Cardiac Motion Estimation Combining Point Set Matching With Surface Structure Feature

Along with the living standards improves,cardiovascular diseases are ranked as the top killer worldwide,which threatens seriously human health.The symptoms of cardiovascular disease are frequent morbidity,Strong concealed and high mortality,so early diagnosis and risk assessment becomes particularly important.The movement of left ventricular(LV)myocardium can reflect the blood supply of cardiac,which provides important basis for diagnosing a variety of cardiovascular disease.Through estimating LV motion,we can obtain the motion trajectories of each myocardial point and then acquire significant transformation function and visualization graph of clinical diagnosis.Point set matching is a common LV motion estimation algorithm,but current algorithm of point set matching only considered the distance between two points,lacking of the consideration of point set shape.In this paper,we present a point set matching method based on surface structure features,including three parts as following:First,a surface feature extraction method based on tensor voting is used in order to describe the structure features of LV.With the points of LV contour as voting points and the approximate feature direction as the voting direction,tensor voting is performed.After voting process,accumulate and decompose all received votes of each points and then the corrected direction is obtained.Experiments on both simulation and real data show that the proposed tensor voting method is effective.Second,for the existing point set matching method only considered the distance between two points and lacking of the consideration of point set shape,a method combining with point set matching and surface structure features is presented and applied to estimate cardiac motion.Our contribution is introducing surface structure description,which describes the curve feature of left ventricle,to point set matching using Gaussian mixture model,presenting a cost function which based on distance and shape of two point sets and calculate the gradient of Quasi-Newton method(QN)to optimizing the cost function so as to obtain the transformation parameters of LV motion and motion trail of LV myocardium.Experiment results show that the presented cost function is feasible.Third,for divergence problem shown in QN method applying to high-dimension parameter space,we propose Stochastic gradient descent method(SGD)to optimize cost function through derivation the gradient of SGD.Regards to the weaker convergence precision of SGD compared to QN,an optimization algorithm combined SGD with QN is proposed in this paper.Controlling transformation parameters by SGD at first,and let it converge to a stable state,then we use QN to improve its convergence accuracy.Experiment results show that the method using combination of SGD and Quasi-Newton method is more robust and more accurate in high-dimension parameter spaceThis paper presents the preliminary study on three aspects: extraction of LV surface structure,the accuracy and robustness of LV point set matching.The experiment results demonstrate that our research can deal well with some key problem in LV motion estimation based on point set matching.