Neural networks can now predict the 3D radiation dose a treatment plan will deliver in well under a second. To fit in GPU memory, though, they usually see the patient either as a coarse, downsampled grid or as small independent tiles. Both shortcuts discard information exactly where dose is hardest to get right: at the thin boundaries between lung, bone, soft tissue and air. At those boundaries, charged particles scatter sideways and over distances that depend on tissue density. We propose to tile the CT volume at its native resolution and to give each tile a surrounding context margin sized by physics rather than by a fixed voxel count. The margin is the distance secondary particles can actually travel through the local tissue, so it is wide in lung and narrow in bone. The goal is a dose predictor that keeps accuracy at density interfaces, runs on a single consumer GPU, and reports its own errors separately for interface and interior regions.
The opportunity comes from two observations in the literature. First, the strongest error signal is local to boundaries: lung stereotactic dose corrections depend on tissue density just inside and just outside the target margin, not on coarse geometric descriptors (Target dose conversion modeling from pencil beam (PB) to Monte Carlo (MC) for lung SBRT). Second, the main open benchmark deliberately coarsened its grid to 3 mm × 3 mm × 2 mm so that groups worldwide could take part (OpenKBP). Accessibility and interface fidelity are currently traded against each other. We argue that the trade is not necessary, and that a whole-volume error average hides the cost of making it.
Core Claim
Our claim has two parts. The first is a measurement claim: voxel-averaged dose error understates the failure of current dose predictors at density interfaces, and models trained on downsampled grids or on fixed-margin tiles show interface-band errors several times their interior errors. The second is a method claim. Monte Carlo practice already pads patches by the maximum particle range to remove edge artifacts: DosePatch sets a buffer equal to the 11.3 mm maximum range of yttrium-90 beta particles (DosePatch). We transfer this physical-buffer principle to a learned predictor and make the buffer density-adaptive. For a tile boundary point $x$ and direction $\hat{u}$, the geometric margin $m(x,\hat{u})$ is the smallest distance satisfying
$$\int_0^{m(x,\hat{u})} \rho(x + s\hat{u})\, ds \;\ge\; R_w,$$
where $\rho$ is the relative electron (or stopping-power) density derived from the CT and $R_w$ is the water-equivalent range of the secondary particles that carry dose sideways. In practice $R_w$ is a few millimetres to centimetres, depending on beam type and energy. Because $\rho$ in lung is roughly a quarter to a third of water, the margin there grows accordingly, and this is exactly where fixed-voxel context falls short.
The closest prior work does one piece of this but not the combination. DosePatch uses a range buffer, but its buffer is fixed and it runs Monte Carlo, taking about 34 hours per patient. The beam-mask predictor uses native-anatomy 3D inputs and random 3×3×3 masking to force local interface learning (Beam mask and sliding window…), but it gives no context-sizing rule. That study reports a lung body gamma pass rate of 92.79% against 95.17% for prostate, which is the gap we target. DoTA handles beam directionality with attention over 2D latent slices (Learning the Physics of Particle Transport via Transformers) and does not model out-of-plane scatter.
Research Program
We will build three components and release them together as a pip-installable tool that runs on a single 24 GB GPU.
The tiler. It computes per-face margins by marching the integral above along the six axis directions, plus the beam direction for proton fields. It uses CT-to-density conversion with material overrides for known artifacts; the iodinated-spacer study shows that water-equivalent override is the right fix there (Systematic study of the iodinated rectal hydrogel spacer…). Margins are rounded up to a small set of discrete sizes so that tiles can be batched.
The predictor. It is deliberately ordinary: a 3D U-Net that takes CT, structure masks and beam-geometry channels in the style of the spatial "plates" in Flexible-Cm GAN. It predicts dose only in the tile core, and the margin acts purely as context. Keeping the network plain lets the context rule, not the architecture, carry the result.
The audit layer. This is the accessibility contribution. Fully abstracted tools hide intermediate failures, so every prediction ships with a map of interface bands. A band is the set of voxels within $R_w$ (in water-equivalent distance) of a density jump above a threshold. The tool reports error or disagreement summaries for band and interior regions separately, so a clinician or student without physics training can see where the output should not be trusted.
Ground truth and datasets
Interface error cannot be measured against downsampled or non-physical references. OpenKBP's reference doses were not produced by Monte Carlo transport. We will therefore generate native-resolution Monte Carlo ground truth with MCsquare, the engine DoTA used at 3 million particles per beam. We will compute proton pencil-beam and full-field doses on thoracic CTs, which have the most heterogeneity, and on pelvic CTs as a homogeneous control.
We will also report on OpenKBP at its native benchmark resolution, so that results connect to a known leaderboard, and on the AIMIS 500-patient lung cohort where we can obtain access. The Mayo proton dataset used by the beam-mask paper is the most direct comparison, but it is not known to be public. If we cannot obtain it, we will retrain that baseline on our data.
Baselines
The naive reference is a whole-volume 3D U-Net trained on a 3 mm-class downsampled grid and trilinearly upsampled to native resolution. This is the OpenKBP regime. Its error inside the interface bands is the quantity we expect to be inflated. As an anchor, LENAS reports an OpenKBP test Dose Score (voxel MAE) of 2.619 Gy for a standard U-Net and 2.601 Gy for ResUNet.
The strongest existing coarse-grid method is the LENAS ensemble, which reaches 2.357 Gy on the same test set (2.565 Gy after distillation). On OpenKBP we will compare against these published numbers directly. On our Monte Carlo data we will retrain a single-model U-NAS-style network under our compute budget. We will not retrain the full ensemble, and we will say so in the results.
The strongest interface-specific baseline is the beam-mask and sliding-window 3D U-Net (Beam mask and sliding window…), reported at 92.79% lung body and 96.55% lung target gamma pass rate (3%/3 mm). We will retrain it on our thoracic data with its 3 mm distal-margin ray-traced masks.
For transport-specific comparison we include DoTA, which reports a 99.59% gamma pass rate on single pencil beams in an unseen lung patient. We will compare against it only in the single-pencil-beam setting it was designed for.
The key ablation is a native-resolution tiled predictor whose margin is fixed in voxels, at two settings. The first is a margin equal to the median adaptive margin. The second is a margin equal to the maximum adaptive margin, which is the lung-sized buffer applied everywhere. This ablation isolates whether density adaptivity matters or whether context size alone does.
Fairness controls apply to all baselines. Every model sees identical training patients. Each gets the same 24 GB GPU and wall-clock budget and an equal hyperparameter search of 20 trials. All are scored with the same band definitions computed from ground-truth density.
Beating the baselines means the following. On the thoracic test set, interface-band MAE must be at least 25% lower than both the downsampled baseline and the median fixed-margin tile. Interior MAE and global gamma must be no worse. On OpenKBP, global Dose Score must stay within the LENAS single-model range.
Why not just…?
Why not just keep the downsampled grid and a stronger model, as OpenKBP entrants did? A coarse grid cannot represent a dose gradient narrower than its voxel, so no model capacity recovers it. The cost appears only if error is measured where the gradient lives. Our first experiment compares band and interior error for the LENAS-style and U-Net baselines. If band error at native resolution is not meaningfully worse than interior error, this proposal's premise fails, and we would report that as a negative result.
Why not just use a large fixed margin everywhere? This is the real threat. A uniform lung-sized buffer may match adaptive margins in accuracy. The max-fixed-margin ablation settles it. If it matches accuracy at acceptable memory and latency, we will drop adaptivity and present the contribution as range-sized fixed padding plus the interface audit, which is still new for learned predictors. Adaptivity is justified only if it gives equal band accuracy with substantially fewer context voxels, and therefore larger cores or lower latency on commodity hardware.
Why not just use a long-range architecture at full resolution, such as a Mamba state-space model? Swin-UMamba gets global context at linear cost through 2D selective scans. Its efficient variant drops to 18.9G FLOPs, but it has been tested only on segmentation. Its scans are 2D, which reintroduces the through-plane loss that slice-wise models show, as documented in Deep convolutional networks for pancreas segmentation in CT imaging. We will run a 3D-adapted Mamba U-Net as an extra arm under the same budget. If it matches tiled prediction at interfaces, then the tiler's value reduces to memory savings.
Why not just add an interface-weighted loss or a DVH-moment loss? Moment losses improved DVH score by 11% on a 128³ lung grid (Domain Knowledge Driven 3D Dose Prediction…). However, a loss cannot supply anatomy the network never sees. We will add an interface-weighted loss to both the baseline and our method to test whether loss and context are complementary or substitutes.
Why not just wrap an LLM agent around an existing predictor? 3DMedAgent and CellAgent show that conversational orchestration works for perception and transcriptomics. But 3DMedAgent compresses volumes into slice-level text notes, and a dialogue layer over a predictor that is silently wrong at interfaces makes the failure easier to miss. A trustworthy, auditable predictor is the prerequisite for any such interface, so we leave the agent out of scope.
Proposed Experiments
The first decisive check is the measurement claim. For each baseline on Monte Carlo thoracic doses, we report the ratio of interface-band to interior MAE and band-restricted 2%/2 mm gamma. We stratify bands by interface type (lung–tumour, bone–lung, air–tissue) and use the pelvis as a control. We predict the ratio will be well above one in thorax and near one in pelvis.
The second check is the method claim. We compare the adaptive tiler against the two fixed-margin ablations and the Mamba arm, plotting band error against context voxels per core voxel. Adaptivity must dominate this frontier for the claim to stand.
The third check tests whether interface error is the clinically relevant signal. OpenKBP-Opt found only a 0.50–0.62 rank correlation between dose-prediction quality and final plan quality after dose mimicking (OpenKBP-Opt). We will pass every model's predictions through the open MaxAbs and MeanAbs mimicking optimizers released there. We then test whether band error predicts post-mimicking DVH quality for structures touching interfaces better than global MAE does.
Finally, we measure peak GPU memory and latency per patient on a 24 GB card. A clinically usable tool should predict a full native-resolution volume in seconds; the beam-mask model runs in under 0.25 s on a downsampled-equivalent workload. We also run a small usability study in which dosimetry trainees use the audit maps to flag the regions where predictions were actually wrong.
References
- Target dose conversion modeling from pencil beam (PB) to Monte Carlo (MC) for lung SBRT
- OpenKBP: The open-access knowledge-based planning grand challenge
- DosePatch: physics-inspired cropping layout for patch-based Monte Carlo simulations to provide fast and accurate internal dosimetry
- Beam mask and sliding window‐facilitated deep learning‐based accurate and efficient dose prediction for pencil beam scanning proton therapy
- Learning the Physics of Particle Transport via Transformers
- Systematic study of the iodinated rectal hydrogel spacer material discrepancy on accuracy of proton dosimetry
- Flexible-Cm GAN: Towards Precise 3D Dose Prediction in Radiotherapy
- LENAS: Learning-based Neural Architecture Search and Ensemble for 3D Radiotherapy Dose Prediction
- Swin-UMamba: Mamba-Based UNet with ImageNet-Based Pretraining
- Deep convolutional networks for pancreas segmentation in CT imaging
- Domain Knowledge Driven 3D Dose Prediction Using Moment-Based Loss Function
- 3DMedAgent: Unified Perception-to-Understanding for 3D Medical Analysis
- CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis
- OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines