[Insight-users] help in choosing right deformable registration algorithms

[Vivien Delmon]

On 10/20/2011 08:03 AM, Ryan Guan wrote:
> Thanks Vivien!
> That's right, the 'deformation' in this case is actually the translation
> of the lower ball, and I think that both algorithms caught this since
> the there is no much vectors around the upper ball.
> I actually also tried Demons and BSpline with another phantom, and the
> motion is closer to a local deformation,
> /Demons:/
> /http://dl.dropbox.com/u/1906263/squash_demons.gif/
> /B-Spline:/
> /http://dl.dropbox.com/u/1906263/squash_bspline.gif/
>
> While when I look at them now, I start to feel that maybe BSpline got
> the motion right, will try to assess the accuracy.
> For Demons, it's not completely unacceptable for the single ball
> (/http://dl.dropbox.com/u/1906263/squash_demons.gif/); but for the
> translation motion, it just couldn't work it out
> (/http://dl.dropbox.com/u/1906263/golf_demons.gif/). Is it because of
> the nature of Demons that prevents it from working for translation?

The thing is that BSpline are naturally smooth since it express the 
motion in certain control point and interpolate it in between. So the 
less control point you have the more smooth motion you get. For Demons 
the motion is locally expressed for each voxel so you need to had a 
regularization to get smooth motion.

Since registration is based on the motion intensity in your case it can 
only find motion on edges of your object then the regularization or the 
BSpline interpolation propagate it around.

In your second example if you want to capture the deformation of the 
squash ball you need to segment it and register it separately or the 
motion will be  influenced by the edge of surrounding objects.

There exists some command line tools to do so in an easier way like :
http://elastix.isi.uu.nl/

> Thanks a lot for your help!

Your welcome,

-- 
Vivien Delmon
PhD student at CREATIS, Lyon
Supported by ANRT and ELEKTA
http://www.creatis.insa-lyon.fr/rio