[Elastix] µCT population average

S. Klein s.klein at erasmusmc.nl
Fri Apr 3 20:00:32 CEST 2015


Hi Fabien, not sure if it's still relevant (i was cleaning up old mails and found this one): when you want to build a population template based on averaging multiple pairwise registration, you should average only transformations which were computed with the same fixed image.

T_ij = transformation result for fixed image i, moving image j
TMEAN_i = 1/N sum_j T_ij
TMEANINV_i = invert (TMEAN_i)
Then bring image i to mean space according to TMEANINV_i, using transformix.

Cheers,
Stefan

On 18-02-2015 10:43, Fabien PERTUY wrote:
Hi Coert,

Thank you for your interest.
Yes I do include an identity transform.
I you consider images A, B and C, and mX_fY the transformation from the registration of moving image X to the fixed image Y.
I tested mean(mA_fA; mA_fB; mA_fC) using WeightedCombinationTransform afterward in TransformiX compared to  mA_fA_fB_fC using MultiMetricMultiResolutionRegistration in ElastiX.

Regards,
Fabien

De : Coert Metz [mailto:coert.metz at gmail.com]
Envoyé : mercredi 18 février 2015 09:34
À : Fabien PERTUY
Cc : elastix at bigr.nl<mailto:elastix at bigr.nl>
Objet : Re: [Elastix] µCT population average

Hi Fabien,

When you average, do you include one identity transform for Tii (so transformation from subject to itself)? This might be the reason it's not working.
See, for example also section 2.2 of http://bigr.nl/website/index.php?page=publications&subpage=publication&id=547

Regards,
Coert Metz


On Wed, Feb 18, 2015 at 9:21 AM, Fabien PERTUY <pertuy at igbmc.fr<mailto:pertuy at igbmc.fr>> wrote:
Hi ElastiX community,

I am a new user of ElastiX interested in registration of µCT scans to generate population average that could be used to create custom atlases.
I tried to reproduce described approaches (including Roy van Pelt’s) based on computing the mean of several transforms. The results I got with these methods were not good at all, although I tried several parameters.
In the mean time I succeeded using multiImageRegistration, registering each image to all others at once.

My main concern is that I cannot figure why I won’t get se same result with one image registration to multiple fixed images (moving_A to fixed_A, fixed_B, fixed_C) and with the mean of the transforms of the same moving image to each fixed images (mean of moving_A to fixed_A, moving_A to fixed_B and moving_A to fixed_C, performed using WeightedCombinationTransform) . Could anyone help me with that ?

The reason why I want to switch to the mean of 1 to 1 transforms instead of a single 1 to multi registration is RAM consumption. For now I work with 100Mo stacks, and I completely fill 32Go RAM with the registration of 1 moving and 20 fixed images. I will probably have to work with over 30 of these images at once and this will heavily use swap instead of my RAM, wasting a lot of time. Moreover I will also have to work with larger data (3-9Go images) and I will probably encounter similar memory problems.

I would also be interested if someone could give me any tip to lower RAM consumption.
You can find a typical parameter file attached, used to register 1 moving to 9 fixed images with bspline transform.

Kind regards,

Fabien Pertuy

_______________________________________________
Elastix mailing list
Elastix at bigr.nl<mailto:Elastix at bigr.nl>
http://lists.bigr.nl/cgi-bin/mailman/listinfo/elastix




_______________________________________________
Elastix mailing list
Elastix at bigr.nl<mailto:Elastix at bigr.nl>
http://lists.bigr.nl/cgi-bin/mailman/listinfo/elastix



--
Stefan Klein
+31 10 7043442
http://www.bigr.nl/people/StefanKlein
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.bigr.nl/pipermail/elastix/attachments/20150403/5f1262e1/attachment.html>


More information about the Elastix mailing list