Skip to content

S4: Motion correction

S4 aligns every functional volume to a common frame and writes the motion-corrected 4D series, the per-frame motion parameters, and a motion and temporal-SNR quality report.

What it does

S4 corrects subject motion in the cord-cropped 4D EPI produced by S3. Cord fMRI is sensitive to small movements because the cord cross-section is a few millimetres across, and its dominant motion is in-plane and can differ slice to slice with the cardiac cycle. Correction runs in two stages: a coarse in-plane bulk alignment of each volume, then a slice-wise realignment regularized along the cord axis. No slice-timing correction is applied; the series is never temporally resampled, following the cord-fMRI field standard (Eippert et al., 2017; Kaptan et al., 2023). Slice-timing metadata is used only later, by RETROICOR in S8.

Algorithm and parameters

Stage 1 estimates a bulk in-plane translation for each volume with FSL flirt v6.0 (2-DOF; Jenkinson & Smith, 2001), registering the volume's axial mean projection to the mean projection of the robust reference from S3, and applies that translation identically to every slice. Stage 2 performs slice-wise in-plane realignment with sct_fmri_moco (SliceReg; De Leener et al., 2017), regularizing the per-slice translations with a polynomial along the inferior–superior axis so that a noisy single slice cannot diverge from its neighbours. The registration metric is restricted to the cord by a cord mask. sct_fmri_moco in the integrated SCT 7.1 builds its own registration target by iterative averaging and takes no external reference, so the robust reference governs Stage 1 and the tSNR comparison rather than the slice-wise stage.

mode : Which stages run. Default 3d+2d (both). Allowed: 3d+2d, 3d, 2d.

stage1_coarse.interpolation_order : Spline order for applying the Stage-1 shift. Default 1 (linear).

stage2_slicereg.poly_order : Degree of the along-Z polynomial regularizing the slice-wise translations. Default 2 (the sct_fmri_moco default).

stage2_slicereg.metric : Registration cost function. Default MeanSquares, appropriate for the same-contrast BOLD-to-BOLD alignment here. Allowed: MeanSquares, MI.

stage2_slicereg.iterations : Maximum optimizer iterations per slice. Default 10.

z_shift_correction.enabled : Detect and correct a bulk inferior–superior shift between runs of the same subject and task. Default false; detection is reported when a run-01 reference of matching shape is available.

Inputs and outputs

S4 reads the cropped BOLD, the robust reference, and the cord segmentation from the S3 run directory. It writes, per run:

derivatives/spineprep/sub-<id>/[ses-<id>/]func/
├── sub-<id>_..._desc-mocoref_bold.nii.gz   # motion-corrected 4D series
└── sub-<id>_..._moco_params.tsv            # per-frame motion magnitude

The signed slice-wise translation fields (moco_params_x.nii.gz, moco_params_y.nii.gz) are retained in the working directory and read by S8 to build the motion confound regressors.

Quality control

The step-local metric is temporal SNR (voxel temporal mean divided by temporal standard deviation) measured inside the cord segmentation, reported before and after correction; effective correction raises cord tSNR (Kaptan et al., 2023). Framewise displacement is reported but does not determine the run's status. It is the sum of the absolute derivatives of the in-plane translations, |Δtx| + |Δty|, a cord adaptation of Power et al. (2012) that drops the through-plane and rotational terms because the slice-wise stage estimates in-plane translation only. The two stages are composed per slice in millimetres, and the absolute temporal difference is taken per slice before averaging across slices, so opposing shifts at different levels cannot cancel.

Motion does not fail a run. Frames are censored downstream in S8 on the intensity metrics (DVARS and reference-RMS, each against a within-run box-plot fence), which is the cord field's practice: Kaptan et al. (2023) censor on those metrics at two standard deviations and carry the slice-wise translations as regressors rather than thresholding displacement. An absolute displacement threshold is not used, for three measured reasons. Power's 0.5 mm was chosen for brain data as a value "well above the norm found in still subjects", and it sits at the median of cord framewise displacement, so it inverts the criterion that produced it. Frames below 0.5 mm show no more residual signal change after correction than frames below 0.25 mm, so censoring there removes usable data. And displacement is not comparable across acquisitions: it scales with repetition time (Jones et al., 2022) and with the definition used, both of which vary across datasets. Framewise displacement is instead supplied to S8 as a nuisance regressor, where its scale does not matter. A run fails only on a technical failure of the correction itself: cord tSNR below 3, or an input defect such as a mask that does not match the series geometry. It is warned when correction lowers cord tSNR, or when a single-frame displacement peak exceeds warn_fd_mm, both of which point the reviewer at a specific run without rejecting it. The reviewer inspects three reportlets: a motion-trace panel (in-plane translation, FD, and DVARS on a shared time axis), a slice-by-time heatmap of the signed slice-wise shift, and a before/after tSNR comparison with a per-slice cord profile.

Limitations

Motion is often sub-voxel in cooperative cohorts, where correction changes the series little and the value is mainly in the QC record. The slice-wise stage estimates in-plane translations only; through-plane and rotational motion are not modelled, consistent with the cord field but a genuine limit on heavily moving runs. Correction quality depends on the S3 cord segmentation, since the registration metric is restricted to it. Framewise displacement built from in-plane translations does not capture out-of-plane motion, so a low FD does not by itself certify a still run.

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). Denoising spinal cord fMRI data: Approaches to acquisition and analysis. NeuroImage.
  • Jenkinson, M., & Smith, S. (2001). A global optimisation method for robust affine registration of brain images. Medical Image Analysis 5(2), 143–156.
  • Jones, S., et al. (2022). A multi-dataset evaluation of frame censoring for motion correction in task-based fMRI. Aperture Neuro. doi:10.52294/ApertureNeuro.2022.2.NXOR2026
  • Kaptan, M., et al. (2023). Reliability of resting-state functional connectivity in the human spinal cord. NeuroImage.
  • Power, J. D., et al. (2012). Spurious but systematic correlations in functional connectivity MRI networks arise from subject motion. NeuroImage 59(3), 2142–2154.

Running S4: see the CLI reference.

Parameters reflect policy/S4_func_motion_correction.yaml, shipped with SpinePrep; verified against the implementation and SCT 7.1 on 2026-07-16. Audit: .claude/specs/s4-algorithm-audit.md.