Radiotherapy plans are still largely tuned by hand. A dosimetrist sets roughly thirty objective weights and dose targets, the planning software computes a dose, the dosimetrist inspects it, adjusts the settings, and runs it again, often for hours. Bayesian optimization is a sample-efficient search method that fits a statistical model of the objective and picks each next trial point where the model predicts improvement. It can automate this loop, but current versions start every patient from scratch. Separately, deep networks trained on past patients can predict a good dose distribution from anatomy alone, yet these predictions are usually imposed as rigid targets that the optimizer must copy even when they are wrong for this patient. We propose to use the prediction as a prior that the optimizer trusts early and abandons once the evidence turns against it. The central mechanism is a trust budget. Each time the optimizer follows the prior instead of its own best guess, it pays a cost it can measure, and it stops following the prior once the running cost exceeds a slowly growing allowance.
The budget has a precise justification. Following a prior amounts to deliberately maximizing the search criterion inexactly, and a recent regret bound for inexact maximization says convergence is preserved whenever the total shortfall grows sublinearly. So the budget carries a guarantee, and it also adapts on its own. When the prior agrees with the data, following it costs almost nothing and it keeps guiding the search. When the prior is wrong, the budget runs out within a few trials. No decay schedule has to be hand-tuned per patient.
Core Claim
Existing prior-injection methods for Bayesian optimization fade the prior on a clock. πBO multiplies the search criterion by a user belief raised to the power $\beta/n$ at iteration $n$. DynMeanBO blends an expert mean function into the Gaussian-process surrogate with a weight that decays exponentially. CoExBO fades learned human preferences through a decaying weight. All three prove that they eventually behave like ordinary Bayesian optimization. All three also leave the decay rate as a heuristic, and the research file flags this as an open problem. A rate that decays too fast throws away a good prior, and one that decays too slowly lets a bad prior waste evaluations. In radiotherapy the prior's quality varies from patient to patient: a dose-prediction network is accurate on typical anatomy and wrong on unusual anatomy. A single global rate is therefore the wrong design.
Our claim is that the decay should be driven by the measured disagreement between the prior and the current surrogate, accounted for as acquisition shortfall. Bayesian Optimization with Inexact Acquisition shows that GP-UCB has cumulative regret $\mathcal{O}(\gamma_T\sqrt{T} + M_T\sqrt{\gamma_T})$. Here $\gamma_T$ is the maximum information gain after $T$ evaluations and $M_T$ is the summed gap between the true maximum of the search criterion and the value at the chosen point. We use this paper's result directly. Prior-guided selection contributes to $M_T$ a quantity we can compute at every step. We cap that contribution explicitly.
Let $\alpha_t$ be the unweighted search criterion at step $t$ and $\pi$ the prior density over planning parameters. Let $\hat x_t$ be the solver's maximizer of $\alpha_t$ and $\tilde x_t$ its maximizer of $\alpha_t(x)\pi(x)$. The prior's cost at step $t$ is $s_t = \alpha_t(\hat x_t) - \alpha_t(\tilde x_t) \ge 0$. With $L_{t-1}$ the cost spent so far, the rule is
$$x_t = \begin{cases} \tilde x_t & \text{if } L_{t-1} + s_t \le c\, t^{\rho} \\ \hat x_t & \text{otherwise,} \end{cases}$$
where $c>0$ sets the size of the allowance and $\rho<1$ its growth. The prior's share of $M_T$ is then at most $cT^{\rho}$. The remaining share comes from the inner solver's own error, and the regret bound above keeps sublinear order provided that part is also sublinear. We only claim the bound under the exact-surrogate assumptions of that paper. The approximate-surrogate case is treated below as an empirical question.
Research Program
Turning a dose prediction into a prior over planning parameters. OpenKBP-Opt evaluated 76 prediction-plus-optimization pipelines. It showed that dose mimicking, meaning optimizing the plan to reproduce a predicted dose, is mathematically equivalent to conventional multi-objective inverse planning under specific conditions. We supply the method from this equivalence. The predicted dose-volume points for each structure become the dose targets in the planning objectives, which gives a single point $x_{\text{KBP}}$ in the roughly 34-dimensional parameter space used by Wang et al.. The prior $\pi$ is a Gaussian centered on $x_{\text{KBP}}$ in normalized coordinates. Its per-coordinate width is set from the prediction model's held-out dose-volume error on validation patients. OpenKBP-Opt treats the prediction as a fixed target for the whole run. We keep it only as long as the budget allows.
The optimizer. The surrogate is a Gaussian process over the plan score. We place a sparsity prior on its lengthscales, following SAAS-BO in Wang et al., because only a few planning parameters typically matter. The search criterion is GP-UCB for the guaranteed version and Expected Improvement as a practical variant. Every evaluation runs the fluence optimization for the proposed parameters on a precomputed per-patient dose-influence matrix and scores the result.
Where approximate inference enters. At 120 evaluations the data are far too few for the matrix cost that motivates sparse Gaussian processes such as S-GP-TS to matter. Using them here would solve a problem the application does not have. The real approximations are two. The first is the surrogate's hyperparameters: Wang et al. sample them with Hamiltonian Monte Carlo, at about 7 hours per patient, and the authors name this cost as a barrier to clinical use. The second is maximizing the search criterion in 34 dimensions. For the first, we replace Monte Carlo sampling with a single most-probable hyperparameter setting or a Laplace approximation. EBUCB constrains this design. It shows that bounding the approximation error in one divergence direction is not enough, because distorted tails can cause persistent mistakes. We therefore measure the calibration of the approximate surrogate's predictive intervals on held-out evaluations, in both tails, instead of assuming it. For the second, we split $M_T$ into the part caused by the prior, which the budget controls, and the part caused by the solver. We estimate the solver part against an expensive many-restart reference maximization. This tests directly whether the linear-size random grid of Kim et al. is still adequate at 34 dimensions, which the research file identifies as unverified.
Baselines
The naive reference points are two. The first is the default institutional template parameters, evaluated once. The second is the quasi-random Sobol search from Wang et al., run for 120 evaluations. Wang et al. found that SAAS-BO beat Sobol search on hot-spot control and homogeneity ($p<0.05$).
The strongest non-sequential competitor is dose mimicking alone, as in OpenKBP-Opt, using its MeanAbs and MaxAbs formulations. On 100 head-and-neck test patients, the best pipeline reached 77.0% satisfaction of region-of-interest clinical criteria. Reference clinical plans reached 69.7% and raw predictions 66.2%. MaxAbs had the lowest median dose error, 2.34 Gy. We rerun these formulations on our test split with the same prediction model and dose-influence matrices, so the 77.0% figure is a calibration target and not the number we compare against. Beating this baseline means exceeding its criteria satisfaction on the same patients within the 120-evaluation budget, and matching it within about the first 20 evaluations.
The strongest uninformed sequential competitor is SAAS-BO with full Monte Carlo inference, together with GPEI, as in Wang et al. They report that SAAS-BO plans were comparable to clinical plans on hot-spot control ($p=0.43$) and homogeneity ($p=0.96$). Both SAAS-BO and GPEI beat clinical plans on conformity and spillage ($p<0.05$). Their wall-clock times were about 7 and about 3 hours respectively. They report no single plan-score number that transfers to our cohort, so we rerun both methods under our score, evaluation budget and hardware.
The strongest prior-informed competitors are πBO with a fixed $\beta$ and DynMeanBO with its exponential mean decay. Both receive the same $\pi$ or the same derived mean function, and both use the same surrogate and inner solver as our method. We tune their decay rates on the 40 OpenKBP validation patients, and we tune our $c$ and $\rho$ on the same patients with the same number of tuning trials. We also run warm-started BO, which places $x_{\text{KBP}}$ and a few perturbations of it in the initial design and then runs unweighted SAAS-BO.
Two ablations isolate the key ingredient. The first is a clock-based schedule whose decay is matched, on validation patients, to the average number of prior-following steps our budget produces. It asks whether adaptivity matters or only the average rate. The second is the budget rule run with an uninformative prior. It should reproduce vanilla BO and confirms that the mechanism causes no harm when there is nothing to trust.
Why not just…?
Why not just warm-start Bayesian optimization at the dose-mimicking plan? This is the most serious threat. Wang et al. already show that BO beats clinical plans on some metrics, and OpenKBP-Opt shows that the dose-mimicking plan is already good. Seeding BO with it may capture most of the benefit. The difference is that a warm start influences only the surrogate's data. It does not steer where the next 30-dimensional query goes, so after the seed points, uninformed exploration begins again. Our experiment compares plan score against the number of evaluations for the two methods on accurate-prior patients. If warm-starting matches the budget rule on both early and final score, we will report that the simpler method suffices and that the budget contributes only its guarantee under bad priors.
Why not just stop at dose mimicking? OpenKBP-Opt's own numbers argue against it. The rank correlation between prediction-model quality and final plan quality was only 0.50 to 0.62, so a better prediction does not reliably yield a better plan. A fixed target also cannot correct predictions that are wrong for a particular patient. We test this with the corrupted-prior condition described below, where mimicking should inherit the corruption and the budget rule should not.
Why not just tune πBO's $\beta$ once on validation patients? If prior quality were uniform across patients, a well-tuned global $\beta$ would be hard to beat, and our advantage would shrink to having a guarantee without tuning. The approach bets that prior quality varies from patient to patient. The matched-schedule ablation settles this. If the clock-based schedule tuned to our average decay performs as well per patient, adaptivity adds nothing, and we will say so.
Why not just use GPEI, which is already faster? GPEI took about 3 hours versus about 7 for SAAS-BO in Wang et al., but it was significantly worse on hot-spot control and homogeneity. The question is whether a cheap approximation of the sparse surrogate keeps SAAS-BO's plan quality. We measure this directly by comparing the Monte Carlo and single-hyperparameter versions, with and without the prior.
Proposed Experiments
The primary testbed is the OpenKBP head-and-neck dataset of 340 patients, split into 200 for training, 40 for validation and 100 for testing (OpenKBP). We train a dose-prediction network on the training split. We compute per-patient dose-influence matrices with the open-source matRad engine used by FatPIN. We define a parameter space of dose targets and weights for the targets and organs at risk that mirrors Wang et al.'s 34 dimensions. The metrics are the OpenKBP-Opt clinical-criteria satisfaction rate, dose-volume errors, a scalar plan score combining conformity, spillage and organ sparing, the number of evaluations needed to reach the dose-mimicking plan's score, and wall-clock time.
Prior quality is controlled in three conditions. In the first, the prior comes from the trained model as is. In the second, it comes from a deliberately weak model trained on 20 patients. In the third, the predicted dose to one parotid gland is systematically shifted, mimicking the gland shrinkage and duct displacement described in the parotid stem-cell sparing study.
Three results would be decisive. First, with the accurate prior, the budget rule should reach dose-mimicking quality early and surpass it by 120 evaluations. With the corrupted prior, it should recover to within noise of unweighted SAAS-BO. Any fixed-rate πBO or DynMeanBO setting should fail in at least one of the two regimes. Second, across patients the budget should be spent quickly exactly where the prior's dose-volume error is large. We check this with a correlation between exhaustion time and the prior's per-patient error. It is the direct test that the decay responds to evidence rather than to the clock. Third, the single-hyperparameter surrogate combined with the budget should match Monte Carlo SAAS-BO in plan score, with well-calibrated two-sided predictive intervals, at a fraction of the approximately 7-hour cost. The measured solver share of $M_T$ should stay sublinear at 34 dimensions. If it does not, we replace the random grid with multi-start local search and report that Kim et al.'s grid result does not transfer to this dimension.
References
- $\pi$BO: Augmenting Acquisition Functions with User Beliefs for Bayesian Optimization
- Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean Decay
- Looping in the Human: Collaborative and Explainable Bayesian Optimization
- Bayesian Optimization with Inexact Acquisition: Is Random Grid Search Sufficient?
- OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines
- High-dimensional Automated Radiation Therapy Treatment Planning via Bayesian Optimization
- Scalable Thompson Sampling using Sparse Gaussian Process Models
- Optimal Regret Is Achievable with Bounded Approximate Inference Error: An Enhanced Bayesian Upper Confidence Bound Framework
- OpenKBP: The open-access knowledge-based planning grand challenge
- Automatic Radiotherapy Treatment Planning with Deep Functional Reinforcement Learning
- Parotid Gland Stem Cell Preservation during Intensity-Modulated Radiotherapy for Nasopharyngeal Carcinoma: Dosimetric Analysis and Feasibility