arXiv:2604.12778v1 [physics.med-ph] 14 Apr 2026
DoseRAD2026 Challenge dataset: AI accelerated photon and proton dose calculation for radiotherapy Fan Xiao1 , Nikolaos Delopoulos1 , Niklas Wahl2,3 , Lennart Volz4 , Lina Bucher5,6,2,3 , Matteo Maspero7 , Miguel Palacios8 , Muheng Li9 , Samir Schulz5,10,2,3 , Viktor Rogowski11 , Ye Zhang9 , Zoltan Perko12,13 , Christopher Kurz1 , George Dedes1,14 , Guillaume Landry1,15,16 , Adrian Thummerer1,17,18,a 1
Department of Radiation Oncology, LMU University Hospital, LMU Munich, Munich, Germany Division of Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ), Heidelberg, Germany 3 National Center for Radiation Research in Oncology (NCRO), Heidelberg Institute for Radiation Oncology (HIRO), Heidelberg, Germany 4 GSI Helmholtzzentrum für Schwerionenforschung, Biophysics Department, Darmstadt, Germany 5 Department of Physics and Astronomy, Heidelberg University, Heidelberg, Germany 6 Institute of Biomedical Engineering, Karlsruhe Institute of Technology, Baden-Württemberg, Germany 7 Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands 8 Department of Radiation Oncology, Amsterdam UMC location Vrije Universiteit Amsterdam, Cancer Center Amsterdam, Cancer Treatment and Quality of Life, Amsterdam, Netherlands 9 Center for Proton Therapy, Paul Scherrer Institute (PSI), Villigen, Switzerland 10 Department of Radiation Oncology, University Hospital Heidelberg, Heidelberg, Germany 11 Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden 12 Delft University of Technology, Department of Radiation Science and Technology, Delft, Netherlands 13 Radformation Inc., New York, US 14 Department of Medical Physics, Faculty of Physics, Ludwig-Maximilians-Universität München, Munich, Germany 15 German Cancer Consortium (DKTK), partner site Munich, a partnership between DKFZ and LMU University Hospital, Germany 16 Bavarian Cancer Research Center (BZKF), Munich, Germany 17 Department of Radiation Oncology, Inselspital, Bern University Hospital and University of Bern, Bern, Switzerland 18 Department of Digital Medicine, Faculty of Medicine, University of Bern, Bern, Switzerland. 2
a
Author to whom correspondence should be addressed: [email protected]
i
Abstract Purpose: Accurate dose calculation is a fundamental component of modern radiotherapy, enabling precise tumor irradiation while sparing surrounding healthy tissue. With the increasing adoption of magnetic resonance imaging (MRI)-guided radiotherapy and growing interest in real-time adaptive radiotherapy, there is a demand for fast and accurate dose calculation techniques on both computed tomography (CT) and MRI. To facilitate research in this area, the DoseRAD2026 dataset and accompanying challenge introduce a publicly available benchmark dataset comprising paired CT and MRI data with corresponding beam-level photon and proton Monte Carlo (MC) dose distributions, enabling the development and systematic evaluation of advanced dose calculation and prediction methods. Acquisition and validation methods: The DoseRAD2026 dataset comprises paired CT and MRI data from 115 patients (75 training, 40 testing) treated on a hybrid MRI-linear accelerator (MRI-linac) for thoracic or abdominal lesions and is derived from the SynthRAD2025 dataset. Image pre-processing included deformable image registration, air-cavity correction and resampling. Ground-truth photon (6 MV) and proton dose distributions were generated using open-source MC dose calculation algorithms and resulted in a training dataset of 40,500 photon beams and 81,000 proton beamlets. Data format and usage notes: The dataset is organized into photon and proton subsets, each comprising paired CT-MRI images, beam-level dose distributions, and JSON files specifying patient-specific beam configurations. Data are provided in compressed MetaImage (.mha) format with standardized naming linking dose files to beam parameters in the JSON files. The dataset is released under a CC BY-NC 4.0 license, with the training set available from April 2026 and the test set withheld until March 2030. Potential applications: The DoseRAD2026 dataset enables benchmarking and development of fast dose calculation methods for radiotherapy applications. Potential use cases include rapid beam-level dose estimation for photon and proton therapy, MRI-based dose calculation in MRI-only and MRI-guided radiotherapy workflows, and support for real-time adaptive radiotherapy strategies.
Keywords: Dose calculation, Dose prediction, Deep learning, MRI-only
ii
I. INTRODUCTION
I.
Introduction
Radiation therapy is indicated for approximately half of all cancer patients, spanning curative, adjuvant, and palliative contexts 1,2 . Photon therapy, such as volumetric modulated arc therapy (VMAT) 3 , enables highly conformal dose distributions around the target volume through continuous gantry rotation combined with dynamic multileaf collimator (MLC) shaping, while sparing surrounding organs at risk (OARs). Proton therapy, such as intensity modulated proton therapy (IMPT) 4 , offers a physically distinct advantage through the Bragg peak 5 , whereby the majority of energy is deposited at a finite tissue depth, affording superior sparing of OARs distal to the tumor. In both modalities, the accuracy of dose calculation is foundational to treatment plan quality 6 . As workflows increasingly move toward online (immediately prior to radiation delivery) 7,8 and ultimately real-time (continuously during radiation delivery) adaptive 9,10 , the computational demand placed on dose engines has grown substantially. Monte Carlo (MC) simulation 11 on high quality computed tomography (CT) is broadly accepted as the reference standard for dose calculation, owing to its accurate physics modelling of particle transport. However, general-purpose MC implementations suffer from long computation times, with plan-level runtimes spanning hours to tens of hours on a single central processing unit (CPU) 12 . Graphics processing unit (GPU) acceleration has substantially narrowed this gap for dose calculations, with recent GPU-based MC engines reporting plan-level runtimes in the range of a few to tens of seconds 13–15 . In parallel, deep learning (DL) methods have emerged as a complementary strategy, demonstrating sub-second dose prediction and showing early promise at MC-comparable accuracy in both proton 16–19 and photon 20–23 therapy. Nevertheless, achieving MC-level accuracy and the computational speed required for online or real-time plan adaptation—ideally in the sub-second or millisecond time-frame—remains an open challenge. Magnetic resonance imaging (MRI) offers superior soft-tissue contrast and eliminates ionizing radiation exposure compared to CT 24 , making it valuable for target and OAR delineation as well as motion management in online adaptive radiotherapy. The clinical introduction of hybrid magnetic resonance imaging-linear accelerator (MRI-linac) systems 25,26 has enabled high-quality soft-tissue imaging prior to and during photon treatment delivery. Meanwhile, prototype systems integrating proton delivery with in-room MRI, as well as MRI-
II. ACQUISITION AND VALIDATION METHODS only proton workflows 27,28 , have also been proposed. However, since MRI does not provide electron density or material composition information, dose calculation in both modalities continues to rely on CT. Current clinical practice bridges this gap via synthetic CT (sCT) generation 29,30 , but this intermediate step introduces additional accuracy potential sources of degradation (difference between sCT and CT), computational latency, and overhead to the dose calculation pipeline, remaining a bottleneck for true real-time plan adaptation. Direct DL-based dose prediction from MRI offers a conceptually cleaner alternative, collapsing the sCT intermediate into a single end-to-end inference step. Early investigations for both photon 31,32 and proton 33,34 modalities have demonstrated dosimetric feasibility, yet progress has been constrained by the absence of large-scale paired datasets providing beam-level dose ground truth on both CT and MRI. To address this gap, we introduce the DoseRAD2026 dataset, the first publicly available external radiotherapy dose dataset providing paired CTs and MRIs (adapted from the SynthRAD2025 dataset 35 ) alongside beam-level photon (6 MV) and proton MC dose distributions generated using open-source MC simulation code. The DoseRAD2026 challenge will utilize this dataset across four tasks: Task 1 (CT to photon dose) targets photon segment (the radiation beam shaped by a given MLC aperture) dose calculation on CT for VMAT delivery in thoracic and abdominal cases, where lung and bowel heterogeneities present a challenging test. Task 2 (MRI to photon dose) extends this to direct photon dose prediction on MRI, leveraging the paired CT-MRI dataset to map MC doses calculated on CT onto MRI. Task 3 (CT to proton dose) addresses proton pencil beam dose calculation on CT across a range of proton energies in thoracic and abdominal cases. Task 4 (MRI to proton dose) targets proton dose prediction directly on MRI, following the same paired data as in Task 2. By releasing beam-level dose labels, DoseRAD2026 is designed to support fast dose calculation both based on CT and also directly on MRI.
II.
Acquisition and validation methods
II.A.
Dataset overview
The DoseRAD2026 dataset comprises paired CTs and MRIs and corresponding photon and proton dose distributions from patients with thoracic and abdominal malignancies.
II.B. Image Dataset
II. ACQUISITION AND VALIDATION METHODS
The imaging data were derived from Center B of the SynthRAD2025 dataset 35 , a singleinstitution subset of the public SynthRAD2025 challenge dataset originally released for synthetic CT generation. This subset was selected for the DoseRAD2026 challenge due to its favorable imaging characteristics. In particular, MRI acquisitions at center B were performed on an MRI-linac using a standardized protocol, ensuring consistent patient positioning between CT and MRI as well as an axial field of view covering the full patient anatomy. In contrast, MRIs from other centers exhibited greater variability in acquisition protocols and imaging characteristics, and were sometimes acquired on diagnostic scanners. For DoseRAD2026, an additional external dataset acquired on a similar MRI-linac is used for testing, but these data will not be released publicly due to data protection restrictions. This article mainly covers the data from SynthRAD2025 Center B.
II.B.
Image Dataset
MRI–CT data from Center B of the SynthRAD2025 dataset originally comprised 140 thoracic and 140 abdominal patients, split 91/14/35 for training, validation and test sets, respectively. For the DoseRAD2026 training set, only the currently public SynthRAD2025 training cases were considered, while for testing, only patients from the currently private SynthRAD2025 test set were used, thereby avoiding any impact on the testing phases of both SynthRAD2025 and DoseRAD2026. Because Tasks 2 and 4 focus on direct MRI-based dose calculation and thus impose strict requirements on CT–MRI alignment, all cases were manually reviewed in three orthogonal planes (axial, coronal, and sagittal) to assess deformable image registration quality (see Section II.B.1.), and cases with obvious misregistration were excluded to ensure high spatial correspondence between modalities. After screening, a total of 115 patients were selected for DoseRAD2026: 53 thoracic (39 training, 14 testing) and 62 abdominal cases (36 training, 26 testing). II.B.1.
Pre-processing
All SynthRAD2025 data had previously undergone several pre-processing steps, including rigid registration, defacing, resampling, and cropping. A detailed description of the SynthRAD2025 pre-processing pipeline is provided by Thummerer et al 35 . For DoseRAD2026, four additional pre-processing steps were performed.
II.B. Image Dataset
II. ACQUISITION AND VALIDATION METHODS
First, CTs were deformably registered to the corresponding MRIs using the ConvexAdam deformable image registration framework 36 . The ConvexAdam hyperparameters were manually tuned and are listed in the Supplementary Materials A.. Second, for abdomen cases, air cavity correction was applied to CTs based on MRI air segmentations. Air cavities in the deformed CTs were first identified by thresholding voxels below -200 HU and then filled with the median HU value of all neighboring voxels of the air cavity. Afterwards, air cavities segmented on the MRIs were transferred to the deformably registered CTs and filled with a constant value of -824 HU 37 . To avoid unrealistically sharp edges at the air-tissue interface, the transition between air and soft tissue HUs was slightly blurred in a 1-voxel band around the air cavity boundary. This mimicked the original appearance of air cavities on CTs. MRI air cavity segmentation was performed using an nnU-Net model 38 trained on a subset of manually annotated cases (20 patients), and all predicted air cavity segmentations were subsequently reviewed and corrected manually. Third, to reduce file sizes and MC photon dose computation times, CTs and MRIs for the photon tasks (Tasks 1 and 2) were resampled to 2 × 2 × 2 mm3 , while images for the proton tasks (Tasks 3 and 4) were not resampled and remained on the same 1 × 1 × 3 mm3 grid as used for SynthRAD2025. Finally, patient body contours were segmented on both CTs and MRIs using TotalSegmentator 39 . The intersection of the two masks was used to define a common foreground region, and voxels outside this region were set to -1024 HU in CT and 0 in MRI. II.B.2.
Image acquisition parameters
All center B MRIs were acquired on a ViewRay MRIdian 0.35 T MRI-linac using a balanced steady-state free-precession sequence (bSSFP, TrueFISP), while CTs were predominantly acquired on a Canon Medical Aquilion/LB CT scanner. Table 1 and Table 2 present detailed imaging parameters of CTs and MRIs of the respective datasets. Per patient imaging parameters are provided as metadata to the dataset release.
II.B. Image Dataset
II. ACQUISITION AND VALIDATION METHODS
Table 1: MRI acquisition parameters for thorax and abdomen datasets (train/test splits). Voxel spacing is reported as in-plane resolution (isotropic); percentages indicate proportion of scans with each value. Thorax Train Manufacturer Model Field strength (T) Sequence Acquisition Flip angle (°) Echo numbers Number of averages Slice thickness (mm)
Test
Abdomen Train
Test
ViewRay MRIdian 0.35 balanced steady-state free-precession (bSSFP, TrueFISP) 2D, breath-hold, no contrast 60 1 1 3
TE (ms)
1.27 (84%) 1.43 (13%) 1.62 (3%)
1.27 (64%) 1.43 (29%) 1.62 (7%)
1.27 (17%) 1.43 (53%) 1.44 (11%) 1.62 (19%)
1.43 (65%) 1.44 (15%) 1.62 (19%)
TR (ms)
3.00 (84%) 3.33 (13%) 3.83 (3%)
3.00 (64%) 3.33 (29%) 3.83 (7%)
3.00 (17%) 3.33 (53%) 3.36 (11%) 3.83 (19%)
3.33 (65%) 3.36 (15%) 3.83 (19%)
Phase enc. steps
225 (32%) 232 (53%) 200 (13%) 207 (3%)
225 (43%) 232 (21%) 200 (29%) 207 (7%)
200 (53%) 225 (14%) 176 (11%) 207 (19%) 232 (3%)
200 (65%) 176 (15%) 207 (19%)
BW (Hz/px)
599 (32%) 604 (53%) 537 (13%) 385 (3%)
599 (43%) 604 (21%) 537 (29%) 385 (7%)
537 (53%) 599 (14%) 534 (11%) 385 (19%) 604 (3%)
537 (65%) 534 (15%) 385 (19%)
Voxel spacing (mm)
1.5 × 1.5 (97%) 1.6 × 1.6 (3%)
1.5 × 1.5 (93%) 1.6 × 1.6 (7%)
1.5 × 1.5 (81%) 1.6 × 1.6 (19%)
1.5 × 1.5 (80%) 1.6 × 1.6 (19%)
Acq. matrix
310 × 360 (53%) 300 × 334 (32%) 266 × 266 (13%) 276 × 276 (3%)
300 × 334 (43%) 310 × 360 (21%) 266 × 266 (29%) 276 × 276 (7%)
266 × 266 (53%) 276 × 276 (19%) 300 × 334 (14%) 234 × 234 (11%) 310 × 360 (3%)
266 × 266 (65%) 234 × 234 (15%) 276 × 276 (19%)
Acq. time (s)
25 (97%) 17 (3%)
25 (93%) 17 (7%)
25 (69%) 23 (11%) 17 (19%)
25 (65%) 23 (15%) 17 (19%)
II.B. Image Dataset
II. ACQUISITION AND VALIDATION METHODS
Table 2: CT acquisition parameters for thorax and abdomen datasets (train/test splits). Voxel spacing is reported as in-plane resolution (isotropic); percentages indicate proportion of scans with each value. Thorax
Manufacturer kVp
Abdomen
Train
Test
Train
Test
Canon Medical Aquilion/LB (97%) Siemens SOMATOM Drive (3%)
Canon Medical Aquilion/LB
Canon Medical Aquilion/LB
Canon Medical Aquilion/LB
120
120
120
120 500
Exposure time (ms)
500
500
500 (94%) 750 (6%)
Tube current (mA)
40 – 305
40 – 140
40 – 440
40 – 190
Rows Columns
512 512
512 512
512 512
512 512
Slice thickness (mm)
3
3
3 (94%) 5 (6%)
3
Voxel spacing (mm)
1.4 × 1.4 (68%) 1.1 × 1.1 (29%) 1.0 × 1.0 (3%)
1.4 × 1.4 (50%) 1.1 × 1.1 (50%)
1.4 × 1.4 (78%) 1.1 × 1.1 (17%) 0.8 × 0.8 (3%) 0.6 × 0.6 (3%)
1.4 × 1.4 (92%) 1.1 × 1.1 (8%)
Data coll. diameter (mm)
700 (68%) 550 (29%) 500 (3%)
700 (50%) 550 (50%)
700 (78%) 550 (17%) 500 (3%) 320 (3%)
700 (92%) 550 (8%)
Recon. diameter (mm)
700 (68%) 550 (29%) 500 (3%)
700 (50%) 550 (50%)
700 (78%) 550 (17%) 418 (3%) 320 (3%)
700 (92%) 550 (8%)
Nr. of slices
103 – 299
117 – 154
20 – 403
87 – 307
II.C. Dose Dataset II.B.3.
II. ACQUISITION AND VALIDATION METHODS
External test dataset
Additional test data were collected at an external institution, but are intended solely for the testing phase of the DoseRAD2026 challenge and will not be made publicly available. This dataset originally comprised 21 paired thoracic CTs and MRIs. Similar to the public data, MRIs were acquired on a ViewRay MRIdian 0.35T MR-Linac using the same bSSFP sequence, a slice thickness of 3 mm and a pixel spacing of 1.6 mm × 1.6 mm. CTs were acquired on a GE Medical Systems Discovery CT590 RT scanner, with a tube voltage of 120 kV, a tube current between 167 and 388 mA, an exposure between 12 and 49 mAs, a pixel spacing between 0.98 and 1.37 mm and a slice thickness of 2.5 mm. These 21 cases underwent the same pre-processing as described above and were manually screened for deformable image registration quality, yielding a filtered external test set of 7 thoracic patients. Since these data were acquired on the same imaging system with the same bSSFP sequence as the training data, they are well suited for use as an external test set for the MRI-based DoseRAD2026 tasks. II.B.4.
Image Dataset Split
Table 3 summarizes the dataset size and split for ease of reference. Table 3: Patient distribution in DoseRAD2026 by anatomical region and dataset split. Training and Testing are from SynthRAD2025 Center B. Region Training Thoracic 39 Abdominal 36 Total 75
II.C.
Testing 14 26 40
External Testing 7 – 7
Total 60 62 122
Dose Dataset
After the pre-processing and manual review, the corrected CTs were subsequently converted into mass density (details see Supplementary Materials B.) and elemental composition maps using a scanner-specific calibration curve 40 . All MC simulations were implemented using the Geant4 toolkit (version 11.00-patch-03) with the QGSP BIC EMV physics list, scoring
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
dose-to-medium of each voxel in units of Gray (Gy). Task-specific simulation details for photon and proton beams are described in Sections II.C.1 and 2. II.C.1.
Photon MC simulation
VMAT plans were generated for all 75 training and 40 test patients using a VMAT implementation 41 in the open-source research treatment planning toolkit matRad 42,43 (v3.2.0). Each plan was optimized based on the patient’s CT and clinically contoured structure set, using a simplified model configured for an Elekta Versa HD Linac with an Agility 160-leaf MLC, in which jaw positions, leaf transmission, and tongue-and-groove effects were not modelled. Optimization followed a three-stage sequential pipeline: fluence map optimization, leaf sequencing via the Siochi algorithm 44 and direct aperture optimization. The resulting plan comprised a single coplanar full arc (gantry −180◦ to 180◦ , couch 0◦ ), yielding 181 control points from matRad, of which 180 were retained after removing the duplicate at 180◦ (coincident with −180◦ ). Each control point was defined by a gantry angle, monitor unit (MU), and MLC leaf positions. 180 beam segments were then derived based on the MLC aperture of the corresponding control point. Each aperture was discretized into a binary mask on a uniform pixel grid (1 mm × 1 mm), where a pixel was assigned a value of one if fully open and zero if blocked. For the 75 training patients, this pool of segments was augmented by randomly shifting each MLC leaf by ±5 mm along its direction of travel two times, expanding the pool from 180 to 540 segments per patient (40,500 in total). VMAT plan segments from the 40 internal testing cases and 7 external testing cases were extracted without augmentation. After segment extraction, Geant4 dose simulations with a 6 MV photon energy spectrum derived from the ELEKTA PRECISE 6MV phase space files provided by the International Atomic Energy Agency 45 were implemented (photon energy distribution in Supplementary Materials B.). The virtual source point-to-isocenter distance was set to 100 cm. At the isocenter plane normal to the beam, a fully open segment mask would correspond to a 40 cm × 40 cm field, and the segment mask was defined a corresponding virtual plane of this dimension located at the opposite of the virtual source point. This served as a fluence source from which particles were emitted with focus on the virtual source point. The dose grid matched the patient CT spacing (2 mm × 2 mm × 2 mm). A uniform 0.35 T magnetic
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
field was simulated in Geant4, aligned parallel to the superior-inferior patient direction and confined within a cylindrical region with a 50 cm diameter 26 . For each training patient, the isocenter from the matRad VMAT plan was shifted along the superior-inferior axis by −2, 0, and +2 cm, defining three arcs. For each gantry angle, the three variants in the 540-segment pool (1 original + 2 MLC-perturbed) were assigned to the three arcs, so that each arc received one unique aperture per angle (180 segments per arc). Segment doses were then simulated at gantry angles every 2◦ for all three arcs, yielding 75 patients × 3 arcs × 180 angles = 40,500 training samples in total. Representative photon segment dose distributions in the training dataset for both Task 1 (CT) and Task 2 (MRI) are illustrated in Fig. 1. For the testing patients, segment doses were simulated according to each testing plan’s control point sequence. Plan doses were then reconstructed by MU-weighted accumulation of segment doses. The dose prescriptions to the planning target volume (PTV) were 70 Gy in 35 fractions for lung patients and 60 Gy in 30 fractions for abdomen patients (2 Gy per fraction in both cases). All MC testing plan doses were scaled so that D95% (PTV) = 0.95×Dprescribed . For both training and test sets, each photon segment dose was simulated at a fixed incident fluence of 1 × 107 photons/cm2 , such that the number of simulated photons varies with segment size.
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
Figure 1: Representative photon segment dose distributions overlaid on CT (Task 1, top row) and MRI (Task 2, bottom row) images for thoracic (left column) and abdominal (right column) cases.
II.C.2.
Proton MC simulation
After the HU-to-mass density and elemental composition conversion, stopping powers were computed internally by Geant4 from the assigned material properties using the Bethe–Bloch formalism, with relative stopping power (RSP) (relative to water) optionally extracted in our Geant4 software (which may be useful for DL model training 46 ). The initial proton energies ranged from 31.73 to 200.80 MeV (85 energy levels). A simplified proton beam model was employed to define the energy-dependent beamlet energy spread (sigma from 0.78 to 6.11 MeV) and spot size (sigma from 4.00 to 7.99 mm), under a single-Gaussian lateral approximation. Proton parameters were derived from matRad’s native ”Generic”
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
proton machine and are listed in Table 4. The proton dose grid matched the patient CT spacing (1 × 1 × 3 mm) and no magnetic field was simulated. For the 75 training patients, a total of 36 gantry angles were simulated in 10◦ increments from 0◦ to 350◦ surrounding the isocenter. Along each gantry direction, proton spots were initialized as parallel pencil beams on a source plane located 100 cm upstream of the isocenter, approximating the extended virtual source-to-axis distances arising from the scanning magnet positions in clinical proton systems. At each gantry angle, 15 rays were arranged on a 5 × 3 beam’s eye view (BEV) grid, with 5 fixed positions of {−40, −20, 0, +20, +40} mm along the superior-inferior axis, and 3 positions along the lateral axis consisting of 0 mm and two randomly sampled offsets from {±5k | k ∈ [1, 6]} mm. Each ray was assigned 2 beamlets of distinct energies sampled from Table 4, yielding 15 × 2 = 30 beamlets per beam. Across all 75 training patients, this setup produced 75 patients × 36 angles × 30 beamlets = 81,000 beamlet dose distributions as the training dose dataset. Representative proton beamlet dose distributions in the training dataset for both Task 3 (CT) and Task 4 (MRI) are illustrated in Fig. 2. For the 40 testing patients, 40 IMPT plans were optimized based on the patient’s CT and clinically contoured structure set using matRad. Plan doses were then calculated as a spot weight-scaled superposition of beamlet doses using Geant4. The PTV dose prescriptions were 70 Gy in 35 fractions for lung patients and 60 Gy in 30 fractions for abdomen patients (2 Gy per fraction in both cases). All MC testing plan doses were scaled so that D95% (PTV) = 0.95 × Dprescribed . For both training and test proton dose simulations, each beamlet dose was simulated with 1 × 106 primary protons.
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
Figure 2: Representative proton beamlet dose distributions overlaid on CT (Task 3, top row) and MRI (Task 4, bottom row) images for thoracic (left column) and abdominal (right column) cases.
II.C. Dose Dataset
II. ACQUISITION AND VALIDATION METHODS
Table 4: Proton beam model parameters: energy, energy spread (σE ), and spot size (σs ) for 85 energy levels (values rounded to 2 decimal places). E (MeV)
σE (MeV)
σs (mm)
E (MeV)
σE (MeV)
σs (mm)
E (MeV)
σE (MeV)
σs (mm)
31.73 36.80 41.38 45.60 49.54 53.25 56.77 60.13 63.35 66.45 69.44 72.33 75.14 77.87 80.53 83.13 85.66 88.13 90.56 92.93 95.26 97.55 99.79 102.00 104.17 106.30 108.41 110.48 112.52
6.11 5.19 4.56 4.09 3.74 3.45 3.21 3.01 2.84 2.70 2.57 2.46 2.35 2.26 2.18 2.10 2.04 1.97 1.91 1.86 1.81 1.76 1.72 1.68 1.64 1.60 1.57 1.53 1.50
7.49 7.62 7.70 7.77 7.82 7.86 7.90 7.93 7.95 7.98 7.99 7.99 7.71 7.56 7.42 7.29 7.16 7.03 6.91 6.80 6.69 6.58 6.48 6.38 6.28 6.19 6.10 6.02 5.93
114.54 116.53 118.49 120.43 122.34 124.23 126.10 127.95 129.78 131.59 133.37 135.15 136.90 138.64 140.35 142.06 143.75 145.42 147.08 148.72 150.35 151.97 153.57 155.16 156.74 158.31 159.86 161.40
1.47 1.44 1.42 1.39 1.37 1.34 1.32 1.30 1.28 1.26 1.24 1.22 1.21 1.19 1.17 1.16 1.14 1.13 1.11 1.10 1.08 1.07 1.06 1.05 1.03 1.02 1.01 1.00
5.85 5.78 5.70 5.63 5.56 5.50 5.43 5.37 5.31 5.25 5.20 5.15 5.10 5.05 5.00 4.95 4.91 4.87 4.82 4.79 4.75 4.71 4.68 4.64 4.61 4.58 4.55 4.52
162.93 164.45 165.96 167.46 168.95 170.43 171.90 173.36 174.81 176.25 177.68 179.10 180.52 181.92 183.32 184.71 186.09 187.46 188.83 190.19 191.54 192.88 194.22 195.55 196.87 198.19 199.49 200.80
0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 0.91 0.91 0.90 0.89 0.88 0.87 0.87 0.86 0.85 0.84 0.84 0.83 0.82 0.82 0.81 0.80 0.80 0.79 0.79 0.78
4.49 4.46 4.44 4.41 4.39 4.36 4.34 4.32 4.30 4.28 4.26 4.24 4.22 4.20 4.18 4.17 4.15 4.14 4.12 4.11 4.09 4.08 4.06 4.05 4.04 4.02 4.01 4.00
II.C.3.
Post-processing
All simulated photon and proton dose distributions were masked to zero outside the body contour derived from the deformed CT. II.C.4.
Computational Resources
The average Geant4 simulation time for a single photon segment dose was approximately 20– 30 hours on a single CPU core, while a single proton beamlet dose required approximately 2–3 hours. All simulations were carried out on the computing clusters of the GSI Helmholtzzentrum für Schwerionenforschung, the Ludwig Maximilian University (LMU) Hospital Munich and Leibniz-Rechenzentrum (LRZ) utilizing approximately 2,000 CPU cores in total.
II.D. Data validation
II. ACQUISITION AND VALIDATION METHODS
II.D.
Data validation
II.D.1.
Image dataset validation
The DoseRAD2026 dataset was derived from the SynthRAD2025 dataset, which already underwent extensive quality control by the initial data providers through visual inspection to ensure completeness and overall data integrity. However, whereas the SynthRAD2025 dataset was designed to represent a heterogenous, multi-center and multi-anatomy cohort, containing a broad range of scanner models and acquisition protocols, the aims of the DoseRAD2026 challenge required special attention to image alignment between CT and MRI. This resulted in the inclusion of only a single center with data from two anatomical regions (thorax and abdomen). All included cases were manually screened and rigorously evaluated for highquality alignment between CTs and MRIs. This assessment was performed by two medical imaging experts, who visually inspected deformably registered image pairs with particular attention to alignment in bony structures, soft tissue interfaces, and external body contours. For these cases, detailed imaging parameters were systematically extracted from the SynthRAD2025 metadata and are provided as metadata to the DoseRAD2026 dataset to facilitate reproducibility and transparency. The data inspection further revealed the presence of metal artifacts, including dental implants, in 17 out of 115 patients (13 in the training set and 4 in the test set). These cases were intentionally retained, as they provide a valuable opportunity to assess the robustness and sensitivity of different dose calculation approaches in the presence of clinically relevant imaging artifacts. A detailed list of the 13 patients with metal implants from the training set is provided in the Supplementary Materials C.. II.D.2.
Dose dataset validation
To assess MC statistical uncertainty, 10 independent simulations with different random seeds were performed on one representative thoracic and one representative abdominal case, for both photon segment doses and proton beamlet doses. The statistical uncertainty was quantified as the voxel-wise standard deviation across the 10 repeated simulations, evaluated within dose regions exceeding 10% of Dmax to exclude low-dose voxels where relative uncertainty is inherently inflated. For photon segment doses, a primary fluence of 1 × 107 photons/cm2
II.E. Code availability
III. DATA FORMAT AND USAGE NOTES
yielded an average statistical uncertainty below 3% in both cases. For proton beamlet doses, simulating 1 × 106 primary protons resulted in an average statistical uncertainty below 2%. All photon dose distributions were verified using BEV extraction based on the beam direction to match the corresponding dose and visual inspection of the segment shape, whereas all proton dose distributions were checked via BEV extraction based on the beamlet direction.
II.E.
Code availability
DoseRAD2026 preprocessing, MC simulation and matRad planning scripts are publicly available and listed in table 5. In addition, analytical pencil beam dose calculation scripts for both photon and proton modalities, built upon matRad’s Python variant pyRadPlan 47 , are provided (link see Table 5). These scripts serve as templates for DoseRAD2026 challenge participants. Table 5: Availability of DoseRAD2026 code and scripts. Component
Repository URL
Pre-processing scripts Geant4 simulation scripts Geant4 toolkit version matRad VMAT planning scripts matRad commit used in this study pyRadPlan pencil beam baseline
https://github.com/DoseRAD2026/preprocessing https://github.com/DoseRAD2026/geant4-dose-sim https://geant4.web.cern.ch/download/11.0.3.html https://github.com/DoseRAD2026/matrad-vmat-example https://github.com/e0404/matRad/tree/71c95e96a1ecc7c7e025b9d75240ea009bdbc063 https://github.com/DoseRAD2026/pyradplan-pb-baseline
III.
Data format and usage notes
III.A.
Data structure and file formats
The DoseRAD2026 dataset is divided into two subsets, corresponding to photon and proton dose calculation tasks, respectively. Each subset comprises an image directory containing paired CTs and MRIs for each patient, a dose directory with dose maps for multiple beam configurations, and a plan JSON file specifying the associated beam parameters. The corresponding folder structures are illustrated in Figure 3 and Figure 4. Each patient is identified by a unique alphanumeric identifier inherited from the SynthRAD2025 dataset.
III.A. Data structure and file formats
III. DATA FORMAT AND USAGE NOTES
The dataset is distributed under a CC BY-NC 4.0 International license. The training set was released in April 2026 and the test set will be released in March 2030, after the five-year embargo period of the SynthRAD2025 test data expires. The data are hosted on Hugging Face, with a persistent DOI provided through Zenodo (https://doi.org/10.5281/zenodo. 19347848).
Figure 3: Folder structure for the photon dataset. B, beam; CP, control point.
Figure 4: Folder structure for the proton dataset. B, beam; R, ray; L, beamlet.
III.B. Usage notes
III.B.
IV. DISCUSSION
Usage notes
Images and dose maps are provided as compressed MetaImage (.mha) files. These files can be accessed using the Insight Toolkit (ITK) framework or its simplified interfaces, such as SimpleITK, which offer bindings for multiple programming languages, including Python, R, Java, and C++. In addition, MetaImage files can be visualized using commonly available image viewers such as 3D Slicer, vv, and ITK-SNAP. The JSON file describes the corresponding beam configuration. For photon data, the JSON contains one or more beams (each associated with an isocenter position), with each beam comprising 180 control points defined by gantry angle and 80-pair MLC leaf positions (mlc left int mm, mlc right int mm) relative to the beam central axis. For proton data, the JSON specifies the isocenter and 36 beam directions (gantry 0◦ to 350◦ , 10◦ steps), with each beam containing rays defined by source and target coordinates, and the associated beamlet energies in MeV. Dose .mha files follow the naming convention Dose B{beam} CP{cp}.mha for photon and Dose B{beam} R{ray} L{beamlet}.mha for proton, directly indexing into the JSON structure. A complete description of all JSON fields is provided in Appendix D..
IV.
Discussion
The DoseRAD2026 dataset is the first publicly available dataset providing paired CT and MRI alongside beam-level MC dose ground truth for both photon and proton therapy, spanning thoracic and abdominal anatomies across four tasks. It aims to address two clinical needs. First, it provides a large-scale benchmark dataset for DL-based acceleration of MC dose calculation on CT, covering photon segment and proton pencil beam delivery in anatomical sites with challenging tissue heterogeneities. Second, by building on the SynthRAD2025 paired CT–MRI data 35 with additional preprocessing steps including ConvexAdam deformable registration, air cavity correction and systematic manual quality control of CT–MRI alignment, DoseRAD2026 supports exploration of direct MRI to beam-level dose calculation, offering a potential pathway toward real-time MR-guided or MR-only adaptive radiotherapy. Together, the four tasks provide a common ground for benchmarking DL dose calculation methods across modalities, anatomies, and radiation types. A key strength of the DoseRAD2026 dataset is the provision of beam-level (rather
IV. DISCUSSION than plan-level) MC dose labels, which substantially increases the number of training samples per patient, enables plan dose accumulation, and allows flexible use of patient-specific beam configurations by varying beam angles, isocenters, segment shapes, and proton energies independently. Dose labels for both photon and proton modalities are generated using Geant4, a well-validated general-purpose MC simulation framework, providing physically consistent ground truth and a broader scope than existing public datasets such as OpenKBP 48 , which covers head-and-neck photon plan doses only. The focus on thoracic and abdominal anatomies introduces clinically relevant tissue heterogeneities, including lung and abdominal air cavities, which present known challenges for analytical dose calculation methods. Full reproducibility is ensured through open-source release of the preprocessing, Geant4 simulation, and matRad VMAT planning pipelines, together with an analytical pencil beam baseline, allowing the community to scrutinise and reproduce the data generation, or adapt them for their own research purposes. A fundamental limitation of DoseRAD2026 is that both photon and proton dose distributions are generated using simplified, non-clinical beam models. For photon plans, a generic machine model is used with binary MLC aperture masks, without modelling jaw positions, leaf transmission, or tongue-and-groove effects, and dose values are not calibrated to absolute clinical output. The VMAT plans are generated using matRad 42 , an open-source research treatment planning toolkit, which ensures reproducibility but may not capture the full complexity of clinically optimized plans. For proton plans, an idealized generic beam model is employed rather than one specific to any clinical delivery system. Consequently, while the dataset is well-suited for benchmarking DL dose calculation algorithms, the dose distributions are not directly representative of clinical deliveries. However, this trade-off is accepted in favor of reproducibility and open-source transparency, the Geant4 MC code and matRad planning toolkit ensures that data generation can be verified and extended by the community. Algorithms developed on DoseRAD2026 may nonetheless serve as a strong starting point for clinical translation: the generalized beam representations and physically consistent MC ground truth support the learning of fundamental dose–anatomy relationships, and fine-tuning on machine-specific data provides a practical pathway toward clinical deployment. Another limitation is that all MRI data originate from a single center and MRI-linac system (0.35 T ViewRay MRidian), which may limit generalisability to other imaging plat-
V. CONCLUSION forms and patient populations. In addition, the dataset relies on deformable registration of CT to MRI to enable supervised learning for direct MRI-based dose prediction. Despite careful validation and quality control, residual registration errors may persist. Since groundtruth MC dose distributions were computed on CT images only and implicitly assumed to be spatially aligned with MRI, such misregistrations can impact MRI-based tasks. Residual geometric MRI distortions following the vendor’s corrections were not investigated in this dataset and masking the scans with a union of CT and MRI patient outline masks may further obfuscate their impact on dose calculations. These factors might limit the direct clinical applicability of models trained solely on the provided data and should be considered when interpreting the results or translating methods to clinical workflows. Beyond the DoseRAD2026 challenge, the dataset will remain publicly available to support continued development and benchmarking of fast CT- and MR-based dose calculation methods. Future extensions may incorporate additional anatomical regions, multiinstitutional data, and more clinically realistic beam models to facilitate systematic evaluation of model generalizability and clinical applicability.
V.
Conclusion
DoseRAD2026 introduces the first publicly available benchmark dataset, providing paired CT and MRI together with beam-level photon and proton MC dose distributions for thoracic and abdominal radiotherapy. The dataset and accompanying challenge establish a common framework for the development and rigorous evaluation of fast dose calculation and dose prediction methods across imaging modalities, anatomies, and treatment modalities. We anticipate that the DoseRAD2026 dataset will serve as a valuable open resource for the community and help accelerate progress toward fast, accurate and robust (DL-based) dose engines for CT- and MRI-based dose calculation in adaptive radiotherapy.
Acknowledgements The authors would like to thank the GSI Helmholtzzentrum für Schwerionenforschung and the Leibniz-Rechenzentrum for the use of their computing resources. Fan Xiao acknowl-
edges financial support from the China Scholarship Council (No. 202308440107). Niklas Wahl and Samir Schulz acknowledge funding from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation)-Project No. 457509854. Niklas Wahl further acknowledges funding from the Cooperation Program in Cancer Research of the Deutsches Krebsforschungszentrum (DKFZ) and Israel’s Ministry of Science, Technology and Space (MOST) as well as from the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - Project No. 443188743. The authors would like to thank the Bayerisches Zentrum für Krebsforschung (BZKF) Lighthouse ”Image-Guidance in Local Therapies” for supporting the data collection for this dataset.
Conflict of Interest The Department of Radiation Oncology of the University Hospital of LMU Munich has research agreements with Brainlab, Elekta and ViewRay. Zoltan Perko is a Senior Applied Scientist at Radformation Inc.
References 1
M. B. Barton, S. Jacob, J. Shafiq, K. Wong, S. R. Thompson, T. P. Hanna, and G. P. Delaney, Estimating the demand for radiotherapy from the evidence: a review of changes from 2003 to 2012, Radiotherapy and oncology 112, 140–144 (2014).
2
Y. Lievens, J. M. Borras, and C. Grau, Provision and use of radiotherapy in Europe, Molecular oncology 14, 1461–1469 (2020).
3
K. Otto, Volumetric modulated arc therapy: IMRT in a single gantry arc, Medical Physics 35, 310–317 (2008).
4
A. Lomax, Intensity modulated proton therapy and its sensitivity to treatment uncertainties 1: the potential effects of calculational uncertainties, Physics in Medicine & Biology 53, 1027–1042 (2008).
5
R. R. Wilson, Radiological use of fast protons, Radiology 47, 487–491 (1946).
6
A. Ahnesjö and M. M. Aspradakis, Dose calculations for external photon beams in radiotherapy, Physics in Medicine & Biology 44, R99–R155 (1999).
7
S. Lim-Reinders, B. M. Keller, S. Al-Ward, A. Sahgal, and A. Kim, Online adaptive radiation therapy, International Journal of Radiation Oncology* Biology* Physics 99, 994–1003 (2017).
8
H. Paganetti, P. Botas, G. C. Sharp, and B. Winey, Adaptive proton therapy, Physics in Medicine & Biology 66, 22TR01 (2021).
9
E. Lombardo et al., Real-time motion management in MRI-guided radiotherapy: Current status and AI-enabled prospects, Radiotherapy and Oncology 190, 109970 (2024).
10
P. J. Keall, I. El Naqa, M. F. Fast, E. A. Hewson, N. Hindley, P. Poulsen, C. Sengupta, N. Tyagi, and D. E. Waddington, Real-time dose-guided radiation therapy, International Journal of Radiation Oncology* Biology* Physics 122, 787–801 (2025).
11
D. Rogers, Fifty years of Monte Carlo simulations for medical physics, Physics in Medicine & Biology 51, R287–R301 (2006).
12
H. M. Paschal, C. N. Kabat, P. Papaconstadopoulos, N. A. Kirby, P. A. Myers, T. D. Wagner, and S. Stathakis, Monte Carlo modeling of the Elekta Versa HD and patient dose calculation with EGSnrc/BEAMnrc, Journal of applied clinical medical physics 23, e13715 (2022).
13
B. Cheng, Y. Xu, S. Li, Q. Ren, X. Pei, K. Men, J. Dai, and X. G. Xu, Development and clinical application of a GPU-based Monte Carlo dose verification module and software for 1.5 T MR-LINAC, Medical Physics 50, 3172–3183 (2023).
14
J. Holmes, H. Feng, L. Zhang, M. K. Fix, S. B. Jiang, and W. Liu, Fast Monte Carlo dose calculation in proton therapy, Physics in Medicine & Biology 69, 17TR01 (2024).
15
Z. Liu, Y. Wang, Y. Han, P. Hu, C. Zheng, B. Yan, and Y. Yang, A GPU-accelerated Monte Carlo dose engine for external beam radiotherapy, Medical Physics 52, e17899 (2025).
16
A. Neishabouri, N. Wahl, A. Mairani, U. Köthe, and M. Bangert, Long short-term memory networks for proton dose calculation in highly heterogeneous tissues, Medical Physics 48, 1893–1908 (2021).
17
C. Wu, D. Nguyen, Y. Xing, A. B. Montero, J. Schuemann, H. Shang, Y. Pu, and S. Jiang, Improving proton dose calculation accuracy by using deep learning, Machine learning: science and technology 2, 015017 (2021).
18
O. Pastor-Serrano and Z. Perkó, Millisecond speed deep learning based proton dose calculation with Monte Carlo accuracy, Physics in Medicine & Biology 67, 105006 (2022).
19
A. Neishabouri, J. Bauer, A. Abdollahi, J. Debus, and A. Mairani, Real-time adaptive proton therapy: An AI-based spatio-temporal mono-energetic dose calculation model (CC-LSTM), Computers in biology and medicine 188, 109777 (2025).
20
C. Kontaxis, G. Bol, J. Lagendijk, and B. Raaymakers, DeepDose: Towards a fast dose calculation engine for radiation therapy using deep learning, Physics in Medicine & Biology 65, 075013 (2020).
21
J. Fan, L. Xing, P. Dong, J. Wang, W. Hu, and Y. Yang, Data-driven dose calculation algorithm based on deep U-Net, Physics in Medicine & Biology 65, 245035 (2020).
22
O. Pastor-Serrano, P. Dong, C. Huang, L. Xing, and Z. Perkó, Sub-second photon dose prediction via transformer neural networks, Medical Physics 50, 3159–3171 (2023).
23
M. Witte and J.-J. Sonke, A deep learning based dynamic arc radiotherapy photon dose engine trained on Monte Carlo dose distributions, Physics and Imaging in Radiation Oncology 30, 100575 (2024).
24
M. A. Schmidt and G. S. Payne, Radiotherapy planning using MRI, Physics in Medicine & Biology 60, R323–R361 (2015).
25
J. J. Lagendijk, B. W. Raaymakers, and M. Van Vulpen, The magnetic resonance imaging–linac system, Seminars in radiation oncology 24, 207–209 (2014).
26
S. Klüter, Technical design and concept of a 0.35 T MR-Linac, Clinical and translational radiation oncology 18, 98–101 (2019).
27
A. Hoffmann et al., MR-guided proton therapy: a review and a preview, Radiation Oncology 15, 129 (2020).
28
T. T. Pham, B. Whelan, B. M. Oborn, G. P. Delaney, S. Vinod, C. Brighi, M. Barton, and P. Keall, Magnetic resonance imaging (MRI) guided proton therapy: a review of the clinical challenges, potential benefits and pathway to implementation, Radiotherapy and Oncology 170, 37–47 (2022).
29
M. A. Bahloul, S. Jabeen, S. Benoumhani, H. A. Alsaleh, Z. Belkhatir, and A. Al-Wabil, Advancements in synthetic CT generation from MRI: A review of techniques, and trends in radiation therapy planning, Journal of Applied Clinical Medical Physics 25, e14499 (2024).
30
M. F. Spadea, M. Maspero, P. Zaffino, and J. Seco, Deep learning based synthetic-CT generation in radiotherapy and PET: A review, Medical Physics 48, 6537–6566 (2021).
31
Z. Liu, K. Men, W. Hu, J. Dai, and J. Fan, An MR-only deep learning inference modelbased dose estimation algorithm for MR-guided adaptive radiation therapy, Medical Physics 52, 4429–4442 (2025).
32
F. Xiao, D. Radonic, N. Wahl, N. Delopoulos, A. Thummerer, S. Corradini, C. Belka, G. Dedes, C. Kurz, and G. Landry, Deep learning-based synthetic-CT-free photon dose calculation in MR-guided radiotherapy: A proof-of-concept study, Medical physics 52, e70106 (2025).
33
M. Li, C. Winterhalter, X. Li, S. Safai, A. Lomax, and Y. Zhang, A proof-of-concept study of direct magnetic resonance imaging-based proton dose calculation for brain tumors via neural networks with Monte Carlo-comparable accuracy, Physics and Imaging in Radiation Oncology 35, 100806 (2025).
34
L. Tian, L. Tsu, P. Vehling, E. Alvarez-Michael, and A. Lühr, Magnetic resonance imaging-based proton dose calculation for pelvic tumors using deep learning, Physics in Medicine and Biology (2026).
35
A. Thummerer et al., SynthRAD2025 Grand Challenge dataset: Generating synthetic CTs for radiotherapy from head to abdomen, Medical physics 52, e17981 (2025).
36
H. Siebert, C. Großbröhmer, L. Hansen, and M. P. Heinrich,
ConvexAdam: Self-
Configuring Dual-Optimization-Based 3D Multitask Medical Image Registration, IEEE Transactions on Medical Imaging 44, 738–748 (2025). 37
P. Godoy Scripes, E. Subashi, S. Burleson, J. Liang, P. Romesser, C. Crane, J. Mechalakos, M. Hunt, and N. Tyagi, Impact of varying air cavity on planning dosimetry for rectum patients treated on a 1.5 T hybrid MR-linac system, Journal of applied clinical medical physics 21, 144–152 (2020).
38
F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation, Nature methods 18, 203–211 (2021).
39
J. Wasserthal et al., TotalSegmentator: robust segmentation of 104 anatomic structures in CT images, Radiology: Artificial Intelligence 5 (2023).
40
S. Schmid, G. Landry, C. Thieke, F. Verhaegen, U. Ganswindt, C. Belka, K. Parodi, and G. Dedes, Monte Carlo study on the sensitivity of prompt gamma imaging to proton range variations due to interfractional changes in prostate cancer patients, Physics in Medicine & Biology 60, 9329 (2015).
41
E. Christiansen, E. Heath, and T. Xu, Continuous aperture dose calculation and optimization for volumetric modulated arc therapy, Physics in Medicine & Biology 63, 21NT01 (2018).
42
H.-P. Wieser et al., Development of the open-source dose calculation and optimization toolkit matRad, Medical physics 44, 2556–2568 (2017).
43
N. Abbani et al., matRad, (2026).
44
R. A. C. Siochi, Variable depth recursion algorithm for leaf sequencing, Medical physics 34, 664–672 (2007).
45
R. Capote, R. Jeraj, C. Ma, D. W. Rogers, F. Sánchez-Doblado, J. Sempau, J. Seuntjens, and J. Siebers, Phase-space database for external beam radiotherapy. Summary report of a consultants’ meeting, (2006).
A. ConvexAdam registration parameters 46
D. Radonic et al., Proton dose calculation with LSTM networks in presence of a magnetic field, Physics in Medicine & Biology 69, 215019 (2024).
47
T. Becher, L. Bucher, A. Kontopoulos, J. Laub, F. Leininger, T. Ortkamp, V. Santiago Aguilar, G. Stanic, S. Schulz, and N. Wahl, pyRadPlan, (2026).
48
A. Babier, B. Zhang, R. Mahmood, K. L. Moore, T. G. Purdie, A. L. McNiven, and T. C. Chan, OpenKBP: the open-access knowledge-based planning grand challenge and dataset, Medical Physics 48, 5549–5561 (2021).
Appendices A.
ConvexAdam registration parameters Table 6: The convexAdam hyperparameters used for MR-to-CT deformable image registration. Detailed description of parameters is provided in Siebert et al. 36 . Parameter disp hwd grid sp convex mind r convex mind d adam mind r adam mind d adam grid sp gauss σ λ iters iters smooth
B.
Description The displacement range in voxels of the discretised search space The grid spacing in voxels used for the effective resolution of features and deformable transformation The MIND radius in voxels The MIND dilation in voxels The MIND radius in voxels The MIND dilation in voxels The grid spacing in voxels for the displacement field resolution σ of Gaussian kernel in voxels for B-spline deformation model The diffusive regularisation weight The number of iterations for Adam optimisation The number of post-smoothing (average) steps with a kernel size of 3
Photon and proton beam parameters
{ " photon ": { " SAD_mm ": 1000 , " m l c _ n u m _ l e a f _ p a i r s ": 80 , " m l c _ l e a f _ t h i c k n e s s _ m m ": 5 , " jaw_x_mm ": [ -200 , 200] , " jaw_y_mm ": [ -200 , 200] , " s o u r c e _ p l a n e _ d i s t a n c e _ m m ": 2000 , " v i r t u a l _ s o u r c e _ d i s t a n c e _ m m ": 1000 ,
Value 6 5 2 2 1 2 2 0.2 1 200 0
B. Photon and proton beam parameters
" _ s o u r c e _ g e o m e t r y _ n o t e ": " Photons sampled from a 400 x400 mm plane at 2* SAD . Beam converges to a focal point at SAD (1000 mm from isocenter ) , this focal point is the virtual point source " , " _ m l c _ a p e r t u r e _ n o t e ": " MLC aperture enforced by rejection sampling against the 400 x400 binary mask (1 mm / pixel ) . History count scales with open area : 10 M histories / cm2 of open MLC aperture ." , " en er g y_ sp ec t ru m ": { " _note ": " Geant4 GPS histogram defining photon energy distribution . 100 points , 0.06 -6.00 MeV , step 0.06 MeV ." , " energy_mev ": [0.06 , 0.12 , 0.18 , 0.24 , 0.30 , 0.36 , 0.42 , 0.48 , 0.54 , 0.60 , 0.66 , 0.72 , 0.78 , 0.84 , 0.90 , 0.96 , 1.02 , 1.08 , 1.14 , 1.20 , 1.26 , 1.32 , 1.38 , 1.44 , 1.50 , 1.56 , 1.62 , 1.68 , 1.74 , 1.80 , 1.86 , 1.92 , 1.98 , 2.04 , 2.10 , 2.16 , 2.22 , 2.28 , 2.34 , 2.40 , 2.46 , 2.52 , 2.58 , 2.64 , 2.70 , 2.76 , 2.82 , 2.88 , 2.94 , 3.00 , 3.06 , 3.12 , 3.18 , 3.24 , 3.30 , 3.36 , 3.42 , 3.48 , 3.54 , 3.60 , 3.66 , 3.72 , 3.78 , 3.84 , 3.90 , 3.96 , 4.02 , 4.08 , 4.14 , 4.20 , 4.26 , 4.32 , 4.38 , 4.44 , 4.50 , 4.56 , 4.62 , 4.68 , 4.74 , 4.80 , 4.86 , 4.92 , 4.98 , 5.04 , 5.10 , 5.16 , 5.22 , 5.28 , 5.34 , 5.40 , 5.46 , 5.52 , 5.58 , 5.64 , 5.70 , 5.76 , 5.82 , 5.88 , 5.94 , 6.00] , " weight ": [2.54238 e -04 , 4.19007 e -03 , 1.64380 e -02 , 3.27511 e -02 , 4.26306 e -02 , 4.56607 e -02 , 4.52659 e -02 , 4.34762 e -02 , 4.47129 e -02 , 3.81616 e -02 , 3.58066 e -02 , 3.37301 e -02 , 3.18390 e -02 , 2.99035 e -02 , 2.81086 e -02 , 2.64351 e -02 , 2.49122 e -02 , 2.34632 e -02 , 2.21145 e -02 , 2.09058 e -02 , 1.97536 e -02 , 1.86973 e -02 , 1.76865 e -02 , 1.67660 e -02 , 1.58873 e -02 , 1.49913 e -02 , 1.42329 e -02 , 1.35083 e -02 , 1.28337 e -02 , 1.22013 e -02 , 1.16182 e -02 , 1.10765 e -02 , 1.05718 e -02 , 1.00526 e -02 , 9.61410 e -03 , 9.18486 e -03 , 8.78546 e -03 , 8.40149 e -03 , 8.03662 e -03 , 7.68496 e -03 , 7.36606 e -03 , 7.06749 e -03 , 6.76740 e -03 , 6.49375 e -03 , 6.22279 e -03 , 5.97915 e -03 , 5.75149 e -03 , 5.51645 e -03 , 5.29513 e -03 , 5.07651 e -03 , 4.89121 e -03 , 4.69359 e -03 , 4.51351 e -03 , 4.33559 e -03 , 4.17315 e -03 , 4.00582 e -03 , 3.85291 e -03 , 3.70279 e -03 , 3.54862 e -03 , 3.41719 e -03 , 3.27277 e -03 , 3.15898 e -03 , 3.02902 e -03 , 2.90728 e -03 , 2.79603 e -03 , 2.67682 e -03 , 2.56908 e -03 , 2.45770 e -03 , 2.36121 e -03 , 2.24374 e -03 , 2.14842 e -03 , 2.05885 e -03 , 1.96066 e -03 , 1.86058 e -03 , 1.76296 e -03 , 1.68080 e -03 , 1.59073 e -03 , 1.50016 e -03 , 1.41159 e -03 , 1.32824 e -03 , 1.25038 e -03 , 1.16362 e -03 , 1.08056 e -03 , 1.00508 e -03 , 9.21253 e -04 , 8.38062 e -04 , 7.57685 e -04 , 6.77143 e -04 , 5.97707 e -04 , 5.16531 e -04 , 4.31229 e -04 , 3.53960 e -04 , 2.77338 e -04 , 1.94645 e -04 , 1.14458 e -04 , 2.88556 e -05 , 5.53396 e -08 , 7.90565 e -09 , 0.0 , 0.0] } }, " proton ": { " source_model ": " point_source " , " spot_profile ": " Gaussian " , " e n e r g y _ s p r e a d _ p r o f i l e ": " Gaussian " , " _ e n e r g y _ t a b l e _ n o t e ": "85 discrete energies . Each entry : ( energy_mev , sigma_energy_mev , sigma_spot_mm ) . sigma_energy is absolute energy spread (1 - sigma ) . sigma_spot is lateral 1 - sigma ." , " energy_table ": [ {" energy_mev ": 31.7290 , " s i g m a _e n e r g y _ m e v ": 6.1130 , " sigma_spot_mm ": 7.4934} , {" energy_mev ": 36.7986 , " s i g m a _e n e r g y _ m e v ": 5.1891 , " sigma_spot_mm ": 7.6244} , {" energy_mev ": 41.3788 , " s i g m a _e n e r g y _ m e v ": 4.5578 , " sigma_spot_mm ": 7.7018} , {" energy_mev ": 45.5978 , " s i g m a _e n e r g y _ m e v ": 4.0937 , " sigma_spot_mm ": 7.7653} , {" energy_mev ": 49.5354 , " s i g m a _e n e r g y _ m e v ": 3.7352 , " sigma_spot_mm ": 7.8180} , {" energy_mev ": 53.2453 , " s i g m a _e n e r g y _ m e v ": 3.4482 , " sigma_spot_mm ": 7.8620} , {" energy_mev ": 56.7659 , " s i g m a _e n e r g y _ m e v ": 3.2122 , " sigma_spot_mm ": 7.8988} , {" energy_mev ": 60.1259 , " s i g m a _e n e r g y _ m e v ": 3.0139 , " sigma_spot_mm ": 7.9294} , {" energy_mev ": 63.3471 , " s i g m a _e n e r g y _ m e v ": 2.8446 , " sigma_spot_mm ": 7.9546} , {" energy_mev ": 66.4468 , " s i g m a _e n e r g y _ m e v ": 2.6979 , " sigma_spot_mm ": 7.9752} , {" energy_mev ": 69.4389 , " s i g m a _e n e r g y _ m e v ": 2.5693 , " sigma_spot_mm ": 7.9916} , {" energy_mev ": 72.3349 , " s i g m a _e n e r g y _ m e v ": 2.4554 , " sigma_spot_mm ": 7.9938} , {" energy_mev ": 75.1442 , " s i g m a _e n e r g y _ m e v ": 2.3537 , " sigma_spot_mm ": 7.7121} , {" energy_mev ": 77.8749 , " s i g m a _e n e r g y _ m e v ": 2.2623 , " sigma_spot_mm ": 7.5646} , {" energy_mev ": 80.5337 , " s i g m a _e n e r g y _ m e v ": 2.1795 , " sigma_spot_mm ": 7.4232} , {" energy_mev ": 83.1266 , " s i g m a _e n e r g y _ m e v ": 2.1041 , " sigma_spot_mm ": 7.2877} , {" energy_mev ": 85.6587 , " s i g m a _e n e r g y _ m e v ": 2.0351 , " sigma_spot_mm ": 7.1577} ,
B. Photon and proton beam parameters
{" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ": {" energy_mev ":
88.1344 , 90.5576 , 92.9319 , 95.2605 , 97.5459 , 99.7909 , 101.9976 , 104.1682 , 106.3045 , 108.4082 , 110.4810 , 112.5242 , 114.5392 , 116.5273 , 118.4897 , 120.4273 , 122.3412 , 124.2323 , 126.1016 , 127.9497 , 129.7775 , 131.5856 , 133.3749 , 135.1458 , 136.8991 , 138.6352 , 140.3548 , 142.0583 , 143.7462 , 145.4189 , 147.0770 , 148.7208 , 150.3508 , 151.9672 , 153.5705 , 155.1611 , 156.7391 , 158.3051 , 159.8591 , 161.4017 , 162.9330 , 164.4532 , 165.9628 , 167.4618 , 168.9506 , 170.4293 , 171.8982 , 173.3576 , 174.8075 , 176.2482 , 177.6799 , 179.1028 , 180.5170 , 181.9227 , 183.3202 , 184.7095 , 186.0907 , 187.4642 , 188.8299 , 190.1880 , 191.5388 , 192.8822 , 194.2185 , 195.5477 , 196.8700 , 198.1855 , 199.4943 ,
" s i g m a _e n e r g y _ m e v ": " s i g m a _e n e r g y _ m e v ": " s i g m a _e n e r g y _ m e v ": " s i g m a _e n e r g y _ m e v ": " s i g m a _e n e r g y _ m e v ": " s i g m a _e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ": " s i g m a _ e n e r g y _ m e v ":
1.9716 , 1.9130 , 1.8586 , 1.8081 , 1.7610 , 1.7169 , 1.6756 , 1.6367 , 1.6000 , 1.5654 , 1.5326 , 1.5015 , 1.4721 , 1.4440 , 1.4173 , 1.3918 , 1.3674 , 1.3441 , 1.3218 , 1.3004 , 1.2799 , 1.2602 , 1.2412 , 1.2229 , 1.2054 , 1.1884 , 1.1720 , 1.1562 , 1.1409 , 1.1261 , 1.1118 , 1.0980 , 1.0845 , 1.0715 , 1.0589 , 1.0466 , 1.0348 , 1.0231 , 1.0119 , 1.0010 , 0.9903 , 0.9800 , 0.9699 , 0.9601 , 0.9503 , 0.9409 , 0.9319 , 0.9230 , 0.9142 , 0.9057 , 0.8975 , 0.8892 , 0.8815 , 0.8736 , 0.8660 , 0.8585 , 0.8514 , 0.8442 , 0.8373 , 0.8304 , 0.8236 , 0.8170 , 0.8107 , 0.8043 , 0.7981 , 0.7919 , 0.7860 ,
" sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ": " sigma_spot_mm ":
7.0329} , 6.9130} , 6.7977} , 6.6868} , 6.5796} , 6.4764} , 6.3771} , 6.2814} , 6.1893} , 6.1006} , 6.0151} , 5.9327} , 5.8533} , 5.7768} , 5.7029} , 5.6316} , 5.5628} , 5.4966} , 5.4328} , 5.3712} , 5.3119} , 5.2547} , 5.1996} , 5.1465} , 5.0953} , 5.0459} , 4.9983} , 4.9525} , 4.9083} , 4.8657} , 4.8247} , 4.7852} , 4.7472} , 4.7105} , 4.6752} , 4.6412} , 4.6084} , 4.5768} , 4.5464} , 4.5172} , 4.4889} , 4.4618} , 4.4356} , 4.4104} , 4.3860} , 4.3626} , 4.3399} , 4.3181} , 4.2970} , 4.2767} , 4.2571} , 4.2381} , 4.2198} , 4.2021} , 4.1849} , 4.1682} , 4.1521} , 4.1364} , 4.1212} , 4.1063} , 4.0919} , 4.0779} , 4.0641} , 4.0507} , 4.0375} , 4.0246} , 4.0119} ,
C. Patient IDs with metal artifacts
{" energy_mev ": 200.7966 , " s i g m a _ e n e r g y _ m e v ": 0.7803 , " sigma_spot_mm ": 3.9994} ] }, " hu_to_density ": { " _note ": " HU - to - density calibration curve used by G4DCM . 10 anchor points linearly interpolated . Both photon and proton simulations use identical tables ." , " entries ": [ {" hu ": -1024 , " density_g_cm3 ": 1.200000 e -03} , {" hu ": -999 , " density_g_cm3 ": 1.210000 e -03} , {" hu ": -200 , " density_g_cm3 ": 8.043754 e -01} , {" hu ": -199 , " density_g_cm3 ": 8.183035 e -01} , {" hu ": -10 , " density_g_cm3 ": 1.006579 e +00} , {" hu ": -9 , " density_g_cm3 ": 9.966749 e -01} , {" hu ": 120 , " density_g_cm3 ": 1.126553 e +00} , {" hu ": 121 , " density_g_cm3 ": 1.095097 e +00} , {" hu ": 3000 , " density_g_cm3 ": 3.027294 e +00} , {" hu ": 4000 , " density_g_cm3 ": 3.698428 e +00} ] } }
C.
Patient IDs with metal artifacts
Metal artifacts were observed in the following training cases: • 1THB016 • 1THB021 • 1THB029 • 1THB031 • 1THB052 • 1THB054 • 1THB067 • 1THB078 • 1THB122 • 1ABB138 • 1THB214 • 1THB217 • 1ABB078
D. Plan JSON field descriptions
D.
Plan JSON field descriptions Table 7: Photon plan JSON field description. Field Per beam beam idx iso center num mlc leaf pairs control points Per control point cp idx gantry angle mlc left int mm mlc right int mm
Unit
Description
– mm – –
Sequential beam index (0-indexed) 3D coordinate of the beam rotation center in patient Number of leaf pairs to shape the segment Discrete beam delivery sampling points along the arc
– deg mm mm
Control point index within the arc (0-indexed) Angle of the photon beam direction around patient Left-edge positions of 80 collimator leaves Right-edge positions of 80 collimator leaves
Table 8: Proton plan JSON field description. Field iso center Per beam beam idx gantry angle rays Per ray ray idx ray source ray target beamlets Per beamlet beamlet idx energy
Unit mm
Description 3D coordinate of the beam rotation center in patient
– deg –
Sequential beam index (0-indexed) Angle of the proton beam direction around patient A group of pencil beamlets from the same BEV grid position within the beam
– mm mm –
Ray index within the beam (0-indexed) Source point of the ray on the source plane Aim point of the ray on the isocenter plane Energy layers for the ray
– MeV
Energy layer index (0-indexed) Proton energy, determining the penetration depth