This collection redistributes public datasets, each under its own license. Evaluation: CC0 (CNP, SOOP); CC BY 4.0 + the TCIA Data Usage Policy (UCSF-PDGM, UPENN-GBM, BraTS 2021); CC BY-SA 3.0 (IXI); CC BY-NC-SA 3.0 (ABIDE I, OpenBHB); CC BY-NC (ADHD-200). OpenBHB additionally asks users to accept the most restrictive data usage agreement of its source cohorts. Pretraining: FOMO300K is CC BY-NC-SA 4.0, each of its collections keeps its own license, and its HCP WU-Minn collections are under the HCP Open Access Data Use Terms (https://www.humanconnectome.org/study/hcp-young-adult/document/wu-minn-hcp-consortium-open-access-data-use-terms). Each dataset's README states its license, terms and required citations.

Brainmarks-sMRI

Public structural brain MRI datasets for building and evaluating sMRI foundation models: evaluation datasets with classification, regression and segmentation targets, and a large pretraining set. Code and benchmark: https://github.com/MedARC-AI/brainmarks-smri.

The datasets come in two kinds:

Hosted: the full data is here.
Manual: you download the data from its source and build the dataset with our scripts. Only the README and the tables (without samples.tsv) are here. Their licenses don't allow redistribution.

Evaluation datasets

Each evaluation dataset has harmonized metadata tables, a fixed dev/test split, and a brain mask and MNI affine for every image. The images are the original releases, unmodified.

Hosted

Dataset	Participants	Images	Targets	License	Size
ABIDE I	1,102	T1w	autism diagnosis, age, sex	CC BY-NC-SA 3.0	7.4 GB
ADHD-200	961	T1w	ADHD diagnosis	CC BY-NC	9.0 GB
BraTS 2021	1,468	T1w, T1c, T2w, FLAIR, tumor mask	tumor segmentation, MGMT	CC BY 4.0	15.8 GB
CNP	265	T1w	psychiatric diagnosis (4-way)	CC0	3.2 GB
IXI	582	T1w, T2w, PD	age	CC BY-SA 3.0	12.8 GB
OpenBHB	3,984	T1w	age (with site debiasing)	CC BY-NC-SA 3.0	32 GB
SOOP	1,715	FLAIR, DWI, ADC, lesion mask	stroke lesion segmentation, discharge mRS, NIHSS	CC0	27 GB
UCSF-PDGM	495	T1w, T1c, T2w, FLAIR, DWI, ADC, tumor mask	IDH, MGMT, 1p/19q, grade, survival, tumor segmentation	CC BY 4.0	15.9 GB
UPENN-GBM	630	T1w, T1c, T2w, FLAIR, tumor mask	survival, IDH1, MGMT, tumor segmentation	CC BY 4.0	25.3 GB

Manual

Dataset	Participants	Images	Targets	Access	Size
OASIS-1	416	T1w	dementia (CDR)	accept the OASIS terms	16 GB
OASIS-2	150	T1w, longitudinal	dementia, conversion	accept the OASIS terms	20 GB
FOMO26 Task 4	40	T2w, nerve and vessel mask	trigeminal nerve and vessel segmentation	download link from the organizers	2.2 GB
FOMO26 Task 5	48	T1w	polymicrogyria	download link from the organizers	1.1 GB
ATLAS R2.1	955	T1w, lesion mask	stroke lesion segmentation	apply for a decryption key	2.9 GB
ADNI	3,070	T1w, FLAIR, longitudinal	diagnosis, conversion, cognition, amyloid and tau PET, CSF, plasma, FreeSurfer volumes	apply for ADNI data access (LONI IDA)	65 GB
A4/LEARN	1,773	T1w, longitudinal	amyloid, tau PET, plasma pTau217, cognition (PACC), treatment arm	apply for A4 data access (LONI IDA)	43 GB

Each dataset folder's README.md has its source, version, license, citation, and split details.

Layout

<dataset>/
  README.md            # source, version, license, citation, splits
  shards/              # zip files of source/ (the original release) and derivatives/ (hosted only)
  tables/
    images.tsv         # one row per image: participant_id, session_id, modality, desc, path
    samples.tsv        # one row per scan session: participant_id, session_id, age, sex, site, targets... (hosted only)
    samples.json       # column descriptions, levels and units
    splits.tsv         # one row per participant: participant_id, split (dev or test)
    derivatives.tsv    # one row per processed image: path, mask method, QC metrics
    shards.tsv         # one row per packed file: path, shard
    source.sha256      # checksums of the original release (source/)
    derivatives.sha256 # checksums of the masks and affines (derivatives/)

All paths in the tables are relative to <dataset>/.
Derivatives: a brain mask on the image grid (<image>_mask.nii.gz) and a 4×4 nibabel-style affine matrix from the image to MNI152NLin6Asym (<image>_mni.txt). derivatives.tsv lists the images that have them.
<dataset>-dev-NNN.zip shards hold whole dev participants: their images, JSON sidecars, brain masks and MNI affines, about 1 GB per shard. <dataset>-meta.zip holds the rest of the release: participant tables, phenotypes and docs. The test participants' shards are not hosted.
Files in the shards keep their original paths under source/ and derivatives/. Unzipping the shards in <dataset>/ restores the dev part of the release, which you can check with sha256sum -c --ignore-missing tables/source.sha256 tables/derivatives.sha256.
For the manual datasets, the checksums let you check your own build against ours.
Splits are by participant: 50/50 dev/test, stratified, with a fixed seed. Benchmark tasks split each half further into train and eval sets. Official splits, where they exist, are the official_split column of samples.tsv.

Usage

hf download medarc/brainmarks-smri --repo-type dataset --include "cnp/*" --local-dir brainmarks-smri

import gzip
import zipfile

import nibabel as nib
import pandas as pd

root = "brainmarks-smri/cnp"
images = pd.read_csv(f"{root}/tables/images.tsv", sep="\t")
samples = pd.read_csv(f"{root}/tables/samples.tsv", sep="\t")
splits = pd.read_csv(f"{root}/tables/splits.tsv", sep="\t")
shards = pd.read_csv(f"{root}/tables/shards.tsv", sep="\t").set_index("path")["shard"]

# The dev participants.
dev = splits[splits.split == "dev"].participant_id

# Read one image from its shard.
path = images["path"][0]
with zipfile.ZipFile(f"{root}/{shards[path]}") as shard:
    image = nib.Nifti1Image.from_bytes(gzip.decompress(shard.read(path)))

To work with plain files instead, unzip the shards: cd brainmarks-smri/cnp && unzip -q 'shards/*.zip'.

Pretraining dataset

Dataset	Subjects	Images	Sequences	License	Size
FOMO300K	47,404	266,450	T1w, T2w, FLAIR, T1c, DWI, ADC, GRE, ASL, ...	CC BY-NC-SA 4.0 + per collection	692 GB

FOMO300K without the collections that share subjects with the evaluation datasets. Each image is masked to the brain, cropped and stored as a sparse int16 npz at native resolution, with its brain mask, MNI affine, QC and thumbnails. tables/images.parquet has one row per image with all the metadata for filtering. tables/shards.parquet gives each file's byte offset in its shard. See fomo300k/README.md for the format and a reading example.

License and citation

Datasets and their derivatives are released under their original licenses. Non-commercial terms apply to ABIDE I, ADHD-200, OpenBHB and FOMO300K. FOMO300K's HCP collections are under the HCP WU-Minn Open Access Data Use Terms. If you use a dataset, use the citation given in its README and follow all dataset-specific acknowledgement conditions.

Reproducing

Every evaluation dataset can be re-downloaded from its original source and checked against tables/source.sha256. The download scripts, table-building and preprocessing code are at https://github.com/MedARC-AI/brainmarks-smri.

Downloads last month: -

Total file size:

818 GB