[Insight-developers] [Insight-registration] vector transformation / reorientation

[Tom Vercauteren]

Hi Brian,

On Mon, Nov 22, 2010 at 23:11, brian avants <stnava at gmail.com> wrote:
> I think Tom's idea is quite nice.   We add a
>
> resampler->SetReorientationStrategy( reorienter );
>
> function to the ResampleImageFilter .... obviously, for most cases, it
> would not be needed.

By the way, would there be any more appropriate name than
ReorientationStrategy? In some specific cases, I guess that the user
might be interested in not only rotating but also re-scaling the
diffusion tensor. That is, what might be wanted is to apply the full
Jacobian matrix: J^-1 . T . J

In this case, reorientation is not exactly what is done. Anyhow, with
some documentation, this might be clear enough.


> Tom, you mentioned "what can be of interest, especially for image
> registration, is the the derivative of this reorientation strategy. It
> should indeed come into the computation of the derivative of the
> objective function."  and also your FS paper.
>
> I am not sure that we can implement the derivative wrt reorientation
> in a general way, though it's possible to implement for some cases,
> e.g. FS and PPD.   Did you have more specific thoughts about this?

Not really, I just thought it would be nice to support it if we can
(much like some computation of a pseudo-Hessian on objective
functions).

If no analytical derivation can be made, we could either return an
identity matrix, compute if from finite difference or automatic
differentiation (
http://en.wikipedia.org/wiki/Automatic_differentiation ) or provide a
way to inform the user that this function is not implemented. For
performance reasons, I wouldn't rely on exceptions but maybe on
something close to boost optional:
http://www.boost.org/doc/libs/release/libs/optional/index.html
http://www.boost.org/doc/libs/release/libs/optional/doc/html/boost_optional/examples.html#boost_optional.examples.optional_return_values


Tom