.A large number of experiments on three different kinds of dynamic contrast-enhanced DMRI data show that LAST-BM can reconstruct complex dynamic information more accurately than some of the latest relevant CS methods.(4)Through human brain dynamic contrast-enhanced DMRI,we can use the contrast dynamic information in the reconstruction results to quantitatively analyze the characteristics of human brain tissue.In order to further improve the accuracy in reconstructing contrast dynamic information,we combined the pharmacokinetic(PK)model with traditional adaptive CS reconstruction model,and proposed an adaptive PK model constrained CS model called APKMC,as well as its solving algorithm in this paper.A series of reconstruction experiments were performed on three human brain tumor contrast-enhanced DMRI data,and the results illustrate the effectiveness of APKMC in reconstructing contrast dynamic information.Through the research of the above four aspects,we have proposed a series of new CS reconstruction methods,which have overcome some shortcomings of the existing methods,to achieve high-quality reconstruction in some complex DMRI imaging scenes such as the imaging process with serious motion disturbance and the reconstruction of rich dynamic spatial-temporal information.