Deterministic, audited preprocessing of the FOMO-MRI/FOMO300K corpus (306,207 brain-MRI scans, 36 source cohorts) into the fixed model-input tensor used by a NeuroJEPA-B encoder: a single-channel 96×108×96 volume (12³ patches, 8×9×8 = 576 tokens), cached at 100×120×100 fp16 with the online centre-crop applied at training time.

This is a derived, access-controlled mirror of FOMO300K. Raw scans, subject manifests and site credentials are not included. Use is gated under the same terms as the source cohorts.

Two preprocessing arms

Every scan is carried through the audited terminal loader (neurojepa_reference_v1: RAS → 0.5/99.5-percentile → pad/crop 180×216×180 → 1 mm → foreground crop margin 4 → resize 100×120×100 → online centre-crop 96×108×96). What differs is what happens before it.

structural/ — the skull-stripped, MNI-aligned arm (T1w, T1ce, T2w, FLAIR, PDw). Offline pipeline, per scan:

Working grid — axes finer than 1 mm are anti-aliased down to 1 mm (coarser axes are left alone; no upsampling of thick slabs). Direction cosines are orthonormalised so ITK accepts every header.
Brain mask — SynthStrip (nipreps, GPU) on every scan. Sources that arrive already stripped are detected by Dice(SynthStrip, source-nonzero) ≥ 0.85 and their own mask is reused, bounded to SynthStrip-brain + 3 vox so retained neck is trimmed.
N4 bias correction, constrained by the brain mask (never shared across contrasts).
Registration — stripped brain to the matching ICBM152 2009c Asymmetric template brain (T1 template for T1w/T1ce; T2 template for T2w/FLAIR/PDw), rigid → 12-DOF affine, masks on all stages. Low-overlap cases get a 24-rotation pose search to correct header/pose disagreement; the better fit is kept.
One resample to template space (image bSpline, mask nearest), then the terminal loader.

native/ — the no-skull-strip comparison arm, run on all 306,207 scans (and the only arm for DWI/b0/SWI/MRA/ASL and other non-structural modalities): canonical RAS, 1 mm physical resample, percentile normalisation, foreground crop, same 100×120×100 cache. No registration, no skull stripping.

The two arms let a downstream study compare whole-head vs skull-stripped-MNI input on identical tensor geometry. Each structural scan keeps its native/ copy too.

Layout

manifest.jsonl              one line per scan: id, dataset, modality, route,
                            status, pack, template_dice, config_version, ...
structural/*.tar           384 shard tars; each holds cache objects
                            (image.npy, foreground.npy, geometry.json) + per-case
                            receipts + brain_mask + affine transforms
structural_harmon/*.tar    reprocessed cases that supersede older-config objects
native/*.tar               96 shard tars + 4 rescue tars (see below)

Each object is content-addressed; geometry.json carries every transform and the cache→native voxel map, so model-space coordinates are invertible to native space.

Status accounting

manifest.jsonl gives every scan a final status. ready = usable; needs_review and needs_profile = produced but flagged; failed/unsupported_geometry = not usable, with the reason recorded. Full counts in RELEASE_REPORT.json.

All 306,207 scans are accounted for (manifest missing_count = 0), no unreadable packs. Final routing:

route	status	count
structural (skull-stripped MNI)	ready	157,418
native (no-strip)	ready	142,753
structural	needs_profile	5,419
structural	needs_review	284
native	needs_review	240
structural	unsupported_geometry	37
native	failed	33
structural	failed	23

300,171 scans (98.0%) are ready on their primary route. Counting the native copy of structural scans that did not pass the structural arm (infant and partial-FOV cohorts kept out of adult MNI by design, plus flagged cases), 305,918 scans (99.91%) have at least one usable image; the manifest row of such a scan carries native_pack. The 289 scans with no usable image are 240 signed-phase GRE images (not anatomical intensity, excluded by design), 33 native failures (DWI volumes concatenated along z, coarse localisers, near-empty volumes) and 16 structural-modality scans that failed on both arms. 9,999 structural scans were reprocessed under the final recipe and supersede their earlier-config objects; 1,764 mid-strip cases were promoted to ready after meeting the template-Dice and retention bars.

Rescued failure classes (each with its evidence in the receipt):

331 ultra-high-res scans (0.2–0.5 mm) that overran the voxel cap — recovered via the 1 mm working grid.
~11,664 GRE/MESE/FLASH/UTE + ASL-M0 parked as "magnitude-or-phase" — released as magnitude by intensity signature (240 signed-phase stay parked).
~2,891 3D ASL frames — released as generic scalar EPI (semantics undetermined).
OASIS2 pose and pre-stripped neck cases — corrected in the structural arm.
6,639 generic-scalar releases (native/rescue-native-generic.v1.tar), processed with per-volume 0.5/99.5-percentile normalisation and tagged so the caveats travel with the data (source_modality keeps the original label):
- 4,799 MP2RAGE_COMPONENT_UNDETERMINED: FOMO repackaging renamed OpenNeuro's inv-1/inv-2/UNI/T1map to generic run-N and renumbered subjects, so the component of each file is unknown. The four MP2RAGE images are mixed, unlabelled.
- 1,802 <map>_UNITS_UNDETERMINED (CBF 971, T1map 435, ADC 159, R1map 105, R2starmap 105, ATT 22, MTRmap 5): units differ across cohorts, so absolute physical values are not preserved; do not use them as quantitative maps. CoRR CBF stores background as -256; that fill was set to 0 before normalisation (recorded per scan). WAND "CBF" maps look like non-quantified ASL/M0 intensity (bright ventricles) rather than calibrated perfusion.
- 38 MP2RAGE_UNI_UNMASKED. This departs from the plan rule that quantitative maps need fixed calibrated bounds; it was chosen deliberately for self-supervised pretraining, where the encoder sees intensity-normalised input either way.

Provenance

Terminal loader: NeuroJEPA reference profile, MONAI 1.5.1, PyTorch CPU.
Template: ICBM152 2009c Asymmetric (TemplateFlow), T1w/T2w/brain-mask at 1 mm, sha256 recorded in RELEASE_REPORT.json.
SynthStrip weights: FreeSurfer synthstrip.1.pt, sha256 37417f80….
Structural recipe: structural_v2_20261017i.

Preprocessing code and audit tooling: see tools/ in RELEASE_REPORT.json notes.

Downloads last month: 73

Total file size:

1.75 TB