[Insight-users] Re: Registration problem

[Luis Ibanez]

Hi Frank,

Thanks for posting your images, they help a lot
to illustrate your problem.

Since the two CT datasets have such large differences
in intensity you may want to use as image metric the

    NoramalizedCorrelation metric
http://www.itk.org/Insight/Doxygen/html/classitk_1_1NormalizedCorrelationImageToImageMetric.html

This metric is invariant to linear transformation
of the image intensities.

Considering the field of view of your images you certainly
want to crop them first to the bounding box of the lungs.

If you use the RegionOfInterestImageFilter for doing the
cropping you will be safe for doing registration with the
cropped images, because this filter computes the origin
of the cropped image with respect to the input image.
Therefore all the physical coordinates of pixels remain
unchanged.


You could still consider segmenting the lungs, but you
probably don't needed unless high speed is a premium
for your application.

If you find that high speed is a real need, then you
may want to consider the following process:


1) Segment the lungs with ConfidenceConnectedImageFilter
    (you will have to truncate the region at the airways
    somehow) The cropping may already be enough for that.

2) Use BinaryMaskToNarrowBandPointSetFilter
http://www.itk.org/Insight/Doxygen/html/classitk_1_1BinaryMaskToNarrowBandPointSetFilter.html
    in order to generate a PointSet with points located
    in a band around the surface of your segmentation.

3) Use the PointSetToImageRegistration
http://www.itk.org/Insight/Doxygen/html/classitk_1_1PointSetToImageRegistrationMethod.html

    in order register this PointSet against the other image.
    Still you will want to use the NormalizedCorrelation
    metric, but this time the PointSet-to-Image Version
http://www.itk.org/Insight/Doxygen/html/classitk_1_1NormalizedCorrelationPointSetToImageMetric.html


This should result in very fast registration since only
a small subset of the points are used, and the metric
in concentrated on the points that make the more
significant contributions to the registration.




Regards,





--------------------
Frank Zhao wrote:

> 
> Hello Luis,
> 
> I have two 3D CT image needed to be registred(We only care about lungs), 
> I attached one slice of them. As you can see, they have different 
> bacground and different intensity value. I directly use them as input 
> for registration, the result is very poor. should I first segment the 
> lung out, and remove background noise( for we only care about lungs)?
> 
> Thank you for your help
> 
> 
> Frank
> 
> _________________________________________________________________
> Overwhelmed by debt? Find out how to ‘Dig Yourself Out of Debt’ from MSN 
> Money. http://special.msn.com/money/0407debt.armx