Joint Segmentation And Motion Tracking Of The Left Ventricle From 3D Cardiac Image Sequences

Heart disease remains the leading fatal disease in the world.Statistics show the Cardio-vascular Disease(CVD) leads to the highest mortality in developed countries.Accurate diagnosis of CVD is the most important work for researchers.Left ventricle(LV),the pumping organ,is the fundamental part of the heart in the literature of the cardiac physiology. CVD often manifests as abnormalities of the ventricular geometry and wall kinematics, which are thus of significant clinical values and offer great technical challenges. With the advancement of non-invasive imaging technologies,especially the magnetic resonance imaging,the acquisitions of the real time and EEG-gated tomographic sequences of in-vivo hearts are more easily available.ventricular geometry retrieved by image segmentation and the myocardium displacement and strain fields achieved by cardiac motion analysis which help to evaluate the global function and local myocardium function of the ventricle play paramount important role in clinical diagnosis.Although segmentation and motion recovery are usually treated as two sequential processes in most of the existing efforts,they can obviously benefit a great deal from each other:on the one hand,using temporal constraints can reduce the influence of noise and lead to improved segmentation;on the other hand, the solution of segmentation can greatly assist in the computation of motion tracking.Following this spirit,we have developed a bio-mechanical model constrained framework performs simultaneous 3D segmentation and motion estimation of the LV.The LV is modeled as an linear elastic body,then dense displacement field can be estimated when the total energy of the elastic body is minimized at equilibrium.Moreover,the LV is modeled as a transversely isotropic,linear-elastic model with fibrous reinforcement. Bio-mechanical model provides plenty of mechanical parameters,such as displacement field and strain field,which help retrieving the possible locations and extents of the infarct area which are caused by heart diseases.Traditionally,force construction of image segmentation only consist of image information corrupted by measuring errors and noises.Here we put forward that the evolution forces imposed on the myocardium are individually constructed for each nodal point through the integration of the data-driven edginess measures,the prior spatial distributions of the myocardial tissues,the temporal coherence of the image-derived salient features,and the cyclic motion characteristics of the heart. We present a method to build the mesh representation of the left ventricle from image sequences at the end of diastole as the initial volume in the our framework.The 2D endo- and epi- contours of the left ventricle are segmented from image slices using a unified framework instead of segmentation separately.Moreover,some of the 3D cardiac image sequence data of interest have higher in-plane resolution than interplane resolution because of the imaging protocol.However,it is more desirable to have roughly equally sampled object representations in all three dimensions to achieve more accurate quantitative analysis and display.For this purpose,we need to interpolate between slices.Since we are initially only interested in the boundary points of the object, interpolation methods based on only the left ventricle contour locations would be more computationally efficient.A shape-based contour interpolation method,using the chamfer distance transformation,has been implemented.As for the numerical method,we use Finite Element Method(FEM) and Meshfree Particle Method(MPM) respectively.The base idea of FEM is to discretize the domain of interests into elements which provide spatial relationships between the sampling nodes. The main computational power of the FEM results from the fundamental idea of replacing a continuous function defined over the entire domain by piecewise approximations over a set of finite number of geometrically simple domains.Meanwhile,the recently developed MPM represent the interested spatial domains with only a set of nodal points but without any mesh constraints which eliminate at least part of the mesh structure by constructing the approximation of the field function and the discrete system equations entirely in terms of the nodes.They can more naturally handle very large deformation and discontinuity.Especially,MPM has advantages in fitting the fiber structure of the left ventricle.Experiments on synthetic and real images show the robustness and accuracy of MPM in the simultaneous framework than FEM.