[Insight-developers] Problems with Affine Registration

[Paul Hughett]

> I don't think the dilating gaussian problem can be solved using
> mutual information. As you have stated in a previous posting, due to
> the symmetry, mutual information can only distinguish between
> different gaussians up to an isotropic scale factor.

Oops! That should have been the
ImageToImageAffineMeanSquaresGradientDescentRegistrationTest program that
                  ^^^^^^^^^^^
isn't handling an isotropic scaling.  I started with the mutual
information metric (which my fingers remembered too well) but switched
to trying the mean squares metric for exactly the ambiguity you
mention.  Unfortunately, the mean squares metric shows the same problem.

> On the real data front, what kind of results are you getting with the
> MR brain data?

I'm not, and I'm beginning to suspect a bug in the ITK affine
registration code; I've tried various optimization parameters without
getting convergence.  I haven't yet tried your new multi-res example
program, since I started several weeks ago from the various test
programs instead.  On the other hand, our in-house registration toolkit
has been happily registering real images using mutual information.


Paul Hughett


 
>> I've been working most recently with the
>> ImageToImageAffineMutualInformationGradientDescentRegistrationTest
 Make that:          MeanSquares
>> program, modified to add a scale factor within the reference gaussian,
>> e.g.
>> 
>>     ri.Set( 200.0 * exp( - 1.1*( x*x + y*y )/(s*s) ) );
>>                            ^^^^
>> This corresponds to roughly a 5% isotropic scaling, and should yield
>> coefficients of [0.95, 0.95] (or maybe [1.05, 1.05]) on the diagonal
>> of the affine transform matrix.  What actually happens is that the
>> diagonal stays within 0.001 or 0.0001 of unity.