DoseRAD2026: Paired CT/MRI Beam-Level Monte Carlo Radiotherapy Dose Dataset: sample notebook
A worked notebook for DoseRAD2026: Paired CT/MRI Beam-Level Monte Carlo Radiotherapy Dose Dataset, with 5 code cells and the output each one printed. On Scigantic it opens in a hosted JupyterLab with the dataset already mounted, so nothing is downloaded first.
- Provider
- Fan Xiao, Nikolaos Delopoulos, Niklas Wahl, et al., LMU Munich Department of Radiation Oncology (LMUK-RADONC-PHYS-RES)
- License
- cc-by-nc-4.0
- Size
- 100,000 files
- Format
- JSON (in sampled files)
- Stored at
- Hugging Face (https://huggingface.co/datasets/LMUK-RADONC-PHYS-RES/DoseRAD2026)
- Access
- Free and public. No account with the data provider is needed.
- DOI
- 10.5281/zenodo.19347848
The notebook
DoseRAD2026: Getting Started
A row here is not a table row, it's a beam. Each of the 75 public training patients ships a CT/MRI pair, a JSON beam configuration, and one Monte Carlo dose grid (.mha) *per beam unit*: 540 photon control-point doses or 1,080 proton beamlet doses per patient. This mirrors the DoseRAD2026 challenge task: predict one of those dose grids straight from CT or MRI.
Nothing below is downloaded. The archive is already FUSE-mounted at /mnt/archive.
import os
root = os.environ.get("SCIGANTIC_MOUNT_PATH", "/mnt/archive")
training_dir = f"{root}/photon/training"
patients = sorted(p for p in os.listdir(training_dir) if os.path.isdir(f"{training_dir}/{p}"))
abdominal = [p for p in patients if "ABB" in p]
thoracic = [p for p in patients if "THB" in p]
print(f"{len(patients)} training patients: {len(abdominal)} abdominal (ABB), "
f"{len(thoracic)} thoracic (THB)")
patients[:5]
Output:
75 training patients: 36 abdominal (ABB), 39 thoracic (THB) ['1ABB006', '1ABB011', '1ABB020', '1ABB021', '1ABB030']
Beam configuration: parsed straight from JSON
Every patient has a .json describing the treatment plan. For photon it's VMAT arcs (Elekta Versa HD, 80-leaf-pair Agility MLC): a list of beams, each with up to 180 control points sweeping the gantry angle.
import json
patient = "1ABB006"
with open(f"{root}/photon/training/{patient}/{patient}.json") as f:
cfg = json.load(f)
beam0 = cfg["beams"][0]
print(f"{len(cfg['beams'])} beams | beam 0: SAD={beam0['SAD']}mm, "
f"{len(beam0['control_points'])} control points, "
f"gantry {beam0['control_points'][0]['gantry_angle']} -> "
f"{beam0['control_points'][-1]['gantry_angle']} deg")
print(f"MLC leaf pairs: {beam0['num_mlc_leaf_pairs']}")
Output:
3 beams | beam 0: SAD=1000mm, 180 control points, gantry -180 -> 178 deg MLC leaf pairs: 80
CT + one control point's Monte Carlo dose
Dose filenames encode the answer key: Dose_B{beam_idx}_CP{cp_idx}.mha maps 1:1 onto beams[beam_idx].control_points[cp_idx] above. A single file is one control point's dose alone, not the summed, weighted plan dose, so per-voxel values are small (a few hundredths of a milligray) compared to a clinical fraction.
import SimpleITK as sitk
ct = sitk.ReadImage(f"{root}/photon/training/{patient}/image/ct.mha")
dose = sitk.ReadImage(f"{root}/photon/training/{patient}/dose/Dose_B0_CP000.mha")
ct_arr, dose_arr = sitk.GetArrayFromImage(ct), sitk.GetArrayFromImage(dose)
print(f"CT: shape={ct_arr.shape}, spacing={ct.GetSpacing()}mm, "
f"HU range=[{ct_arr.min():.0f}, {ct_arr.max():.0f}]")
print(f"dose: shape={dose_arr.shape} (same grid as CT), "
f"peak={dose_arr.max():.2e} Gy for this one control point")
Output:
CT: shape=(246, 246, 249), spacing=(2.0, 2.0, 2.0)mm, HU range=[-1024, 2834] dose: shape=(246, 246, 249) (same grid as CT), peak=5.83e-05 Gy for this one control point
Overlay the slice with the most dose signal
import numpy as np
import matplotlib.pyplot as plt
z = dose_arr.sum(axis=(1, 2)).argmax()
fig, ax = plt.subplots(figsize=(6, 6))
ax.imshow(ct_arr[z], cmap="gray", vmin=-200, vmax=400)
dose_slice = np.ma.masked_less(dose_arr[z], dose_arr.max() * 0.03)
ax.imshow(dose_slice, cmap="jet", alpha=0.6)
ax.set_title(f"{patient}, photon beam 0, control point 0, slice {z}")
ax.axis("off")
plt.show()
Proton is a separate story, not a variant
proton/training/{patient}/ has its own independently-resampled CT/MRI pair (1x1x3mm³, not photon's 2x2x2mm³, don't mix the two) and dose files named Dose_B{beam}_R{ray}_L{beamlet}.mha, one per pencil-beam energy layer. Beamlet energies (31.7-200.8 MeV) join against proton/training/beam_parameters.json's energy_table by exact float value.
with open(f"{root}/proton/training/{patient}/{patient}.json") as f:
proton_cfg = json.load(f)
ray0 = proton_cfg["beams"][0]["rays"][0]
print(f"proton beam 0, ray 0: {len(ray0['beamlets'])} beamlets, "
f"energies {[round(b['energy'], 1) for b in ray0['beamlets']]} MeV")
Output:
proton beam 0, ray 0: 2 beamlets, energies [45.6, 126.1] MeV
Next steps
- Only the 75-patient training split (864 GB) is mounted here. The preliminary-test (40 patients) and external-test (7 patients) splits stay private until March 2030, so submit predictions on grand-challenge.org rather than expecting to find them mounted.
- To build a full-beam dose, sum a beam's control points (photon) or a ray's beamlets (proton); this dataset ships the unweighted per-unit contributions, not the plan sum.
- License is CC BY-NC 4.0: research use only.
- Ask the AI panel to walk any other patient, beam, or the HU-to-density calibration table in beam_parameters.json.
Run it
Open this notebook to run it, see the dataset page for the full description, or browse the full catalog.