[vtkusers] Image Resampling

[David Gobbi]

Hi Andrew,

The "masking" part is what makes this especially tricky.  Also, the NN
interpolator is very specifically a "nearest neighbor" interpolator,
its code would have to be changed drastically in order for it to
instead look for the "nearest neighbor that has a value of X or Y".
It would need a whole new inner loop to perform the search.

I have a class that doesn't do quite what you need, but it might
provide you with ideas: it is an interpolator specifically written for
label images.  It models the voxel positions as random variables
(Gaussian) and, for each voxel in the resampled image, it finds the
highest-probability label for that voxel.  It does not, however, allow
you to ignore specific label values.

http://github.com/dgobbi/AIRS/blob/master/ImageRegistration/vtkLabelInterpolator.h

 - David


On Thu, Sep 4, 2014 at 7:00 AM, Andrew Crozier
<andrew.crozier at medunigraz.at> wrote:
> I and a colleague are working on an image segmentation problem, where we use
> marching cubes to generate the surface of a segmentation, apply a custom
> smoothing operation to the surface mesh with numpy in python, then re-render
> as a vtk image at an arbitrary higher resolution before writing out for the
> rest of our workflow.
>
> The original data set is, however, not simply a binary data set, but both
> the 'ones' and 'zeros' are further subdivided into integer 'tag' values. We
> would like to map said tags to the resampled image, but I am trying to
> figure out the most efficient way to do that in VTK.
>
> -----------------
>
> Say, for example, we have an image with four 'tags'; 10, 11, 12 and 13. 10
> and 11 are assigned a zero value for the surface extraction and smoothing,
> and 12 and 13 are assigned ones. The surface is extracted, smoothed, and
> resampled image generated, which now also has ones and zeros as values. I
> would now like to reassign the appropriate tags, such that zeros always
> become either 10 or 11 and ones become either 12 or 13.
>
> Conceptually, I consider this problem as requiring something like a 'masked
> nearest neighbour' approach, meaning that each zero pixel in the resampled
> image should be assigned the tag of the spatially closest 10 or 11 pixel in
> the original image. I say masked nearest neighbour as it seems equivalent to
> a nearest neighbour interpolation (such as by using vtkImageInterpolator)
> but where pixels with a 12 or 13 value are ignored.  The equivaltent mapping
> would of course be done for 'one' pixels in the resampled image, finding the
> closest in the original with a 12 or 13 value.
>
> -----------------
>
> I could, or course, implement this algorithm in pure python as part of my
> script, but I am keen not to do so unless necessary as the resampled data
> sets can be quite large. I have fairly limited experience using vtk and
> would appreciate any suggestions or pointers on how to achieve the desired
> result.
>
> Best wishes,
> Andrew