[Insight-users] Similarity Metrics - Clarification - Regions of Consideration

[M. Wirtzfeld]

Good evening,

I am a novice ITK user and I am current working on an image- 
registration project.  I would like to clarify the regions over which  
similarity metrics are calculated.


Based on very general descriptions of both the  
itk::MeanSquaresImageToImageMetric and the  
itk::NormalizedCorrelationImageToImageMetric classes ("Detailed  
Description" section of the Doxygen documentation) and the more  
detailed description of the  
itk::MattesMutualInformationImageToImageMetric, it appears that these  
metrics are calculated using a region of the fixed-image and a  
corresponding moving-image region from an applied transform and  
interpolator.

The ITK Software Guide, page 415, Section 8.10, "Metrics", states:   
"Typically, the metric samples points within a defined region of the  
fixed image. For each point, the corresponding moving image position  
is computed using the transform with the specified parameters, then  
the interpolator is used to compute the moving image intensity at the  
mapped position."


With this in mind, and considering the entire fixed-image area/region  
unless noted otherwise, three cases come to mind:

1.)  No initial overlap (even with an applied initial tranformation)  
of the fixed-image and moving-image.  In this obvious case, no metric  
can be determined.

2.)  Partial overlap of the fixed-image and moving-image.   
Considering an initial identity-transform with no rotation/ 
translation/etc., only the common area/region of the fixed and moving  
images are used to determine the metric.  If an initial rotation/ 
translation/etc. is applied along with the identity-transform, the  
common area/region could be smaller or larger depending on the  
resulting movement.

3.)  Fixed-image bounded by moving-image.  As with 2., consider an  
initial identity-transform.  Either because the fixed-image area/ 
region is smaller than the moving-image area/region or a region-of- 
interest has been applied to the fixed-image, two possibilities  
exist.  a.)  No initial rotation/translation/etc. - all fixed-image  
area/region is used for metric determination.  b.)  Large rotation/ 
translation/etc. partial fixed-image area/region is used depending on  
amount of movement.  Of course, a non-identity transform could also  
change number of pixels/voxels considered.


The ITK Software Guide (Section 8.10) indicates the SetFixedImageMask 
() and SetMovingImageMask() methods of the itk::ImageToImageMetric  
class can also be used to refine metric calculations.   With respect  
to the SetFixedImageMask() method, an  
itk::RegionOfInterestImageFilter could also be applied to define a  
smaller, fixed-image working area/region?  How is the  
SetMovingImageMask() typically applied?


Thank you,

Michael.


M. Wirtzfeld
Robarts Imaging
London, ON

"There are three kinds of lies: lies, damned lies, and statistics." -  
Benjamin Disrael





-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.itk.org/pipermail/insight-users/attachments/20080715/98e69f0d/attachment.htm>