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

cs.CVNov 3, 2026

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

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

Organizations: The University of Texas Health Science Center at Houston · Georgia Institute of Technology · University of Houston–Clear Lake · ETH Zurich

Abstract

Automatic tools can now generate whole-brain MRI parcellations at cohort scale, but generated-label releases often provide little evidence about which cases are reliable enough to reuse. We introduce OpenBrain, a confidence-aware release of 35,838 provenance-verified OpenNeuro T1w cases from 607 source datasets. OpenBrain is organized as an auditable case-record release rather than a directory of standalone segmentation files: each record links a primary parcellation, two committee parcellations, source/license provenance, raw QC-7 measurements, percentile risks, a combined risk score R, and a confidence tier. QC-7 is a reference-free reliability panel over label geometry, image-label consistency, and committee disagreement, so users can filter, weight, audit, or re-threshold generated labels without rerunning inference. We evaluate this released evidence under a frozen protocol. Across external cohorts, lower QC-7 risk corresponds to higher generated-label quality; in fixed-budget training, lower-risk selections improve paired macro Dice over random selection by +3.2 to +3.7 percentage points. OpenBrain also exposes raw axes, committee outputs, provenance, and datasheet documentation, so confidence claims remain inspectable after download. It does not turn generated labels into manual annotations or clinical-grade segmentations. Its contribution is a large open generated-parcellation substrate, a reusable case-level risk-evidence layer, and an external validation protocol for release-time QC. Dataset: https://huggingface.co/datasets/lanqz7766/OpenBrain. Code: https://github.com/lanqz7766/openbrain.

Figures & tables

Figure 1: Release-time contrast between a standard generated-label release and OpenBrain. OpenBrain packages each parcellation with provenance, committee outputs, QC-7 risk fields, confidence tiers, and validation evidence.

Benchmark / resource	Scale	MRI / cohort	Labels	W	M	O	Q	Primary role
Brain extraction, tissue, and anatomical segmentation benchmarks
NFBS [ 20 ]	125 scans	NKI T1w	Brain masks	$\times$	$\circ$	$\checkmark$	$\circ$	Skull stripping
CC-359 [ 21 ]	359 scans	Multi-vendor T1w	Brain masks	$\times$	$\checkmark$	$\checkmark$	$\circ$	Skull stripping
Mindboggle-101 [ 6 ]	101 subj.	Multi-source T1w	DKT cortex	$\circ$	$\checkmark$	$\circ$	$\times$	Cortical anatomy
Hammers atlas [ 7 , 8 ]	30 subj.	Adult T1w	83 ROIs	$\checkmark$	$\times$	$\circ$	$\times$	Atlas anatomy
LPBA40 [ 9 ]	40 subj.	Adult T1w	56 ROIs	$\checkmark$	$\times$	$\circ$	$\times$	Reg./seg. eval.

Table 1: Positioning of OpenBrain against public brain MRI segmentation resources. W: whole-brain labels; M: multi-cohort or acquisition diversity; O: open redistribution; Q: released QC/risk fields or release audit.

Figure 2: OpenBrain case-level release record. Each source-scoped T1w case links generated parcellations, committee outputs, QC-7 checks, percentile risks, combined risk $R$ , and a confidence tier.

Figure 3: Retention-quality curves for QC-7 risk signals. Columns are external cohorts and rows group QC families. Curves report retained-case

\Delta

macro Dice after selecting the lowest-risk

K\%

of cases.

Cohort	Overall $\Delta$ macro (pp)	Def-A $\Delta$ small-struct (pp)
IXI	$+1.3$ [+1.2, +1.4]	$+0.9$ [+0.8, +1.0]
BrainLat	$+3.7$ [+3.5, +3.9]	$+1.5$ [+1.3, +1.7]
Mindboggle-101	$+1.5$ [+1.0, +2.0]	$+1.2$ [+0.3, +2.0]

Table 2: QC-7 retention at

K{=}25\%

for overall 96-label macro Dice and Def-A small-structure Dice. Values are percentage-point deviations from the full-cohort baseline; colored brackets are bootstrap 95% CIs.

Paired $\Delta$ macro Dice (pp; per-case)
Selection	Mindboggle-101	BrainLat	IXI
	( $n{=}101$ )	( $n{=}394$ )	( $n{=}563$ )
Reference	manual labels	FreeSurfer-derived	FreeSurfer-derived
Lower-risk	0.746 [0.733, 0.758]	0.719 [0.709, 0.730]	0.769 [0.764, 0.775]
Random	0.710 [0.696, 0.722]	0.687 [0.678, 0.697]	0.733 [0.727, 0.738]
Higher-risk	0.456 [0.450, 0.461]	0.458 [0.454, 0.463]	0.490 [0.487, 0.492]

Table 3: Fixed-budget training-selection results on three external cohorts. Models are trained on 1,000-case lower-risk, random, or higher-risk OpenBrain selections and evaluated with 96-label macro Dice; colored brackets are bootstrap 95% CIs. The reference row gives the basis of each cohort’s evaluation labels.

Panel of Figure 4 — Figure 4: QC-7 axis correlation and fixed-budget training-selection results. Panel A shows partial correlations among risk axes. Panel B compares nnU-Net models trained on 1,000-case lower-risk, random, and higher-risk selections.

Figure 5: Qualitative audit of low- and high-risk cases. Rows separate manual-GT cohorts from no-GT training-pool selections; tiles show image/label views, entropy, QC-7 bars, and risk position.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark / resource	Scale / split	MRI / cohort	Labels	Source	W	M	O	Q	Gap vs. OpenBrain
Brain extraction, tissue, and anatomical segmentation benchmarks
NFBS [ 20 ]	125 scans	NKI T1w	Brain masks	Manual correction	$\times$	$\circ$	$\checkmark$	$\circ$	Mask-only
CC-359 [ 21 ]	359 scans	Multi-vendor T1w	Brain masks	Consensus masks	$\times$	$\checkmark$	$\checkmark$	$\circ$	Mask-only
Mindboggle-101 [ 6 ]	101 subj.	Multi-source T1w	DKT cortex	Manual edits	$\circ$	$\checkmark$	$\circ$	$\times$	Small; cortical-heavy
Hammers atlas [ 7 , 8 ]	30 subj.	Adult T1w	83 ROIs	Manual	$\checkmark$	$\times$	$\circ$	$\times$	Small atlas cohort
LPBA40 [ 9 ]	40 subj.	Adult T1w	56 ROIs	Manual protocol	$\checkmark$	$\times$	$\circ$	$\times$	Narrow healthy cohort

Appendix

Table 4: Detailed positioning of OpenBrain against public brain MRI segmentation benchmarks. W: whole-brain labels; M: multi-cohort or acquisition diversity; O: open redistribution; Q: automated label-quality scoring or release audit.

Source	$n$	Population	Age range
ABIDE I	100	autism + typical	6.5–39 y
CoRR	418	healthy	6–62 y
ADNI	100	50 AD + 50 HC	55–96 y
SALD	486	healthy	19–80 y
total	1,104	mixed	6.5–96 y

Appendix

Table 5: ModernSurfer training cohort summarized from upstream model provenance.

Figure 6: Development provenance for the twelve original candidate QC signals. Each panel reports retained-case

\Delta

macro Dice after ranking by one candidate signal. Status labels mark retained, dropped, replaced, and hard-gate signals under the frozen QC-7 release policy.

Paired $\Delta$ Def-A small-struct Dice (pp; per-case)
Selection	Mindboggle-101	BrainLat	IXI
	( $n{=}101$ )	( $n{=}394$ )	( $n{=}563$ )
Lower-risk	0.812 [0.802, 0.822]	0.812 [0.802, 0.822]	0.852 [0.846, 0.856]
Random	0.803 [0.790, 0.814]	0.809 [0.799, 0.818]	0.842 [0.836, 0.847]
Higher-risk	0.557 [0.544, 0.570]	0.537 [0.529, 0.544]	0.582 [0.576, 0.587]
Lower-risk $-$ Random	$+0.9$ [ $+0.3$ , $+1.6$ ]	$+0.4$ [ $-0.5$ , $+1.2$ ]	$+1.0$ [ $+0.5$ , $+1.5$ ]

Appendix

Table 6: Def-A small-structure Dice for the fixed-budget training selections across three external cohorts. Paired differences are reported in percentage points; colored brackets are bootstrap 95% CIs.

Selection	Manifest-tier composition	$R$ range	Interpretation
Lower-risk	720 HighConf / 247 Standard / 33 Flagged	$0.082$ – $0.997$	Lower-risk selection from the low-risk end
Random	258 HighConf / 508 Standard / 234 Flagged	$0.123$ – $0.997$	Fixed-size random baseline without confidence ranking
Higher-risk	0 HighConf / 205 Standard / 795 Flagged	$0.461$ – $0.997$	Higher-risk selection from the high-risk end

Appendix

Table 7: Composition of lower-risk, random, and higher-risk training selections

Dataset	Committee model	$n$	Mean Dice $\pm$ SD	Reported $M_{1}$
IXI	$M_{1}$ ModernSurfer	563	0.8895 $\pm$ 0.0176	0.8895 $\pm$ 0.0176
IXI	$M_{2}$ SwinUNETR	563	0.8424 $\pm$ 0.0532	—
IXI	$M_{3}$ UNETR	563	0.8051 $\pm$ 0.0739	—
BrainLat	$M_{1}$ ModernSurfer	394	0.8651 $\pm$ 0.0631	0.8798 $\pm$ 0.0282 ∗
BrainLat	$M_{2}$ SwinUNETR	394	0.7936 $\pm$ 0.1164	—
BrainLat	$M_{3}$ UNETR	394	0.7236 $\pm$ 0.1570	—

Appendix

Table 8: Per-committee-member mean macro Dice (

\pm

SD) on the 78-label FastSurfer-aligned grouping.

Cohort	$n$ kept	96-label kept mean	96-label $\Delta$	110-label $\Delta$	Gap
IXI	141	0.9028	+1.3	+1.6	+0.3
BrainLat	98	0.9004	+3.7	+3.6	-0.0
Mindboggle-101	25	0.8327	+1.5	+1.4	-0.1

Appendix

Table 9: Label-space sensitivity at the primary

K\!=\!25\%

operating point. Effects are percentage-point changes in retained-case macro Dice relative to the all-case baseline.

Resource / accounting set	Cases	Sources	Label or audit basis	Role
OpenNeuro T1w records audited	38,458	667	BIDS path + source licence/provenance audit	candidate pool
Strict hosted OpenBrain release	35,838	607	generated WB labels + complete QC-7 manifest	released cases
Excluded during audit	2,620	62	manual-review, missing, non-CC0, missing/ambiguous path	not hosted
IXI validation	563	1	expert-curated anatomical reference	external validation
BrainLat validation	394	1	expert-curated anatomical reference	external validation
Mindboggle-101 holdout	101	1	manual labels + expert repair	confirmatory holdout

Appendix

Table 10: Release provenance, exclusion accounting, and validation resources.

Figure 7: Marginal and partial correlations among QC-7 percentile-risk axes. Partial correlations condition on the other five axes. Both panels use hosted release cases with complete $r_{1},\ldots,r_{7}$ values.

Risk composition	BrainLat	IXI	Mindboggle-101	Mean
Full QC-7 ( $r_{1}$ – $r_{7}$ )	$+3.68$	$+1.31$	$+1.51$	$+2.17$
Drop $r_{5}$	$+3.58$	$+1.26$	$+1.35$	$+2.06$
Drop $r_{6}$	$+3.61$	$+1.33$	$+1.51$	$+2.15$
Drop $r_{7}$	$+3.70$	$+1.26$	$+1.32$	$+2.10$
Collapse $r_{5}$ – $r_{7}$ to one term	$+3.61$	$+1.15$	$+1.34$	$+2.04$
$r_{1}$ – $r_{4}$ only	$+2.88$	$+0.79$	$+0.87$	$+1.51$

Appendix

Table 11: Sensitivity of the

K{=}25\%

retention gain to the committee axes in

R

. Values are retained-case

\Delta

macro Dice in percentage points on the same external predictions; no model is retrained.

Field	Coverage	Full release	HighConf vs. Flagged contrast
Sex	72.9% (26,114)	F/M/missing: 39.9% / 33.0% / 27.1%	F: 52.0% vs 35.0%; M: 23.6% vs 38.8%; missing: 24.3% vs 26.2%
Age	63.8% (22,854)	known mean $29.2\pm 15.0$ y; missing 36.2%	known mean $26.6\pm 9.2$ vs $32.0\pm 18.4$ y; missing 31.7% vs 40.4%
Health	29.0% (10,409)	healthy/ill/missing: 20.3% / 8.8% / 71.0%	healthy: 32.6% vs 10.0%; ill/path.: 6.1% vs 10.5%; missing: 61.3% vs 79.4%
Manufacturer	74.9% (26,833)	Siemens/Philips/GE/missing: 52.8% / 14.2% / 7.9% / 25.1%	Philips+GE: 10.1% vs 31.1%; missing: 31.2% vs 16.8%
Field strength	71.0% (25,446)	3T/1.5T/7T-UHF/other/missing: 60.9% / 4.8% / 1.6% / 3.7% / 29.0%	1.5T: 1.3% vs 8.9%; 7T-UHF: 0.0% vs 4.4%; other: 0.0% vs 12.7%

Appendix

Table 12: Metadata coverage and tier composition in the strict OpenBrain release. Tier contrasts compare HighConf and Flagged cases as an imbalance audit, not as demographic fairness evidence.

Figure 8: Metadata-stratified QC-risk audit. Panels summarize metadata coverage, subgroup mean-risk offsets, sampling-arm composition shifts, and the interpretation boundary for this audit.

Figure 9: Subgroup distributions of QC-risk score

R

. Each row shows random-pool case-level risks; the black segment marks the interquartile range and the central point marks the median.

Figure 10: Mindboggle-101 worst example. Section A shows manual GT, the three committee predictions, GT–model differences, committee vote entropy, and pairwise model differences at the worst-disagreement cuts. Section B shows the corresponding QC-7 bars and raw $Q_{k}$ values.

Brain MRI

Brain MRI uses magnetic resonance images to map brain anatomy and pathology for diagnosis, measurement, prediction, and image-guided analysis.

See all 133 papers

NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding
May 19, 2026Mohammad H. Abbasi, Favour Nerrise, Shaurnav Ghosh +12Healthcare 3D Medical Imaging
Stanford University · 1Stanford University, Stanford, CA, USA
PerFact: Perception-Derived Fact Prompting for 3D Brain MRI Report Generation
Aug 18, 2026Jianyu Sun, Zhenxuan Zhang, Guang Yang +1Radiology Report Generation Medical VLMs
Department of Bioengineering, Imperial College London, London, UK · Imperial-X, Imperial College London, London, UK · School of Biomedical Engineering & Imaging Sciences, King’s College London, London, UK
Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data
Jun 24, 2026Shangkun Li, Jie Xu, Yi Guo +2Medical Image Anomaly Detection VLM Reasoning
College of Biomedical Engineering, Fudan University, Shanghai, China

May 19, 2026J/K move · Enter open · S save

NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding

Mohammad H. Abbasi, Favour Nerrise, Shaurnav Ghosh, Ridvan Yesiloglu+11

Stanford University · 1Stanford University, Stanford, CA, USA

Healthcare 3D Medical Imaging Medical Visual Question Answering Medical Image Grounding 3D Visual Question Answering Medical Image Benchmarks+1

Open paper AskPDF arXiv

We present NeuroQA, a large-scale benchmark for visual question answering in 3D brain magnetic resonance imaging (MRI), with 56,953 QA pairs from 12,977 subjects across 12 datasets. It spans ages 5-104 and five clinical domains: Alzheimer's, Parkinson's, tumors, white matter disease, and neurodevelopment. Unlike prior medical Visual Question Answering (VQA) efforts that operate on 2D slices or rely on narrow diagnostic labels, NeuroQA pairs every item with a full 3D volume. It evaluates 11 clinically grounded reasoning skills across Yes/No, multiple-choice, and open-ended formats. Of the 203 templates, 131 are image-grounded (answerable from a 3-plane viewer) and 72 are image-informed (ground truth from quantitative volumetry or clinical instruments). To remove text-only shortcuts, we apply answer-distribution refinement, reducing closed-format text-only accuracy from

>

80% to 44.6%; image necessity is assessed separately through an image-grounding protocol released with the benchmark. A 38-rule deterministic pipeline and two rounds of expert review verify every QA pair against FreeSurfer measurements, metadata, or radiology report fields, with zero same-subject contradictions across templates. We conduct a clinician evaluation in which two clinicians independently assess 100 frozen test items on a three-plane viewer. On closed-format (Yes/No + multiple-choice) test-public items, the best zero-shot vision-language model and a supervised 3D CNN baseline reach 47.5% and 43.7% accuracy respectively, both below the 49.4% text-only majority-template floor. NeuroQA adopts a two-tier release with public QA pairs for open-access datasets and reproducible generation scripts for datasets restricted by data use agreements (DUAs), plus subject-level splits, a held-out private test set, and an online leaderboard.

Abstract

Brain MRI

Explore similar work

NeuroQA: A Large-Scale Image-Grounded Benchmark for 3D Brain MRI Understanding

PerFact: Perception-Derived Fact Prompting for 3D Brain MRI Report Generation

Enhancing Brain MRI Anomaly Detection and Reasoning with ROI Rethink and Synthetic Data