Acknowledge source terms and privacy precautions

Requests are manually reviewed for acceptance of the source-specific terms and privacy precautions listed below. No co-authorship or collaboration is required.

Log in or Sign Up to review the conditions and access this dataset content.

BEM-QC

MRI-derived three-shell geometries with reproducible geometric quality control.

Release v0.1.1

The current geometry release contains 1,539 retained model records in ten source partitions. The historical source-policy-filtered technical audit covered 1,890 records: 1,593 accepted by geometric QC and 297 QC-excluded. A subsequent MRI-reference audit withdrew one previously accepted artifact because its saved BEM coordinate frame no longer matched its MRI reference after recon-all conformation. The retained audit cohort is now 1,592; 53 of those records remain deferred from release. Audit records are not verified unique people.

This is an explicit withdrawal, not an anatomical correction or silent replacement. See CHANGELOG.md and withdrawn_records.json for the affected opaque artifact ID and reason. Do not use that artifact from older snapshots. Older commits remain available for reproducibility; retained artifact files are byte-for-byte unchanged.

Source partition Records License
AOMIC-ID1000 375 CC0-1.0
AOMIC-PIOP1 138 CC0-1.0
AOMIC-PIOP2 130 CC0-1.0
NKI-Rockland 839 CC-BY-NC-4.0
FullHead-Stroke-Pilot 43 CC-BY-NC-SA-4.0
HyKid-Pathology 3 CC-BY-NC-4.0
RestingEEG 8 CC0-1.0
ABIDE-I-Dataset 1 CC-BY-NC-SA-3.0
ds000221 1 CC0-1.0
ds003505 1 CC0-1.0

Access

Files are manually gated. Request access through the dataset page using your Hugging Face account. Requests are reviewed against the same published criteria: acceptance of the applicable source-specific terms and acknowledgement of the privacy precautions below. No collaboration, co-authorship or project-specific scientific endorsement is required. Gating does not replace or alter upstream licenses, and it does not make facial geometry anonymous.

These are human-derived head surfaces, potentially including identifying facial shape. Handle them as sensitive derivatives: do not attempt identification or linkage to identified persons. Removed headers and opaque artifact IDs do not prove anonymization. This release includes no MRI voxels, recon-all directories, participant crosswalk, clinical metadata, raw execution logs or BEM matrices.

Files

data/<source>.zip contains five files per artifact:

  • native.npz: vertices (3, N, 3) and shared triangle indices (F, 3).
  • solver.npz: the exact saved solver geometry, with explicit unit normalization.
  • geometry.json: layer order, frame and units, source/export coordinate transform, canonical mesh hashes and conductivities.
  • qc.json: sanitized scientific QC records, including thresholds, flags, backend versions, unavailable fields and the independent CGAL profile.
  • source.json: source version, source-specific license evidence and release state.

Both NPZ files use float64 vertices and int64 triangle indices. The surface order is inner skull, outer skull, scalp. Native and solver resolutions differ. Native coordinates are FreeSurfer surface RAS in millimetres. Embedded solver coordinates are FIFF MRI (frame 5) in metres and are multiplied by 1,000 for export. This is unit normalization, not spatial registration or a verbatim copy of metre values. The original metre-array hash and canonical millimetre-array hash are both stored. Scanner RAS and surface RAS must not be treated as interchangeable.

record_manifest.json lists all 1,539 artifacts. release.json records the release inventory and archive checksums. Each ZIP also has an external member-level SHA-256 manifest. Load NPZ files with numpy.load(..., allow_pickle=False).

The retained per-artifact source.json files preserve their original v0.1.0 provenance. Current membership is defined by the v0.1.1 record and withdrawal manifests, not by an individual artifact's historical version field.

Retrospective MRI-reference evidence

The metadata extension retrospective_mri_reference_20260915_v1 adds per-record evidence to record_manifest.json, without changing any geometry archive or the historical qc.json records. The current 1,539 bundles comprise 843 matched construction replays and 696 records with unavailable replay evidence. The one confirmed mismatch is recorded only in withdrawn_records.json and is not part of the current geometry release. These are constructor/reference checks, not independent anatomical validation or prospective reference-gate passes.

Each manifest entry now includes:

Field Meaning
reference_evidence_state reference_replay_matched or reference_replay_unavailable; the withdrawal manifest uses reference_mismatch
reference_screening_version Version of the retrospective screening evidence
reference_replay_available Boolean; missing evidence is never counted as a pass
reference_replay_result matched, unavailable, or mismatch in the withdrawal manifest
reference_screening_report_sha256 SHA-256 of the primary per-record screening report
reference_evidence_scope retrospective_constructor_replay
prospective_reference_gate not_assessed for these historical records

reference_evidence_summary.json records the counts and input-manifest hashes. Ready-to-use opaque-ID lists are provided as subsets/reference_replay_matched.json (843 IDs) and subsets/reference_replay_unavailable.json (696 IDs). For MRI-referenced reuse, preferentially select the matched subset, then verify the image-to-surface transform for your own inputs. Replay-unavailable records remain available for geometry-only analyses subject to their other QC evidence.

import json
from pathlib import Path

records = json.loads(Path("record_manifest.json").read_text())
matched = [r for r in records
           if r["reference_evidence_state"] == "reference_replay_matched"]
assert len(matched) == 843
ids = {r["record_id"] for r in matched}
archives = sorted({r["archive"] for r in matched})

The standard-library helper select_reference_subset.py additionally validates the evidence fields and writes a selected record manifest. Archives remain source-partitioned: download the required ZIP archives after access approval, verify their checksums, and retain only artifacts/<record_id>/ members from the selected IDs. Per-record HTTP downloads are not provided by this archive layout.

The geometry version remains v0.1.1. DOI 10.57967/hf/10500 describes the registered revision 9337c2314e2d5c0bbb1ecbef1d1e2c3c9477e9a0; this metadata extension is a later repository commit, not a silently rewritten DOI snapshot. Record the extension commit alongside the DOI when using its evidence fields. The parent geometry archive hashes and withdrawal remain unchanged.

Scope and limitations

Strict QC checks computational geometry, not anatomical boundary accuracy. No independent anatomical ratings or clinical source-localization ground truth are claimed. The deposit contains accepted meshes only; QC-excluded meshes are not included. The accompanying manuscript reports aggregate exclusion evidence.

Input-file identity, mesh identity and QC state are supported; exact historical constructor/repair attribution is incomplete. The present portable workflow has stronger prospective attempt logging than some historical artifacts.

Three intact nested compartments cannot represent every skull opening, implant or complex tissue configuration. The surfaces do not define cortical sources, sensors, a complete forward operator, or a clinically validated model. Users must check numerical resolution for their own sources and sensors. Recheck QC after remeshing or transformation. Prevent participant/acquisition leakage in ML splits; an artifact identifier is not a verified participant identifier.

Attribution and versioning

Dataset authors, in citation order:

  1. Ruslan Kalimullin: Center for Bio- and Medical Technologies, Moscow, Russia.
  2. Nikolay Koshev: Center for Bio- and Medical Technologies, Moscow, Russia; LLC "LIFT Center", Moscow, Russia.

Machine-readable author and affiliation metadata are provided in CITATION.cff. These credits do not replace upstream dataset attribution.

See DATA_LICENSES.md, per-record source evidence and the upstream dataset citations. The MIT license of the processing software does not license these meshes. Cite this repository at the exact release commit/tag together with the upstream sources used. DOI metadata, when assigned, are available through the repository's DOI badge. This Hugging Face deposit is not a separate mirror archive.

Downloads last month
52