Release v1.0 · 2026

An open fMRI corpus for brain foundation models.

Open-fMIND is a uniformly preprocessed, openly accessible corpus of parcellated fMRI timeseries and functional connectivity matrices, built from public OpenNeuro studies. Every run is processed with fMRIPrep and parcellated under 52 atlases, ready for self-supervised pretraining and connectivity analyses at scale.

322
OpenNeuro datasets
11,379
subjects
45,731
functional runs
52
atlases, 15 families
2.38M
timeseries (runs × atlases)
11.4M
volumes
6,044
hours of scanning
4.35TB
in 1,821 shards
Connectivity space

The latent organization of 44,241 runs

Each point is one preprocessed run, placed by UMAP from its full Schaefer-400 connectivity (79,800 Fisher-z edges). Runs do not form one homogeneous cloud: twelve dense clusters hold 36% of runs, rest runs gather in a band, and higher motion concentrates in one region shared by many datasets. No cluster corresponds to a single dataset.

Loading 44,241 runs…
Drag to pan · Ctrl/Cmd + scroll or +/− to zoom · click a point to follow its dataset

UMAP (15 neighbours, minimum distance 0.1) of 44,241 preproc runs from 288 datasets; datasets with fewer than 5 subjects are left out. Dataset colours follow the median position of each dataset's runs (hue by x, lightness by y), so neighbouring datasets have similar colours. Clusters: HDBSCAN (minimum cluster size 500, minimum samples 50). Motion: mean framewise displacement, capped at the 95th percentile. Axes are arbitrary. Same embedding as Fig. 2 of the paper.

Data

Two versions of the same runs

quasiraw

fMRIPrep output projected onto each atlas, nothing else; shipped with the fMRIPrep confounds, frame-aligned, so you can apply your own denoising.

preproc

The same timeseries after a fixed recipe: confound regression, scrubbing, detrending, band-pass filtering and z-scoring.

quasirawpreproc
Datasets322320
Subjects (counted per dataset)11,37911,320
Runs45,73145,019
Resting-state runs8,7568,704
Task runs36,97536,315
Atlases5252
Timeseries (runs × atlases)2.38 M2.34 M
Volumes (3D time points)11.39 M10.92 M

Every run is available for every atlas, with one connectivity matrix per timeseries. preproc counts the volumes kept after scrubbing. Each file has a JSON sidecar with the full provenance (download, fMRIPrep, parcellation, denoising, connectivity), and qc.tsv gives per-run motion.

Pipeline

How the data was made

  1. Source data

    Brain BOLD datasets from OpenNeuro, at the snapshot recorded in each dataset description.

  2. fMRIPrep 25.2.3

    One job per subject, without surface reconstruction; output in MNI152NLin2009cAsym at 2 mm.

  3. Parcellation

    Mean signal per region with nilearn (NiftiLabelsMasker, or NiftiMapsMasker for probabilistic atlases). This is quasiraw.

  4. Denoising

    One nilearn.signal.clean call: 24 motion parameters plus white-matter and CSF signals, scrubbing above 0.5 mm FD, detrending, 0.01–0.1 Hz band-pass, z-scoring. Runs with fewer than 50 volumes left are dropped. This is preproc.

  5. Connectivity

    Correlation per run with nilearn's ConnectivityMeasure (default settings), stored as float32.

Atlases

52 parcellations, 7 to 1,024 regions

Every atlas is applied to every run under identical processing, so representations are comparable across granularities and atlas families.

Schaefer100–1000 · 7 & 17 networks
Schaefer + Tian432, 450
Yan (homotopic)100–800 · Yeo7 / Yeo17 / Kong17
DiFuMo64–1024 · probabilistic
BASC64–444
Smith ICA10, 20, 70 · probabilistic
Yeo7, 17
MSDL39 · probabilistic
Tian subcortex16, 32, 50
Pauli16
AAL166
AICHA384
Glasser (HCP-MMP1.0)360
Destrieux150
Harvard-Oxford48 cortical, 21 subcortical
Use

Pretraining on Open-fMIND

A Brain-Semantoks encoder pretrained on Open-fMIND alone matches the published checkpoint pretrained on UK Biobank on HCP-YA, with frozen features and identical folds and classifiers.

0.945
HCP-YA sex, AUROC
UK Biobank 0.945 · random init 0.877
0.415
HCP-YA cognition, Pearson r
UK Biobank 0.409 · random init 0.380
62.0%
ABIDE diagnosis, balanced accuracy
UK Biobank 63.1%
Access

Quick start

Files are packed into ~3 GB tar shards that keep their BIDS paths, split by atlas and by rest / task. A parquet index locates every file, so you can filter, stream whole shards, or fetch a single run with an HTTP range request.

Choose runs with the index, stream with WebDataset

import gzip, json
import pandas as pd, nibabel as nib, webdataset as wds

base = "https://huggingface.co/datasets/cbarbano/open-fmind/resolve/main"
idx = pd.read_parquet(f"{base}/index/preproc_timeseries.parquet",
                      filters=[("atlas", "==", "schaefer400"), ("task_type", "==", "rest")])

urls = [f"{base}/{s}" for s in sorted(idx.shard.unique())]
for sample in wds.WebDataset(urls, shardshuffle=False):
    img = nib.Nifti1Image.from_bytes(gzip.decompress(sample["nii.gz"]))
    ts = img.get_fdata().squeeze().T      # (n_timepoints, n_rois)
    meta = json.loads(sample["json"])     # RepetitionTime, ROILabels, TaskType, ...

Download one atlas and rebuild the BIDS tree

hf download cbarbano/open-fmind --repo-type dataset \
   --include "preproc/timeseries/atlas-schaefer400/*" --local-dir open-fmind
cat open-fmind/preproc/timeseries/atlas-schaefer400/*/shard-*.tar | tar -xf - -i

Licenses. Derived from public OpenNeuro datasets, each under its own license (251 CC0, 34 PDDL, 30 not stated, 5 CC-BY / CC-BY-SA, 1 CC-BY-NC, 1 other). LICENSES.tsv lists the license, authors, DOI and requested acknowledgement of every source dataset; please cite the datasets you use. Built on OpenNeuro, fMRIPrep and nilearn.

Cite

Citation

Barbano, C. A., Marraffini, G., Wassermann, D. Open-fMIND: A Large-Scale Open-Access fMRI Corpus for Pretraining Brain Foundation Models. arXiv preprint arXiv:2611.21215, 2026.

@article{barbano2026openfmind,
  title   = {Open-fMIND: A Large-Scale Open-Access fMRI Corpus for Pretraining Brain Foundation Models},
  author  = {Barbano, Carlo Alberto and Marraffini, Giovanni and Wassermann, Demian},
  journal = {arXiv preprint arXiv:2611.21215},
  year    = {2026}
}