Runtime Benchmarks

Wall-clock timing for MRSIPrep's mni-norm mode (default settings, ANTs registration backend) across a range of --nthreads values, on two subjects acquired at different field strengths, to give a sense of how runtime scales with acquisition resolution as well as thread count.

Hardware

Compute (this benchmark)

Component	Spec
CPU	Intel Core i9-14900K, 24 cores / 32 threads, up to 6.0 GHz
RAM	125 GB
GPU	NVIDIA RTX 5000 Ada Generation, 32 GB VRAM (not used by mni-norm mode; SynthSeg here ran CPU-only)
OS	Linux 5.15 (Ubuntu)
Docker	29.4.1
MRSIPrep image	mrsiup/mrsiprep:cpu

MRI scanners

	3 Tesla	7 Tesla
Scanner	Magnetom TrioTim	Magnetom Terra.X
Coil	32-channel	32-channel

MRSI Acquisition: ECCENTRIC

Both datasets were acquired with the ECCENTRIC FID-MRSI sequence (Klauser et al., 2024, Imaging Neuroscience, "ECCENTRIC: A fast and unrestrained approach for high-resolution whole-brain metabolic imaging at ultra-high magnetic field"), a compressed-sensing-accelerated concentric-ring k-space trajectory designed for fast, high-resolution whole-brain MRSI.

Metabolite acquisition

Parameter	3 Tesla	7 Tesla
Field of view	220 × 220 × 130 mm³	220 × 220 × 110 mm³
Slab thickness	95 mm	100 mm
Nominal voxel size	5.0 × 5.0 × 5.2 mm³	3.4 × 3.4 × 3.5 mm³
Scan resolution	44 × 44 × 25	64 × 64 × 31
TR	457 ms	400 ms
TE₁ / TE₂	0.78 ms / 65 ms	0.68 ms
Flip angle	45°	35°
Spectral bandwidth	1320 Hz	2280 Hz
Vector size	512 points	688 FID points
Spatial encoding	ECCENTRIC trajectory, circle radius 0.25 k_max	ECCENTRIC trajectory, circle radius 0.25 k_max
Acceleration factor	2.5	2.5
Total acquisition time	6 min 54 s	11 min 52 s

Water reference

Matched spatial coverage, lower resolution — used for coil combination, field correction, and metabolite intensity normalization.

Parameter	3 Tesla	7 Tesla
Field of view	220 × 220 × 130 mm³	220 × 220 × 110 mm³
Nominal voxel size / resolution	10.0 × 10.0 × 10.0 mm³	10 × 10 × 10 mm³
Scan resolution	22 × 22 × 13	—
TR	460 ms	404 ms
TE₁ / TE₂	0.72 ms / 65 ms	0.59 ms
Flip angle	45°	35°
Acquisition time	1 min 21 s	59 s

Reconstruction and quantification

Both MRSI acquisitions were reconstructed using a compressed-sensing SENSE low-rank framework with total-generalized-variation regularization and simultaneous lipid suppression. Metabolite quantification was performed with LCModel.

MRSIPrep Benchmark Method

Two single-subject mrsiprep runs, one per dataset, repeated at --nthreads 8, 12, 16, and 32 (--nproc 1 throughout — one subject per run, so --nthreads is the only varying parameter). Each run used a fresh --work-dir (no Nipype caching carried over between thread-count variants), so every number below reflects genuine full-pipeline computation, not a partially cached rerun. All 8 runs (2 subjects × 4 thread counts) were executed strictly sequentially, one at a time, with no other run or concurrent load on the machine, to rule out cross-run resource contention affecting the timings.

• 3 Tesla subject — a real MRSI acquisition with an MP2RAGE anatomical.

• 7 Tesla subject — a real MRSI acquisition with an MP2RAGE anatomical.

Both runs: --mode mni-norm --metabolites NAANAAG,GPCPCh,CrPCr,GluGln,Ins --ref-met CrPCr, default --synthseg-mode robust, default ANTs registration backend.

Resolution and useful-voxel counts

The two subjects differ substantially in both anatomical (T1w) and MRSI grid resolution — this is the main driver of the runtime difference below, since ANTs registration and SynthSeg both operate on the full-resolution T1w volume, not the coarser MRSI grid.

	3 Tesla	7 Tesla	Ratio (7T / 3T)
T1w voxel size (mm)	1.00 × 1.33 × 1.33	0.66 × 0.60 × 0.60	—
T1w volume shape	160 × 192 × 192	256 × 396 × 416	—
T1w total voxels	~5.9 M	~42.2 M	~7.2×
MRSI voxel size (mm)	5.00 × 5.00 × 5.25	3.44 × 3.44 × 3.55	—
MRSI useful (non-zero, in-brain) voxels	15,315	32,638	~2.1×

Results

Stacked bar height = total wall-clock elapsed time (label above each bar); segments show each pipeline step's share. "Container startup / other overhead" covers Docker startup and the CLI's own preflight input-check, which aren't wrapped in a named, timed pipeline step.

Interpretation

• Runtime is essentially flat from 8 to 32 threads for both subjects when each run is isolated from other load: 3T ranges 300–318s (~6% spread) and 7T ranges 1227–1235s (under 1% spread), with no consistent downward trend past 8 threads. ANTs registration (the largest single segment at both field strengths) and SynthSeg tissue segmentation show effectively no benefit from more than ~8 threads, so for batch processing it is generally better to use ~8 threads per subject and increase --nproc (running more subjects in parallel) rather than allocate more threads to each individual subject.

• The 7T subject takes about 4× longer than the 3T subject (~20.5 vs. ~5.1 minutes). This is mainly due to the 7T T1w image having ~7.2× more voxels, which increases registration and segmentation costs. The MRSI grid has only ~2.1× more usable voxels, making anatomical — not MRSI — resolution the main runtime driver. Tissue segmentation (mri_synthseg) and MRSI-T1w-MNI registration (ANTs rigid+affine+SyN) together account for the large majority of total runtime at both field strengths.

Registration Frameworks

--registration-backend offers ANTs (default: rigid+SyN for MRSI→T1w, rigid+affine+SyN for T1w→MNI — see the transform-stage table below, since the two ANTs stages are not symmetric) and FSL (FLIRT affine, with an FNIRT deformable stage on by default — --no-fsl-deformable for FLIRT-only). This section compares four MRSI→T1w registration configurations — ANTs (Rigid+SyN), ANTs (Rigid+Affine), FSL FLIRT-only, and FSL FLIRT+FNIRT — each against both supported T1w registration targets, brain (skull-stripped) and brain+CSF (skull-stripped T1w with the CSF compartment re-added, since CSF also produces real MRSI signal that a brain-only target would otherwise clip at the boundary).

Transform used at each registration stage, by backend:

Backend	MRSI → T1w	T1w → MNI
ANTs (Rigid+SyN) — mrsiprep's default	Rigid + SyN (deformable) — no separate Affine stage, antsRegistrationSyN[sr]	Rigid + Affine + SyN (deformable), antsRegistrationSyN[s]
ANTs (Rigid+Affine)	Rigid + Affine — a genuine second registration run (antsRegistration, transform="a"), since mrsiprep's default MRSI→T1w stage never computes an Affine stage to reuse	Rigid + Affine (the SyN warp dropped from the default [s] run above — this stage already has a real Affine, so no extra registration needed here)
FSL FLIRT-only	Affine only (FLIRT, 12 DOF, corratio cost)	Affine only (FLIRT, 12 DOF)
FSL FLIRT+FNIRT	Affine (FLIRT) + deformable warp (FNIRT, --fsl-deformable)	Affine only (FLIRT, 12 DOF) — FNIRT is not used for this stage

Note that --fsl-deformable only adds a deformable (FNIRT) stage to the MRSI→T1w registration; the T1w→MNI stage is always FLIRT affine-only under the FSL backend, in both the FLIRT-only and FLIRT+FNIRT variants compared here.

The two ANTs stages are asymmetric, which is why "ANTs (Rigid+Affine)" needed a real second registration run, not just a transform dropped from the default pipeline. ANTs' default MRSI→T1w stage (antsRegistrationSyN[sr]) runs a Rigid stage followed directly by SyN — antsRegistrationSyN.sh's own sr code has no separate Affine stage at all (MRSI and T1w start from roughly the same subject geometry, so mrsiprep's default skips the extra affine correction there). Dropping SyN from that default leaves Rigid-only, not Rigid+Affine. ANTs (Rigid+Affine) in this comparison therefore required actually running antsRegistration with transform="a" at the MRSI→T1w stage — genuine new registration compute, not a reuse of an already-computed transform. ANTs' T1w→MNI stage (antsRegistrationSyN[s]) already runs the full three-stage Rigid → Affine → SyN pipeline by default, since subject anatomy genuinely differs from the MNI template in scale and shape, not just position.

Method

Full mrsiprep --mode mni-norm runs (not isolated registration calls) on the same 3 Tesla and 7 Tesla subjects used above, varying --registration-backend/--fsl-deformable and --registration-t1-target (6 combinations × 2 subjects = 12 runs).

Leakage is reported as signal-weighted mass outside the brain mask: at each quality-passing voxel (resampled CRLB ≤ 20, mrsiprep's own default --crlb-max) the resampled CrPCr signal magnitude is summed; the percentage reported is the fraction of that total signal mass that falls outside the reference brain mask. This is deliberately not a raw voxel count of "covered" voxels outside the mask — a boundary voxel carrying negligible signal and a voxel carrying real, substantial signal in the wrong location both count as "1 voxel" under a voxel-count metric, which conflates thin resampling/registration boundary artifacts with genuine misregistration. Weighting by signal magnitude fixes that: a one-voxel-thick sliver of near-zero smeared or dilated boundary contributes almost nothing to the percentage, while real signal displaced into clearly wrong anatomy does. (Two earlier, superseded versions of this metric were tried and rejected: "resampled signal nonzero" as the coverage criterion overstated leakage via linear/spline interpolation smear at the boundary; a corrected mask-based voxel-count metric fixed that but then over-penalized FNIRT specifically, because nearest-neighbor-resampling a binary mask through FNIRT's nonlinear warp locally dilates the mask's footprint independent of true registration accuracy. Weighting by actual signal magnitude avoids both problems at once.)

This is computed against three reference masks: the T1w brain-only mask, the T1w brain+CSF mask, and the MNI152 template's own standard brain mask (in MNI space) — summing how much "covered" signal mass falls outside each reference mask.

Results

3 Tesla subject, all four backends × both targets, axial slice, same intensity scale throughout:

7 Tesla subject, same layout and intensity-scale convention (its own scale, since 7T signal levels differ from 3T):

Signal extending past the skull ring, or with a jagged/scalloped rather than smooth outer edge, indicates voxels that have leaked beyond the true brain boundary during registration — visible here for both FSL variants and, most noticeably, for ANTs (Rigid+Affine) at both field strengths, which shows the coarsest, most scalloped outer edge of any of the four configurations.

Signal-weighted MNI-space leakage and total mni-norm wall-clock runtime, by backend (brain target; brain+CSF leakage is within ±0.3 points of these and shows the same pattern, not shown). Runtime is averaged across the brain/brain+CSF targets — ANTs (Rigid+Affine) is not a full mni-norm run (see Method above), so it isn't a like-for-like total-pipeline number; instead its bar reports the registration-only wall-clock time (both stages combined, timed directly via mrsiprep.interfaces.ants.register(), nthreads=16 matching the other runs' default, hatched in the figure), which should be read as a lower bound on how much a full mni-norm run would take, not a directly comparable total:

Conclusions

• ANTs (Rigid+SyN), the mrsiprep default, produced the lowest signal-weighted leakage at both 3T and 7T (0.34% and 0.44%), clearly outperforming FSL FLIRT-only (2.1% and 0.97%) and FLIRT+FNIRT (5.4% and 2.7%). This agrees with the visual overlays, where ANTs yielded the sharpest anatomical alignment.

• ANTs (Rigid+Affine) performed unexpectedly poorly, with 4.30% leakage at 3T and 4.73% at 7T. Its additional affine degrees of freedom may overfit the low-resolution MRSI reference when not followed by a deformable SyN stage. This non-default configuration therefore cannot currently be recommended.

• Adding FNIRT to FLIRT increased both leakage and runtime. At 7T, FNIRT required 53.0 minutes, compared with 11.2 minutes for FLIRT-only and 27.0 minutes for ANTs (Rigid+SyN). Overall, ANTs (Rigid+SyN) provides the best accuracy–runtime balance. For FSL users, FLIRT-only is preferable to FLIRT+FNIRT based on these results.

• The choice between brain-only and brain+CSF registration targets had only a minor and inconsistent effect across field strengths, indicating that the registration backend is the dominant factor.