SynthMRSI-VBA-Project: synthetic MRSI ground-truth benchmark dataset for MRSIPrep voxel-based-analysis validation

A small, synthetic MRSI dataset used to validate MRSIPrep's voxel-based-analysis (VBA) detection pipeline end-to-end. 32 subjects, single session each. T1w anatomical images are real, subsetted from two CC0 OpenNeuro datasets (ds000001, ds000117). MRSI signal (5 metabolites: NAANAAG, GPCPCh, CrPCr, GluGln, Ins) is model-synthesized, conditioned on each subject's own T1w, with a deterministic region-specific abnormality injected into a group=1 subset (16 of 32 subjects) and no injection in the group=0 subset. CRLB, SNR, and FWHM quality maps are real empirical acquisition-quality measures copied from real MRSI acquisitions, not synthetic. This is not real patient MRSI data.

This release contains only the raw, unprocessed inputs (T1w anatomicals and raw/synthesized MRSI signal maps under derivatives/mrsi-orig, MRSIPrep's documented raw-input convention) plus the population-level ground-truth injection masks used to score VBA detection performance (derivatives/mrsiprep/group/synthmrsi/mni/). Per-subject, per-registration-backend processed derivatives are not included; they are fully regeneratable from the raw files here using MRSIPrep (https://github.com/MRSI-Psychosis-UP/MRSIPrep). See the included README for full details.

License note: this dataset (T1w images and MRSI data, as released here) is CC0, consistent with the CC0 license of its OpenNeuro source datasets. The MRSIPrep pipeline/code and Docker image that produced and can preprocess this data are separately licensed under the CHUV academic non-commercial research license and are not covered by this CC0 release.