Skip to content

S6: Functional-to-anatomical registration

S6 registers the functional data to the subject's own anatomy and stores the transform in both directions, so later steps can move data between functional and anatomical space without recomputing the alignment.

What it does

S6 aligns the distortion-corrected mean functional image from S5 to the anatomical cord reference from S2, and writes a forward warp (functional to anatomy) and its inverse. S7 (template normalization) and S8 (confound extraction) reuse these warps rather than recomputing them, so the functional series stays in its native space and is never resampled onto the anatomical grid. The registration is driven by the cord segmentation rather than image intensity, because a cord-cropped EPI has an intensity cost surface dominated by the air around the cord, on which an intensity-only registration diverges. Because the cost sees only the cord outline, the same recipe works whether the anatomy is T1w, T2w or T2star.

Algorithm and parameters

S6 is a single sct_register_multimodal call (De Leener et al., 2017) with a three-stage, cord-segmentation-driven parameter string, run with the functional image as the moving image and the anatomy as the destination, -i funcref -d anat. Before registration the anatomy is cropped to a dilated cord region (sct_crop_image -m anat_seg -dilate 10x10x10), which restricts the cost surface to cord context.

The three stages are built from SCT's standard registration primitives: centermassrot (slicewise center-of-mass and rotation alignment, for oblique acquisitions), columnwise (right-left scaling and anterior-posterior deformation along the cord axis), and bsplinesyn (slicewise nonlinear B-spline refinement, 20 iterations). All three use the cord segmentation as the cost (type=seg) with the MeanSquares metric.

This composition is SpinePrep's own, not a published recipe copied verbatim. SCT's default template chain is two stages (centermassrot then bsplinesyn); the inserted columnwise stage and the higher iteration count are deliberate tuning for cord-cropped EPI. Registering the functional image to the subject's own anatomy, rather than to the template directly, is a two-hop design (functional to anatomy here, anatomy to PAM50 in S7) shared with CoSpine (Wei et al., 2025) and Eippert et al. (2017); some pipelines instead register the template to the functional mean directly.

step{1,2,3}.algo : The three stage algorithms above. Defaults centermassrot, columnwise, bsplinesyn.

step3.iter : B-spline refinement iterations. Default 20.

anat_crop.dilate : Dilation of the cord region the anatomy is cropped to. Default 10x10x10.

interpolation : Resampling kernel. Default spline.

Inputs and outputs

derivatives/spineprep/sub-<id>/[ses-<id>/]func/
├── sub-<id>_..._from-bold_to-anat_xfm.nii.gz     # forward warp (+ .json sidecar)
├── sub-<id>_..._from-anat_to-bold_xfm.nii.gz     # inverse warp
├── sub-<id>_..._space-anat_desc-mean_bold.nii.gz # mean BOLD in anat geometry (QC)
└── sub-<id>_..._desc-tsnr_funcref.nii.gz         # tSNR reference (used by S7)

The forward-warp JSON sidecar carries the reproducibility receipt: the policy hash, source path, registration method and parameters, and software versions.

Quality control

The reported metric is 3D cord Dice, the overlap between the EPI cord warped into anatomical space and the anatomy cord segmentation. This is a convergence check, not an independent validator: the registration is driven to maximize that same overlap, so a high Dice mainly confirms the optimizer reached its objective. It is structurally blind to two things a reviewer should check on the reportlet directly. One is a cord aligned in cross-section but shifted along its axis (Dice on a smooth cord is nearly invariant to axial shifts); the other is intensity mismatch inside the cord. The 0.85 pass level is SpinePrep's operating point on cord-cropped EPI, not a threshold reported in the literature.

Three metrics support it, all observability-only (they never fail a run): HD95 (95th-percentile Hausdorff distance) catches a few cord voxels sitting far off even when Dice is high, though it is quantized to the EPI slice thickness and sensitive to single end-slice dropout; ASD (average symmetric surface distance) gives the mean boundary disagreement; and a centerline round-trip drift reports how far the cord centerline moves under forward-then-inverse warp, which is non-zero even for a good registration because bsplinesyn optimizes the two directions separately.

The reviewer inspects two reportlets: bold_on_anat (BOLD against anatomy with the cord contour overlaid, axial and sagittal; if the contour sits off the cord, or the cord rises or falls along Z relative to anatomy, the registration is wrong) and cord_dice_per_slice (per-Z Dice, where a few low slices point to an HD95 outlier and uniformly middling Dice points to a global mis-registration).

Limitations

Cord Dice cannot by itself certify the registration, for the reasons above; the visual overlay is the real check, and an independent level- or intensity-based metric is a planned addition. The nonlinear refinement is applied to the cord segmentation rather than the EPI intensities, which insulates it from the susceptibility-driven "twisted warp" failure that motivates avoiding nonlinear warps of cord EPI (Vahdat et al.); but because each slice is optimized independently, an under-segmented or near-circular cord can still admit non-physical slice-to-slice deformation. Registration quality depends on the S3 and S2 cord segmentations it consumes.

References

  • De Leener, B., et al. (2017). SCT: Spinal Cord Toolbox, an open-source software for processing spinal cord MRI data. NeuroImage 145, 24–43.
  • Eippert, F., et al. (2017). Investigating resting-state functional connectivity in the cervical spinal cord at 3T. NeuroImage.
  • Kaptan, M., et al. (2023). Reliability of resting-state functional connectivity in the human spinal cord. NeuroImage 275, 120152.
  • Wei, Z., et al. (2025). CoSpine: a simultaneous brain and spinal cord fMRI dataset. Scientific Data.

Running S6: see the CLI reference.

Parameters reflect policy/S6_func_to_anat_registration.yaml, shipped with SpinePrep; verified against the implementation and Kaptan et al. (2023)'s published code on 2026-07-16. Audit: .claude/specs/s6-algorithm-audit-v2.md.