A quality-controlled re-release of FOMO-300K for training 3D brain-MRI models. It is derived from an earlier jvol-compressed copy (quality 60, upstream jvol format), which this release replaces.

Every volume was re-encoded in the chunked, checksummed medrs 0.3 .jvol format. The corpus was then audited volume by volume. The audit combined image-content checks with visual review of every flagged case, and removed, relabelled or fixed what it found.


Volumes	288,781 (from 295,132; 6,351 quarantined)
Sites / subjects	37 / 57,665
Size	~478 GB in 545 tar shards
Format	medrs 0.3 `.jvol`, lossy quality 70, 32³ chunks (crop-first reads)
Split	subject-disjoint train/val (259,456 / 29,325) with duplicate-aware subject linking

Quick start

pip install "huggingface_hub>=0.30" "medrs>=0.3.0"
python fetch_fomo_clean.py --out fomo_clean --modality T1w T2w FLAIR   # or omit --modality for everything

import medrs
img = medrs.load("fomo_clean/T1w/PT021_IXI_sub-002_ses-01_T1w.jvol")
x = img.to_numpy()            # float32, voxel order as stored
affine = img.affine           # RAS world, float64
crop = medrs.load_cropped("fomo_clean/T1w/....jvol", (40, 40, 40), (96, 96, 96))  # decodes only the chunks needed

Files unpack to <modality>/<site>_<subject>_<session>[_run]_<suffix>.jvol. The directory name is the modality. Diffusion is split by b-value (dwi_bval1000/, …).

Files

file	contents
`shards/<modality>-NNNNN.tar`	~2 GB uncompressed tars (the `.jvol` payload is already compressed)
`manifest.csv`	one row per volume: `rel, site, subject, modality, shape, spacing_mm, size_bytes, obliquity_deg, lr_unverified, relabelled_from, oasis2_reoriented, split, shard`
`split.json`	train / val lists (relative paths) and split metadata
`SHARDS.json`	per-shard file count, bytes, sha256 (checked by `fetch_fomo_clean.py`)
`quarantine.json`	every removed volume: class, evidence, date
`relabels.json`	every relabelled volume: old and new modality, evidence
`ENCODER.json`	exact encoder (medrs version and commit, quality, chunk shape)
`QC_REPORT.md`	the full audit log: methods, thresholds, what was looked at, and what was retracted
`source_metadata/`	carried over unchanged from the source: top-level `LICENSE`, per-site `LICENSE.txt`, data use agreements (`DUA.pdf`: BraTS, GSP, HCP, HCP Test-Retest, OASIS-1, OASIS-2), and `participants.tsv` / `mapping.tsv` / `mri_info.tsv`

Provenance and fidelity

The release has been through two lossy steps relative to the original FOMO-300K NIfTI:

original NIfTI → upstream jvol, quality 60 (the earlier compressed copy, now retired);
upstream jvol → medrs 0.3 .jvol, quality 70, 32³ chunks (this release).

Step 2 was calibrated on 331 volumes. Its median error in percentile-normalised intensity is 0.0051 RMSE relative to the upstream jvol; the worst case is 0.0077. For comparison, typical tokenizer reconstruction error is ~0.05. Per-volume decode, checksum, shape and affine checks passed on all files. For lossless work, start from the original FOMO-300K.

What the audit changed

Removed (moved to quarantine, listed in quarantine.json):

class	n	what
`duplicate_copy`	4,963	the same scan stored under two subject IDs (shared noise pattern, checked by eye). Mostly MSD-BrainTumor copies of BraTS, OpenNeuro re-uploads, and SLIM/CoRR. One copy per group is kept.
`phase_or_field_map`	634	wrapped phase, SWI phase, B0 field maps, MP2RAGE phase
`non_brain_spine`	396	OpenNeuro ds005025, a cervical-spine study
`derived_map`	218	edge/gradient-like derived maps in MGH-Wild (not acquisitions)
`blank_or_corrupt`	121	blank, noise-only, block-corrupted, signal-free or partial volumes
`flattened_4d_series`	19	HBN diffusion series stored as one 3D volume concatenated along z

Relabelled (476, relabels.json): T2w→T1w 165, T2w→FLAIR 215, T1w→T2w 60, T2w→scan 36 (a non-standard WAND contrast). Each was confirmed by eye against reference T1/T2 images. Only volumes that were viewed were changed.

Fixed: 748 OASIS-2 T1w volumes were stored upside-down under RAS headers: in each session, the second half of the runs. They were re-encoded upright from the source, and a full re-score confirms it.

Checked, no change needed:

Orientation was checked by registration to MNI152 on 3 volumes from each of 17,400 acquisition protocol groups. It detects 93% of planted orientation errors and falsely flags 3%. No errors beyond OASIS-2 were found.
Intensity checks (wraparound, ceilings, saturation, non-finite values) found no corruption in real images.
Noisy high-b diffusion was kept: it shows anatomy, and noise level is a real property of the data.

Split: subjects are hashed into train/val (stable under corpus edits). Subjects joined by duplicate or same-person evidence (9,225 links) always share a side.

Known limitations

Left-right cannot be verified from image content. lr_unverified=True marks OASIS-1/2 (ANALYZE-derived, the usual source of L-R errors).
Obliquity: 31% of voxel grids are tilted > 10° against the head; the header is correct. Volumes are stored on their native grid, so use obliquity_deg or the affine to filter or resample.
Mislabels may remain: the contrast audit covered T1w/T2w; other modalities were reviewed only through low-score sampling.
Data properties, not errors: many gre volumes are dark and bias-dominated. Thin-slab acquisitions (≤ ~40 slices) are common in clinical sites.
Duplicates were searched within identical voxel grids. Copies that were re-sampled or cropped differently are not detected.

License and attribution

Distributed under the same CC BY-NC-SA 4.0 licence as FOMO-300K. All original dataset licences, data use agreements and citation requirements apply. The per-site LICENSE.txt files and data use agreements (DUA.pdf) are included in source_metadata/; read the agreement for any site you use. See also the original dataset page.

Please cite the original FOMO-300K paper:

@article{Cerri2026large,
  title={A large-scale heterogeneous 3D magnetic resonance brain imaging dataset for self-supervised learning},
  author={Cerri, Stefano and Munk, Asbj{\o}rn and Llambias, Sebastian N{\o}rgaard and Ambsdorf, Jakob and Machnio, Julia and Nersesjan, Vardan and Hedeager Krag, Christian and Liu, Peirong and Rocamora Garc{\'\i}a, Pablo and Mehdipour Ghazi, Mostafa and Boesen, Mikael and Benros, Michael Eriksen and Iglesias, Juan Eugenio and Nielsen, Mads},
  journal={arXiv preprint arXiv:2506.14432},
  year={2026},
  url={https://arxiv.org/abs/2506.14432}
}