[Insight-users] Offset CenteredTransformInitializer

Mon Sep 3 19:05:52 EDT 2007

Hi Luis,

Thanks for your answers, really appreciate it. The reason I asked
about setting the center is that it had quite an effect in registration
problems which i made myself. An MRT1 image and an MRA image,
and quite a big rotation (10 mm, 10 degrees).
So based on this I figured that setting the centre is important in a real
registration problem (however here you don't know the true center).
My problem is the registration of pre-operative MRI images (T1, T1 contrast,
T2, MRA, FLAIR and DTI). The images are taken in the same session so the
registration shift is really small (maybe max. a few milimeter)
My supervisor always uses (0,0,0) as center, this being the center 
of the MRI field (most homogeneous). Ive tried this and it works good but Im
trying to find out if choosing a different center gives better results. 
regards,
Lars Nygard

----- Original Message ----
From: Luis Ibanez <luis.ibanez at kitware.com>
To: Lars Nygard <lnygard at yahoo.com>
Cc: insight-users at itk.org
Sent: Monday, September 3, 2007 3:14:00 PM
Subject: Re: [Insight-users] Offset CenteredTransformInitializer

Hi Lars,

Setting the center is not "Mathematically" necessary.

In fact, introducing the notion of "center of rotation"
makes that the ITK transforms are over-determined.

However, having that center explicitly set, serves a
practical purpose. It helps the optimizer have an
easier walk in the parametric space of the Transform.

In a practical image registration problem, you have to
choose the initialization based on the characteristics
of both images.

For example

                if you have two MRI images, one is T1, the
other is T2, and they fully contain the head of a patient,
the you will be fine using Geometry. While Moments will
not work well due to the fact that the intensities will
be concentrated in different regions in the T1 and T2
modalities.

If you have two CT images of the head, then you are
probably safe using Moments, because, being a single
modality, chances are that the center of mass (assimilating
intensity to mass) will be located in the same
anatomical position. Note however that IF one of the
images has a contrast agent and the other doesn't, then
the assumption on the intensity distribution breaks and
you are better off using Geometry again.

If you have an abdominal CT scan and you are trying
to register a local 3D ultrasound of the liver, then
none of Geometry, nor Moments will do you any good,
since both the field of view and the modalities are
different between the images. In this case, your best
option is the landmark initializer.

As you see,
(just as in the rest of image analysis)
there is not a single recipe that works for all the cases.

    If you have a particular scenario in mind,
    and share it with us,
    then we will be happy to suggest you a suitable
    initialization mechanism.

      Regards,

         Luis

-----------------
Lars Nygard wrote:
> Hi Luis,
> 
> My program is working as expected but I have some questions about
> setting the center. As said I get the best results when I use the geometric
> center. This is not very surprising I just figured because this is also the center
> I used when creating the registration problems.
> The momentsOn calculates a center which is far off.
> My question is how do you determine the center in a real registration problem?
> Is there a unique point around which rotatations occur?  
> Is it possible to constrain the registration problem with different rotation
> axis?
> greets,
> Lars
> 
> ----- Original Message ----
> From: Luis Ibanez <luis.ibanez at kitware.com>
> To: Lars Nygard <lnygard at yahoo.com>
> Cc: insight-users at itk.org
> Sent: Tuesday, August 28, 2007 1:48:54 PM
> Subject: Re: [Insight-users] Offset CenteredTransformInitializer
> 
> 
> Hi Lars,
> 
>     Please help us solve the mystery.
> 
> Share with use the following information:
> 
> 1) Coordinates of the Fixed image origin
> 2) Fixed image spacing
> 3) Fixed image number of pixels
> 4) Coordinates of the Moving image origin
> 5) Moving image spacing
> 6) Moving image number of pixels
> 
> 7) Center of rotation computed by the
>     CenteredTransformInitializer.
> 
> 
> Also, notes that the MomentsOn will set
> the center of rotation in the center of
> mass of your pixel intensity values.
> 
>     What is the content of your image ?
>       a brain ? is it centered ?
>     or is it a microscopy image with a huge
>       organelle on the corner of the image ?
> 
> 
> With this information at hand we shoul be
> able to figure out what is going on in your
> application.
> 
> 
> 
>     Thanks
> 
> 
>        Luis
> 
> 
> ------------------
> Lars Nygard wrote:
> 
>>Hi,
>>
>>Im getting wrong results using the CenteredTransformInitializer in both
>>MomentsOn and GeometryOn. When I manually set the center of the transform
>>as the center of the fixed image (origin + Spacing*Dimension/2) everything goes OK. 
>>But when I use the  initializer Im ending way besides the correct registration result.
>>I was wondering, is the Offset used when the CenteredTransformInitializer calculates
>>the center (in GeometryOn)??
>>And does anyone can explain why Im getting good results with setting the center manually
>>while the results aren't good when using MomentsOn??
>>Thanks for any feedback,
>>greets,
>>Lars
>>
>>
>>
>>
>>      
>>_________________________________________________________
>>Alt i én. Få Yahoo! Mail med adressekartotek, kalender og
>>notisblokk. http://no.mail.yahoo.com
>>_______________________________________________
>>Insight-users mailing list
>>Insight-users at itk.org
>>http://www.itk.org/mailman/listinfo/insight-users
>>
> 
> 
> 
> 
> 
> 
>       
> _________________________________________________________
> Alt i én. Få Yahoo! Mail med adressekartotek, kalender og
> notisblokk. http://no.mail.yahoo.com
> 

_________________________________________________________
Alt i én. Få Yahoo! Mail med adressekartotek, kalender og
notisblokk. http://no.mail.yahoo.com