Source-linked AI summary
BEAM3R: Beam's-eye-view architecture with Mamba-3 for implicit dose reconstruction
Chen Cheng, Michael Ferraro, James Grover, David E J Waddington, Emily Hewson
TL;DR
Rapid photon and proton dose calculation is needed because Monte Carlo accuracy is computationally costly for time-sensitive radiotherapy, motivating the DoseRAD2026 challenge. BEAM3R uses BEV CNN-sequence models with task-specific depth modelling and synthetic CT generation for MRI inputs, achieving similar CT-based performance but larger degradation for MRI-based proton prediction.
Problem
Monte Carlo is the radiotherapy dose-calculation gold standard, but its computational cost limits use in time-sensitive applications such as online adaptive and dose-guided radiotherapy.
Method
BEAM3R uses a shared BEV CNN-sequence framework with task-specific photon and proton cores and inputs, while MRI pipelines first generate synthetic CTs with a SwinUNETR-based conditional GAN.
Results
CT-based photon and proton models achieved approximately 96% local gamma pass rates, while replacing CT with sCT reduced rates by 7.3% for photon and 21.4% for proton prediction.
Takeaways & Limitations
BEAM3R provides similar CT-based photon and proton performance, whereas sCT accuracy is a major limitation for MRI-based proton dose calculation.
Takeaways & Limitations
The sCT model may depend on scanner and acquisition characteristics, while simplified RSP and range approximations may not generalise beyond the challenge energies and anatomy.
Abstract
from arXiv · showhide
To enable accurate and rapid photon control point and proton beamlet dose calculation in the DoseRAD2026 challenge, we present BEAM3R, a dose estimation framework operating in beam's-eye-view (BEV). Our core innovation combines a Mamba-3 state-space depth-sequence core with physics-based transport conditioning to model long-range depth transport without expensive 3D convolutions. BEAM3R shares a 2D CNN encoder-decoder architecture for photon and proton dose tasks, processing per-plane BEV slices. Proton beamlets are conditioned on water equivalent thickness and remaining range, encoding the parameters determining Bragg peak position. Photon models use a bidirectional Mamba-3 core to capture dose contributions from materials downstream of the calculation point, while the proton model uses a forward core with learned energy-prefix tokens and a Bragg-peak refinement module. To reduce interpolation artifacts and support high spatial resolution, we introduce axial grid alignment of BEV lattices with CT slices and an implicit super-resolution representation via sub-pixel phase packing, evaluated by a differentiable Triton-accelerated resampler that reconstructs packed cubic B-spline coefficients directly in CT space. For MRI-based tasks, synthetic CTs (sCT) are generated by a patch-based conditional GAN with a SwinUNETR backbone. On the preliminary DoseRAD2026 test set, CT-to-photon and CT-to-proton models achieved 1%/1 mm local gamma pass rates of 96.8% and 96.0%, with stratified plan-level MAEs of 0.0041 and 0.0079. Substituting sCT reduced gamma pass rates to 89.7% for photon and 75.4% proton plan level doses, with stratified plan-level MAEs of 0.0093 and 0.0336. Standardised runtimes were 23.4 s and 18.4 s for CT-to-photon and CT-to-proton prediction, increasing to 39.7 s and 42.8 s for the corresponding MRI-based pipelines.
1 Introduction
BEAM3R targets rapid beam-level photon and proton dose calculation by combining BEV sequence modelling with task-specific inputs and a shared CNN framework. The approach is designed for CT and MRI-based challenge tasks, where MRI workflows first generate synthetic CTs.
- Monte Carlo is the radiotherapy dose-calculation gold standard, but its computational cost limits time-sensitive applications.
- The DoseRAD2026 challenge benchmarks rapid photon and proton beam-level dose calculation from CT and MRI.Beam-level modelling supports beam-to-anatomy dose relationships for plan optimisation and dose accumulation workflows.
- BEV slice-sequence models represent transport along a common beam-depth axis, reducing the need to encode treatment geometry explicitly.BEAM3R formulates 3D dose calculation with a 2D CNN encoder, depth-sequence core, and 2D CNN decoder.
- BEAM3R modifies a BEV CNN-sequence architecture for beam coverage, resampling accuracy, model capacity, and downstream depth context.Photon and proton models share the framework but use task-specific sequence cores and inputs, while MRI tasks first generate synthetic CTs with a SwinUNETR-based conditional GAN.
2 Methods
BEAM3R uses a shared BEV CNN-sequence framework for photon control-point and proton beamlet dose prediction, with task-specific inputs, sequence processing, and transport conditioning. The pipeline combines aligned BEV-to-CT reconstruction, synthetic CT generation for MRI tasks, staged training, and physics-informed losses.
- Data and model selection: Patient-level splits separated validation patients from training, with separate models trained for each task and anatomical cohort.Hyperparameters and checkpoints were selected using held-out validation cohorts.
- Architecture: BEAM3R builds on a shared BEV CNN-sequence architecture with a strided encoder, depth-sequence core, and slice-wise decoder for photon control points and proton beamlets.The framework predicts photon dose from one control point and proton dose from one beamlet while using task-specific components.
- Beam-specific inputs: Proton inputs combine CT, Gaussian spot fluence, cumulative WET, and remaining range to encode transport conditions associated with Bragg-peak position.WET uses an energy-dependent stopping-power estimate, while range uses the CSDA approximation over 31.729–200.797 MeV.
- Depth-sequence modelling: Mamba-3 sequence cores propagate information along beam depth in a fixed-size recurrent state with linear-time cost, using SISO blocks for both dose models.The encoder rearranges per-slice latent features into depth-ordered sequences before sequence processing.
- Reconstruction: The decoder predicts four phase-packed cubic B-spline coefficient channels, representing a two-fold finer dose field along depth and height without materialising a high-resolution BEV volume.The BEV lattice is snapped to CT slice positions to reduce resampling error, while packed coefficients support CT-space reconstruction.
- MRI pipeline: MRI tasks first generate synthetic CTs with a SwinUNETR-based conditional GAN, whose outputs are passed to the corresponding CT-based dose model.The generator operates on 64 × 160 × 160 patches, with body-masked optimisation and subject- or population-based standardisation.
- Training: Dose models use staged optimisation with MSE and masked MAE objectives, while proton training adds a non-negativity penalty within the valid sampling region.The mask selects voxels receiving at least 10% of the maximum reference dose; training uses Adam, batch size 4, gradient clipping, and bfloat16 mixed precision.
3 Results
On the preliminary DoseRAD2026 test set, CT-based photon and proton models achieved similar beam-level accuracy, while MRI-based proton performance degraded substantially. Development comparisons favoured larger BEV extent, model capacity, bidirectional depth modelling, and interpolation-reducing methods, but pairwise results require within-pair interpretation.
- Challenge performance: 96.8% and 96.0% local gamma pass rates were achieved by the CT-based photon and proton models, respectively.Both models had masked MAE of 0.0084.
- Challenge performance: 75.4% MRI-to-proton local gamma pass rate followed a decrease from 96.0% for CT-to-proton prediction.MRI-to-proton masked MAE increased from 0.0084 to 0.0326.
- Challenge performance: 39–43 s MRI runtimes exceeded the approximately 22–23 s CT runtimes because of synthetic-CT inference.Local RTX A6000 runtimes were 16.7 s for 180 photon beams and 8.176 s for 300 proton beamlets.
- Ablation experiments: Ablation results should be interpreted within pairs because cohorts and secondary metrics varied across comparisons.The reported table uses different metric sets, including BEV-space and CT-space measures.
- Ablation experiments: 8.5% lower masked MAE and 30.0% lower IDD distance resulted from increasing thorax BEV depth from 256 to 384 planes.Doubling model capacity reduced masked MAE by 15.5%, while adding the reverse-depth Mamba branch reduced masked MAE by 5.3% and IDD distance by 20.9%.
- Qualitative error analysis: The qualitative examples show largest discrepancies at peak dose deposition, proton range shifts, and MRI-related errors coinciding with inaccurate soft-tissue and bone Hounsfield units.The proton range shift was consistent with the spatial difference map, while MRI sCT examples showed striation artifacts.
4 Discussion
Performance was approximately 96% for CT-based photon and proton prediction but decreased after replacing CT with synthetic CT, especially for proton prediction. The discussion links this sensitivity to proton range dependence and reports architectural benefits alongside important limits on interpreting development experiments.
- Performance across input modalities: MRI-to-proton gamma pass rate decreased 21.4% relative to CT-to-proton, compared with a 7.3% decrease for MRI-to-photon.CT-based models had local gamma pass rates of approximately 96% and normalised masked MAE below 0.9% of the reference beam maximum.
- Performance across input modalities: Proton prediction is more sensitive to synthetic-CT errors because integrated stopping-power errors can shift the Bragg peak.Synthetic-CT errors also affect proton inputs through WET and remaining-range volumes derived from the synthetic CT.
- Architectural design choices: Supersampling, increased BEV extent, model width, and bidirectional depth processing were associated with the largest observed development improvements.Grid snapping and CT-space training were used to reduce interpolation error between BEV and CT space.
- Qualitative error analysis: Figure 2 compares CT and MRI-input photon and proton predictions using dose maps, difference maps, and IDD curves.MRI panels additionally compare MR input, synthetic CT, and real CT; the examples illustrate failure modes from held-out validation data.
- Architectural design choices: Development effect sizes are not directly comparable or attributable to single components because cohorts, metric implementations, and additional settings varied.This limits interpretation of the reported associations across architectural comparisons.
- Limitations: Validation cohorts contained only two or three patients per task, preventing reliable confidence intervals or formal patient-level significance testing.The limitation applies to the validation evidence supporting the reported comparisons.
- Limitations: Proton fine-tuning was not completed on the full training cohort, potentially limiting anatomical generalisability.The authors also note possible dependence of synthetic-CT performance on scanner and acquisition characteristics.
5 Author Contributions
The listed contributors jointly covered supervision, resources, conceptualisation, methodology, software, validation, analysis, investigation, data curation, writing, and visualisation.
- D.E.J.W. contributed writing review and editing, supervision, and resources.
- E.H. contributed writing review and editing and supervision, while C.C. contributed conceptualisation, methodology, software, validation, analysis, investigation, data curation, writing, and visualisation.
- M.F. contributed conceptualisation, methodology, software, validation, analysis, investigation, writing, and visualisation.
- J.G. contributed conceptualisation, methodology, software, validation, investigation, and writing the original draft.
6 Other Information
The paper acknowledges scholarship, grant, equipment, and project-grant support and reports no competing interests.
- The authors acknowledge Australian Government scholarships, NHMRC and University of Sydney support, and a Cancer Council NSW Project Grant.
- The authors declare no interests to disclose.