Research On Cardiac Motion Estimation Based On Sparse Constraint

Cardiovascular disease has become one of the primary diseases seriously threatening human health.Combination of computer technology and medical imaging makes significant development of analysis of cardiac pathology.Magnatic Resonance(MR)imaging has an excellent spatial resolution and does not require contrast injection,which can directly scan the heart along with different directions,such as cross-section,coronal plane,sagittal plane,and oblique section.Moreover,MR imaging is good at intensity contrast,which can accurately observe the changes in cardiac structure.The cardiac image sequences are capable of describing the spatio-temporal motion states of the interested regions,edges,and contours.The aim of cardiac motion estimation is to detect the displacement information of the heart from the cardiac image sequences.Image registration is an essential technique for cardiac motion estimation.It employs a spatial transformation model to deform the source image to match the target image by estimating the transformation coefficients or the dense deformation fields.Image registration can be used to estimate the displacement of the heart between different time points,which describes the changes of the cardiac anatomical structure with time and the cardiac motion model during a cardiac cycle.By employing the image registration technology,the motion model of the left ventricular myocardium can be estimated,and the real-time changes of the interesting regions and shapes in different phases during the cardiac cycle can also be estimated,which provides clinical diagnostic parameters to assist doctors in diagnosis and treatment.In this paper,the left ventricular motion estimation using sparse constraints for MR cardiac imaging is studied.To solve the existed problems of image registration,sparse constraints are introduced into the transformation model to improve the robustness of the transformation model.Furthermore,the sparse constrained transformation model is combined with the point corresponding relationship estimation and point set matching algorithms,to solve the problems of point mismatching and deformation field topology non-preserving,and improve the accuracy of left ventricular motion estimation.The contributions of this paper are as follows:(1)A sparse constrained transformation function using multiple supports radial basis functions is proposed.Since the landmark location errors exist inevitably in the corresponding relationship estimation in image registration,which requires that the spatial transformation model is robust against the location errors.That is,the deformation field is expected to be accurate when corresponding errors exist in corresponding relationships.The L1 norm regularization is introduced in the transformation model to constrain the elastic deformation coefficients.The influence of outliers is attenuated,and the optimal support for each radial basis function is selected with the sparsity of L1 norm.Besides,L1 regularization improves the robustness of the transformation model.Furthermore,this paper analyzes the close form of the bending energy of the transformation model in the reproducing kernel Hilbert spaces and introduces the bending energy constraint into the transformation model to make the deformation field smooth.Finally,this paper combines the sparse constrained multiple supports radial basis transformation model with the robust point set matching algorithm.It simultaneously estimates the corresponding relationship between point sets and the spatial transformation function,which registers images automatically and can be used to estimate the motion fields of the left ventricle.(2)A multi-level deformable graph matching method for left ventricular motion estimation is presented.Since existed graph matching algorithms are not applicable to cases with large elastic deformation,this paper combines graph matching algorithm with sparse constrained transformation function to estimate correspondence relationship between Left ventricle myocardium at different time.By employing the graph matching algorithm to estimate the candidate corresponding point pairs between left ventricular myocardium,the spatial deformation function is estimated by using the sparse constrained transformation to reduce mismatched point pairs further.Next,the source point sets are mapped as a new source point set,and the graph matching algorithm is used to estimate the corresponding relationship between the new point sets again.During the iterative procedure,the corresponding relationship and the spatial transformation are more and more accurate.Moreover,a multi-layer framework is proposed to perform graph matching algorithm and spatial transformation alternately from bottom to top until an accurate matching result is obtained.This multi-level deformable graph matching method is validated in improving the accuracy of left ventricular motion estimation.(3)A point set matching algorithm based on surface structure features is proposed,using the surface structure features of the myocardial to describe the point features.The surface feature of left ventricular is introduced into the point matching algorithm based on a Gaussian mixture model.A cost function is proposed by combining the sparse constrained transformation using multiple supports radial basis functions with surface structure features in the point matching based on a Gaussian mixture model.The traditional thin-plate spline transformation is replaced with sparse constrained transformation function using multiple supports radial basis functions,to attenuate the influence of outliers and construct complex deformation fields.The solution of this model is derived in detail.Experimental results show it can effectively estimate the displacement field of the left ventricle.(4)The effectiveness of our proposed methods is validated on several public MR cardiac data sets,and the experimental results show that the proposed methods can effectively estimate the left ventricular motion field.On the other hand,the estimated motion field can be used to calculate the strain coefficient of the left ventricular myocardium,to provide auxiliary parameters for clinical diagnosis and treatment.