Binary and softsegs produce the same registration results (is thresholding performed on the segmentation?)

[plbenveniste]

Opening this issue to investigate if there is a thresholding during image registration using algo=centermass.
The investigation comes from the fact that we saw that the registration output of a registration done using a soft SC seg and the output of the registration done using a binary SC seg are identical.

This is related to issue 91 and PR 95.

[joshuacwnewton]

It looks like there is a rounding step specifically for the PCA computation used by centermass and centermassrot here:

spinalcordtoolbox/spinalcordtoolbox/registration/algorithms.py

Line 1434 in 0481a03

data2d = data2d.round().astype(int)

I think that this is essentially equivalent to binarizing the image at 0.5? (<=0.5 --> 0, >0.5 --> 1)

Note

It turns out that for sct_register_to_templtae, there is also another binarization step that gets performed on all segmentations (regardless of algo):

spinalcordtoolbox/spinalcordtoolbox/scripts/sct_register_to_template.py

Lines 451 to 457 in 0481a03

	# binarize segmentation (in case it has values below 0 caused by manual editing)
	printv('\nBinarize segmentation', verbose)
	ftmp_seg_, ftmp_seg = ftmp_seg, add_suffix(ftmp_seg, "_bin")
	img = Image(ftmp_seg_)
	out = img.copy()
	out.data = binarize(out.data, 0.5)
	out.save(path=ftmp_seg)

But this does not occur for sct_register_multimodal.

If you have any other questions, feel free to ask! :)