[Insight-users] Re: Registration Issues

[Karthik Krishnan]

On Sat, 2005-09-17 at 19:33 -0400, Avinash Aghoram Ravichandran wrote:
> Hey,
> I am trying to register MR to MR using Mutual information , Linear
> Interpolator and Gradient Decent Algorithm . I tested the algorithm by
> giving hte fixed image and hte moving image as the same and then
> initializting it with a transform of [5,5,5,0,0,0]. The optimizer
> seems to do very weird things the solution I converge is at 23,50,23.

What you might want to do to understand this wierdness is:

1. Use the itkExhaustiveSearchOptimizer to plot the metric space. The
optimizer walks through every point in the region specified. At every
step, the optimizer invokes an IterationEvent, so that you can plot the
metric space. (as is done in all the examples). This might give you an
idea of how the gradients of the metric values look like in the
vicinity, so you know exactly why the gradient descent optimizer is
doing what its doing.

In fact we put this optimizer in precisely for this purpose, cause
optimizers sure can do wierd things :)

In your case, the most likely reason is that you are driving it too far
away from the registered potion and there may be other features that
look more interesting for Mutual information in the vicinity. What might
help here is Multi-resolution resolution. (Or smoothing the datasets
sufficiently prior to registration. Muti-resolution registration is
somewhat akin to smoothing at different scales).


Hope this helps.
Regards
Karthik
> 
> 
>  I have smoothened the images using the gaussian kernel and also
> normalized it. Is there something that I am missing ?
> 
> Avinash
>