[Insight-developers] Bug found in SignedDanielssonDistanceMap

[Padfield, Dirk R (GE Global Research)]

Hi Brad,

Thank you for finding this bug and letting us know.  After some searching, I found that the problem is a result of an error in the initialization of the voronoiMap variable in the itkDanielssonDistanceMapImageFilter.  There is no problem with itkSignedDanielssonDistanceMapImageFilter since all it does is call itkDanielssonDistanceMapImageFilter twice, one time with an inverted image, and then subtract the two unsigned distance maps to get a signed distance map.  The itkSignedDanielssonDistanceMapImageFilter sets the "SetInputIsBinary" flag of itkDanielssonDistanceMapImageFilter to true, so that the following code is called to initialize the voronoiMap in lines 170-186 of the PrepareData method of itkDanielssonDistanceMapImageFilter:

  if ( m_InputIsBinary )
    {
    VoronoiPixelType npt = 1;
    while ( !ot.IsAtEnd() )
      {
      if ( it.Get() )
        {
        ot.Set(npt++);
        }
      else
        {
        ot.Set(0);
        }
      ++it;
      ++ot;
      }
    }

Here "it" is an iterator over the input image, and "ot" is an iterator over the voronoiMap.  In this loop, the voronoiMap values are ever increasing as the variable "npt" is incremented. This causes the voronoiMap values to wrap-around when "npt" is incremented beyond the value that can be held by the type.  In itkDanielssonDistanceMapImageFilter.h, the type of the VoronoiMap variable is a template parameter, and, if it is not set, it defaults to the type of the InputImage: 
template< class TInputImage, class TOutputImage, class TVoronoiImage = TInputImage >

So if, for example, the input is "unsigned char", which is reasonable, but the image has more than 255 pixels total, which is usual, this npt will lead to wrap-around errors.  I believe this variable should not be incremented.  In that case, the voronoi map resulting after running the whole filter will have values all 1, which makes sense because by stating that "SetInputIsBinary", we are saying there is only one label, and thus the voronoi partition has only one region.  I tested this on a number of images, and it worked correctly.

We should probably also change the two calls to "SetInputIsBinary" from "true" to "false" in itkSignedDanielssonDistanceMapImageFilter.hxx.  As I mentioned, having these set to "true" results in a voronoi map that is all one label.  However, if they are set to "false", this yields a more informative voronoi map (actually, the second one will become binary anyway because of the inversion process, so the flag for that one doesn't matter).  For example, I have attached to this message the voronoi map resulting from turning those flags off for the test image that you used.  If you overlay it with the original image, you will see that it indeed creates the correct voronoi partition of the image from the zero values of the original image.

So, I recommend that we get rid of the "++" of the "npt" variable, and this will solve the problem.  We should also add a regression test to test this case in ITK itself (since you found it with a test in SimpleITK).  I am willing to make these changes and submit a patch.

Any thoughts?

Dirk



Date: Thu, 4 Oct 2012 07:20:33 -0400
From: Bradley Lowekamp <blowekamp at mail.nih.gov>
Subject: [Insight-developers] Bug found in SignedDanielssonDistanceMap
To: ITK Developers <insight-developers at itk.org>
Message-ID: <65179107-1A18-4C78-8334-33CB59CF483D at mail.nih.gov>
Content-Type: text/plain; charset="us-ascii"

Hello,

I had updated SimpleITK's ITK version to the 4.2.1 release candidate and then to 4.2.1 with out a problem.

However, yesterday I upgraded sitk's next branch to the latest ITK master, to get fixes for a label map contour filter, and I am getting wide spread failures for the SignedDanielssonDistanceMap image filter:

http://open.cdash.org/testSummary.php?project=40&name=BasicFilters.SignedDanielssonDistanceMap&date=2012-10-04

Here is the failing comparison to the baseline:

http://open.cdash.org/testDetails.php?test=161367787&build=2595127

While the errors appear to be random, I check a mac, linux and windows system and they appear to be failing in the same fashion.

I am not sure were this change occurred. Does anyone know what may have changed?

Thanks,
Brad
-------------- next part --------------
A non-text attachment was scrubbed...
Name: VoronoiMap.png
Type: image/png
Size: 2759 bytes
Desc: VoronoiMap.png
URL: <http://www.itk.org/pipermail/insight-developers/attachments/20121005/0749a75e/attachment.png>