Negative value(s) in WM segmentation causes fail of snr_single computation in compute_cnr.py

[valosekj]

When processing some of my data, I encountered an issue with the computation of the snr_single. The issue occurs when the WM segmentation contains negative values.

Description

If the GM segmentation exceeds the SC segmentation:

The WM segmentation created by the sct_maths command:

gm-challenge/process_data.sh

Lines 125 to 126 in f8fd2b8

	# Generate white matter segmentation
	sct_maths -i ${file_1_seg}${ext} -sub ${file_1_gmseg}${ext} -o ${file_1}_wmseg${ext}

includes -1 value(s):

and also all voxels which were zero now have small negative values close to zero:

This results in a negative variance in the compute_cnr.py script:

gm-challenge/gmchallenge/compute_cnr.py

Lines 93 to 94 in f8fd2b8

	variance = np.average((values - average) ** 2, weights=weights)
	return np.sqrt(variance)

for example:

which in turn causes that return np.sqrt(variance) returns NaN and computation of the noise_single_slice and snr_single_slice results also in NaN:

gm-challenge/gmchallenge/compute_cnr.py

Lines 112 to 114 in f8fd2b8

	noise_single_slice = \
	[weighted_std(data1[..., iz], weights=mask[..., iz]) for iz in range(nz) if np.any(mask[..., iz])]
	snr_single_slice = [m / s for m, s in zip(mean_in_roi, noise_single_slice)]

Workaround

The possible workaround which fixes the issue for me is the thresholding of the WM segmentation, for instance, like this:

sct_maths -i ${file_1}_wmseg${ext} -thr 0 -o ${file_1}_wmseg${ext} 

SCT version

$ sct_check_dependencies -short                                                     

--
Spinal Cord Toolbox (git-HEAD-45b1c1759f991332697b4120160303e963f7a3c7)

sct_check_dependencies -short
--

SCT info:
- version: git-HEAD-45b1c1759f991332697b4120160303e963f7a3c7
- path: /usr/local/lib/sct_latest
OS: linux (Linux-5.10.0-0.bpo.9-amd64-x86_64-with-debian-10.12)
CPU cores: Available: 16, Used by ITK functions: 16
RAM: Total: 128744MB, Used: 15793MB, Available: 111367MB 

[jcohenadad]

Great catch @valosekj ! And excellent detective work 🕵️

Issue opened here: spinalcordtoolbox/spinalcordtoolbox#3843