Skip to content

S7: Normalization

Step Code: S7_template_normalization Depends on: S2 (anat-to-PAM50 warps), S6 (BOLD-to-anat warps), S5 (functional reference) Required by: S8 (Confounds), S9 (Functional Derivatives), S10 (QC & Release)

Purpose

S7 links each functional run to the PAM50 spinal cord template (De Leener 2018), so that template masks and atlases can be brought into the run's own space and so that group analysis has a common coordinate system. It does this by composing warps, not by resampling the data:

  • It fuses S2's anat-to-PAM50 warp with S6's BOLD-to-anat warp into a single composite BOLD-to-PAM50 transform (and its inverse).
  • It uses the inverse to pull the PAM50 cord, tissue masks and level atlas into native functional space.

The 4D BOLD is never resampled into PAM50

The per-subject GLM stays in native functional space, matching SCT batch_processing.sh, the CoSpi lineage, and Eippert 2017. Resampling 4D BOLD into template space would cost signal-to-noise and disk for no per-subject benefit. Only S9 pushes contrast/statistic maps to PAM50 for group inference.

S7 trusts S2's anat-to-PAM50 warp as the single source of truth for the rootlet-versus-disc registration decision; it does not redo that step.

Algorithm

  1. Compose warps. The S2 (anat↔PAM50) and S6 (BOLD↔anat) warps are concatenated (sct_concat_transfo) into the composite BOLD↔PAM50 transform.
  2. Optional EPI-level refinement (disabled by default). A second sct_register_multimodal pass against the functional reference, initialised from the composed warp, is supported but off by default (see below). Its reference is PAM50_t2s, the best contrast match for T2*-weighted BOLD.
  3. Warp the atlas to native space. sct_warp_template -a 1 brings the full PAM50 pack into native functional space: cord, CSF, WM and GM masks plus the spinal-level atlas.

Refinement is off by default

The SCT batch_processing.sh fMRI block includes a refinement pass; S7 disables it because it added no information over S6's already-thorough cord-driven registration and, on the pain dataset, dropped cord Dice from 0.82 (compose-only) to 0.68. When enabled, the pass runs slicereg (seg, MeanSquares, smooth 2) then bsplinesyn (image, MeanSquares, iter 5, gradStep 0.5) — the SCT recipe verbatim. Set refinement.enable: true to opt in.

This is a cervical pipeline: PAM50 coverage is used for levels C1-T1. The shipped spinal-level atlas spans the whole column, but only the cervical levels are used in practice.

Step Metric and QC

The step-local truth metric is cord Dice in native functional space — the overlap between the PAM50 cord (warped through the composite transform into native space) and the S6 cord segmentation. It tests whether the composite S2∘S6-and-refine warp actually lands the template cord on the functional cord (Cohen-Adad 2014).

The primary gate is the median per-level cord Dice, computed level by level along the cord. This is deliberately coverage-independent: a brain-plus-cord acquisition images only a few cervical levels, which caps whole-volume overlap near 0.84 even for a perfect registration, so a per-level median is the fair measure of registration quality where the cord is actually present. The whole-volume cord Dice is kept as a legacy fallback and observability value for when per-level Dice is unavailable.

A cord-restricted forward-then-inverse round-trip drift is also reported, as observability only — bsplinesyn optimizes forward and inverse separately, so some drift is intrinsic even at high Dice.

Diagnostic Reportlets

  • pam50_on_func — the functional reference with the warped PAM50 cord contour overlaid, in sagittal and axial views. This is the direct visual check that the template cord lands on the functional cord.
  • cord_dice_per_level — a per-vertebral-level cord Dice bar chart, colored by the PASS/WARN bands. A single low level flags a broken edge slice; uniformly low bars flag a global composition problem.

Outputs

derivatives/spineprep/{dataset}/sub-{id}/func/
├── sub-{id}_..._from-bold_to-PAM50_xfm.nii.gz    # composite warp (+ .json sidecar)
├── sub-{id}_..._from-PAM50_to-bold_xfm.nii.gz    # composite inverse
├── sub-{id}_..._space-PAM50_desc-funcref.nii.gz  # funcref in PAM50 (QC overlay only)
├── sub-{id}_..._desc-PAM50cord_mask.nii.gz       # PAM50 cord in native func
├── sub-{id}_..._desc-PAM50csf_mask.nii.gz        # + wm / gm masks
└── sub-{id}_..._desc-PAM50spinallevels.nii.gz    # spinal-level atlas in native func

derivatives/spineprep/{dataset}/sub-{id}/figures/
├── sub-{id}_..._desc-S7_pam50_on_func.png
└── sub-{id}_..._desc-S7_cord_dice_per_level.png

The warps are stored as SCT-native displacement fields (.nii.gz), matching the S2 and S6 convention, with a JSON sidecar carrying the reproducibility receipt.

QC Status Logic

status = PASS if:
  - median per-level cord Dice >= per_level_pass_min (0.90)
    (fallback when per-level Dice is unavailable: whole-volume Dice >= 0.80)

status = WARN if:
  - median per-level cord Dice below the PASS gate but not failing, OR
  - any single level Dice < per_level_broken_below (0.50) — that level is excluded

status = FAIL if:
  - whole-volume cord Dice < fail_dice_below (0.65), OR
  - sct_register_multimodal exits non-zero

On the 11-run validation cohort, S7 lands 10 PASS and 1 WARN; the single WARN is a genuinely borderline run (Dice 0.797 against the 0.80 whole-volume gate) rather than a failure — the metric is calibrated correctly. All runs ran with refinement disabled.

References

  1. PAM50 template: De Leener et al. NeuroImage 165:170-179 (2018).
  2. Native-GLM convention: Eippert et al. (2017).
  3. Cord registration validation (Dice): Cohen-Adad et al. (2014).
  4. Rootlets-based PAM50 registration: Valošek et al. (2025).
  5. SCT batch_processing.sh fMRI block — canonical refinement recipe.

Last updated: July 2026