README.md

---
license: cc-by-nc-sa-4.0
pretty_name: DiSCo replay pack
tags:
  - diffusion-mri
  - monte-carlo
  - replay-pack
  - tractography-phantom
---

# DiSCo as a replay pack

The DiSCo substrate (Rafael-Patino, Girard, Truffet, Pizzolato, Caruyer, Thiran, *The diffusion-simulated
connectivity (DiSCo) dataset*, Data in Brief 38 (2021) 107429, doi:10.1016/j.dib.2021.107429; data
doi:10.17632/fgf86jdfg6.3) walked once and stored as a replay pack. Any acquisition a human scanner can play is
then a replay of the same walk, voxel by voxel on the dataset's own 40³ grid of 25 µm voxels, and every replay
carries a Monte-Carlo certificate.

DiSCo published one acquisition of this substrate. This dataset is the substrate itself in replayable form:

- The same 12,196 strands, two tubes per strand: the listed inner diameter, and the outer tube at 1/0.7 of it.
- The dataset's diffusivity, 0.6 × 10⁻⁹ m²/s, in both pools: intra water inside the inner tube and extra water
  outside the outer one.
- 321,138,027 walkers walked for 100 ms on 31,815 of the 64,000 voxels; the rest hold no water.
- Every tier stored: positions in 256 bands, occupancy, wall contact (8 bands), and the myelin sheath's field
  along each path (12 channels, 32 modes).

The time-segment layout of the same walk, at DiSCo's own seven timing classes, is a separate dataset:
[SubstrateCommons/disco-time-segments](https://huggingface.co/datasets/SubstrateCommons/disco-time-segments).
It is smaller and faster, but answers only those timings.

## What the pack is certified for

This pack is certified for **any acquisition in the class below**, not for a list of protocols.

| | |
|---|---|
| Scanner class | Connectome 2.0: G_max 0.5 T/m, slew 600 T/m/s |
| Waveforms | any within the class: PGSE, OGSE, stimulated echo, arbitrary; up to 16 refocusing pulses; TE ≤ 100 ms |
| Relaxation | T₂ and T₁ per pool, surface relaxivity ρ₂ ≤ 3 µm/s |
| Field | B₀ of 1.5, 3 or 7 T in any direction, myelin χ up to 2 × 10⁻⁷ (isotropic and anisotropic) |
| Floor | 1.003 %: the per-voxel standard error that the walker count was planned for |
| Numerical error | ≤ 1.66 × 10⁻³ in the complex signal, under half the floor: truncation, quadrature, field read and quantisation together |

The certificate lives in `disco/manifest.json` under `meta.fidelity`, term by term. It is inherited from an
admission walk measured on the same engine, recipe and container (`inherited_from`). The per-voxel floor is
measured from each voxel's own walkers at replay.

## Use me

Install dmipy-sim from its main branch, where the reader lives. The dataset is public, so no login is needed:

```
pip install "git+https://github.com/dmrai-lab/dmipy-sim@main" "huggingface_hub>=0.25"
```

Open the pack by reference. Nothing is downloaded whole: every byte read is one the replay uses, fetched as HTTP
range reads of the columnar layout under `disco/`.

```python
import numpy as np
from dmipy_sim.replay import ReplayPack
from dmipy_sim import sequences
from dmipy_sim.spec.tissue import Tissue

pack = ReplayPack.open("hf://SubstrateCommons/disco-replay/disco")
seq = sequences.pgse([[1, 0, 0], [0, 0, 1]], 0.0102, 0.0167, bvalues=[1e9, 3.094e9], TE=0.0535)   # SI: s/m², s

print(pack.plan(seq))                                       # bands, tiers and bytes, decided before any transfer
view = pack.view(K=32, voxels=[(20, 20, 20)])               # one voxel's rows at 32 bands: an ordinary ReplayPack
S = view.replay(seq)                                        # bare diffusion, the dataset's own physics

wm = Tissue(T2={"intra": 0.05, "extra": 0.055, "myelin": 0.01}, rho2=1e-6, chi_iso=-1e-7, chi_aniso=-1e-7)
view = pack.view(K=32, modes=8, contact=True, voxels=[(20, 20, 20)])   # the tiers this physics needs
S_3T = view.replay(seq, tissue=wm, scanner=3.0)             # relaxation, wall contact and the sheath field at 3 T
```

`pack.plan(seq)` says how many bands and field modes an acquisition needs, and so how many bytes it reads. A
clinical scanner reads only a prefix of the bands.

## Files

| path | what | size |
|---|---|---|
| `disco/manifest.json` | the pack's header: substrate, walk, envelope, codec, certificate, provenance | 0.8 MB |
| `disco/index.json` | the rows of each voxel and pool | 3.4 MB |
| `disco/columns/` | the columns, split into parts of ≤ 1.3 GB | 294 GB |
| `substrate/` | the strand files the walk read, and their source note | |
| `disco/reference/DiSCo_gradients.*` | DiSCo's own gradient table, from which the time-segment classes were taken | |

## How it was made

- **Walk:** the CUDA backend of dmipy-sim (commit `3b1207b`). Blocks of 99.5 µs with adaptive sub-steps, own-tube
  wall interaction, and the strand field read every 33 µs. The tiers were accumulated in the walk as band
  projections; no trajectory was stored.
- **Planning:** walker counts per voxel and pool from a pilot at the class envelope, so that each voxel reaches the
  floor under any combination of the tiers.
- **Assembly:** 704 shards, merged in block order into one pack.
- **Checks:** the merged pack was replayed against the time-segment layout at 16 voxels and the seven timing
  classes. The worst disagreement was 2.2 × 10⁻⁵, against a combined bound of 3.0 × 10⁻³.
- **Earlier version:** the previous revision of this dataset holds the first DiSCo walk, a JAX walk on a save grid
  with fewer certified tiers. It stays in the history.

## Licence

The replay pack is distributed under **CC BY-NC-SA 4.0**, or a commercial licence on request
(`pack_license: CC-BY-NC-SA-4.0 OR LicenseRef-Commercial` in the header). The DiSCo substrate it was walked in
is the DiSCo authors' work, CC BY 4.0 (`license` in the header). Cite both.

## Citation

- Rafael-Patino J, Girard G, Truffet R, Pizzolato M, Caruyer E, Thiran J-P. The diffusion-simulated connectivity
  (DiSCo) dataset. Data in Brief 38 (2021) 107429.
- Fick RHJ. Replayable Monte Carlo MRI: from persistent tissue microphysics to queryable virtual phantoms (in preparation, 2026). dmipy-sim, https://github.com/dmrai-lab/dmipy-sim.