Using the atlas to extract MTR in white matter

Now that the atlas has been warped to the MT space, it can be used to extract MTR for specific white matter regions. First, we will extract MTR for the white matter as a whole.

sct_extract_metric -i mtr.nii.gz -f label/atlas -method map -l 51 -o mtr_in_wm.csv
Input arguments:
  • -i : Image to extract values from.

  • -f : File or folder used to pick out specific regions from the input image. In this case, we supply the folder of the white matter atlas that was transformed to the space of the MTR data.

  • -method : We specify map to choose the Maximum a Posteriori method, which helps to account for the partial volume effect. This method is most suitable for small regions. This method should only be used with the PAM50 white/gray matter atlas (or with any custom atlas) as long as the sum across all overlapping labels equals 1 in each voxel part of the atlas.

  • -l : The IDs of the label (or group of labels) to compute metrics for. In this case, label 51 is a combined label that represents all of the WM tracts together. You can see the full list of labels to choose from by running sct_extract_metric -list-labels.

  • -o : The name of the output file.

mtr_in_wm.csv: MTR values in white matter

Slice (I->S)

Label

Size [vox]

MAP()

STD()

0

white matter

0

None

None

1

white matter

48.128699333405

56.7691774857722

7.13099896095076

2

white matter

63.3587044639798

41.9298789113495

5.4096594947144

3

white matter

58.324215639338

55.8216325391571

4.36523840112145

4

white matter

66.8988296465732

52.8017466468748

4.69537815140102

5

white matter

72.5729277595079

53.2903651621953

5.32202535964913

6

white matter

74.921209288943

52.9356588028125

4.66731745304554

20

white matter

72.8836498521706

62.0087038623367

5.50098090540239

21

white matter

0

None

None

The label volume fraction is indicated as “Size [vox]”, which gives you a sense of the reliability of the measure. In this example, for each slice, the metric was computed based on 50-70 voxels. As demonstrated in [De Leener et al., Neuroimage 2017; Appendix], having at least 30 voxels results in an error smaller than 2%, while having at least 240 voxels results in an error smaller than 1% (assuming an SNR of 10).

Warning

Be careful to always check the associated volume fraction of your metrics (indicated by the Size [vox] column). This is especially relevant if you are restricting the metric computation to a subset of vertebral levels or axial slices; if the number of voxels included in the computation is too low, your quantifications will be unreliable. If you publish, we recommend including the volume fraction associated with all estimated metrics.