disco/reference/README.md

# DiSCo reference volumes, replayed from the pack

Pre-replayed images of the DiSCo substrate on the dataset's own 40^3 grid (25 um voxels; the NIfTI affine is
the identity, as in the dataset's files, so these overlay `DiSCo_DWI.nii.gz` directly). Every volume is S/S0
per voxel in float32, 0 where the pack holds no walkers (the free-water voxels farther than 46 um from every
strand, and the volume outside the strands' reach). `disco_replay_floor.nii.gz` is the certificate's per-voxel
Monte-Carlo floor beside them (pass 1: median 0.035 on the mask).

| file | what | tissue | bands / modes |
|---|---|---|---|
| `disco_replay_bare.nii.gz` | DiSCo's 364-measurement protocol, bare diffusion (the dataset's own simulation) | none | 64 / 0 |
| `disco_replay_3T.nii.gz` | the same protocol with white matter at 3 T | T2 50/55 ms, rho 1.16 um/s, chi -0.1 ppm | 64 / 32 |
| `disco_replay_7T.nii.gz` | the same at 7 T | T2 47/47 ms, rho 1.16 um/s, chi -0.1 ppm | 64 / 32 |
| `disco_acq_clinical.nii.gz` | b = 1000, 30 directions, delta/Delta 12/24 ms | none | plan |
| `disco_acq_research.nii.gz` | b = 1000/2000/3000, 90 directions each, 17/30 ms | none | plan |
| `disco_acq_c2.nii.gz` | b = 6000, 60 directions, 8/20 ms (275 mT/m) | none | plan |
| `disco_acq_ogse.nii.gz` | cosine OGSE 50 Hz, b = 1000, 60 directions | none | plan |

The protocol is the dataset's as it simulated it: one echo time of 53.5 ms, ideal pulses, the three upper
shells at the ActiveAx ex vivo timings (10.2/16.7, 7.6/45.9, 17.7/35.8 ms; Alexander et al. 2010), the
b = 1000 shell at the first shell's timing (the dataset does not state it). Gradient tables are beside each
volume (`.bvals` in s/mm^2, `.bvecs` as 3 x N). The tissue values are the biophysical catalogue's
(dmipy-sim `canonical_white_matter`); the volumes are one tissue choice, not a ground truth.

`*_timing.json` records the plan each volume was read with (bands, field modes, tiers, bytes per row, bytes
and seconds streamed from this dataset's `disco/` layout). `disco_replay_compare.json` is the bare volume
against the dataset's noise-free images (per shell: median and rms |dS/S0|, correlation) -- the replay is
higher by ~1 % of S0 on every shell because the dataset's tubes are triangle meshes inscribed in the nominal
circles (their intra volume is 0.953 of the pack's); the residual scatter is pass 1's floor. The layout these
were read from includes the repair of dmipy-sim#295: 332 voxels whose pool the plan had clipped at two walkers
(their certified floors ran to 1.0; the top shell's correlation with the dataset was 0.84 before, 0.98 after)
were rewalked at (share / 0.008)^2 walkers and appended with union weights (`columnar.repair` in the manifest). `tract/`
holds the scoring records: CSD + probabilistic tracking + the 16 x 16 streamline-count matrix against the
dataset's connectivity matrices (its own images through the same pipeline score 0.904 noise-free and 0.905 at
its SNR-30 release; the replayed protocol at SNR 30 scores 0.912 +/- 0.003 over three seeds).

Regenerated after every pass of the fill; the pack manifest these were read from is recorded in each timing
record's `uri` and in `manifest_sha256`.

manifest_sha256: 294dd9af218d8bb06728aed0a0e4d71e174415b3185a0c01b983b4ecf7c4e946