4D CT Lung Motion Estimation By Robust Point Matching

Lung cancer is one of the major cancer diseases that threaten human health,and image-guided radiation therapy is an important means for treatment of lung cancer.Breathing motion is an important factor that leads to image artifacts and to tumor position uncertain in image-guided radiation therapy.Accurate lung motion model can help to define the marginal for precise treatment,calculate the dose distribution of radiation and improve gating prediction model,etc.Therefore,to identify lung motion model against individual,through studying the characteristics of lung motion,is an important direction in the current image-guided radiation therapy.Image registration is an important technology of the lung motion estimation.This article focuses on four-dimensional(4D)image registration algorithm based on point matching in the 4D CT lung image.The main research results are as follows.First,a lung motion estimation algorithm based on dynamic point matching is proposed for the fixed point set of the traditional point matching algorithm that can not further improve the registration accuracy.The algorithm firstly builds a fuzzy corresponding matrix of points according to the relations of point set shape and the correlation coefficient of the local image information around the point,which has constructed a correspondence between the virtual target point set and the source point set.The virtual target point is shifted according to the corresponding relation of the image content around the virtual target point,and constrained least squares model is established accordingly,thereby reverse to resolve the target point set,which is shifted through the original target points,in this way can better demonstrate the corresponding relation of image content between the source point set and target point set.Also,an approximate solution algorithm to solve the constrained least squares model is proposed,which can reduce the solution time effectively and improve the efficiency of point shifting.Through multi-iterations,the corresponding relation of image content between the source point set and target point set is improved gradually,thus enhances the registration accuracy of point matching algorithm in the solution of image registration.This algorithm not only has good performance in the lung parenchyma motion estimation,but also in the entire chest image sliding motion including lung parenchyma and chest wall.The spatial accuracy and computing speed of the algorithm are improved greatly compared with other classic algorithms.In the sliding motion estimation of the entire chest,it can effectively estimate sliding motion of lung boundary.Second,a lung motion estimation algorithm based on point matching and spatiotemporal tracking is proposed for the correlation of time dimension that is poor in the spatial registration algorithm of the discrete time point.The algorithm,on the basis of the robust point matching,obtains the mapping position of the feature point in the different phases.Then,it builds feature point trajectory fitting model of least squares constraint of L1 regularization in time dimension.The trajectory will introduce the correlation of time dimension into 4D image registration,so we can get a more stable trajectory.Further,take the location of the point of trajectory fitting as the initial position of the spatial mean-shift tracking of the target point to track target point,to make local image information around the target point and its corresponding source point more similar.Repeat this process above until all target points no longer need to shift.When the algorithm is used in small lung motion estimation,the registration accuracy is higher than other existing algorithms.Meanwhile,when the algorithm is used in evaluation of lung data set with large lung motion,this algorithm not only can ensure better spatial accuracy,at the same time,the topology-preserving performance of spatial deformation domain keeps better than the existing algorithms due to the introduction of time dimension trajectory fitting.Third,a point matching algorithm based on L1 regularization and topology-preserving is proposed for the deformation function of the point matching algorithm is disturbed easily by the outliers that deformation fields are more prone to be abnormal.For point matching,it constructs a regularization-constrained least squares model,which introduce regular restraint respectively in the fields of the robust of elastic transformation,the stability of affine transformation,the bending energy of spatial transform and topology preservation,and has given the regular optimization model of point matching.In the case that there are many outliers,this model also can be used to solve the deformation function which is stable,with less bending energy and topology preservation.The function can be used in the lung motion model estimation,which has solved the problem of traditional point matching that deformation fields were more prone to be abnormal when there were many outliers.Further,this algorithm provides the regularization-constrained least square model with precise solution model,and even the fast solving algorithm.When there are a few outliers,the bending energy is as good as existing point matching algorithms,while the comprehensive evaluation of registration accuracy and volume-preserving performance is better than the existing algorithms.When there are many outliers,the algorithm has higher registration accuracy,at the same time,the measurement of shear strain shows that deformation fields using this algorithm have a better topology-preserving performance.Fourth,a 4D deformation model algorithm based on spatio-temporal radial basis function is proposed for the landmark corresponding relationship based traditional deformation model that only obtains spatial information and lacks the time dimension information.It introduces time dimension into traditional radial basis function,to construct the spatio-temporal radial basis function,and on this basis,build the spatio-temporal deformation model based on spatio-temporal radial basis function,it can estimate the spatial location of any point at any time in three-dimensional(3D)data.In this paper,the properties of the spatio-temporal deformation model are discussed in detail including separability,solvability,spatial smoothing,temporal smoothing,etc.Theoretical analysis shows that the spatio-temporal deformation model can decouple to a series of one-dimensional transformations and a 3D transformation,so as to greatly reduce the computational time of the 4D deformation model.Positive definite radial basis function can guarantee the solvability of the model.Meanwhile,the selected spatial radial basis function determines the smoothness of the model in the spatial domain;the selected temporal radial basis function determines the smoothness of the model in the temporal domain.This spatio-temporal deformation model can describe the motion model of dynamic organs,determine the motion trajectory of any point in the spatial domain,cause the deformation result to remain stable over time and obtain more accurate,more biological significant deformation model.