[Insight-users] Is there way to speed up cross-correlation based deformable registration?

[Aviv Hurvitz]

Hi Jian,

Well, as you can see, Elastix is a stand-alone program and not a library. I
suspect it won't be easy for you to integrate it into your program on the
source-code level.

What I actually do is run Elastix as a separate program when I need it, and
then I parse its output file to get the final transform parameters. In my
case, this was much easier than trying to get my classes to compile with it.

Regarding your speed problem, I think the largest factors really are the
number of samples and the pyramid level. Consider skipping the last pyramid
level, or limit the number of iterations. If you got a good registration on
the second-to-last level, maybe that's enough?

BTW, to be clear, I'm not an Elasitx developer.

Hope it helps.

- Aviv


On Thu, Jun 5, 2008 at 5:52 PM, Jian Yang <yaland1977 at gmail.com> wrote:

> Hi Aviv, I still have several questions about this issue:
>
> 1. How to specify samples for the cross-correlation based
> registration? Could you please give me piece of sample code?
>
> I compiled the elastix program and I found it is very much
> powerful.The sampler methods developed in this package are exactly
> what I need, but I still don't know how to made the codes integrate
> into my program. Could you please give me some advises?
>
>
> 2. I computed the metric gradient, but it does not seem to work well.
> I think probably because the image is too big for the computation.
>
>
> 3. I used multiple resolution strategy  for the images (pyramid
> strategy) and for the B-spline transform (multiple grid size), but
> when iterations  reach higher level, it runs still very slow.
>
>
> So, I think probably the best way to speed up the cross-correlation
> based large data sets registration is firstly specify the computation
> samples (random sample, grid sample etc.), and then use multiple
> resolution strategies.
>
>
> Thank you very much for your reply. Any help would be greatly appreciated.
>
>
> Jian
>
>
>
> Hi Aviv, Thank you very much for your help.
>
> Jian
>
>
> Date: Mon, 2 Jun 2008 22:02:09 +0300
> From: "Aviv Hurvitz" <aviv.hurvitz at gmail.com>
>
> Subject: Re: [Insight-users] Is there way to speed up
>       cross-correlation       based deformable registration?
> To: insight-users at itk.org
> Message-ID:
>
>       <e25acca30806021202l50d68fa3j4a494ac5b4685b15 at mail.gmail.com>
> Content-Type: text/plain; charset="iso-8859-1"
>
> See if you can use the elastix program (which is based on ITK code).
>
> http://elastix.isi.uu.nl/index.php
>
> It has several speed improvements:
> 1. You can specify samples.
> 2. It utilizes the sparse Jacobian of the B-spline transform for faster
>
> computation of the metric gradient.
> 3. You can use a multiple resolution strategy, both for the images and for
> the B-spline transform itself.
>
> - Aviv
>
>
> On Mon, Jun 2, 2008 at 9:20 PM, Jian Yang <yaland1977 at gmail.com> wrote:
>
>
> > Hi all,
> >
> > I'm currently working on deformable registration of CT images. I'm using
> > cross-correlation (DeformableRegistration6.cxx) and mutual information
> > (DeformableRegistration15.cxx) metrics with a B-Spline Deformable
>
> Transform
> > in a 300*208*241 3D image.
> >
> > It seems that the example DeformableRegistration15.cxx  works much faster
> > than example DeformableRegistration6.cxx. I guess example
>
> > DeformableRegistration6.cxx  is time consuming since it is performed on
> the
> > whole image. And the Mattes Mutual Information metric works faster
> because
> > it allows to take samples (SetNumberOfSpatialSamples) in the image
> instead
>
> > of taking all voxels. So, I am wondering is there way to do the same
> thing
> > in the cross-correlation metric to make it faster?
> >
> > Thank you very much for your reply.
>
> >
> >
> > Best Regards,
> >
> >
> > Jian
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20080606/803aa801/attachment-0001.htm>