[Insight-users] demons \w MI and BSpline optimization

Tom Vercauteren tom.vercauteren at gmail.com
Mon Apr 16 09:15:01 EDT 2007


Hi Ruben,

There is an insight journal submission by De Craene et al. about your
first question:
http://hdl.handle.net/1926/183
I didn't try it myself but it seems to answer your needs.

I think that the tricky thing when you change the similarity metric in
the demons, is the way you compute the derivatives.

You could try to compute the derivatives explicitly or you could
replace, as Luis suggested, the computation of the force by a two step
procedure:  1) find the intensity relationship from the current
displacement field 2) Compute the derivative of the metric by keeping
the intensity relationship fixed. Cf. "Multimodal Brain Warping Using
the Demons Algorithm and Adaptative Intensity Corrections" by Guimond
et al.

You can also take a look at:

"Free-Form Registration Using Mutual Information and Curvature
Regularization" by D'Agostino et al. which is an extension to mutual
information of an algorithm closely related to the demons

"Iconic Feature Based Nonrigid Registration: The PASHA Algorithm" by
Cachier et al. which extends the demons algorithm to use local
statistics.

Hope this helps.

Best,
Tom Vercauteren

------------------------
Ruben Schilling wrote:
> Hi all,
>
> I have two questions:
>
> (1)
> I am wondering whether there is an example out using some variant of
> mutual information with some demons algorithm. I didn't find anything
> in the software guide nor the examples. It seems, that it would be
> interesting to compare its performance to the BSpline + mutual
> information registration.
>  From the literature I don't see why it wouldn't be possible, since the
> demons only rely on similarity measure and distance to compute  the
> forces, which does not limit one to e.g. squared differences. Is  there
> a reason in optical flow, why we couldn't use demons with  mutual
> information? Can't we replace the intensity similarity of  optical flow
> with mutual information similarity?
> Has anyone tried this yet? Is there a reason why there is no example  (I
> thought it would be a rather general demand to have it).
>
>
> (2)
> I am optimizing currently also the DeformableRegistration8.cxx  example
> (under consideration of DeformableRegistration6.cxx). Are  there any
> general guidelines in optimizing the LBFGSB optimizer used?  I find it
> quite non-trivial, because I don't have the necessary  background on
> non-linear optimization theory. Especially for using  the BSplines I
> find it hard, I can imagine setting e.g. the right  scalings in an
> affine transform, but for a BSpline grid the  parameters are the
> displacements of the nodes, right? Are there any  general
> hints/guidelines how to procede here?
>
>
>
> To make my point clear, I am actually trying to solve a difficult
> practical problem with multi-modalities and I am not intending to  start
> any debate on which elastic transform is better. I am combining  the
> elastic transform with a quite successful rigid and affine "pre-
> registration", but still the remaining problem is difficult enough,
> this is why I would like to explore all techniques available (yes, I  do
> also consider FEM but haven't started on it for now).
>
>
> I would be grateful for comments on these matters.
>
> best regards
> Ruben
>
>
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>


More information about the Insight-users mailing list