Methods Overview

SpinePrep implements an end-to-end preprocessing pipeline for spinal cord fMRI, designed for robustness against the unique challenges of spinal imaging: physiological noise, small anatomical targets, and susceptibility artifacts.

Why Spinal Cord fMRI Needs Specialized Preprocessing

Spinal cord fMRI presents fundamentally different challenges compared to brain imaging:

• Small target anatomy: The spinal cord cross-section (~50-80 mm²) is roughly 1/1000th the size of a brain slice, requiring sub-millimeter precision in registration and segmentation.
• Physiological noise: Cardiac pulsation, respiration, and CSF flow cause signal fluctuations that can exceed the BOLD signal of interest.
• Susceptibility artifacts: The cord's proximity to vertebral bodies, lungs, and air-tissue interfaces creates severe B0 inhomogeneities.
• Motion complexity: Head, neck, and swallowing motions introduce non-rigid deformations that standard brain motion correction cannot handle.

Brain fMRI pipelines such as fMRIPrep are built for a different problem and are not designed for these challenges. SpinePrep addresses them by automating the spinal-cord-specific methods of the Spinal Cord Toolbox (SCT) and related tools — it integrates and orchestrates those methods rather than introducing new algorithms of its own.

Comparison with fMRIPrep

SpinePrep follows fMRIPrep's design philosophy — BIDS-native, containerised, QC-first — but wraps the spinal-cord-specific methods of SCT and related tools rather than the brain algorithms fMRIPrep uses. The contribution is the integration, automation, reproducibility, and standardised QC, not the underlying algorithms.

Pipeline Architecture

flowchart LR
    S0[S0: Setup] --> S1[S1: Input Verify]
    S1 --> S2[S2: Anat Cord Ref]
    S2 --> S2B[S2B: Func Denoise<br/>optional]
    S2B --> S3[S3: Func Init & Crop]
    S3 --> S4[S4: Motion Correction]
    S4 --> S5[S5: Distortion Correction]
    S5 --> S6[S6: Func to Anat Reg]
    S6 --> S7[S7: Template Normalization]
    S7 --> S8[S8: Confounds & Physio]
    S8 --> S9[S9: Functional Derivatives]
    S9 --> S10[S10: QC & Release]
    style S2B stroke-dasharray: 4 4

S2B is an optional thermal-noise denoising step that is OFF by default; when disabled it passes the raw data straight through to S3. The chain now runs S1–S10 contiguously, with S10 (QC Aggregation & Release) as the final step. An earlier step S10 — analyst-owned region-of-interest timeseries and connectivity analysis — was removed on 2026-06-11 and its number reused for the release step. That analysis belongs to the analyst downstream, not to the preprocessing release.

The pipeline performs no slice-timing correction (STC), following the field standard for cord-only fMRI — every reference cord pipeline omits it (Eippert 2017; Barry 2014; Kaptan 2023; Spinal Cord Toolbox). The reasons: resting-state and block-design signal is low-frequency (below 0.1 Hz), where STC barely matters; event-related task timing is instead handled with a temporal-derivative regressor in the GLM; and STC's temporal interpolation interacts badly with the slice-wise motion and physiological correction that dominate the cord noise budget (the same reason the HCP pipeline skips it). Slice-timing metadata is used only by RETROICOR in S8 (to phase the cardiac and respiratory regressors), never to resample the BOLD series. This is a defensible trade-off, not a hard rule: at these TRs STC would give a small benefit for event-related designs, which the GLM temporal derivative recovers.

Design Principles

Step Summaries

The analyst-owned ROI timeseries and connectivity analysis that was formerly step S10 was removed from the active pipeline on 2026-06-11 and is no longer part of the preprocessing release. Its step number has been reused for the QC aggregation & release step above.

For detailed technical descriptions of each step, see the per-step pages in this section.