[Insight-developers] Some question on ITK registration framework

[Guorong Wu]

Hi all,
     I am recently developing an registration algorithm (HAMMER) in ITK. 
Basically, We want to start the work by following the Demons 
implementation in ITK,  since ITK has a good registration framework. But 
after we have implemented the registration function, we have the problem 
in registration filter.
    For the registration methods in ITK, most of them assume that the 
calculation of each point is not only INDEPENDENT with its nearby 
points, but also the operation is same (compute the difference and the 
gradient). Therefore, it takes advange of multi-thread to speed up. 
Besides, the calculation of update deformation field is performed on 
each point, regardless of back ground points or image points.
    However, HAMMER is different. First, the calculation of deformation 
is the result of feature matching, instead of gradient-based. In each 
iteration, the update is performed point by point, where the update on 
particular point will be affected by the result of previously updated 
one. Second, a set of salient point will be selected to update the 
correspondence, not all. The deformation of other points are just 
interpolated from these salient points.
    Therefore, we consider Demons implementation may be not the best 
class to derive from or we need to rewrite the framework which take the 
same role of itk::DenseFiniteDifferenceImageFilter. And we are not sure 
if there is any filter in ITK which is similar with our case.We highly 
want your suggestion to solve our problem. Thanks.

Best
GR