Registering 2D RGB images

[kretep]

Hi, I'd like to register two 2D RGB images (16-bit unsigned int, in TIFF format). This doesn't seem to work out of the box. I tried extracting the red channel and registering those, but the 2D 16-bit pixel type doesn't seem to be available. I also tried converting the RGB image to a 3-layered TIFF image and that seems to work in that it performs the registration, but also in the through-stack dimension. Is it possible to constrain the transform to 2D? Or is there a different approach that I might be missing?

[N-Dekker]

Hi Peter! Thanks for your question! Do you use the elastix executable or the library interface? Or one of the Python wrappers, SimpleElastix, SimpleITK, or ITKElastix?

[kretep]

Hi Niels, at the moment I'm trying the executable. (I also considered the wrappers, but for various reasons they didn't seem to work out-of-the-box on my Apple Silicon and I prefer not having to build the binaries myself.)

[N-Dekker]

OK, so then can you possibly try "float" as internal pixel type, for your 2D 16-bit unsigned images? In your parameter txt files, as follows:

(FixedInternalImagePixelType "float")
(MovingInternalImagePixelType "float")

Internally, elastix will then convert your pixel data to 32-bit float, but it should still work fine. At least, it should!

[kretep]

Thanks, that seems to work in that the registration runs, but what do I then use as the ResultImagePixelType? I tried various combinations using only one-channel images for now.

When I use "unsigned short", the image seems completely black. When applying transformix on the same image, the image also seems black, but I can retrieve (some) information by dramatically increasing the contrast.

When I do the registration on 8-bit images and use "unsigned char", the result does show the image. But my goal is to preserve all the information that is in the 16-bit images... any ideas?

[N-Dekker]

Maybe the final interpolation used when generating the result caused the output image to look black. Which value do you have for ResampleInterpolator? I'm not sure, but maybe you can try nearest neighbor:

(ResampleInterpolator "FinalNearestNeighborInterpolator")

You can probably just run transformix, it should not be necessary to rerun the registration.

Please let us know if that helps!

[kretep]

I had another go at debugging it and it turned out that when I extracted the color channels, it converted to 8-bit images by default instead of 16-bit. So after making sure I got the 16-bit inputs, it now works using float as the internal pixel type.