[Insight-users] Using MI-metrics with image couples that have huge size differences

[Luis Ibanez]

Hi Ralf,

Given the small spatial coverage of the rotation angiography
image it actually makes more sense to use it as the fixed
image.

In your current use of the CT as fixed image you will be
spending a lot of time in mapping pixels that end up being
outside of the range of the rotation angiography image.

However, if you prefer to insist in using the CT as fixed
image, then here are two suggestion that may help:


1) Use the FixedImageRegion: to reduce the region from which
    samples are taken. This may require a bit of user interaction
    for defining the ROI of the CT that corresponds to the angiography
    image.


2) You can disable the check of the MI metric, and just force it
    to continue despite the low ratio of samples that map inside
    the moving image.  Just be aware that there is a tradeof between
    the number of actual samples and the number of bins on your MI
    histogram. You want to have enough samples to populate the histogram.
    If you end up with a reduced number of samples, you may be forced
    to reduce the number of bins in your histogram too.

    If you want to disable the exception, just comment out line 154
    of the file


          Insight/Code/Algorithms/
             itkMutualInformationImageToImageMetric.txx



With an appropriate ROI in (1) you may not need to do (2).


3) Note that if you change your mind and you use the rotation
    angiography image as fixed image, you can always get the
    inverse of the transform that results from that registration
    and that will be (in theory) equivalent to the registration
    that you are currently performing.



     Regards,


        Luis



====================
Floca, Ralf wrote:
> I have used itk in different registration task so far and it worked
> fine. Working on my current task I came across a problem using mutual
> information metrics (e.g. Mattes MI). I have two images: Image 1
> (modality: CT angiography; 160x160x112 mm) and Image 2 (modality:
> rotation angiography; size 47x47x47 mm).
> 
> If I define Image 1 as the fixed image (which would be more suitable
> regarding other parts of the application), the registration normally
> fails in the first step, because to many samples are mapped outside the
> moving image. Looking at the sizes of the image, it becomes clear that
> the chances for that behavior are pretty high. Image 2 (even if the
> transform is initialized with total coverage) covers only around 4% of
> Image 1, therefore with randomly drawn samples it is pretty probable
> that less than 25% samples will be drawn successfully. So the metric
> exception will be thrown.
> 
> A workaround would be to switch both images; but I would appreciate
> another way, if possible. Modifying the metric in a generic way, to be
> aware of the moving image limitations, seems not possible, because on
> the level of itk::Transform there is no possibility to map fixed image
> samples onto moving image samples.
> 
> If anybody can give me a hint which metric would be suitable for my task
> or how I can manage the problem with the MI metrics in a sound way, I
> would be very thankful.
> 
> Best regards,
> Ralf o Floca
> 
> ---------------------------------------------------------------------
> Ralf Floca
> University of Heidelberg                           
> Institute for Medical Biometry and Informatics
> Department of Medical Informatics
> Im Neuenheimer Feld 400            
> D-69120 Heidelberg, Germany                Tel. : +49 (0)6221 56-7484  
> http://www.klinikum.uni-heidelberg.de/mi   Fax  :               -4997
> 
> ralf.floca at med.uni-heidelberg.de           Sekr.:               -7483
> ---------------------------------------------------------------------
> 
>  
> _______________________________________________
> Insight-users mailing list
> Insight-users at itk.org
> http://www.itk.org/mailman/listinfo/insight-users
>