Radiotherapy plans are judged against clinical criteria. A typical criterion reads "spinal cord maximum dose below 45 Gy." Each criterion is met, narrowly missed, or badly violated, and the patient on the treatment couch never sits exactly where the planning CT says they do. Today's automated planners optimize a single nominal scenario and report how many criteria that one scenario satisfies. We propose instead to tune the planner's objective weights so as to maximize the probability, under simulated setup and anatomical perturbations, that each clinical criterion lands in an acceptable tier.
The surrogate is the key change. Instead of a Gaussian process (GP), the usual smooth probabilistic regression model in Bayesian optimization (BO), we learn the full distribution of clinical metrics with a conditional normalizing flow, an invertible neural network that turns Gaussian noise into samples of a target distribution. We train it on perturbed simulations, as in neural likelihood estimation. Because flow samples are a differentiable function of the noise and the planning parameters, the reparameterization trick that makes BoTorch-style acquisition functions differentiable carries over unchanged. Candidate plans can then be found with exact gradients rather than derivative-free search.
The opportunity is concrete. On the OpenKBP head-and-neck benchmark, the best of 76 published planning pipelines satisfies 77.0% of clinical criteria in the nominal scenario (OpenKBP-Opt). Nobody in the file reports how that figure holds up once the patient shifts by a few millimetres. In that regime, metrics near steep dose gradients plausibly become skewed or bimodal: an organ edge either slips into the high-dose region or it does not. Those are exactly the distributions that a GP, which summarizes each metric by a mean and a variance, describes worst.
Core Claim
Our hypothesis is specific. When the objective is "probability of reaching clinical tier k under perturbation," a flow surrogate of the joint metric distribution should give better robust Pareto fronts per simulator call than a GP on summary moments. We also expect it to beat models that predict each tier's satisfaction independently, and to win only where metric distributions are non-Gaussian.
Write $\theta$ for the planning weights. Write $y_j$ for clinical metric $j$ (for example cord maximum dose) after a random perturbation. Write $t_{j,k}$ for the threshold separating tier $k$ from tier $k-1$. The objectives are tier-exceedance probabilities:
$$ q_{j,k}(\theta) = \Pr\big(y_j \text{ meets } t_{j,k} \mid \theta\big) \approx \frac{1}{M}\sum_{m=1}^{M} \sigma\!\left(\frac{t_{j,k} - T_\phi(\epsilon^m;\theta)_j}{\tau}\right). $$
Here $T_\phi$ is the flow with parameters $\phi$. The $\epsilon^m$ are $M$ fixed base noise draws. The function $\sigma$ is a logistic sigmoid with temperature $\tau$ that smooths the tier indicator.
The ordinal structure comes for free from this construction. All tiers of criterion $j$ are read off the same sampled metric, so $q_{j,1} \ge q_{j,2} \ge \dots$ by construction. This is the non-crossing guarantee that Noncrossing Ordinal Classification needs quadratic or mixed-integer programming to enforce. Fixing the base draws makes the estimate deterministic and smooth in $\theta$, which is the sample average approximation used in BoTorch and Maximizing acquisition functions for Bayesian optimization. We then group criteria into at most four objectives (target coverage, serial organs, parallel organs, low-dose spill) and maximize differentiable expected hypervolume improvement over them (qEHVI). The cap of four keeps us below the five-to-six-objective point where its box decomposition explodes.
Position Relative to Prior Work
Three works motivate the gap. High-dimensional Automated Radiation Therapy Treatment Planning via Bayesian Optimization wraps GP-based BO around a commercial planning system over 34 weights. It collapses all criteria into one fixed linear score, ignores geometric uncertainty, and needs about 7 hours per patient with SAAS-BO, a GP variant with a sparsity prior fitted by MCMC; standard GP Expected Improvement needs about 3. We keep its framing of planning as black-box weight tuning. We replace the scalar score with tier probabilities under perturbation and replace its surrogate.
OpenKBP and OpenKBP-Opt provide the benchmark, the dose-influence matrices and the clinical-criteria evaluation. They also show that voxel dose error ranks pipelines only loosely against plan quality, with a rank correlation of 0.50–0.62. That finding motivates optimizing clinical criteria directly rather than dose error. Both papers are deterministic and nominal-scenario only.
Ordinal Bayesian Optimisation shows that ordinal latent models help on discontinuous objectives. It does so by discarding absolute scale and searching over discrete cells, so it has no input gradients and handles only a single objective. We keep absolute dose scale, because tier thresholds are in Gy, and recover gradients.
One body of work supplies the method. Sequential Neural Likelihood learns $p(\text{output}\mid\text{parameters})$ with autoregressive flows over sequential rounds of simulation. In SNL those rounds are proposed from the current posterior. In our setting they are proposed by the acquisition function, so the same likelihood-learning machinery is pointed at optimization rather than inference.
Another constrains the design. The Poisson Ordinal Network shows that a one-parameter ordinal head ties its variance to its mean and fits overdispersion poorly. That warns against compact parametric tier models and is part of why we model the metric distribution itself. Finally, Automatic Posterior Transformation reports that neural estimators need far fewer simulations than SMC-ABC, a sequential Monte Carlo variant of approximate Bayesian computation, which rules out rejection-style approaches at our budgets.
Research Program
The simulator is built on OpenKBP's 340 head-and-neck patients (200 train, 40 validation, 100 test).
Simulator. For given objective weights $\theta$ (on the order of 20–35 dimensions, one weight and one dose level per structure objective), an inner convex fluence optimization runs over the patient's dose-influence matrix. The resulting plan is then re-evaluated under sampled rigid setup shifts and small structure-mask dilations or erosions. Each call returns the clinical metrics for one perturbation draw. Running a call with many draws is cheap relative to re-solving the fluence, so every solve yields a small batch of perturbed metric vectors.
Surrogate and training. The flow conditions on $\theta$ and is trained on all collected pairs. Its training can be warm-started across patients by also conditioning on a structure-geometry embedding, which gives amortization in the BayesFlow sense (BayesFlow 2.0). To quantify epistemic uncertainty, meaning uncertainty from limited data rather than from perturbations, we use a deep ensemble of flows. The acquisition integrates over ensemble members as well as over base noise, because a single flow has no notion of what it has not seen.
Analysis. We will characterize where metric distributions are non-Gaussian, measuring per-criterion skewness and bimodality across patients. This tells us in advance where the method should and should not help.
Baselines
The naive reference is the plan produced by the best OpenKBP-Opt pipeline: prediction followed by dose mimicking, with no weight tuning. It is re-evaluated under our perturbation model. Its reported nominal clinical-criteria satisfaction is 77.0%, against 69.7% for reference plans (OpenKBP-Opt). No robust (perturbed) satisfaction number is reported on Lacuna for this pipeline, so we will compute it. A second naive reference is Sobol random search over $\theta$ at the same simulator budget.
The strongest GP competitor, and the one we expect to be hardest to beat, is the moment-matching baseline. A multi-output GP models the per-criterion mean and log standard deviation of each metric across perturbation draws. Tier probabilities come from Gaussian tails, and qEHVI with L-BFGS optimizes the same four objectives. It shares our simulator calls, perturbation draws, objective grouping and acquisition optimizer, so the only difference is the surrogate.
We also run the scalarized approach of High-dimensional Automated Radiation Therapy Treatment Planning via Bayesian Optimization. That means GP Expected Improvement and SAAS-BO on a weighted score, here computed from robust criterion satisfaction. On its rectal dataset the file reports only significance tests and runtimes (about 7 hours versus 3 for 120 iterations), not score values, so it is rerun on OpenKBP.
As a non-surrogate competitor we run multi-objective evolutionary search directly on the simulator, in the style of MO-RV-GOMEA from BRIGHT. A variant of Ordinal BO on the scalarized robust score tests whether rank-only modeling suffices.
Three ablations isolate the ingredients. The first keeps the flow but predicts each tier probability with an independent classifier head, which allows crossing; this tests the shared-latent ordinal construction. The second keeps the flow but maximizes the acquisition with CMA-ES, an evolutionary derivative-free optimizer, rather than gradients; this tests whether differentiability matters. The third uses a single flow without the ensemble; this tests epistemic uncertainty.
All methods get the same simulator-call budget (100, 200 and 400 fluence solves per patient), the same 32 perturbation draws per solve, the same initial Sobol design, and equal hyperparameter tuning on the 40 validation patients. "Beating" the moment baseline means a higher mean final hypervolume of tier-probability objectives over the 100 test patients, assessed with a paired Wilcoxon test. It also means at least 3 points higher robust satisfaction of the "acceptable" tier, the same order as the 3.5-point gap that OpenKBP-Opt treats as meaningful between references and predictions.
Why not just…?
Why not just fit GPs to the continuous dose metrics and threshold them afterward? This is the moment baseline above, and it is a real threat. If metric distributions under perturbation are close to Gaussian, it will match us at lower cost, and we will report that. The bet is that for organs abutting steep gradients the distribution is skewed or bimodal, so Gaussian tails misstate exactly the tier-boundary probabilities that matter. The pre-registered non-Gaussianity analysis predicts per criterion where the gap should appear. The claim is supported only if gains concentrate on those criteria.
Why not just model each tier's satisfaction probability directly with a classifier per tier? Independent per-tier models can cross. They can say a stricter tier is more likely than a looser one, which Noncrossing Ordinal Classification shows is a common failure that needs heavy machinery to fix. Our ablation with independent heads measures how often crossing occurs and whether it misleads the acquisition.
Why not just optimize the acquisition with CMA-ES or random restarts? For GP surrogates, exact gradients were an order of magnitude faster and more reliable than CMA-ES, and qEHVI was over 10x faster than PESMO, another multi-objective acquisition method (qEHVI). Whether that transfers to flow surrogates, whose acquisition surfaces may be rougher, is untested. The CMA-ES ablation answers it directly in both wall-clock time and final hypervolume.
Why not just do scenario-based robust fluence optimization with fixed weights? Dose is linear in fluence, so minimax over a set of shift scenarios is convex and needs no surrogate. It remains a strong competitor and enters as an extra baseline using clinical template weights. It does not, however, choose the trade-off between structures. The weight-tuning layer above it stays a black-box problem, and our method can wrap a robust inner solver as easily as a nominal one. If robust-inner plus template weights matches our Pareto front, the tuning layer is not earning its cost.
Decisive Checks
Four results would settle the question. The first is the robust-versus-nominal gap for the OpenKBP-Opt reference: if perturbation barely changes criteria satisfaction, the problem is not worth solving. The second is the flow-versus-moment-GP hypervolume difference, stratified by measured non-Gaussianity: no stratified effect falsifies the core claim. The third is the crossing ablation. The fourth is wall-clock time per patient relative to the 3–7 hours reported for GP-based tuning. A method that wins on hypervolume but needs much longer than the 7 hours reported for SAAS-BO has no clinical path, so we report time-to-target-hypervolume as a co-primary metric.
References
- OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines
- Noncrossing ordinal classification
- BoTorch: A Framework for Efficient Monte-Carlo Bayesian Optimization
- Maximizing acquisition functions for Bayesian optimization
- Differentiable Expected Hypervolume Improvement for Parallel Multi-Objective Bayesian Optimization
- High-dimensional Automated Radiation Therapy Treatment Planning via Bayesian Optimization
- OpenKBP: The open-access knowledge-based planning grand challenge
- Ordinal Bayesian Optimisation
- Sequential Neural Likelihood: Fast Likelihood-free Inference with Autoregressive Flows
- Poisson Ordinal Network for Gleason Group Estimation Using Bi-Parametric MRI
- Automatic Posterior Transformation for Likelihood-Free Inference
- BayesFlow 2.0: Multi-Backend Amortized Bayesian Inference in Python
- Better and faster catheter position optimization in HDR brachytherapy for prostate cancer using multi-objective real-valued GOMEA