Output Spaces & Preprocessing Steps

fMRIPrep produces preprocessed BOLD data in three output spaces. Each space applies a different chain of spatial transforms, but all share the same upstream steps (NORDIC denoising, slice timing correction, head motion estimation, fieldmap estimation, coregistration estimation). The spaces differ only in which transforms are applied at the final resampling step.

Shared upstream processing

These steps are computed once per run and are identical across all output spaces:

Step	Tool	Description
NORDIC denoising	MATLAB `nordic_denoise.m`	Thermal noise suppression on complex-valued data, applied before fMRIPrep. Produces a denoised BOLD NIfTI in `derivatives/nordic/bids_input/`.
Slice timing correction	AFNI `3dTshift`	Temporal interpolation to align all slices to the middle of the TR (t = TR/2). See STC note below.
Head motion correction (HMC)	FSL MCFLIRT	Per-volume rigid-body (6 DOF) alignment to a reference volume (boldref). Produces `from-orig_to-boldref_desc-hmc_xfm.txt`.
Susceptibility distortion correction (SDC)	SDCFlows (pepolar)	B0 fieldmap estimated from AP/PA EPI pairs; stored as B-spline coefficients (`desc-coeff_fieldmap.nii.gz`).
Coregistration	FreeSurfer `bbregister`	Boundary-based registration of boldref to the subject’s T1w anatomical. Produces `from-boldref_to-T1w_desc-coreg_xfm.txt`.
Spatial normalization	ANTs SyN	Non-linear warp from T1w to MNI152NLin2009cAsym template. Computed during anatomical preprocessing.
Confounds estimation	fMRIPrep	Framewise displacement, DVARS, CompCor components, motion parameters, cosine drift regressors. Identical across spaces.

Output spaces

`func` — Native acquisition grid

Property	Value
Grid	Run’s original EPI grid (124 × 124 × 69)
Voxel size	~1.8 × 1.8 × 2.0 mm (acquisition resolution)
Transforms applied	HMC + SDC
Transforms NOT applied	Coregistration, spatial normalization
STC	See note below
BIDS filename	`{prefix}_desc-preproc_bold.nii.gz` (no `space-` entity)
Primary use	Within-subject voxelwise analyses (GLMsingle), single-trial beta estimation

The func space preserves the native acquisition grid with zero interpolation-induced smoothing beyond the single HMC+SDC resampling step. Each run has its own grid — cross-run alignment requires the cached coregistration transforms.

`MNI152NLin2009cAsym:res-2` — Volumetric standard space

Property	Value
Grid	MNI template at 2 mm isotropic (97 × 115 × 97)
Voxel size	2.0 × 2.0 × 2.0 mm
Transforms applied	HMC + SDC + coregistration + spatial normalization
STC	Yes (AFNI 3dTshift)
BIDS filename	`{prefix}_space-MNI152NLin2009cAsym_res-2_desc-preproc_bold.nii.gz`
Primary use	Group analyses, atlas-based ROI analyses, cross-subject comparisons

All runs are in a common volumetric space. Standard neuroimaging atlases (Schaefer, Glasser, Harvard-Oxford, etc.) apply directly without per-subject resampling.

`fsaverage6` — Surface standard space

Property	Value
Mesh	FreeSurfer fsaverage6 (40,962 vertices per hemisphere)
Resolution	~3.1 mm average vertex spacing
Transforms applied	HMC + SDC + coregistration + cortical surface projection
STC	Yes (AFNI 3dTshift)
BIDS filename	`{prefix}_hemi-{L,R}_space-fsaverage6_bold.func.gii`
Primary use	Surface analyses (ISFC, ISC, hyperalignment), cross-subject cortical patterns

Data is sampled onto the cortical surface via FreeSurfer’s volume-to-surface projection, then resampled to the fsaverage6 template mesh. Subcortical structures are not represented.

Summary matrix

	func	MNI res-2	fsaverage6
NORDIC denoising	✓	✓	✓
Slice timing correction	*	✓	✓
Head motion correction	✓	✓	✓
Susceptibility distortion correction	✓	✓	✓
Coregistration to T1w	—	✓	✓
Spatial normalization (MNI)	—	✓	—
Surface projection (fsaverage6)	—	—	✓
Interpolation steps	1	1	1 + surface
Cross-run alignment	via cached xfms	intrinsic	intrinsic
Cross-subject alignment	no	yes	yes

* See STC note below.

Slice timing correction

STC is applied in the MNI and fsaverage6 spaces (produced by fMRIPrep, which runs AFNI 3dTshift when SliceTiming metadata is present).

For the func space post-hoc backfill (existing runs resampled via resample_bold_to_func.py), STC is intentionally skipped:

GLMsingle fits per-voxel HRFs that absorb slice-timing delays
STC’s temporal interpolation can reduce tSNR for single-trial estimation
The acquisition TR (1.5 s) is short enough that the HRF shape change across slices is modest relative to per-voxel HRF fitting flexibility

For future fMRIPrep runs (new subjects/sessions), fMRIPrep applies STC to all output spaces including func. To produce non-STC’d func output for new data, either use --ignore slicetiming or run the post-hoc script.

Follow-up: An empirical comparison of GLMsingle voxel reliability with and without STC is planned to validate this decision.

Per-run transform files

Each BOLD run in fmriprep_nordic/{sub}/{ses}/func/ includes cached transforms that can be used to move between spaces:

File	Description
`*_from-orig_to-boldref_desc-hmc_xfm.txt`	Per-volume rigid HMC transforms (ITK format, one affine per volume)
`_from-boldref_to-B0map_xfm.txt`	Rigid alignment from boldref to fieldmap space
`*_from-boldref_to-T1w_desc-coreg_xfm.txt`	Boundary-based coregistration (boldref → T1w)
`*_desc-hmc_boldref.nii.gz`	Motion-corrected reference volume
`*_desc-coreg_boldref.nii.gz`	Coregistered reference volume

Fieldmap coefficients are in fmriprep_nordic/{sub}/{ses}/fmap/:

File	Description
`_fmapid-_desc-coeff_fieldmap.nii.gz`	B-spline SDC coefficients
`_fmapid-_desc-epi_fieldmap.nii.gz`	EPI reference aligned with coefficients
`_fmapid-_desc-preproc_fieldmap.nii.gz`	Preprocessed fieldmap in Hz (for visualization)

Two fieldmaps per session: one for encoding runs (B0mapencodingXXX), one for retrieval/resting runs (B0mapretrievalXXX). The mapping from BOLD run to fieldmap is encoded in the from-boldref_to-B0map* filename.

The T1w → MNI warp is computed during anatomical preprocessing and stored in fmriprep_nordic/{sub}/anat/.

Fieldmap-to-run mapping

Each session acquires two pairs of AP/PA EPI fieldmaps:

Fieldmap ID pattern	Applies to
`B0mapencodingXXXXX`	TBencoding (all runs), TBmath
`B0mapretrievalXXXX`	TBretrieval (all runs), TBresting

The fieldmap ID is embedded in both the transform filename (from-boldref_to-B0map{id}_xfm.txt) and the coefficient filename (fmapid-B0map{id}_desc-coeff_fieldmap.nii.gz), so the mapping is self-documenting.