itkFEMRegistrationFilter.h File Reference

Go to the source code of this file.

Classes
class	itk::fem::FEMRegistrationFilter< TMovingImage, TFixedImage >::FEMOF
class	itk::fem::FEMRegistrationFilter< TMovingImage, TFixedImage >
	FEM Image registration filter. The image registration problem is modeled here with the finite element method. Image registration is, in general, an ill-posed problem. Thus, we use an optimization scheme where the optimization criterion is given by a regularized variational energy. The variational energy arises from modeling the image as a physical body on which external forces act. The body is allowed to deform so as to minimize the applied force. The resistance of the physical body to deformation, determined by the physics associated with the body, serves to regularize the solution. The forces applied to the body are, generally, highly non-linear and so the body is allowed to deform slowly and incrementally. The direction it deforms follows the gradient of the potential energy (the force) we define. The potential energies we may choose from are given by the itk image-to-image metrics. The choices and the associated direction of descent are : Mean Squares (minimize), Normalized Cross-Correlation (maximize) Mutual Information (maximize). Note that we have to set the direction (SetDescentDirection) when we choose a metric. The forces driving the problem may also be given by user-supplied landmarks. The corners of the image, in this example, are always pinned. This example is designed for 2D or 3D images. A rectilinear mesh is generated automatically given the correct element type (Quadrilateral or Hexahedral). Our specific Solver for this example uses trapezoidal time stepping. This is a method for solving a second-order PDE in time. The solution is penalized by the zeroth (mass matrix) and first derivatives (stiffness matrix) of the shape functions. There is an option to perform a line search on the energy after each iteration. Optimal parameter settings require experimentation. The following approach tends to work well : Choose the relative size of density to elasticity (e.g. Rho / E ~= 1.) such that the image deforms locally and slowly. This also affects the stability of the solution. Choose the time step to control the size of the deformation at each step. Choose enough iterations to allow the solution to converge (this may be automated). More...
Namespaces
namespace	itk
namespace	itk::fem