[Insight-users] Fundamental question about image registration

Luis Ibanez luis.ibanez at kitware.com
Fri Jul 10 11:50:18 EDT 2009


Hi Matthias,

This is an engineering question, more than a scientific one.

These means that there is not necessarily a "right" answer,
but rather that you have to make a compromise among all the
requirements and conditions of your registration problem.


Factors that you should consider are:


A) What are you going to do with the outcome of the
    registration process ?

    1) Do you use the Transforms for guiding a surgical instrument ?
    2) Are you going to resample the moving image into the
       coordinate system of the fixed image ?


B) Do you have restrictions on the computation time that
    the registration should use to complete ?


For example,
In the case of the Ultrasound to CT image that you mention,
Your Ultrasound is not really a 3D image.  :-)

     Your pixels will have anisotropy ratio of 1:20

You should rather think of it as a collection of 2D image,
in which case it makes more sense to use the Ultrasound image
as the fixed image (or rather a collection of fixed points)

This will require less computation than projecting all the
pixels from the CT scan.

Note that if you are using invertible transforms, you can
always use the inverse transform in order to map the Ultrasound
image to the CT one at the end.



For the case of the PET CT image, you probably want to use
the PET image as the fixed image, again for using less computation,
but, what is most important here is to *select* the pixels that
should be taken into account, and to *exclude* all the pixels that
are not relevant for your registration. In particular, you may
want to exclude the high intensity pixels from the PET image
and use only the low intensities, since those are the ones that
tend to contain traces of information about the anatomy.


It may sound contradictory, since, after all what it interesting
from the PET image is the high signal values, but, by the same
token, those high signal values are the features that do not
exist in the CT image, (and the very reason why you need to
register them), so, the high signal features do not have
correspondences  in the CT and will not contribute to the
registration process.


That said,

Your overall image analysis problem may have additional
restrictions and conditions that may swivel the choice of
the fixed vs moving image on way or the other.



PS. I'll be surprised to find a publication on any
     practical aspects of our trade, such as the important
     question that you raised here.


     Journals and conferences in our field are too busy
     looking for "original" methods that can not be
     reproduced.




     Regards,



         Luis



---------------------
Matthias Keil wrote:
> Dear all,
> 
> me and a colleague came up with a fundamental question about how to 
> choose the fixed and moving image for the registration.
> 
> My colleague is working with a PET Volume with a pretty bad resolution 
> and a CT volume with a small voxel size.
> 
> In my case I deal with an Ultrasound volume with 0.1132 mm resolution in 
> plane and 2 mm plane thickness and a CT volume with 0.5 mm in each 
> direction.
> 
> The question is now would you choose the volume with better resolution 
> as the fixed image or would you use it to move around?
> 
> I tend to use the CT images (better resolution) as moving images.
> 
> Is there a rule of thumb or a scientific answer to our problem? Maybe 
> even a publication?
> 
> Thanks in advance for your help!
> Matthias
> 


More information about the Insight-users mailing list