OpenBrain

An Auditable Generated-Label Release for Whole-Brain MRI Parcellation

Qizhen Lan · Yu-Chun Hsu · Yuxiang Wei · Lijing Zhu · Zenan Sun · Liang He · Lishan Yu · Xiaoqian Jiang

_{✉️ Contact: Qizhen.Lan@uth.tmc.edu}

35,838 skull-stripped T1w MRI scans from 607 OpenNeuro datasets, each with a 110-region parcellation, two committee parcellations, and seven quality checks. Labels are model output, not manual annotation. All data are CC0.

Scans	Source datasets	Regions	Quality checks	Size
35,838	607	110	7	226 GB

Files

One tar per OpenNeuro dataset. Inside, one folder per case with five files.

cases/ds000001.tar
  ds000001__sub-01__ses-DEFAULT__sub-01_T1w/
    t1w_brain.nii.gz                 skull-stripped T1w, 1 mm
    brain_mask.nii.gz                SynthStrip mask
    parcellation_modernsurfer.nii.gz 110-region parcellation (primary label)
    parcellation_swinunetr.nii.gz    committee model 2
    parcellation_unetr.nii.gz        committee model 3
metadata/cases.csv        one row per case: source dataset, archive, DOI, license, age/sex/scanner when known
metadata/quality.csv      Q1–Q7, r1–r7, R, tier
metadata/labels.json      region ids and names
metadata/q2_fences.csv    which volume fence each case used
metadata/training_selections.csv  the paper's three training selections, release part (906 / 935 / 931 cases)
metadata/sources.csv  raw_paths.csv  excluded.csv  archives.csv  checksums.csv
sample/                   64 cases, unpacked, one per dataset

Download

pip install -U huggingface_hub
hf download lanqz7766/OpenBrain --repo-type dataset --local-dir openbrain --include "metadata/**"   # 60 MB
hf download lanqz7766/OpenBrain --repo-type dataset --local-dir openbrain --include "sample/**"     # 64 cases
hf download lanqz7766/OpenBrain --repo-type dataset --local-dir openbrain cases/ds000001.tar        # one dataset
hf download lanqz7766/OpenBrain --repo-type dataset --local-dir openbrain                           # everything, 226 GB

Load a case

# pip install git+https://github.com/lanqz7766/openbrain
from openbrain import load_case
case = load_case("ds000001__sub-01__ses-DEFAULT__sub-01_T1w")   # downloads cases/ds000001.tar once, then extracts
case.t1w, case.parcellation, case.quality                          # nibabel images and the quality row

To pick cases by quality first:

import pandas as pd
q = pd.read_csv("openbrain/metadata/quality.csv")
ids = q.loc[q.tier == "HighConf", "case_id"]

Quality fields

Field	Meaning
Q1	labeled voxels outside the brain mask
Q2	regions whose volume is an outlier for its source dataset
Q3	labeled voxels in fragments far from their region
Q4	intensity variation inside regions
Q5	overlap between the primary and the two committee parcellations
Q6	voxel-level disagreement among the three
Q7	region-volume disagreement among the three
r1–r7	each Q as a percentile over all cases; higher is riskier
R	mean of r1–r7
tier	HighConf lowest quarter of R · Standard middle half · Flagged top quarter

_{A low-risk (left) and a high-risk (right) case}

Definitions, provenance accounting and limitations are in the paper and in docs/datasheet.md.

Citation

@inproceedings{lan2026openbrain,
  title     = {OpenBrain: An Auditable Generated-Label Release for Whole-Brain MRI Parcellation},
  author    = {Lan, Qizhen and Hsu, Yu-Chun and Wei, Yuxiang and Zhu, Lijing and Sun, Zenan and He, Liang and Yu, Lishan and Jiang, Xiaoqian},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026},
  eprint    = {2611.05348},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url       = {https://arxiv.org/abs/2611.05348}
}