BEAMSTER

Brain mEtAstases segMentation for STEreotactic Radiotherapy — a retrospective MRI dataset with expert segmentations. Re-host of the author deposit figshare 10.6084/m9.figshare.29365844 v1 (CC BY 4.0), from University Hospital Ostrava, Czech Republic, Oct 2019 - Sep 2024.

140 patients, one contrast-enhanced T1w 3D MPRAGE volume each, with a binary brain-metastasis mask drawn by a board-certified radiation oncologist (13 years' experience) for stereotactic radiotherapy planning and independently verified by a board-certified neuroradiologist.

Cohorts

subset	cases	lesions	mean lesion vol	note
Dataset_A	113	216	6.7 cc	treated Oct 2019 - Apr 2022, 1/3/5 fractions
Dataset_B	27	44	0.5 cc	treated Jul 2021 - Sep 2024, deliberately enriched with very small lesions

There is no official train/val/test split — train.jsonl carries all 140 rows with split: "train". subset is a cohort label, not a split.

Do not use A vs B as a train/test partition. Their accrual windows overlap by 9 months at the same institution, and the paper never states the cohorts are patient-disjoint. The de-identified IDs make this unverifiable.

Layout

dataset/Dataset_A/Dataset_A_YYY.nii.gz        ce-T1w volume
dataset/Dataset_A/Dataset_A_YYY_segm.nii.gz   binary metastasis mask
dataset/Dataset_B/...                          same convention
Spreadsheets/Table_clinical_data.{csv,xlsx}    140 rows of clinical data
Spreadsheets/image_voxel_parameters.{csv,xlsx} per-case dimensions + voxel size
README_original.txt                            the deposit's own README, verbatim
train.jsonl                                    one row per case (see columns below)

train.jsonl columns: case_id, subset, image, mask, shape, spacing_mm, orientation, lesion_volume_cc, n_components, fg_voxels, split, plus the clinical fields sex, age, scanner, primary_origin, rt_dose_gy, n_fractions, isodose_pct, vital_status, survival_days, prior_surgery, treatment_status. vital_status (1 case) and survival_days (15 cases) are nullable.

Verified properties

• Image and mask share an identical shape and affine in all 140 cases — no resampling is needed to pair them.
• Masks are strictly {0, 1} but stored as float32 — cast to uint8 in a loader.
• All volumes are 512x512 in-plane, z = 201-506. 72 distinct voxel spacings; slice thickness is 1.0 mm (137 cases), 1.25 mm (2), 0.625 mm (1).
• Orientation is LPS for 139 cases and LAS for Dataset_B_001. Image and mask agree within every case, so this only matters to code that hardcodes axis order.
• No empty masks. ID sequences have gaps: Dataset_A spans 001-123 (10 missing), Dataset_B spans 001-030 (3 missing).

Caveats that affect evaluation

1. Masks are a GTV/PTV mixture. Contours follow ICRU 50 and the exported structure is "GTV or PTV, depending on availability" — so an unknown subset of masks carries a planning margin and is systematically larger than the tumour itself. No column in any released table records which case got which.
2. Annotation is partial by design. Only lesions selected for irradiation were contoured. Other metastases visible in the same scan are unlabelled background, so a correct detection can be scored as a false positive.
3. Not native geometry. Every volume was rigidly registered to the radiotherapy planning CT and resampled into CT coordinate space and resolution; acquisition was 0.9-1.0 mm isotropic. Volumes are also AFNI-defaced. The planning CT is not released.
4. Connected components != lesions. The masks contain 335 26-connected components against the paper's 260 lesions, and no size threshold reconciles the two. Total segmented volume does match the paper (within 4%), so the masks are faithful — but instance-level metrics will not reproduce the published lesion counts.
5. Extreme class imbalance. 55.8% of components are under 1 cc; the smallest foreground fraction is 2.8e-06 (Dataset_A_074, ~200 voxels in 75M). Evaluating on a single sample will report a near-zero Dice that looks like a bug but is not.
6. Single observer. One annotator; the paper explicitly states inter-observer variability was not assessed. There are no multi-rater or consensus tiers.

Overlap with other datasets

None found. Single-institution Czech data; the BraTS-METS 2023, BraTS-METS 2025 Lighthouse and UCSF-BMSR papers contain no reference to Ostrava, Czech sites, CyberKnife or MultiPlan. No cross-reference ID column exists — IDs are de-identified Dataset_X_YYY only.

Citation

Nohel M, Reguli S, Kaplanova R, Jackaninova J, Chmelik J, Knybel L. BEAMSTER: Brain mEtAstases segMentation for STEreotactic Radiotherapy, A Retrospective MRI Dataset with Expert Segmentations. Scientific Data (2026). doi:10.1038/s41597-026-07777-0

Data: doi:10.6084/m9.figshare.29365844 — CC BY 4.0, redistribution permitted with attribution, commercial use permitted.