README.md

metadata

license: cc0-1.0
pretty_name: >-
  MASiVar sub-cIs1 as a replay phantom: the MT-CSD FOD field, tissue fractions
  and an 84-region parcellation
tags:
  - diffusion-mri
  - replay
  - phantom
  - connectome

MASiVar sub-cIs1 as a compositional replay phantom

The input of the brain Space (rfick/brain-zero, dmrai-lab/disco-space#8): one real scan reconstructed once, offline, by dmipy-fit alone into the arrays a compositional replay phantom composes with replay packs. Nothing here is a diffusion-weighted image: the subject's DWI is not redistributed (it is on OpenNeuro under CC0); this dataset is what multi-tissue CSD made of it, plus a parcellation of the subject's T1.

Source

• Subject sub-cIs1, session ses-s1Ax1, the first scan, of MASiVar (OpenNeuro ds003416, version 2.0.2, DOI 10.18112/openneuro.ds003416.v2.0.2; Cai et al., MASiVar: Multisite, Multiscanner, and Multisubject Acquisitions for Studying Variability in Diffusion Weighted MRI, Magnetic Resonance in Medicine 2021, DOI 10.1002/mrm.28926). Licence of the source: CC0.
• The scan: 3 T, 2.5 mm isotropic, five shells b = 1000 / 1500 / 2000 / 2500 / 3000 s/mm² with 96 directions each and one b = 0 per shell (485 measurements), taken from the dataset's PreQual v1.0.0 derivative (denoised, susceptibility-, eddy- and motion-corrected), the five per-shell files merged. The sidecars record no echo time or pulse timing.
• The T1: ses-s1Ax2, 1 mm.

What is in it

file	content
manifest.json	the subject, the source DOIs and licence, the grid (shape, voxel size, affine), the protocol (b-values, image-frame directions), the FOD basis, the reconstruction record (tool, commit, responses, solver), the parcellation record
fod_wm.npy	(X, Y, Z, 45) float32: the white-matter FOD as real spherical harmonics to order 8 in the image frame, as dmipy-fit's CSD emits them (tournier07 ordering), 0 outside the mask
fractions.npy	(X, Y, Z, 3) float16: the geometric WM / GM / CSF fractions of the multi-tissue fit, clipped to [0, 1]
mask.npy	(X, Y, Z) bool: the brain mask (median-Otsu on the mean b = 0)
labels.npy	(X, Y, Z) int16: the 84 Desikan-Killiany + subcortical regions (MRtrix fs_default order: left 34 cortical then 8 subcortical, then right), 0 elsewhere
regions.json	the 84 regions (id, FreeSurfer label, name, hemisphere, lobe) and the 14 lobar groups (7 lobes × 2 hemispheres)
stop_mask.npy	(X, Y, Z) bool: where a streamline may continue (cerebral and cerebellar white matter, brain stem); leaving it ends the streamline
mean_b0.npy	(X, Y, Z) float32: the mean b = 0 image, the page's background
checks/	msmt_sample.npz, a fixed-seed sample of 2,000 voxels' signal, coefficients and fractions (the reference for dmipy-fit#39's fast solver); sanity_summary.json; an axial slice with the FOD's principal directions

How it was made

• Reconstruction, dmipy-fit only (tools/build_brain_asset.py in dmrai-lab/disco-space): three-tissue response functions estimated from the data (three_tissue_response_dhollander16, Dhollander 2016, selected from the brain mask eroded by 3 passes as MRtrix3's dwi2response dhollander -erode 3, the WM response by Tournier 2013's iterative single-fibre selection), then multi-shell multi-tissue CSD (MultiCompartmentSphericalHarmonicsModel, Jeurissen 2014, solver="csd_msmt_torch": the batched interior-point solver of dmipy-fit#40, every voxel to a KKT tolerance of 1e-13) at order 8 with the responses' S0 folded in, so the fractions are geometric. The whole brain (71,052 voxels × 485 measurements) took 24 s on a CPU. The dmipy-fit commit is in the manifest. Revision 1 of this dataset held the same fit by CsdCvxpyOptimizer (OSQP, 87 minutes on a CPU), whose inexact solutions left negative FOD lobes; revision 2 replaced it with the interior-point solver; revision 3 corrects the response selection: revision 2 selected its WM response from the un-eroded mask by FA, which took 300 noise voxels of FA ≈ 1 at the brain's edge and outside it (131 unlabelled by SynthSeg), so its WM response had 0.77 of white matter's b = 0 and every white-matter voxel's fractions came out 1.2 to 1.3 times too large (median sum 1.24 in cerebral WM). With the eroded mask and Tournier 2013, 282 of the 300 selected voxels lie in cerebral WM and the fit's b = 0 is 0.997 of the measured (IQR 0.992-1.003) in every tissue class.
• The sum of the fractions. Multi-tissue CSD does not constrain WM + GM + CSF to one: each response carries one image-wide b = 0 amplitude while the measured b = 0 varies voxel by voxel (receive field, partial volume, T2), so the sum follows the measured b = 0 (Pearson 0.81 in interior white matter). Median sum 0.90 (IQR 0.80-0.98 in cerebral WM, 0.82-1.04 in cortical GM), 20,177 of 71,052 voxels above one, mostly in the cerebellum (median 1.4, where b = 0 is 1.6-1.8 times the cerebral white matter's) and the brain stem. No smooth field explains it (a cubic polynomial or a 30 mm Gaussian explains 0.3 of the variance of its logarithm), so a consumer that needs shares of a voxel divides by the sum where it exceeds one (the brain Space does, and counts the voxels it did it to), as MRtrix3's mtnormalise would after its smooth correction.
• Parcellation (tools/build_parcellation.py): SynthSeg --parc --robust (Billot et al.) on the T1, a rigid mutual-information registration of the T1 onto the mean b = 0, and a plurality vote of 64 sub-points per diffusion voxel. All 84 nodes are present; per-label volume 578 to 20,656 mm³ (median 4,844).

Use

import json, numpy as np
from huggingface_hub import snapshot_download
d = snapshot_download("SubstrateCommons/masivar-brain", repo_type="dataset")
man = json.load(open(f"{d}/manifest.json")); fod = np.load(f"{d}/fod_wm.npy"); fr = np.load(f"{d}/fractions.npy")

The brain Space composes these with the replay packs of SubstrateCommons/cactus-axons (white matter) and SubstrateCommons/grey-matter-spheres (grey matter) through dmipy-sim's phantom route.

Please cite Cai et al. 2021 and the OpenNeuro dataset when you use this.