Cross-Site/Cross-Sequence Regional Profile Reproducibility

The benchmarks on the other two pages validate MRSIPrep's registration and detection performance on data it was built and tuned against — a real 3T/7T ECCENTRIC pair, and a synthetic ground-truth cohort. This page tests something harder: whether MRSIPrep-derived regional metabolite profiles reproduce on entirely independent real-world data — different subjects, different scanners, different sites, and different MRSI sequences, processed with the same pipeline and parcellation.

Because absolute metabolite units are not expected to agree across different sequences and field strengths — scanner-, sequence-, and reconstruction-dependent scaling and referencing choices affect the overall signal level independently of true tissue concentration — this is deliberately framed as a test of whether the relative spatial pattern of regional metabolite levels reproduces across acquisitions, not whether absolute values agree.

Dataset

Two independently acquired real-world datasets, merged into one group-level MRSIPrep connectivity/profile archive.

	Lausanne3T-FID	Lausanne3T-ECCENTRIC
n	12	15

Method

For each of the 5 metabolites (CrPCr, GluGln, GPCPCh, NAANAAG, Ins) and each of the 203 parcels of the chimeraLFMIHIFIS scale-3 parcellation:

Per-subject point estimate: the median across the $K_{\mathrm{pert}}=50$ CRLB-perturbed draws in the archive's metab_profiles_subj_list array (shape (subjects, parcels, nperm, metabolites)) — MRSIPrep's uncertainty-propagated regional profile estimation, see the main docs' connectivity/profile documentation.
GM-only filtering: excludes only brain-stem-midbrain (203 → 202 parcels) as not cleanly gray matter. This parcellation scheme contains no white matter parcels at all — cortex (ctx-), cerebellum (cer-), hippocampus (hipp-), subcortical nuclei (subc-), and thalamic nuclei (thal-) are all retained. A GM+WM follow-up is left for future work once a scheme with WM parcels is used.
Sub-parcel merging: parcels sharing the same anatomical root differ only by the Lausanne scheme's finer sub-parcellation index (e.g. ctx-lh-superiorfrontal_1 through _8 are the same gyrus at finer granularity, not eight distinct regions) — averaged into a single root-region value per subject. This collapses 202 GM parcels into 82 root regions.
Group profile: for each metabolite and root region, the mean (and 95% CI, across subjects) within each cohort.
Cross-cohort comparison: Spearman correlation between the two cohorts' 82-region mean profiles, per metabolite — a threshold- independent, scale-free test of whether the two acquisitions rank regions the same way, deliberately insensitive to any absolute scale difference between them.

Results

Metabolite	Spearman ρ	p-value	n roots
GPCPCh	0.93	2.1×10⁻³⁵	82
CrPCr	0.90	3.2×10⁻³¹	82
NAANAAG	0.89	5.0×10⁻²⁹	82
Ins	0.88	7.0×10⁻²⁸	82
GluGln	0.85	9.4×10⁻²⁴	82