[Insight-users] MultiResolution with LevelSetMotionRegistrationFilter

[Martin Urschler]

Stefanie Nowak wrote:
> I try to combine the itkLevelSetMotionRegistrationFilter with the 
> itkMultiResolutionPyramidImageFilter to accelerate the registration 
> process.
> 
> For the first level this works quite fine, but in the next level after 
> invoking the update of the LevelSetMotionRegistrationFilter, I get a 
> RegionOfInterestException in the itkDataObject.
> 
> To my opinion the problem is that I try to start the registration with 
> the deformation field of the coarser level as an initial deformation for 
> the finer level.


Stefanie Nowak wrote:

 > Hallo,
 >
 >
 >
 > I try to combine the itkLevelSetMotionRegistrationFilter with the 
itkMultiResolutionPyramidImageFilter to accelerate the registration process.
 >
 > For the first level this works quite fine, but in the next level 
after invoking the update of the LevelSetMotionRegistrationFilter, I get 
a RegionOfInterestException in the itkDataObject.
 >
 > To my opinion the problem is that I try to start the registration 
with the deformation field of the coarser level as an initial 
deformation for the finer level.


hi

did you upsample the deformation field of the coarser level before you 
put it into the finer level? the deformation field needs to have the 
same size as the fixed image...
you don't write anything about that.
remember that, in contrast to the rigid case where you initialize the 
finer level with the transform parameters of the coarser level, the 
nonlinear registration case stores its transform in the deformation 
field itself which has to be upsampled accordingly between scales

regarding to this issue i also have a small question to the community:
what do you think of upsampling 3d deformation fields using the 
VectorLinearInterpolate? As far as i understood it 
VectorLinearInterpolate just does a linear interpolate for each 
coordinate which is imho not quite correct, instead some sort of 
quaternion interpolation would be necessary for the 3d case to be 
correct, wouldn't it?
I know this doesn't fit into the nD framework, but I just wanted to ask 
if everyone is using the VectorLinearInterpolate for multi-resolution 
deformable registration?

regards,
Martin