Elaborate strategy to deal with other contrasts

[jcohenadad]

TODO

[valosekj]

In progress in Google Slides report and in #5

[valosekj]

Below is a summary of different spinal cord segmentation methods (sct_propseg, sct_deepseg_sc) with various settings (different contrasts, with or without init centerline) across contrasts.

The version of the script I used is available here.

The full QC is stored: ~/duke/projects/canproco/canproco_other_contrasts_SC_seg_2022-02-03.

STIR (available only for Calgary)

• -c t2 works better -c t1
• sct_deepseg_sc works better than sct_propseg
• no big difference when using the init centerline (registered from T2w space) compared to no provided centerline

--> Conclusion: use sct_deepseg_sc -c t2 without init centerline

Example:

PSIR (available for Edmonton, Vancouver, Montreal, and Toronto)

• -c t1 works better -c t2
• sct_deepseg_sc with 3D kernel works better than sct_deepseg_sc 2D. But still, results are worse compared to sct_propseg
• sct_propseg works better than sct_deepseg_sc, but still, results are not perfect and will need manual corrections
• no big difference when using the init centerline (registered from T2w space) compared to no provided centerline

--> Conclusion: use sct_propseg -c t1 without init centerline

Examples:

T2star

• sct_deepseg_sc -c t2s works well but some slices will need manual corrections

Example:

MT-T1

• sct_deepseg_sc -c t1 works well

Example:

MTon

• sct_deepseg_sc -c t2s works well

Example:

MToff

• no method works - not surprising - same conclusion done within the contrast-agnostic project

Example:

[valosekj]

TODO: Try the contrast-agnostic model