[Insight-users] Should ITK warn about using ResampleImageFilter for downsampling?

[Zachary Pincus]

In answering (I hope correctly to at least a second approximation) the 
last question about interpolation in ITK, I came up with a question of 
my own.

Specifically, it seems that the interpolation functions and the 
resampling framework in ITK are not set up to properly handle 
decimation/downsampling. This is of course reasonable the framework is 
designed around *interpolation* functions and registration tasks which 
should not require significant up- or downsampling.

However, if someone for some reason uses a resample filter to 
substantially downsample (shrink) an image, they might be in for a real 
surprise at the quality of the output. This is because the resampling 
framework only draws one interpolated sample for each output pixel.

For example, if you use the linear interpolation function to downsample 
a 2D image by a factor of 4, then each output pixel should represent 16 
pixels in the input image. However, the output value will be 
interpolated between only four of them. (1)

Now, this state of affairs seems pretty reasonable. There should be no 
need for substantial resampling during most registration tasks, right? 
Moreover, there is a "ShrinkImageFilter" should a specific need arise 
(and for use by the pyramids algorithm). However, it would probably be 
best to warn users that the resampling filters are really only 
appropriate for upsampling, resampling, and only mild (less than a 
factor of two or so) downsampling.

Alternately, simple box, tent, and gaussian interpolator/decimators 
appropriate for up- and downsampling could be written. However, this 
would also necessitate a change in the resampling framework... 
Currently the resampler only asks for the interpolation at the center 
of the pixel region in the transformed space. To properly implement 
downsampling, the interpolation functions need to be passed the entire 
Voroni region "belonging" to a pixel in the output image to decimate 
over. This seems like a bad idea to me. So probably there should be a 
warning about not using the resampling filter to do too severe 
downsampling of images.

Zach


(1) The BSpline and WindowedSinc filters have larger support, so they 
can be used for some downsampling (decimation) tasks. However, a user 
would need to carefully adjust the support for these filters to match 
the amount of decimation going on -- which would be hard to do in a 
registration task, as well as computationally demanding.