S2B: Functional Denoise (optional)

Step Code: S2B_func_denoise Depends on: S1 (Input Verify) Required by: S3 (Func Init & Crop) — consumes the denoised series when present Default state: OFF (opt-in)

Purpose

S2B optionally removes thermal noise — the random, structureless electronic noise the scanner adds to every voxel — from the raw functional time series before any other functional processing. Thermal noise is a large share of the moment-to-moment fluctuation in the tiny spinal cord, so suppressing it can meaningfully raise the temporal signal-to-noise ratio (tSNR), the basic measure of how stable a voxel's signal is over time.

The step is disabled by default. When off, it is a clean passthrough: it emits a PASS with no outputs and S3 simply reads the raw BOLD. It is opt-in because, on thin cord data, aggressive denoising can also remove real signal (see Caveats), so it ships behind QC that the human must eyeball.

Algorithm

S2B runs MP-PCA (Marchenko–Pastur Principal Component Analysis) via MRtrix3's dwidenoise. The method slides a small 3D patch through the image, decomposes each patch's voxel-by-time matrix into principal components, and discards the components whose strength is consistent with pure noise (predicted by random- matrix theory). What remains is the signal estimate; the difference is the removed noise.

• Tool: dwidenoise (MRtrix3), the reference MP-PCA implementation.
• Why first: MP-PCA assumes independent, identically-distributed noise across voxels. That assumption holds only on the rawest, non-interpolated data, so S2B runs at the very head of the functional chain — before S3 cropping and S4 motion correction, both of which resample and correlate neighbouring voxels.
• Patch size (extent): unset by default, so dwidenoise auto-sizes the smallest isotropic patch that exceeds the volume count (e.g. 7×7×7 for ≤343 volumes), keeping voxels ≥ volumes as the method requires.
• Failure handling: any dwidenoise error falls back to the raw series; the chain never breaks on an optional step.

Literature basis

• MP-PCA: Veraart et al., NeuroImage 142:394–406 (2016).
• Spinal-cord precedent: Kaptan, Eippert et al., NeuroImage (2023) report a whole-cord tSNR gain; on the in-house balgrist cohort we measured roughly +50% in-cord tSNR.

Caveats (why it is opt-in)

• Magnitude MRI data has a Rician (non-Gaussian) noise floor; MP-PCA does not correct for it.
• Low-rank denoising can cause activation "spreading" (blurring of true signal into neighbours).
• Patch mixing of cord, CSF and surrounding tissue in the very thin cord is unstudied for fMRI — which is exactly why the residual-structure QC below exists.

QC Reportlets

S2B emits three reportlets per run. One look should tell you whether denoising helped or quietly removed signal.

1. Noise sigma map (desc-S2B_noise_sigma.png) — where, and how much, thermal noise was estimated across the image. A healthy map is flat and anatomy-free.
2. tSNR before vs. after (desc-S2B_tsnr_before_after.png) — temporal SNR on matched slices and colorbar, annotated with the median tSNR gain. tSNR should visibly rise.
3. Residual-structure check (desc-S2B_residual_structure.png) — the temporal standard deviation of the removed signal (raw minus denoised). This must look like structureless noise. If cord or CSF anatomy is visible here, real signal was removed (over-denoising).

Step Metric

The step-local truth metric is the in-cord median tSNR gain (%) from before to after denoising. The step also reports a residual-structure correlation — the correlation between the removed-noise map and the mean image — which gates over-denoising directly.

A negative tSNR gain warns (denoising did little or hurt); a residual that correlates strongly with anatomy fails the run.

Outputs

<workfolder>/denoise/<run_id>/
├── desc-denoised_bold.nii.gz    # denoised 4D series (S3 reads this when present)
└── denoise_noise_map.nii.gz     # estimated noise (sigma) map

derivatives/spineprep/{dataset}/sub-{id}/figures/
├── sub-{id}_..._desc-S2B_noise_sigma.png
├── sub-{id}_..._desc-S2B_tsnr_before_after.png
└── sub-{id}_..._desc-S2B_residual_structure.png

When denoising is disabled, no outputs are written and S3 falls back to the raw BOLD.

Policy

Key knobs in policy/S2B_func_denoise.yaml:

• enabled: false — opt-in master switch.
• nthreads: 1 — dwidenoise's own thread pool (it ignores OMP_NUM_THREADS).
• extent: null — auto-size the patch.
• qc_thresholds.warn_residual_corr: 0.4, fail_residual_corr: 0.6, warn_min_tsnr_gain_pct: 0.0.

References

1. MP-PCA: Veraart et al. NeuroImage 142:394–406 (2016). DOI
2. Cord tSNR gain: Kaptan, Eippert et al. NeuroImage (2023).

Last updated: June 2026