[Insight-developers] Modifying MultiResolutionPDEDeformableRegistration->SetInitialDeformationField

[Tom Vercauteren]

Hi all,

In order to finalize the transfer of the diffeomorphic demons into
ITK, I'd like to modify MultiResolutionPDEDeformableRegistration but
I'd like some feedback.

Currently an initial deformation field can already be passed to
MultiResolutionPDEDeformableRegistration with the
SetInitialDeformationField method.

The tricky thing is that this initial deformation field is used as is
at the coarsest scale of the pyramid. For this deformation field to be
well-behaved at the coarsest level, the caller typically needs to know
the pyramid schedule and smooth the initial field accordingly. This is
cumbersome and not user-friendly.

This behavior is also not documented in the file. I thus guess that
most users will simply call SetInitialDeformationField with a
deformation field that has the same characteristics (spacing, size,
origin) than the input images and would not apply any prior smoothing
and downsampling.

I think that a more convenient way of doing things is to ask for a
deformation field that has the same characteristics than the input
images. Then MultiResolutionPDEDeformableRegistration should take care
of smoothing and downsampling the initial deformation field according
to the pyramid schedule. This is how I implemented
SetInitialDeformationField in the IJ submission (At that time
MultiResolutionPDEDeformableRegistration->SetInitialDeformationField
had not been implemented yet).

I see a few options here:

1) Change the behavior of
MultiResolutionPDEDeformableRegistration->SetInitialDeformationField
to include smoothing and dowsampling. This may not be really compliant
with ITK's backwards compatibility policy though.

2) Add a flag to
MultiResolutionPDEDeformableRegistration->SetInitialDeformationField
to allow both options

3) Leave MultiResolutionPDEDeformableRegistration->SetInitialDeformationField
unchanged and add a new method to use an initial deformation field
with smoothing and dowsampling


What do you think would be the best option?

Regards,
Tom