A radiotherapy dose surrogate is a neural network that takes a patient's CT scan and the drawn outlines of the tumor and nearby organs, and predicts the 3D radiation dose that a slow treatment-planning optimizer would produce. Such surrogates are trained once, at a large upfront cost: every training example needs a full planning-solver run. They are then evaluated on one fixed set of outlines per patient. In practice, two clinicians outlining the same patient disagree by millimetres, and those millimetres decide whether an organ is spared. A surrogate that gives the right dose for one clinician's outlines can give the wrong dose change when a second clinician redraws them, because it never saw the same patient drawn two ways. This proposal asks how a fixed budget of solver runs should be split. One option is to spend it all on more patients. The other is to spend part of it on realistic re-drawings of the same patients, so the surrogate learns how dose actually moves when the outline moves.
The core claim is that, at a matched solver budget, a surrogate trained partly on a few realistic observer-style contour variants per patient will predict how dose changes between observers more accurately than one trained on more patients with fixed contours. This should hold without losing accuracy on the standard benchmark. The comparison that makes this testable is a new metric: how well the surrogate's predicted dose difference between two contour sets matches the solver's own difference. No benchmark currently measures this.
Core Claim
Current surrogates are judged on whether they reproduce the planned dose for one fixed set of contours. Clinical use, however, depends on how the dose responds when the contours change. We define, for a patient with two contour sets $c$ and $c'$, the sensitivity error
$$ S(c,c') = \frac{\lVert (f(c') - f(c)) - (D(c') - D(c)) \rVert_1}{\lVert D(c') - D(c) \rVert_1}, $$
where $f$ is the surrogate's predicted dose volume, $D$ is the dose produced by rerunning the planning optimizer on those contours, and the norm is summed over voxels, optionally restricted to organs at risk. A surrogate with $S$ near 1 has essentially learned nothing about contour dependence. It may still score well on ordinary dose error, because it can copy patient-level structure. We hypothesize three things. First, current surrogates have large $S$ on realistic observer variation. Second, $S$ falls sharply when a fraction of the solver budget goes to observer-style variants. Third, the gain depends on the variants being realistic, meaning spatially correlated and concentrated at low-contrast boundaries, rather than uniform erosion or dilation.
Relation to Prior Work
The surrogate side of the problem is set by the OpenKBP benchmarks. OpenKBP released 340 contoured head-and-neck cases with reference doses produced by an automated planning pipeline. OpenKBP-Opt paired 19 prediction models with 4 dose-mimicking optimizers, which are optimizers that turn a predicted dose into a deliverable plan. Both provide one standardized contour set per patient, which is exactly the assumption we relax. Because their reference doses come from an automated pipeline, the same pipeline can be rerun on perturbed contours. This makes OpenKBP the natural place to treat solver calls as a countable budget.
DiffDose conditions a diffusion model, a generator that denoises random noise into samples, on CT and contours. Chen et al. add a distance-to-target field computed from the drawn tumor outline, and lowered full-body dose error from 5.07% to 4.89%. Both treat contours as noise-free. The distance field in particular may amplify sensitivity to outline jitter, because shifting the tumor edge rewrites distances across the whole volume. We test this directly rather than assume it.
The variability side provides realistic perturbations, not the method itself. Dushatskiy et al. simulated observer styles with erosion, dilation and shifts on 32 prostate cases, and used a cheap proxy objective to separate styles. Their uniform perturbations serve as our "unrealistic variant" control. Pionono and the Latent Doctor Model represent each annotator as a low-dimensional latent vector, but only on 2D pathology. They never ask what downstream physics does with the sampled annotations. DiFF is the closest in spirit: it scores multi-rater labels by a downstream task. That task is a scalar classifier, though, and DiFF changes the labels rather than teaching a physical surrogate to respond to them. Our work points the other way: observer variation is an input distribution the surrogate must handle, and the cost of covering that distribution is paid in solver calls.
QUBIQ supplies 3D multi-rater CT and MRI masks, which we use to fit the variant generator. It also reports a constraint: ensembles captured rater disagreement better than single probabilistic models, but were expensive in 3D. This is a reason to keep the deployed surrogate a single network. SegTHOR and You et al. motivate the work by documenting clinically consequential contour disagreement for organs at risk and target volumes.
Research Program
Variant generator. We first build a generator of observer-style variants for 3D organ and target masks. Using the multi-rater 3D tasks in QUBIQ (kidney, pancreas and prostate), we fit a model of how rater surfaces differ. The model describes a displacement along the surface normal that is spatially smooth. Its amplitude is conditioned on local image contrast and on the structure type, and it has a per-rater offset that captures systematic "conservative" versus "aggressive" tendencies, following the inter- and intra-observer split of Pionono. The generator is deliberately simple, a smooth random field on the surface, so that its realism can be checked by comparing variance profiles against held-out QUBIQ raters. We then apply it to OpenKBP target and organ structures.
This transfer from abdomen and pelvis to head and neck is the weakest link. We address it by collecting a small set of real re-contours, described under evaluation.
Budgeted dataset construction. We define the budget $B$ as the number of runs of the OpenKBP automated planning pipeline. Each run takes one patient and one contour set and returns a reference dose. We then compare allocation rules at equal $B$:
- all runs on distinct patients with original contours;
- a fixed fraction of runs on $k$ generated variants per patient;
- the same, but with uniform erosion and dilation variants;
- an acquisition rule that runs the solver on variants where a cheap committee of surrogate checkpoints disagrees most about the dose change in organs at risk.
The acquisition rule is an ablation, not the headline. The headline is the allocation question itself.
Distance-field analysis. We train the surrogate with and without the distance-to-target input of Chen et al. and measure $S$ for each. This settles whether that input trades standard accuracy for contour fragility.
Baselines
The naive reference point is a standard OpenKBP-style dose surrogate, a 3D U-Net trained on the 200 training patients with their single contour set. At test time we simply feed it whatever contours an observer drew. This is the "do nothing" case: it shows how large $S$ is today, and our method must keep its ordinary dose error while lowering $S$.
The strongest existing method is the best reported OpenKBP-Opt pipeline, a prediction model paired with a dose-mimicking optimizer. OpenKBP-Opt reports 77.0% ROI clinical-criteria satisfaction for this pipeline, versus 66.2% for raw predictions and 69.7% for reference plans. Its MaxAbs optimizer had the lowest median dose error, 2.34 Gy. We do not have every challenge model. We therefore reimplement a top-ranked architecture with matched parameter count, training epochs and hyperparameter search budget. We pass its predictions through the same MaxAbs optimizer for every arm. All arms use identical patients for testing and an identical solver budget $B$ for training-label generation.
Beating this baseline means a statistically significant reduction in $S$ on observer-swapped test contours, by paired test across patients. It also means non-inferior dose score and criteria satisfaction on the original OpenKBP test contours. No reported result was found on Lacuna for dose-surrogate sensitivity under contour perturbation, so the baseline's $S$ will be measured here.
Three ablations isolate the key ingredient. The first uses uniform erosion and dilation variants in the style of Dushatskiy et al., at the same budget, to test whether realism matters. The second is "label-preserving augmentation": contours are perturbed but the original dose is kept as the target, which costs no extra solver calls. This tests whether new solver labels are needed at all. The third removes the distance-field input.
Why not just…?
Why not just train on more patients? More patients improve average dose error, and it is plausible that a surrogate generalizes across patients well enough to also get contour sensitivity right. This is the main threat. The equal-budget comparison settles it. If the all-patients arm matches the variant arms on $S$, the claim fails, and we will report that budget is better spent on patient diversity. Our reason for doubt is that between-patient variation is dominated by anatomy, while within-patient contour changes are small and locally confined. A network has no training signal tying the second to dose unless it sees the same patient drawn twice.
Why not just augment contours without rerunning the solver? It is free, and it is the obvious move from segmentation practice. But it teaches the surrogate that dose does not depend on contours. That is precisely the wrong behavior when an observer enlarges a target next to the esophagus or optic chiasm. We expect this arm to make $S$ worse while appearing robust, and we include it to show that this failure is real.
Why not just sample contours at test time and push them through the existing surrogate? Propagating multi-rater samples, for example from a Pionono-style latent model, gives a dose distribution at the cost of only a few forward passes. However, this distribution is only as good as the surrogate's response to contour changes, which is the quantity we suspect is wrong. We run exactly this propagation with the naive and the variant-trained surrogates. We then compare the spread of predicted dose against the spread of solver doses for the same contour samples.
Why not just add a fixed safety margin? Margins are already standard clinical practice, but they are a planning policy, not a model of what dose a given outline produces. The file's evidence from SegTHOR and You et al. suggests that margins neither remove clinician inconsistency nor protect adjacent organs. Our evaluation does not depend on how margins are set.
Proposed Experiments and Decisive Checks
All dose training and testing uses OpenKBP: 200 training, 40 validation and 100 test patients. We sweep $B$ from the 200 original runs up to several times that number. In each arm the extra runs go either to synthetic patient diversity (resampling and augmenting existing patients' CT with original contours) or to contour variants. A cleaner variant of this comparison caps the number of distinct patients and spends the budget either way.
Evaluation contours come from two sources. The first is generated variants from held-out generator settings. The second, which is decisive, is a small real re-contouring study: several clinicians independently redraw targets and key organs at risk on a subset of OpenKBP test patients, and we rerun the planning pipeline on each set. We report the following measures:
- $S$ restricted to organs at risk;
- dose score and DVH-point error on the original contours;
- ROI criteria satisfaction after MaxAbs optimization;
- the rank correlation between surrogate-predicted and solver-predicted ordering of observers by organ dose. This is the clinically relevant question of whether the surrogate can tell which clinician's outline will overdose the cord.
The generator is validated on held-out QUBIQ raters by comparing per-structure surface-distance distributions.
The proposal succeeds if three things hold on real re-contours: variant-trained surrogates lower $S$ relative to both the naive surrogate and the all-patients arm at equal $B$; realistic variants beat uniform ones; and original-contour accuracy is non-inferior. It fails cleanly if the all-patients arm matches it. In that case the useful finding is that existing surrogates already track contour changes and the extra solver cost is unnecessary.
References
- OpenKBP: The open-access knowledge-based planning grand challenge
- OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines
- Diffusion-based Radiotherapy Dose Prediction Guided by Inter-slice Aware Structure Encoding
- Combining distance and anatomical information for deep-learning based dose distribution predictions for nasopharyngeal cancer radiotherapy planning
- Observer variation-aware medical image segmentation by combining deep learning and surrogate-assisted genetic algorithms
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- The Latent Doctor Model for Modeling Inter-Observer Variability
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- Segthor: Segmentation of Thoracic Organs at Risk in CT Images
- Deep Learning-Aided Automatic Contouring of Clinical Target Volumes for Radiotherapy in Breast Cancer After Modified Radical Mastectomy