Cardiac MRI:From Image Reconstruction To Motion Analysis

Heart disease is a vital disease that threaten human health at present,and it is the number one killer of human health and life.Moreover,with the influence of various factors,its incidence rate still shows an upward trend.Today,cardiac magnetic resonance imaging can provide a full range of structural and functional information for the heart,and has become an important tool for the diagnosis and treatment of heart disease.A large number of theoretical studies and clinical experiments have shown that heart disease is directly related to the shape change of the ventricle atrium and the movement status of the myocardium.Robust detection of cardiac dynamic parameters(such as displacement field and strain field)based on image sequences is essential to improve the early diagnosis rate and treatment effect of heart disease.This paper focuses on the simultaneous acquisition of multi-parameter dynamic image sequences of the heart and the robust extraction of myocardial motion and deformation from the image sequences.Because the traditional magnetic resonance multi-parameter quantitative imaging is too long and the imaging quality is easily affected by the patient's breathing,this paper chooses the magnetic resonance fingerprinting technology with shorter imaging time for multi-parameter image reconstruction.Magnetic resonance fingerprinting technology can create a unique time signal for each voxel of the relaxation time of each tissue under different pseudorandom imaging parameters.Different from the traditional reconstruction method to extract the image features of the mixed parameter map,the Transformer network proposed in this paper can directly map the fingerprinting sequence,which realizes the separation and reconstruction of the tissue parameter image from the highly under-sampled image series.This method realizes the sequence-to-sequence reconstruction through the encoder-decoder structure,in which the prediction process of the entire target sequence can be obtained one by one in the forward propagation.Instead of using convolutional neural networks and recurrent neural networks that are mostly used in magnetic resonance fingerprinting image reconstruction,we use the attention mechanism to extract information,therefore,the Transformer network has the advantages of both network architectures in terms of time dependence and network training rate.On the other hand,recovering the motion field and deformation parameters of the entire myocardium from the image sequences is a morbid inverse problem,and it is necessary to add suitable constraints to obtain the only optimal solution.This paper proposes a graph variation model using sparse regularization.This method uses graph variation to mine the spatial similarity between different regions in the image sequence,and establishes a connection between similar regions that does not consider the spatial distance in the entire image sequence.It overcomes the limitation of considering only the similarity with the adjacent regions and retains texture details and fine structure.Simulation experiments and real experiments verify the accuracy and robustness of the method.