OpenBrain packages automatically generated brain MRI labels with evidence intended to help researchers decide which records to reuse. In a preprint posted to arXiv on November 3, 2026, the authors describe 35,838 provenance-verified T1-weighted scan/session records from 607 OpenNeuro datasets. Peer review has not been verified, and the supplied evidence contains no independent replication.
Whole-brain parcellation assigns anatomical labels to regions in an MRI scan. Automated tools can produce these labels far faster than experts can review them, creating a practical problem: a plausible-looking output file offers little guidance about its reliability. OpenBrain attaches a primary parcellation, two additional model outputs, source and licensing information, and quality measurements to each record. These records represent scans or sessions, not necessarily distinct people.
The seven-check QC-7 panel examines label geometry, image-label consistency and disagreement among three models. Its checks cover issues such as labels extending beyond a brain mask, unusual regional volumes, disconnected fragments and inconsistent image intensities within labeled regions. Each measurement becomes a percentile risk, and the seven risks are averaged into score R. These are relative rankings, not calibrated probabilities of anatomical error. The release divides records into 8,960 HighConf, 17,918 Standard and 8,960 Flagged cases, while retaining measurements that users can inspect or use to change thresholds.
The authors tested whether lower-risk cases had better labels on IXI, with 563 cases; BrainLat, with 394; and Mindboggle-101, with 101. IXI and BrainLat informed development. The authors report fixing the quality checks and evaluation choices before testing Mindboggle, the only manually labeled holdout. The main outcome was macro Dice, an overlap measure averaged across 96 anatomical labels, excluding 14 sparse or ill-defined classes from the released 110-class parcellations.
Keeping the lowest-risk 25% increased average Dice relative to each full cohort by 1.3 percentage points on IXI, 3.7 on BrainLat and 1.5 on Mindboggle. Reported bootstrap 95% confidence intervals were 1.2–1.4, 3.5–3.9 and 1.0–2.0 percentage points, respectively; the Mindboggle subset contained 25 cases. This shows that selection identified cases with better labels, rather than improving an individual segmentation.
In a separate comparison, nnU-Net models trained on lower-risk selections of 1,000 cases outperformed models trained on equally sized random selections. Reported paired gains were 3.6 percentage points on Mindboggle (95% CI 2.8–4.4), 3.2 on BrainLat (2.6–3.8) and 3.7 on IXI (3.2–4.2). Benefits were smaller on a 16-region small-structure subset: 0.4–1.0 percentage points, with the BrainLat difference not statistically distinguishable.
The training selections also differed in source composition, so these comparisons cannot isolate the causal contribution of label quality or individual checks. IXI and BrainLat use FreeSurfer-derived references sharing the primary generator’s training-label convention; the manual holdout is comparatively small. Available details do not establish variability across training runs or whether uncertainty estimates account for repeated participants or shared sources.
OpenBrain excludes records with incomplete quality evidence, and demographic metadata remain incomplete. Its results therefore do not establish demographic fairness or generalization to excluded cases and other acquisition settings. The resource offers inspectable evidence for research on generated labels; its confidence tiers do not turn those labels into manual annotations or establish suitability for diagnosis or treatment planning.
Cover illustration generated with AI; conceptual, not research evidence.