The threshold parameter of `sct_deepseg` doesn't work with the model `seg_sc_contrast_agnostic` for soft segmentation. about spinalcordtoolbox

[naga-karthik]

yes! But it was a bad quick fix given the urgency at the time. Surely, a cleaner way exists ...

@joshuacwnewton let me know if you're looking into it or should I. In case you are looking into it, please note that the default threshold we decided was 0.1. That means, if a user explicitly requires softseg then SCT should output a result with 0.1 threshold. This threshold value was chosen based on comparing different thresholds and their downstream effects on the spinal cord CSA! (details here)

[simonqueric]

Thank you for finding this issue! That's very strange... I'm happy to investigate!

In the short term, does -thr 0.01 work for you?

Yes this threshold works on the branch master. Notice that the soft-segmentation works on the branch nk/4481-add-monai-postprocessing.

[joshuacwnewton]

Thank you for finding this issue! That's very strange... I'm happy to investigate!

In the short term, does -thr 0.01 work for you?

[jcohenadad]

I think @naga-karthik fixed it here

[joshuacwnewton]

Taking a look at this now! My apologies for the delay.

[joshuacwnewton]

I was able to fix this issue and generate a softseg, and I'll be pushing a fix ASAP.

Off-topic

However, I noticed when I test with t2.nii.gz, there are negative values in the output softseg? Also, the max value is > 1?

This surprised me, as I expect softsegs to be in the range [0, 1]. Have you experienced this from the contrast agnostic model in the past?

[jcohenadad]

However, I noticed when I test with t2.nii.gz, there are negative values in the output softseg? Also, the max value is > 1?

could this be caused by interpretation when the prediction is resampled back to the input image resolution? But this would only happen with spline interpolation of order 2+ (which should not be used for segmentation anyway)

[joshuacwnewton]

@jcohenadad You're exactly right, opening a new issue. :)