The next MRI dataset already exists—across our institutions.
Our CortexMAE work found strict power-law improvement as open fMRI pretraining data grew. Scaling Neuro gives US academic labs a low-friction, institution-governed path to build that shared commons together.
Built by academic neuroimaging researchers behind MindEye, CortexMAE, and Brainmarks
One year can exceed decades of shared data.
Sources + assumptions
UK Biobank 100k · OpenNeuro MRI 50,674 · HCP 1,113. Six centers use BU’s 27,577 MRI exams/year. Capacity comparison only: exams are not participants, and not every exam is brain imaging, consented, or shareable. The dotted curve is a conceptual extension of the strict data-scaling trend reported in CortexMAE, not a performance projection.
One approved folder becomes a continuous contribution.
Point the tool at the folder where DICOMs already land. No labels or manual curation required.
Approve
Only approved projects enter.
Watch
Detect new series automatically.
Strip
Remove identifying metadata locally.
Clear
Process structural faces; hold uncertainty.
Share
Encrypt, transfer, and return access.
A bucket humans can actually understand.
Browse by participant, compare metadata, and inspect a scan instantly. Synthetic public previews · real NIfTI stays local
Functional: clearedConsent + metadata de-identification.
Structural: clearedLocal face processing + privacy check.
Compute: beside dataAvoid needless duplication.
Select one of the synthetic collection previews. Local development can also load a private NIfTI test volume.
Point it at one folder. Let every future scan count.
Keep the scanner workflow. One portable tool clears approved functional and structural scans locally, then shares only what passes.
# the downloadable preview is safe: it reads and uploads nothing
$ bash neuro-sync-preview.sh --deface auto ./new_session_dicoms
✓ Source folder accepted: ./new_session_dicoms
✓ Plan: classify → de-identify → privacy-check → upload
✓ Laptop mode: CPU-only · one series at a time · resumable
i Preview mode only. No files were opened or transmitted.# proposed experience — one file; privacy pack fetched only if needed
$ ./neuro-sync ./my_lab_dicoms --project LAB_04
[device] CPU-only · 1 series at a time · safe to stop and resume
[watch] ./my_lab_dicoms → 4 new series
↑ task-rest_bold EPI 2.0mm 480 vol shared
↑ task-movie_bold EPI 1.6mm 610 vol shared
↑ dwi_dir98 DWI 1.5mm 99 vol shared
↑ T1w_mprage face → processed → QC pass shared
[deid] stripped PHI · kept scanner·field·TR·TE
[up] s3://scaling-neuro/ (encrypted) completeconsent only process projects approved for open research sharing
watch the DICOM directory for newly written series
detect modality and whether structural facial anatomy is present
clear process faces locally; no-face FOVs pass unchanged
verify check face removal and brain preservation
hold anything uncertain stays on this machine
budget one scan at a time; checkpoint every completed series
convert DICOM → research-ready NIfTI
deid remove names, MRNs, dates; keep TR/TE, field
upload to encrypted, redundant S3 (server-side enc.)
rule nothing leaves the building without a privacy pass