All primary datasets are freely downloadable with no institutional affiliation required. Large datasets should be cached in S3 after initial download rather than stored in the git repository.
The primary brain reference is Lüsebrink 2021 (in-vivo 450 µm T2 SPACE), not MGH ds002179. MGH is retained as an optional dataset for future cortical-ribbon work — see ADR-001 in decisions.md for the rationale.
Single-subject ultrahigh-resolution in-vivo MRI from Lüsebrink et al. 2021, acquired at 7T as a "human phantom" reference. T2 SPACE at 0.45 mm isotropic gives bright-CSF contrast across the full brain with physiologically intact subarachnoid space — what MGH cannot provide.
| Field | Value |
|---|---|
| Resolution | 0.45 mm isotropic (raw T2 SPACE); T1w MP2RAGE also available |
| Coverage | Whole brain, in vivo, physiologically intact CSF |
| Format | NIfTI (BIDS-organized) |
| Subject | sub-yv98 (the author), session ses-3777 (primary T2 SPACE session) |
| Size (pipeline subset) | ~325 MB — raw T1w + T2w + bias-corrected derivatives + sidecars |
| Access | Free, no account required; public s3://openneuro.org/ds003563/ |
| Download | OpenNeuro ds003563 |
| Paper | Lüsebrink et al., Scientific Data 2021 (Open Access) |
Use in pipeline:
- Phase 2 (segmentation):
sub-yv98_T2w_biasCorrected.nii.gzis the default SynthSeg input. Bias-corrected and contrast-agnostic segmentation recovers ventricles, extraventricular CSF, and cortical structures. - Phase 4 (manual refinement): the raw T2 SPACE volume is the primary reference image in Slicer. Bright-CSF contrast makes cisterns, foramina, and the cerebral aqueduct dramatically easier to trace than on a T1-weighted volume.
- Phase 8 (LBM microstructure): in-vivo sulcal SAS geometry provides defensible boundary widths for the raycast-based SAS-width sampling. Fine-structure permeability calibration still comes from Rossinelli SRµCT, not from the MRI.
Pipeline file layout after Phase 1 download:
raw/lusebrink_2021/sub-yv98/anat/
├── sub-yv98_T1w.nii.gz # 89 MB raw MP2RAGE (co-registered reference)
├── sub-yv98_T1w.json
├── sub-yv98_T1w_biasCorrected.nii.gz # 90 MB bias-corrected T1w
├── sub-yv98_T2w.nii.gz # 71 MB raw 450 µm T2 SPACE
├── sub-yv98_T2w.json
└── sub-yv98_T2w_biasCorrected.nii.gz # 73 MB SynthSeg input (brain_input default)
Session id is stripped from filenames during upload so downstream paths key only on the subject id.
The standard open-access dataset for spinal cord MRI, with the PAM50 standardized spinal cord + canal atlas.
| Field | Value |
|---|---|
| Subjects | 260 healthy adults across 42 centres (multi-subject); for single-subject, ~20 "subjects" are the same person at different scanners |
| Atlas | PAM50 template: spinal cord + CSF canal segmentations, C1 to S5 |
| Format | NIfTI (BIDS-organized), distributed via git-annex |
| Access | Free, no account required |
| Download | Single subject (GitHub + git-annex) · Multi-subject (GitHub) |
| Paper | Cohen-Adad et al., Scientific Data 2021 (Open Access) |
Important — Zenodo caveat: The Zenodo mirror contains only a 215 KB zip of git-annex pointer files, not the MRI data itself. Use the GitHub repo + git annex get (the download_spine_generic.sh script handles this). The git-annex remotes computecanada-public and amazon-private hold the actual blobs.
Use in pipeline: Phase 3 input for spinal canal segmentation. Canal mask minus cord mask = spinal subarachnoid space. Default subject: sub-douglas.
14 healthy adults imaged with CISS, DESS, and T2-TSE sequences targeting the cauda equina and lumbosacral nerve roots.
| Field | Value |
|---|---|
| Coverage | Lumbosacral spine (L1-S2), spinal SAS, nerve roots, dura |
| Sequences | CISS, DESS, T2-TSE (CSF-optimized) |
| Access | Open access (Scientific Data / CC licence) |
| Download | Paper + Data |
| Code | GitHub (Blender scripts) |
Use in pipeline: Supplements the Spine Generic dataset for the L1-S2 region where the thecal sac and cauda equina are difficult to resolve.
The highest spatial-resolution whole-brain MRI dataset available (post-mortem, 7T, 100 µm isotropic). Not the default brain input — ex-vivo fixation drains the subarachnoid space and collapses sulci, so its "CSF" compartment is a deformation artifact rather than a physiological fluid volume. See ADR-001 in decisions.md.
Retained in the dataset catalogue because its 200 µm cortical-ribbon detail will matter for future OCT-catheter SLAM validation work and any cortical-surface morphometrics where the pial boundary geometry needs to be as tight as possible.
| Field | Value |
|---|---|
| Resolution | 100 µm isotropic native; 200 µm and 500 µm derivatives included |
| Coverage | Whole brain, ex vivo (no spinal cord); one subject (sub-EXC004) |
| Format | NIfTI (.nii.gz) |
| Size | ~95 GB total; ~3.2 GB for the 200 µm derivatives + 74 MB 500 µm MNI quick-test |
| Access | Free, no account required; public s3://openneuro.org/ds002179/ |
| Download | OpenNeuro ds002179 |
| Paper | Edlow et al., Scientific Data 2019 (Open Access) |
How to download (ad-hoc — not in default Phase 1):
# Submit directly to Batch with a separate run_id:
aws batch submit-job --profile neurobotika --region eu-central-1 \
--job-name mgh-download-ad-hoc \
--job-queue neurobotika-cpu \
--job-definition neurobotika-download-mgh \
--parameters subject=sub-EXC004,s3_dest=s3://neurobotika-data/runs/mgh-ref-only/raw/mgh_100umOr run locally:
./pipeline/01_data_acquisition/run_downloads.sh \
--dataset mgh --subject sub-EXC004 \
--s3-dest s3://neurobotika-data/runs/mgh-ref-only/raw/mgh_100umThese datasets are not used for anatomy reconstruction but as quantitative validation targets for the Phase 8 LBM microstructure simulation pipeline.
Cervical CSF velocity waveforms measured with 4D phase-contrast MRI in healthy subjects. Provides the independent test tier targets for the Level 2 Brinkman simulation.
| Field | Value |
|---|---|
| Subjects | 10 healthy adults |
| Measurement | CSF velocity waveforms at C2–C3, C5–C6 (anterior and posterior SAS) |
| Key data | Peak cervical velocity 2–5 cm/s; anterior-posterior ratio 1.5–3×; phase lag ~40–60 ms/vertebral level |
| Access | Open access |
| Paper | Yiallourou et al., PLoS ONE 2012 |
Use in pipeline (Phase 8): Independent test tier — Level 2 Brinkman simulation outputs are compared against these waveforms as the final validation step, never used during parameter calibration. Target: NRMSE < 0.2 against published waveform data.
High-resolution SRμCT morphometry and DNS of CSF dynamics in the optic nerve subarachnoid space (ONSAS). The primary source for quantitative calibration and validation targets.
| Field | Value |
|---|---|
| Resolution | 1.625 μm/pixel (SRμCT-derived geometry) |
| Key morphometry (2023) | Trabecular thickness PDF (peak 40–60 μm, max 200 μm); separation PDF; volume fraction ~35%; surface area amplification 3.2–4.9× |
| Key DNS results (2024) | Pressure gradient 0.37–0.67 Pa/mm for 0.5 mm/s flow; exponential κ–VF scaling; 17× mass transfer decrease without microstructure |
| Access | Open access |
| Papers | Rossinelli et al. 2023, Fluids Barriers CNS · Rossinelli et al. 2024, Fluids Barriers CNS |
Use in pipeline (Phase 8): Calibration tier — thickness/separation PDFs are the quantitative morphometric targets (V6a/b, Wasserstein distance). DNS pressure gradient is the V5 target. Mass transfer amplification is the V7 target. κ–VF exponential scaling verified via V8 sweep-level fit.
20 um histological 3D reconstruction. Useful as a cross-validation reference for tissue boundaries adjacent to CSF spaces.
| Field | Value |
|---|---|
| Resolution | 20 um histological; volumetric at 100-400 um (MNI-registered) |
| Format | NIfTI, MINC, STL, gii |
| Access | Free (non-commercial research/education) |
| Download | BigBrain Project or EBRAINS |
These datasets are large. Do NOT commit them to the git repository. The recommended workflow:
- Download to
data/raw/locally for development - Upload to S3 (
s3://neurobotika-data/raw/) for persistence and sharing - Pipeline scripts support both local paths and S3 URIs
The data/ directory is in .gitignore. See architecture.md for the full S3 bucket layout.
Phase 8 validation data: The Yiallourou 2012 extracted waveform data and Rossinelli 2023 digitized PDFs (once available as .npy files) should be stored at:
s3://neurobotika-data/validation/
├── yiallourou_2012_waveforms.npy # 4D PC-MRI velocity waveforms (independent test)
├── rossinelli_2023_thickness_pdf.npy # thickness PDF (calibration target V6a)
└── rossinelli_2023_separation_pdf.npy # separation PDF (calibration target V6b)
Paths are referenced via reference_pdf_path / reference_cld_path in pipeline/08_microstructure_generation/config.yaml.