[Insight-users] Registration of 3D images

Luis Ibanez luis.ibanez@kitware.com
Thu, 17 Oct 2002 14:22:47 -0400


Hi Ghassan,


As you already said,

Performance of registration methods depends on many factors.
In general, you have to go through a process of fine-tunning
parameters if you want to get the best out of a registration
method.

Here is however, a nominal time that will give you an
idea of the order of magnitude for the registration time.



This is running an existing example in ITK:

       Insight/Examples/MultiResMIRegistration

Which performs registration of 3D images using

   - Multiresolution in a pyramid with 5 levels
   - MutualInformation metric
   - Affine transform (which in 3D == 12 Degrees of Freedom)
   - GradientDescent optimizer 2500 iterations per level

Applying this method to two BrainWeb images in our ftp site:

   - brainweb165a10f17.raw
   - brainweb165a10f17Rot10Tx15.raw

Where the second image is a version of the first one
manually rotated 10 degrees and translated 15 mm.

The dimensions (in pixels) of these images are

        181 x 217 x 181

with pixel spacings = 1.0 x 1.0 x 1.0  mm


The full processing took (as timed with Unix 'time'):

    186.81s user 1.48s system 88% cpu 3:33.13 total


That is: roughtly 3 minutes.


This was executed in a single processor Pentium 4
at 1.8Ghz with 512Mb RAM. Under Linux Redhat 7.1,
compiled with gcc3.0.2.  The program was compiled
for debugging (without any optimization).


The results were:

Overall transform matrix:
       0.984805    0.173661     -5.92032e-05
      -0.173661    0.984805     -9.64365e-05
       4.15563e-05 0.000105253   1.0

Overall transform offset:
     -14.9961  0.0267664  -0.0157536


You can verify that 0.1736 is about the sin(10 deg)
and the translation of -14.99 is quite close to the
expected -15.0 value.





Please let us know if you have further questions.


   Thanks



    Luis


===============================================================

Hi,

I am interested in getting a rough idea on how long it takes to register 
two 3D
volumes using ITK. I understand that this depends on many factors. Could 
those
of you who have tried it provide me with feedback on their experience, 
perhaps
by stating: time needed, computer/OS specs., spatial transformation/DOF,
dimension of images, etc, in a line or two,

Your feedback is very much appreciated,

/Ghassan

Ghassan Hamarneh, PhD
Hospital for Sick Children
Toronto, Canada