Template-based muscle registration for voxel-wise muscle analysis

mm_register_to_template registers an input image and its segmentation to a predefined MuscleMap anatomical template.
This enables standardized voxel-wise analyses, such as spatial parametric mapping of intramuscular fat, across subjects.

This step is typically performed after segmentation with mm_segment.

1. Basic usage 

mm_register_to_template -i image.nii.gz -s image_dseg.nii.gz -r abdomen

This command:

  • loads the input image (-i)
  • loads the corresponding segmentation (-s)
  • registers the image and each label to the selected regional template (-r)
  • outputs warped images, segmentations, and transformation fields
Warning
The segmentation must match the input image dimensions and orientation. Always use the segmentation produced for that exact image.

2. Requirements

2.1 Spinal Cord Toolbox

This tool depends on Spinal Cord Toolbox (SCT) and expects:

  • SCT version 6.5
Warning
MuscleMap’s template registration workflow is developed and tested with Spinal Cord Toolbox 6.5. Other versions may not behave identically.

The SCT installation directory must be available via:

export SCT_DIR=/path/to/spinalcordtoolbox

3. How the registration works

For each label in the segmentation (each muscle or structure with label > 0), the following steps are performed:

  1. Label extraction
    A binary mask is created for the current label.

  2. Centerline estimation
    A slice-wise center-of-mass is computed to estimate a label-specific centerline.

  3. Initial affine alignment
    Three landmark points (≈10%, 50%, 90% along the centerline) are used to compute a constrained affine transform, which is applied to:
    • the input image (linear interpolation)
    • the label mask (nearest-neighbour interpolation)

  4. Nonlinear registration to the template
    Using sct_register_multimodal, the affine-initialized image and label are nonlinearly registered to the template.

  5. Warp concatenation and application
    Final transformation fields are concatenated and applied to generate outputs in template space.

4. Outputs

For each label, the following files are generated (naming simplified):

  • *_dseg_label-<L>.nii.gz — binary label mask in native space
  • *_label-<L>_affine.nii.gz — image after initial affine alignment
  • *_dseg_label-<L>_affine.nii.gz — label after affine alignment
  • warp_*_affine2<region>_template.nii.gz — nonlinear warp field
  • warp_*2<region>_template.nii.gz — concatenated warp (native → template)
  • *_label-<L>2<region>_template.nii.gz — image in template space
  • *_dseg_label-<L>2<region>_template.nii.gz — label in template space

5. Quality control

Always visually inspect registration quality:

fsleyes template.nii.gz image_label-<L>2abdomen_template.nii.gz &

Check for:

  • correct anatomical alignment
  • absence of flips or large shifts
  • plausible deformation of each label
Warning
Registration may fail for very small or noisy labels. Inspect representative labels before group analyses.

6. Voxel-wise correlation analysis (Randomise example)

After registration, template-space metric maps can be analysed voxel-wise.

6.1 Merge subjects into a 4D file 

fslmerge -t all_subjects_metric_4D.nii.gz sub-01_metric.nii.gz sub-02_metric.nii.gz

6.2 Create a template-space mask 

fslmaths mask.nii.gz -bin mask_bin.nii.gz

6.3 Run Randomise 

randomise   -i all_subjects_metric_4D.nii.gz   -o stats_metric_corr   -m mask_bin.nii.gz   -d design.mat   -t design.con   -n 5000   -T
Note
Intermediate reflects a voxel with intermediate voxel signal not clearly corresponding to either fat or muscle.