DoseRAD2026 Dataset¶

Overview¶

The DoseRAD2026 dataset provides paired multimodal imaging and high-accuracy Monte Carlo–simulated dose distributions for benchmarking fast 3D radiation dose calculation methods.

It supports all four challenge tasks:

• Photon dose calculation on CT
• Photon dose calculation on MRI
• Proton dose calculation on CT
• Proton dose calculation on MRI

All tasks are based on a unified dataset of paired MRI–CT volumes, enabling consistent cross-modality evaluation.

Cohort¶

The training dataset consists of:

• 75 registered MRI–CT pairs
  • 36 abdominal cases
  • 39 thoracic cases

Each case includes:

• A planning MRI volume
• A deformably registered CT volume

MRI–CT pairs are spatially aligned to ensure voxel-level correspondence. Imaging data are provided in 3D .mha format with consistent metadata.

Beam Configurations¶

For each case, multiple independent radiation beams are provided. Each beam must be evaluated separately.

Photon Beams¶

Photon beams are defined by:

• Multi-leaf collimator (MLC) apertures
• Gantry and collimator angles
• Isocenter location
• Beam energy and delivery parameters

Beam shapes are generated from arbitrary MLC configurations representative of VMAT-style treatments.

Proton Beams¶

Proton beams are defined by:

• Source position and angle
• Sets of energies originating from the same source (rays)

Configurations reflect idealized proton delivery scenarios.

Ground Truth Dose¶

For every beam, a corresponding 3D dose distribution is provided.

Ground truth dose distributions are generated using:

• Full-physics Monte Carlo particle transport simulations
• Tissue-dependent interaction modeling

All dose grids:

• Are beam-specific
• Are spatially aligned with CT and MRI volumes
• Serve as the reference standard for evaluation

Data Organization¶

Each case includes:

• CT volume
• MRI volume
• Beam parameter files
• Corresponding Monte Carlo dose distributions

Data are organized in a standardized folder structure compatible with automated evaluation pipelines. Detailed format specifications are provided in the documentation.

Training and Test Split¶

The dataset is divided into:

• A public training set (images, beam definitions, ground truth dose)
• A private preliminary test set for initial evaluation
• A private test set for final evaluation

The private test sets will remain inaccessible to participants to ensure objective benchmarking.