[Insight-developers] Patch for some filters that don't propagate direction cosines

[Peter Cech]

On Tue, Sep 12, 2006 at 09:41:03 -0500, Kent Williams wrote:
> Hans ran into an issue with itk::JoinSeriesImageFilter and 
> itk::ExtractImageFilter -- they did not propagate the direction cosines 
> from the input to the output.
> 
> I have a feeling that this is also the case for a number of the other 
> filters -- we only notice that they don't set the output Directions when 
> we try to use them and they're always identity.
> 
> I can write a simple template function that does the Direction Cosine 
> assignment between the input and output images, taking into account the 
> dimensionality of both -- that's what I do in-line with this patch.  
> Then it can be called from any filter.
> 
> My question is where I would put this function:
> 
> 1. As a member function of itkImageBase?
> 2. As a free-standing template function in itkImageBase?
> 3. As a member function of itkImageToImageFilter.h
> 4. As a free-standing ""

I would say 4 as it is unlikely to be used outside image filters and
does not depend on any member of itkImageToImageFilter. 

> 5. Something else?

Namespace itk::ImageToImageFilterDetail
http://www.itk.org/Doxygen/html/namespaceitk_1_1ImageToImageFilterDetail.html


The following patch does not work:

> Index: itkExtractImageFilter.txx
> ===================================================================
> RCS file: /cvsroot/Insight/Insight/Code/BasicFilters/itkExtractImageFilter.txx,v
> retrieving revision 1.21
> diff -c -r1.21 itkExtractImageFilter.txx
> *** itkExtractImageFilter.txx	27 Jan 2006 17:07:07 -0000	1.21
> --- itkExtractImageFilter.txx	12 Sep 2006 14:40:32 -0000
> ***************
> *** 200,205 ****
> --- 200,228 ----
>           }
>         }
>   
> +     //
> +     // Copy the direction cosines from the input to the output.
> +     // On extract the output is always smaller or equal dimension
> +     // as the input
> +     typedef typename InputImageType::DirectionType InputDirectionType;
> +     typedef typename OutputImageType::DirectionType OutputDirectionType;
> +     InputDirectionType inputDir = inputPtr->GetDirection();
> +     OutputDirectionType outputDir = outputPtr->GetDirection();
> +     unsigned inputdim = InputImageType::GetImageDimension();
> +     unsigned outputdim = OutputImageType::GetImageDimension();
> + 
> +     for(unsigned i = 0; i < inputdim; i++)
> +       {
> +       for(unsigned j = 0; j < inputdim; j++)
> +         {
> +         if(j < outputdim && i < outputdim)
> +           {
> +           outputDir[i][j] = inputDir[i][j];
> +           }
> +         }
> +       }
> +     outputPtr->SetDirection(outputDir);

Call to outputPtr->SetDirection(...)

> + 
>       // set the spacing and origin
>       outputPtr->SetSpacing( outputSpacing );
>       outputPtr->SetDirection( outputDirection );

Another call to outputPtr->SetDirection(...). It overwrites direction
with stuff prepared in an if-else block just preceding the first lines
of the patch.

Anyway, this patch is computing output direction incorrectly - it
ignores collapsed dimensions (or assumes the are always the last ones).
The current code in CVS is buggy, but I think the attached patch fixes
it.

Regards,
Peter Cech
-------------- next part --------------
Index: itkExtractImageFilter.txx
===================================================================
RCS file: /cvsroot/Insight/Insight/Code/BasicFilters/itkExtractImageFilter.txx,v
retrieving revision 1.21
diff -u -3 -d -r1.21 itkExtractImageFilter.txx
--- itkExtractImageFilter.txx	27 Jan 2006 17:07:07 -0000	1.21
+++ itkExtractImageFilter.txx	12 Sep 2006 17:00:03 -0000
@@ -190,10 +190,15 @@
           {
           outputSpacing[nonZeroCount] = inputSpacing[i];
           outputOrigin[nonZeroCount] = inputOrigin[i];
+          int nonZeroCount2 = 0;
           for (unsigned int dim = 0; dim < OutputImageDimension; ++dim)
             {
-            outputDirection[nonZeroCount][dim] =
-              inputDirection[nonZeroCount][dim];
+            if (m_ExtractionRegion.GetSize()[dim])
+              {
+              outputDirection[nonZeroCount][nonZeroCount2] =
+                inputDirection[nonZeroCount][dim];
+              ++nonZeroCount2;
+              }
             }
           nonZeroCount++;
           }