Pancreatic cancer is usually found late, and the AI systems meant to find it earlier are trained on the wrong patients. Retrospective CT archives are dominated by large, obvious tumors, because those are the cases that reached imaging. The lesions that matter for early detection are small, and they look almost like normal pancreas on CT. A detector trained on such an archive has absorbed a hidden assumption about how big a tumor usually is, and that assumption is wrong for the population a screening tool would serve.
We propose to state that assumption explicitly and change it at test time. Several detectors are trained, each specialized to one tumor-size range. Their outputs are combined by Bayesian model averaging, which means weighting each detector by how probable it is that this detector's size range describes the scan. The prior over size ranges is set to the deployment population rather than the training archive, so the combined detector can be pointed at small tumors without retraining. The claim is that the uncertainty that causes missed early tumors is uncertainty about which size range produced the scan, not uncertainty about network weights. Bayesian averaging over the right set of models should therefore help where averaging over weights has repeatedly failed.
Core Claim
The closest existing system is the PANORAMA winner, PanDx. It already trains four detectors on tumor-size quartiles, a scheme it calls distribution-aware stratified ensembling, and averages them with fixed weights. This raised validation average precision from 0.7983 to 0.8247 and reached a patient-level area under the ROC curve (AUROC) of 0.9263 and an average precision (AP) of 0.7243 on the official test set. PanDx shows that size specialization matters. Its averaging, however, is uniform and silently inherits the training cohort's size mix.
Our change is narrow and testable. We treat the size range as a model index $m$, with one value per expert plus one "no tumor" value. The per-scan weight on each expert is
$$P_{\text{target}}(m \mid x) \propto q(m \mid x)\,\frac{\pi_{\text{target}}(m)}{\pi_{\text{train}}(m)}$$
The symbols are as follows. $q(m \mid x)$ is a small classifier, trained on the training data, that predicts the size range of scan $x$. $\pi_{\text{train}}(m)$ is the size-range frequency in the training cohort. $\pi_{\text{target}}(m)$ is the frequency expected in the deployment population.
The final malignancy score averages the experts' scores under these weights. This is the standard prior-shift correction for classifiers, applied to the index of a model average instead of to the class label. Its value is that a clinic can move the detector toward the small-tumor regime by changing a handful of prior numbers, and every change has a stated meaning.
Why the obvious Bayesian move is the wrong one
The tempting intersection idea is to wrap a pancreas detector in a Bayesian neural network or a deep ensemble and use the resulting uncertainty to flag doubtful scans. Two papers in the file argue against expecting that to work.
Adams & Elhabian benchmarked nine scalable approximations to weight-space averaging on abdominal CT, including deep ensembles, SWAG (which fits a Gaussian to the weights visited by stochastic gradient descent), MC Dropout and Rank-1 Bayesian networks. None of them signaled uncertainty on atypical pancreas tumors; every method stayed overconfident. Izmailov et al. show that even exact weight-space averaging can fall up to 25% behind a single network under covariate shift. Their explanation is that weight directions the training data never constrains are filled with prior noise.
We read these two results as constraints. A small tumor that looks like normal tissue is not far from the training data in feature space. It sits inside the region the data constrains, so weight posteriors have no reason to disagree on it. The missing information is a prior over lesion size, and that has to be put into the model space directly. Wilson argues that marginalization curbs overconfidence. We accept the argument but hold that it only helps if the space being marginalized over contains the hypothesis the model is wrong about.
Research Program
All experts share one backbone: a two-stage coarse-to-fine ResU-Net on portal-venous CT, following PanDx. Each expert is trained on data enriched for one size quartile, using lesion volumes computed from the PANORAMA segmentation masks.
The size classifier $q(m \mid x)$ is a lightweight head on the coarse-stage features. It is temperature-calibrated on held-out folds, because the correction is only as good as these probabilities.
We then build controlled size shifts inside the public PANORAMA data, since the official test set is hidden. In the mild setting, test folds are resampled toward small lesions. In the severe setting, the smallest quartile is removed from training entirely, which imitates an archive of late-diagnosed patients. Here the target prior is the only route by which a small-tumor regime can gain weight. In the severe setting the smallest-range expert must be trained on the next quartile and given extra weight, and we report how much of the gain survives.
Two extensions are kept secondary. The first replaces the closed-form weights with the amortized, input-dependent weights of Input Adaptive BMA, which beat uniform averaging and mixture-of-experts on non-imaging tasks. The second wraps the averaged score in conformal prediction, a method that guarantees a stated error rate on new cases. This would give a "safe negative" call with a guaranteed false-negative rate. We would compare it against the posterior-weighted conformity scores of CBMA.
The Fisher-weighted merging of the experts into a single network is left out of scope. Merging fixes the weights and so cannot express per-scan reweighting.
Baselines
The naive reference is a single two-stage ResU-Net trained on the natural size distribution. It is the model a hospital would deploy by default, and it shows how large the size-bias problem actually is. The broader single-network reference is the Medical Segmentation Decathlon, whose leading pipeline reached only 0.21 Dice on pancreatic tumors (Dice measures mask overlap, with 1.0 a perfect match). We do not compare against that number directly, because it is a segmentation score on a different dataset.
The strongest existing method is PanDx's uniform size-stratified ensemble. It uses the same experts we use, so any gain from our method comes only from the weighting. The strongest reported number is PanDx's validation AP of 0.8247 with stratified ensembling. Our re-implementation must first reproduce this to within fold noise before any comparison is trusted.
The weight-space baselines are a four-member deep ensemble, SWAG, and MC Dropout. All three are trained with the same total GPU-hours as the four experts and given the same hyperparameter search budget. They test the central claim directly. If they match our size-shifted sensitivity, our diagnosis of the failure is wrong.
A learned gate is a softmax network over the experts, trained end to end. It tests whether the Bayesian structure adds anything beyond flexible combination.
The key ablation is our method with $\pi_{\text{target}} = \pi_{\text{train}}$. This reduces to likelihood-weighted averaging without prior correction, and it isolates the prior-shift term.
Beating the baselines means two things. First, on the small-lesion shifted test folds, sensitivity at 95% specificity on the smallest quartile must exceed uniform PanDx-style averaging by more than the fold-to-fold standard deviation. Second, whole-population AUROC must not fall by more than that same margin.
Why not just…?
Why not just use uniform size-stratified averaging, as PanDx does? This is the most serious threat. If the size classifier is poorly calibrated, uniform averaging may be as good. The ablation that sets the target prior equal to the training prior, together with a sweep of $\pi_{\text{target}}$, settles the question. If small-lesion sensitivity does not respond to the prior in the predicted direction, the method adds nothing over PanDx and we will report that.
Why not just oversample small tumors when training one network? Oversampling fixes a single size prior at training time. It cannot be changed afterwards and gives no per-scan weighting. We train a single network for each oversampling ratio matched to each target prior. If those single networks match our method at every prior, the argument for averaging reduces to convenience, which is still a real deployment benefit but a weaker scientific claim.
Why not just train a deep ensemble or a Bayesian network and threshold its uncertainty? Adams & Elhabian found that this family stays overconfident on pancreas tumors. We nonetheless rerun it under equal compute on our size-shift splits, because their stress test used large tumors and not small ones.
Why not just learn a gating network? A learned gate inherits the training size mix exactly as the experts do. It has no separate prior that can be adjusted, so it should degrade under the severe shift. The severe-shift experiment tests this prediction.
Proposed Experiments
The primary experiment is five-fold cross-validation on the public PANORAMA scans. We report patient-level AUROC and AP, sensitivity at 95% specificity broken down by size quartile, expected calibration error, and the false-negative rate among scans called confidently negative. Each metric is reported under the natural, mild-shift and severe-shift test distributions.
The decisive check is a dose-response curve. As $\pi_{\text{target}}$ moves mass toward small lesions, smallest-quartile sensitivity should rise steadily while specificity falls by a predictable amount. Weight-space baselines and learned gates, which have no such control, should show no comparable curve.
A secondary transfer check applies the experts to the multi-phase cohorts of Xia et al., where alignment-ensemble tumor Dice of 72.27% and 64.38% are reported, if those data are accessible. The question is whether target priors fit on PANORAMA still help at a new site. If the data are not available, we state that no external validation was possible.
References
- PanDx: AI-Assisted Early Detection of Pancreatic Ductal Adenocarcinoma on Contrast-Enhanced CT
- Benchmarking Scalable Epistemic Uncertainty Quantification in Organ Segmentation
- Dangers of Bayesian Model Averaging under Covariate Shift
- The Case for Bayesian Deep Learning.
- Input Adaptive Bayesian Model Averaging
- CBMA: Improving Conformal Prediction through Bayesian Model Averaging
- Merging Models with Fisher-Weighted Averaging
- The Medical Segmentation Decathlon
- Detecting Pancreatic Ductal Adenocarcinoma in Multi-phase CT Scans via Alignment Ensemble