openmind

OpenMind: preprocessed brain MRI (NeuroSpin)

Preprocessed versions of the OpenMind dataset (Wald et al., 2025): head & neck 3D MRI pooled from ~800 OpenNeuro datasets. Here the brain T1w, T2w and FLAIR images were preprocessed at NeuroSpin using brainprep and packed as WebDataset shards, ready for large-scale (e.g. self-supervised) deep learning.

All folders share the same sample identifiers: (dataset_id, subject, session, run), with dataset_id the shard folder name, so derivatives of the same scan can be joined across folders.

Repository layout

README.md
val_split.csv                 # recommended validation split, for all folders (see below)
participants.tsv              # per scan: age, sex, bmi, race, handedness, health_status
quasiraw-v2.1.1/              # see below
cat12vbm-12.8.2/              # see below

Each preprocessing folder holds one folder of WebDataset shards per OpenNeuro dataset and an index.json.gz with the byte offsets of every file of the shards, used by nidl for random access.

Validation split

val_split.csv (columns dataset_id, participant_id, session, run) defines a held-out validation set made of whole datasets (no subject or dataset is shared with training), stratified on age, sex and BMI. It holds ~10% of the images. Only the rows listed in the csv belong to validation. The others belong to training.

The split is given by scan identifiers, shared by all folders: it applies to quasiraw-v2.1.1 and cat12vbm-12.8.2 alike.

Usage with nidl

nidl provides an OpenMind PyTorch Dataset that reads the shards by byte offset (from index.json.gz): random access, any sampler and any number of DataLoader workers, modality pairing and the subject-level train / val split. By default the images are streamed from this repository with HTTP range requests (nothing is written to disk but the index and the split); streaming=False downloads the needed shards first.

pip install "git+https://github.com/neurospin-deepinsight/nidl"

from torch.utils.data import DataLoader
from nidl.datasets import OpenMind

root = "openmind"  # local mirror of this repository, files are fetched on demand

# Quasi-raw: one T1w image per item, subject-level split from <root>/val_split.csv
train = OpenMind(root, preprocessing="quasiraw", modality="t1", split="train", apply_mask=True)
val = OpenMind(root, preprocessing="quasiraw", modality="t1", split="val", apply_mask=True)
print(train[0].shape)  # (1, 182, 218, 182)

# Quasi-raw: paired T1w / T2w acquired in the same session
paired = OpenMind(root, modality=("t1", "t2"), group_by="session", split="train")
sample = paired[0]  # {"t1": array, "t2": array}

# CAT12 VBM: grey matter map and ROI morphometry
vbm = OpenMind(root, preprocessing="cat12vbm", split="train", return_morphometry=True)
gm, morphometry = vbm[0]
print(gm.shape, morphometry["total_volumes"]["TIV"])  # (1, 113, 137, 113) ...

loader = DataLoader(train, batch_size=8, shuffle=True, num_workers=8)

nidl does not read the CAT12 quality ratings (qc.json): use the webdataset example of the cat12vbm-12.8.2 section to get them.

quasiraw-v2.1.1

Content

Preprocessing

brainprep subject-level-quasiraw (v2.1.1 container), applied per modality, with the matching MNI152 template (T1 template for T1w, T2 template for T2w / FLAIR):

1. Reorientation to the MNI152 orientation.
2. Brain mask with FreeSurfer SynthStrip.
3. N4 bias field correction.
4. Resampling to 1 mm isotropic.
5. Linear (affine, 9 DOF) registration to the MNI152 1 mm template.
6. The registration is applied to the bias-corrected image and the brain mask; the image is then multiplied by the mask (skull-stripped).

"Quasi-raw" means minimal preprocessing: no non-linear warp, no tissue segmentation, no intensity standardization across scans. Intensities are those of the bias-corrected scan.

Layout

quasiraw-v2.1.1/
├── index.json.gz                     # byte offsets for random access (nidl)
└── <dataset_id>/                     # one folder per OpenNeuro dataset, e.g. ds000001
    └── <dataset_id>-NNNNNN.tar       # WebDataset shard(s), ≤ ~3 GB each

Each sample is one image of one (subject, session, run, modality). Its key is sub-<label>_ses-<label>_mod-<t1|t2|flair>_run-<id> (ses-DEFAULT when the source has no session) and it holds three files:

Different subjects can have the same subject's id across different OpenNeuro datasets: a subject is thus uniquely identified by the pair (dataset_id, subject), with dataset_id given by the shard folder name.

The run is the BIDS run- index of the source file when it is unambiguous, otherwise a hash of the source file name (brainprep convention).

Streaming with webdataset

Without downloading (pip install webdataset nibabel huggingface_hub):

import gzip, io, json
import nibabel as nib
import webdataset as wds
from huggingface_hub import HfFileSystem, hf_hub_url

repo = "neurospin/openmind"
fs = HfFileSystem()
shards = sorted(fs.glob(f"datasets/{repo}/quasiraw-v2.1.1/*/*.tar"))
urls = [hf_hub_url(repo, s.split(f"{repo}/", 1)[1], repo_type="dataset") for s in shards]

def load_nifti(b):
    fh = nib.FileHolder(fileobj=io.BytesIO(gzip.decompress(b)))
    return nib.Nifti1Image.from_file_map({"header": fh, "image": fh})

def decode(s):
    meta = json.loads(s["meta.json"])
    mod = meta["modality"]
    return {
        "dataset_id": s["__url__"].rsplit("/", 2)[-2],
        "meta": meta,
        "image": load_nifti(s[f"{mod}.nii.gz"]).get_fdata(dtype="float32"),
        "mask": load_nifti(s[f"{mod}_mask.nii.gz"]).get_fdata(dtype="float32") > 0,
    }

ds = wds.WebDataset([f"pipe:curl -s -L {u}" for u in urls], shardshuffle=True).map(decode)
sample = next(iter(ds))
print(sample["dataset_id"], sample["meta"], sample["image"].shape)

Quality control

No manual QC beyond the source curation was applied to this release: a small fraction of images may have failed skull-stripping or registration. The source OpenMind metadata includes a per-modality, per-dataset image quality score (1–5) that can be used to filter.

cat12vbm-12.8.2

Content

43,076 T1w images were processed: 125 are missing from this release because CAT12 failed on them or their group-level measures could not be computed.

Preprocessing

brainprep subject-level-vbm and group-level-vbm (v2.1.0 container) running CAT12 12.8.2 (r2166, SPM12 standalone), with the brainprep CAT12 batch (cross-sectional, DARTEL registration):

1. Bias correction, affine registration to MNI and tissue segmentation (grey matter, white matter, CSF) with CAT12.
2. Non-linear registration (DARTEL) to the CAT12 MNI152NLin2009cAsym template, resampled at 1.5 mm.
3. The released image is the modulated normalized grey matter map (mwp1): GM tissue probability multiplied by the Jacobian of the warp, so that voxel values sum to the subject's GM volume.
4. ROI volumes are computed by CAT12 in native space for each atlas; total intracranial (TIV), grey matter, white matter and CSF volumes come from the CAT12 report.

Layout

cat12vbm-12.8.2/
├── index.json.gz                     # byte offsets for random access (nidl)
└── <dataset_id>/                     # one folder per OpenNeuro dataset, e.g. ds000001
    └── <dataset_id>-NNNNNN.tar       # WebDataset shard(s), ≤ ~3 GB each

Each sample is one T1w image of one (subject, session, run). Its key is sub-<label>_ses-<label>_mod-mwp1_run-<id> and it holds four files:

run follows the same convention as quasiraw-v2.1.1: the BIDS run- index of the source file when it is unambiguous, otherwise a hash of the source file name (brainprep convention). A sample of this folder and the quasiraw-v2.1.1 T1w sample of the same scan thus share (dataset_id, subject, session, run) in meta.json: 42,850 of the 42,951 samples have such a quasi-raw counterpart (the other 101 scans failed quasi-raw preprocessing, mostly because their raw image contains NaN voxels).

Morphometry

morphometry.json holds one entry per atlas. Features are named <tissue>_<ROI name> with <tissue> in Vgm (grey matter), Vwm (white matter) and Vcsf (CSF) volumes, in cm³:

Streaming with webdataset

See Usage with nidl for random access. Streaming without downloading (pip install webdataset nibabel huggingface_hub):

import gzip, io, json
import nibabel as nib
import webdataset as wds
from huggingface_hub import HfFileSystem, hf_hub_url

repo = "neurospin/openmind"
fs = HfFileSystem()
shards = sorted(fs.glob(f"datasets/{repo}/cat12vbm-12.8.2/*/*.tar"))
urls = [hf_hub_url(repo, s.split(f"{repo}/", 1)[1], repo_type="dataset") for s in shards]

def load_nifti(b):
    fh = nib.FileHolder(fileobj=io.BytesIO(gzip.decompress(b)))
    return nib.Nifti1Image.from_file_map({"header": fh, "image": fh})

def decode(s):
    return {
        "dataset_id": s["__url__"].rsplit("/", 2)[-2],
        "meta": json.loads(s["meta.json"]),
        "gm": load_nifti(s["mwp1.nii.gz"]).get_fdata(dtype="float32"),
        "morphometry": json.loads(s["morphometry.json"]),
        "qc": json.loads(s["qc.json"])["group_stats"],
    }

ds = (wds.WebDataset([f"pipe:curl -s -L {u}" for u in urls], shardshuffle=True)
      .map(decode)
      .select(lambda s: s["qc"]["qc"] == 1))  # keep scans passing the CAT12 QC
sample = next(iter(ds))
print(sample["dataset_id"], sample["meta"], sample["gm"].shape,
      sample["morphometry"]["total_volumes"]["TIV"])

Quality control

qc.json holds the CAT12 image quality ratings of the scan: noise to contrast ratio (NCR), inhomogeneity to contrast ratio (ICR) and the overall image quality rating (IQR), on the CAT12 grading scale (lower is better). qc is 1 when NCR < 4.5 and IQR < 4.5, the thresholds of OpenBHB (Dufumier et al., 2022), and 0 otherwise; 41,353 of the 42,951 images pass. No visual QC was applied.

Phenotypes / metadata

participants.tsv (tab-separated) gives the phenotypes of every scan of this repository, one row per scan (66,661 rows, 33,827 subjects, 790 datasets), identified by dataset_id, participant_id, session and run: it joins with the meta.json of any folder on (dataset_id, subject, session, run). Missing values are n/a.

Scanner information and the per-dataset image quality scores are in the original OpenMind release: see MIC-DKFZ/OpenMind.

License and citation

The original OpenMind dataset is released under CC-BY 4.0 and is built from OpenNeuro datasets, most of which are released under CC0. This derivative is released under CC-BY 4.0. If you use it, please cite OpenMind, the OpenNeuro datasets you use, brainprep and, for cat12vbm-12.8.2, CAT12:

@article{wald2025openmind,
  title   = {An OpenMind for 3D medical vision self-supervised learning},
  author  = {Wald, Tassilo and others},
  journal = {arXiv preprint arXiv:2412.17041},
  year    = {2025}
}

@misc{brainprep,
  title  = {{BrainPrep source code}},
  author = {Grigis, Antoine and Victor, Julie and Dorval, Loic and Duchesnay, Edouard},
  url    = {https://github.com/brainprepdesk/brainprep}
}

@article{gaser2024cat,
  title   = {{CAT}: a computational anatomy toolbox for the analysis of structural {MRI} data},
  author  = {Gaser, Christian and Dahnke, Robert and Thompson, Paul M. and Kurth, Florian
             and Luders, Eileen and {the Alzheimer's Disease Neuroimaging Initiative}},
  journal = {GigaScience},
  volume  = {13},
  year    = {2024}
}

Contact

Benoit Dufumier, NeuroSpin (CEA, Université Paris-Saclay).