zimmeryWo/MC_proton_simulation_DoseRAD2026

Water-Phantom Proton Monte Carlo (DoseRAD2026)

Full 3D Monte-Carlo energy-deposit distributions for 85 proton energies in a water phantom, 10⁹ primaries each. This is the reference data behind the machine look-up table of team DoseHappens's DoseRAD2026 Grand Challenge entry (submitted algorithm codename MAALGO): the analytic pencil-beam engine's depth-dose and lateral-spread curves are fitted to these volumes, and then refined by backpropagating the dose engine against them.

Code that consumes this dataset: https://github.com/LFetty/doserad2026-maalgo

Contents

Energies	85, from 31.7290 to 200.7966 MeV
Primaries per energy	10⁹
Volume grid	1500 × 500 × 500 voxels at 0.2 mm isotropic = 300 × 100 × 100 mm
Voxel type	MET_DOUBLE (float64)
Size per energy	3.0 GB (2.79 GiB)
Total	~238 GiB, 235 files

Each energy contributes:

1e9_<energy>MeV__edep.mhd    # MetaImage header (428 B)
1e9_<energy>MeV__edep.raw    # the volume (3.0 GB)
1e9_<energy>MeV_Statistics.txt   # GATE run statistics (JSON)

Note the double underscore before edep — the loaders match on it. Statistics files are present for 65 of the 85 energies.

The stored quantity is energy deposit per voxel (GATE DoseActor, edep), not dose. For a uniform water phantom the conversion is a single global scalar (divide by the voxel mass), which is what the fitting code does.

Simulation setup

Simulated with OpenGATE (GATE 10) on Geant4.

• Phantom — a G4_WATER box inside a 1 m³ G4_AIR world.
• Beam — protons along the phantom's long axis. Each energy was simulated with that energy's own machine parameters taken from the challenge beam model (beam_parameters.json): a Gaussian energy spread and a Gaussian spot size, both varying with energy. For example 31.7290 MeV used energy_sigma = 6.1130 MeV and spot_sigma = 7.4934 mm. These are therefore not bare pencil beams — the machine's energy spread and finite spot are baked into every volume, which is what makes them the right reference for fitting a clinical machine LUT.
• Scorer — DoseActor attached to the phantom, edep active, hit_type = "random", edep uncertainty off.
• Cost — these are long runs; the 101.9976 MeV job recorded 18.1 days of simulation time for its 10⁹ primaries.

Geometry

From the MetaImage headers:

DimSize       = 1500 500 500
ElementSpacing= 0.2 0.2 0.2
Offset        = -149.9 -49.9 -49.9
ElementType   = MET_DOUBLE

The 1500-voxel axis is depth — 300 mm of water along the beam. The two 500-voxel axes are the lateral plane, centred on the beam axis (±50 mm). Readers that return [z, y, x] arrays (SimpleITK, for instance) will therefore see depth as the last numpy axis.

Usage

Both consumers glob *__edep.mhd and parse the energy out of the filename, so point them at a directory laid out exactly as this repo:

hf download zimmeryWo/MC_proton_simulation_DoseRAD2026 \
    --repo-type dataset --local-dir mc_1e9

# stage 1 — analytic per-energy fit of depth dose and the double-Gaussian lateral shape
python scripts/export_proton_lut_fast.py \
    --edep-dir mc_1e9 \
    --double-fit-mode direct --kernel-width-mm 74 \
    --output-mat-path lut_fast_3d_1e9.mat

# stage 2 — differentiable calibration: backpropagate the dose engine against the same MC
python scripts/optimize_lut_water.py --all \
    --mc-dir mc_1e9 \
    --base-lut lut_fast_3d_1e9.mat \
    --out-lut  lut_fast_3d_1e9_opt.mat

scripts/benchmark_pb_vs_mc_water.py --mc-dir mc_1e9 scores the analytic engine against these volumes per energy (MAE, peak error, local gamma).

Downloading a subset

At ~238 GiB, you probably want a few energies rather than all 85:

hf download zimmeryWo/MC_proton_simulation_DoseRAD2026 --repo-type dataset \
    --include "1e9_164.4532MeV*" --local-dir mc_subset

The fitting scripts run per energy, so a subset is enough to reproduce individual curves; only a full LUT rebuild needs the whole set.

Related

• Code — https://github.com/LFetty/doserad2026-maalgo
• MR→synthetic-CT model — https://huggingface.co/zimmeryWo/MRI-sCT_converter-DoseRAD2026

Citation

If you use this dataset, please cite the DoseRAD2026 challenge and link this repository.