Knowledge-based radiotherapy planning trains a model on past treatment plans so it can predict, for a new patient, how much radiation each organ should receive. When those past plans come from several hospitals, the model also learns each hospital's habits. One clinic may tolerate more dose to the parotid glands to get better tumor coverage, while another does the opposite. This is a batch effect, a systematic difference between groups of data caused by who produced it rather than by the patients. It differs from the batch effects familiar from genomics or MRI in one respect. Two hospitals' plans for the same anatomy are usually not "true signal plus site noise." They are two different points on the same trade-off curve between tumor coverage and organ sparing. Subtracting site means from dose statistics, as standard batch correction does, can therefore produce dose targets that no deliverable plan can reach.
We propose to model the site effect where it actually lives, in the preference weights a planner implicitly used. For each historical plan we recover those weights by inverse optimization. We then fit an empirical-Bayes hierarchical model: the weights follow from anatomy, plus a site-level preference shift, plus plan-level noise. Finally, a dose predictor is conditioned on the harmonized weights. The result separates "what this anatomy allows" from "what this hospital prefers." A new site can be characterized from a handful of plans, and harmonized predictions stay on the achievable trade-off surface by construction.
Core Claim
The claim is that multi-institutional variation in planning style should be corrected as a shift in the objective-weight vector, not as a shift in dose values. Shrinking site shifts toward a shared prior, which is ComBat's "borrowing strength" idea moved into weight space, is what makes this work when a site contributes few plans. Our testable prediction is that this beats both naive pooling and per-site conditioning on held-out institutions when only 5–20 plans per new site are available. We also predict that dose-level post-hoc correction produces a measurable fraction of infeasible or dominated targets, while weight-level correction does not.
Three lines of prior work supply the pieces, but none combines them. Models for predicting objective function weights in prostate cancer IMRT supplies the method for recovering weights. It shows that anatomy features predict inverse-optimized weights 73–74% closer to ground truth than population averages. It also reports that about 20% of patients are "weight-insensitive," meaning many weight settings give nearly the same dose. That study is single-institution and has no notion of site. We reuse its inverse-optimization step. We also turn its insensitivity finding into an argument for shrinkage: weakly identified weights should be pulled toward the site mean rather than trusted.
The ComBat family supplies the statistical mechanism. Image harmonization: A review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization reviews these methods, including covariance-aware variants. Fed-ComBat: A Generalized Federated Framework for Batch Effect Harmonization in Collaborative Studies shows the same estimates can be computed across sites without sharing records. All of this work applies location and scale corrections to measured features. Here the corrected quantity is a latent preference on a simplex, the set of non-negative weight vectors that sum to one. We avoid correcting dose values directly because those values are coupled by physics.
Demo: Generative AI helps Radiotherapy Planning with User Preference is the closest existing work, and it motivates this proposal. Its Flexible Dose Proposer names institutional bias as a problem and conditions a dose decoder on a preference vector through adaptive instance normalization, a layer that rescales internal features according to a conditioning signal. That preference vector comes from a human moving sliders. The model never estimates what preference each training plan encoded, and it never separates site preference from anatomy. We estimate that vector for every training plan and model its site structure.
A second line of work motivates the concern about models learning site identity. Batch Effects In Brain Foundation Model Embeddings found that learned representations predict acquisition site with 85–94% accuracy. It also found that applying ComBat after the fact to frozen embeddings lowered site predictability without recovering biological signal. This constrains our design: correction should happen in an interpretable variable before representation learning, not on embeddings afterward.
Research Program
The first stage recovers preferences. For each plan $i$ with delivered dose $d_i$ and anatomy $a_i$, we solve an inverse-optimization problem. It finds non-negative weights $w_i$ over $K$ planning objectives, summing to one, that make $d_i$ as close to optimal as possible under the forward planning model, as measured by its duality gap. For implementation we use the dose-mimicking formulation of OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines. That paper proved dose mimicking equivalent to multi-objective inverse planning, so its optimizers define the forward model. We record the curvature of the duality gap around each recovered $w_i$ as a per-plan measure of how well the weights are identified.
The second stage harmonizes those preferences. Let $z_i$ be $w_i$ mapped to unconstrained additive log-ratio coordinates, a standard transform that removes the sum-to-one constraint. We fit
$$z_i = g_\theta(a_i) + \gamma_{s(i)} + \epsilon_i, \qquad \gamma_s \sim \mathcal{N}(\mu, T), \quad \epsilon_i \sim \mathcal{N}(0, \Sigma_i).$$
The terms are defined as follows. $g_\theta$ is a small regressor from overlap-volume-histogram features, which describe how closely each organ sits to the target. $\gamma_s$ is the preference shift of site $s$. $T$ is the between-site covariance. $\Sigma_i$ is the per-plan noise, inflated for weight-insensitive plans using the curvature from stage one.
The posterior mean of a site's shift is a precision-weighted blend of its own plans and the population mean $\mu$. This is ComBat's shrinkage applied to preferences. It is what lets a new site with five plans get a stable estimate. A "harmonized" training target replaces $\gamma_{s(i)}$ with a chosen reference, either $\mu$ or a target site's posterior. The plan's dose is then regenerated by running the forward optimizer at the new weights, so every harmonized target is a deliverable plan.
The third stage trains the predictor. A 3D U-Net takes CT and contours, together with the weight vector through adaptive instance normalization. It is trained on the original plans with their recovered, shrunk weights. At deployment, a clinic supplies its posterior shift $\gamma_s$ or a consensus $\mu$. The predicted dose is then passed through the OpenKBP-Opt dose-mimicking models to produce a plan.
We also run a scanner-signature check. A linear probe on the predictor's bottleneck features measures how well site can be decoded once weights are supplied, following Batch Effects In Brain Foundation Model Embeddings. Residual site decodability flags leakage through image appearance, for example differences in CT intensity calibration.
Evaluation needs ground-truth site effects, which no public benchmark has. OpenKBP: The open-access knowledge-based planning grand challenge deliberately used a single automated pipeline to avoid exactly this variation. We therefore build a semi-synthetic multi-site benchmark on the OpenKBP patients and add a real multi-site validation.
The semi-synthetic benchmark uses the 200 training, 40 validation and 100 test head-and-neck cases. We partition patients into 8–12 simulated institutions. Each institution draws per-plan weights from its own Gaussian in log-ratio space, with shifts on the scale of realistic parotid-versus-coverage trade-offs. Plans are then generated with the OpenKBP-Opt optimizers.
For real data, we will seek a multi-institution head-and-neck or prostate cohort from collaborating centers. If pooling is not permitted, site hyperparameters will be estimated federatedly in the style of Fed-ComBat. This dependency is real. If no partner data materialize, the claim rests on the semi-synthetic benchmark alone, and we will say so.
Baselines
The naive reference is a single 3D U-Net trained on all sites pooled, with no site information. LENAS: Learning-based Neural Architecture Search and Ensemble for 3D Radiotherapy Dose Prediction reports this baseline at an OpenKBP Dose Score, the mean absolute voxel error, of 2.619. It reports its best ensemble at 2.357. We expect pooling to regress toward an average style that matches no site.
The strongest dose-prediction competitor is the LENAS-style ensemble with heavy augmentation. It will be trained with the same augmentation and a matched GPU-hour budget. Clearly beating it on held-out-site Dose Score and DVH Score would support the proposal.
The strongest site-aware competitor conditions the same U-Net on a learned per-site embedding, in the spirit of the site-code conditioning in Learning Site-specific Styles for Multi-institutional Unsupervised Cross-modality Domain Adaptation. A new site's embedding is fit from its few available plans. It shares our backbone, optimizer and tuning budget, so it isolates whether the explicit, shrunk weight representation matters.
A post-hoc correction baseline applies ComBat to predicted DVH endpoints or to dose maps in regions of interest, removing site location and scale. It tests whether correcting in dose space suffices.
A slider-style baseline follows the Flexible Dose Proposer. It uses the same preference conditioning but receives only the population-mean weights, which is what a user would provide without per-plan estimates. An oracle variant receives the true simulated weights and serves as the upper bound.
The ablation that isolates our key ingredient keeps everything but sets $T \to \infty$. Each site's shift is then estimated only from its own plans, with no shrinkage. We apply the same ablation to the identifiability-weighted noise $\Sigma_i$.
For plan quality after optimization, the reference number is from OpenKBP-Opt. Its best pipeline satisfied 77.0% of clinical criteria, against 69.7% for reference plans and 66.2% for raw predictions. That was achieved on single-style synthetic data. On multi-site data we compare all methods through the same four mimicking models.
Why not just…?
Why not just augment and ensemble? Augmentation makes the model robust to nuisance transformations of the image. Planning style is not a transformation of the image: the same CT legitimately maps to different doses at different sites. Ensembling pooled models averages styles rather than choosing one. The semi-synthetic benchmark settles this directly. If the pooled ensemble matches our method on held-out-site DVH error, the preference model is unnecessary.
Why not just apply ComBat to dose-volume endpoints after prediction? ComBat corrects each endpoint independently. Dose to adjacent organs, however, is coupled through beam physics and the trade-off curve. Shifting parotid mean dose down and target coverage up at the same time can yield a combination no plan achieves. Batch Effects In Brain Foundation Model Embeddings also found that post-hoc correction removed site signal without recovering the underlying signal. We measure the infeasibility directly. For each corrected target, we run dose mimicking and record the gap between target and achieved DVH, and the fraction of targets dominated by the achieved plan. Our method's targets are feasible by construction, so this comparison is decisive only if ComBat's feasibility gap turns out to be non-trivial. If it is small, the post-hoc approach is a serious cheap competitor and we will report that.
Why not just condition on a site code or fine-tune per site? This is the real threat. With many plans per site, a learned site embedding should capture style as well as our model does, and possibly better, because it can represent style variation that weights miss. Our advantage is predicted to appear only in the few-plan regime and in interpretability, since the shifts are in clinical units a physicist can inspect or override. The experiment sweeps the number of plans from the new site, $k \in \{2, 5, 10, 20, 50\}$. We expect our method to win at small $k$ and converge to the site-embedding baseline at large $k$. If it never wins, the shrinkage story fails.
Why not just let planners set sliders, as in the Flexible Dose Proposer? Sliders fix inference but not training. Without per-plan weight labels, the training data's styles remain entangled with anatomy, so a slider setting means something different depending on which site's patients resemble the input. The oracle-weights variant tells us how much per-plan labels are worth. The population-weights variant tells us how much is lost without them.
Proposed Experiments
The first decisive check is recovery on the semi-synthetic benchmark. We report the error between estimated and true site shifts $\gamma_s$ in log-ratio space, as a function of plans per site. We report it separately for weight-insensitive and well-identified patients. This verifies that inverse optimization plus shrinkage recovers known preferences and that curvature-weighted noise helps where Models for predicting objective function weights in prostate cancer IMRT predicts identification is poor.
The second check is held-out-site generalization. Using leave-one-site-out evaluation with $k$ adaptation plans, we report OpenKBP Dose Score and DVH Score against that site's own plans for all baselines. We also report clinical criteria satisfaction after each of the four OpenKBP-Opt mimicking models.
The third check is feasibility of harmonized targets. For ComBat-corrected targets versus our regenerated targets, we report the achieved-versus-target DVH gap after mimicking. OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines found only 0.50–0.62 rank correlation between prediction quality and final plan quality. We test whether multi-site style mismatch explains part of that gap by computing the same rank correlation with and without harmonization.
The fourth check is a leakage probe. We measure site-classification accuracy from bottleneck features, with and without weight conditioning, in a confounded split where site correlates with tumor stage. Weight conditioning should absorb style, so residual site decodability indicates image-level artifacts that planning-style correction cannot fix.
On real multi-institution data, if obtained, we repeat the second check. As a clinically interpretable output, we also report the estimated between-site covariance $T$, which shows which trade-offs hospitals disagree on most.
No reported result was found on Lacuna for any multi-site dose-prediction benchmark with real inter-institutional plans. The comparisons above therefore establish the first such numbers, rather than beating a published one.
References
- Models for predicting objective function weights in prostate cancer IMRT
- Image harmonization: A review of statistical and deep learning methods for removing batch effects and evaluation metrics for effective harmonization
- Fed-ComBat: A Generalized Federated Framework for Batch Effect Harmonization in Collaborative Studies
- Demo: Generative AI helps Radiotherapy Planning with User Preference
- Batch Effects In Brain Foundation Model Embeddings
- OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines
- OpenKBP: The open-access knowledge-based planning grand challenge
- LENAS: Learning-based Neural Architecture Search and Ensemble for 3D Radiotherapy Dose Prediction
- Learning Site-Specific Styles for Multi-institutional Unsupervised Cross-Modality Domain Adaptation