Synthetic Lesions Should Not Buy Certainty: Provenance-Aware Evidence for Imbalanced Segmentation

When a segmentation model must find something rare, such as a small kidney tumor occupying a fraction of a percent of a CT volume, practitioners often add synthetic examples of the rare class: pasted or generated lesions with automatically assigned labels. Synthetic examples can raise recall. They also carry a quieter cost that has not been measured. A model trained on fabricated foreground learns to be confident about patterns it has only seen in fabricated form, so its uncertainty estimates stop telling the clinician where it is guessing. This proposal treats that as the central problem. It makes the model's confidence depend on where the training signal came from, so that evidence learned from synthetic voxels is capped and only evidence learned from real annotations can make the model fully certain.

The tool for this is evidential deep learning, a way of training a network to output "amounts of evidence" for each class instead of plain probabilities. Those amounts define a Dirichlet distribution, which is a distribution over class probabilities, and from it comes an explicit "I have not seen enough" score. That score equals the number of classes divided by the total evidence, so it shrinks as evidence grows. Because the score is tied to the quantity of evidence, it can be budgeted: synthetic voxels can be allowed to contribute some evidence, but not unlimited evidence. Two more ingredients complete the method. The overlap-based Dice term of the usual hybrid loss is restricted to real voxels. The background prior is rebalanced so that the overwhelming majority class does not swamp the rare one.

Core Claim

Synthetic minority data degrades segmentation uncertainty in a specific, measurable way. Lesion-like healthy tissue and implausible synthetic patterns both end up receiving low uncertainty, which breaks the link between uncertainty and error. The claim is that bounding the evidence that synthetic voxels can contribute, and keeping them out of the size-invariant Dice term, keeps most of the recall gain from synthesis and restores that link. Two further predictions follow. Per-lesion uncertainty becomes a usable detector of bad synthetic samples. And the "collateral damage" to the majority class, meaning confident false positives in background, shrinks.

The prior work in the research file sits next to this idea but does not occupy it. One group of papers supplies the evidential segmentation machinery but trains only on real data. Cross-Slice Attention and Evidential Critical Loss adds focal-style modulation to a voxel-wise evidential loss for 10,000:1 prostate lesion imbalance. Mutual Evidential Deep Learning and Evidential Prototype Learning fuse evidence across networks for semi-supervised segmentation. None of the three asks what happens when some labels are fabricated. A second group motivates the proposal: synthesis distorts boundaries and harms majority classes. Generative Adversarial Minority Oversampling constrains generated points to convex combinations of real minority points so they stay near the decision margin. Influence of minority class instance types on SMOTE shows that oversampling outlier seeds amplifies noise. Bansal et al. and Card et al. describe how synthetic ambiguity bleeds into majority regions. All of these work on whole-image or tabular classification and say nothing about uncertainty. The closest conceptual ancestor is Exploiting Synthetically Generated Data with Semi-Supervised Learning. It treats synthetic points as unlabeled within a kernel semi-supervised SVM so that their labels cannot corrupt the boundary. We keep that intuition but soften it: synthetic labels are partially trusted rather than discarded, and the trust is expressed directly in the quantity that later drives the uncertainty score.

Two papers constrain the design. Hybrid-EDL reports that evidential uncertainty can paradoxically fall under distribution shift. Synthetic data is a form of shift, so we should expect, and must test for, exactly this failure. Evidential Prototype Learning notes that evidential objectives can degrade boundary overlap when dense labels are plentiful. That is why we keep a Dice term rather than replacing it. Finally, we do not claim that "Dice over evidential expected probabilities" is new on its own. It is a natural construction, and its novelty is not the point. The new part is which voxels it is allowed to see.

Research Program

The model is a standard 3D U-Net in the nnU-Net style. Its output layer is replaced by a non-negative evidence head producing $\alpha_i \in \mathbb{R}^K$, the Dirichlet parameters for voxel $i$ over $K$ classes, with total evidence $S_i=\sum_k \alpha_k$ and uncertainty $u_i = K/S_i$. Every training voxel carries a provenance flag, $r_i=1$ if its label comes from a real annotation and $r_i=0$ if it lies inside a synthetic lesion or its pasted margin. The training objective is

$$ L = \sum_{i:\,r_i=1} \lambda_c\, L_{\text{EDL}}(\alpha_i, y_i) \;+\; \sum_{i:\,r_i=0} \mathrm{KL}\big(\mathrm{Dir}(\alpha_i)\,\|\,\mathrm{Dir}(\mathbf{1}+E\,y_i)\big) \;+\; \beta\, L_{\text{Dice}}\big(\{\bar p_i\}_{r_i=1}, \{y_i\}_{r_i=1}\big). $$

Here $y_i$ is the one-hot label and $L_{\text{EDL}}$ is the usual evidential cross-entropy with its regularizer toward an uninformative Dirichlet. The per-class weight $\lambda_c$ follows the class-pooled, frequency-aware prior of Class-Balanced Evidential Deep Learning, adapted from whole-sample to voxel counts. $\bar p_i = \alpha_i/S_i$ is the expected class probability. $E$ is the evidence budget for synthetic voxels. The middle term pulls synthetic voxels toward a Dirichlet with exactly $E$ units of evidence on the labeled class, so synthetic examples can teach the model where tumors might be but cannot by themselves drive $u_i$ toward zero. The Dice term sees only real voxels. Dice is size-invariant, so a pasted ten-voxel lesion would otherwise carry as much objective weight as a real organ. Dice also pushes predictions toward extreme probabilities, as Calibrating the Dice Loss documents, and applied to synthetic voxels it would undo the budget.

The program has three parts. The first is a diagnostic study. It measures, with ordinary softmax compound-loss models, how much synthetic lesions improve tumor recall and how much they damage the correlation between uncertainty and error, especially in background regions. If that damage is negligible, the proposal's premise fails, and the study says so. The second part is the method above. It includes a sweep over $E$, and a variant that sets $E$ per synthetic lesion from how much the model's real-data evidence agrees with the pasted label. The third part is a controlled-noise analysis. A known fraction of synthetic lesions is deliberately corrupted, through anatomically impossible placement, shuffled intensities, or labels dilated into background, to test whether per-lesion mean uncertainty or evidence conflict separates bad synthetic samples from good ones. This would replace the dataset-specific combinatorial instance filtering of Skryjomski and Krawczyk with a signal the network computes during training.

Baselines

The naive reference point is the same U-Net trained with the standard cross-entropy plus Dice compound loss on real data only, with no synthesis. Every claimed gain in recall must be measured against it, and every claimed loss in uncertainty quality must be measured against its softmax-entropy uncertainty.

The strongest existing segmentation pipeline is nnU-Net with compound loss. It won KiTS19 with kidney Dice 0.974, tumor Dice 0.851 and composite Dice 0.912, as reported in the KiTS19 challenge paper. Those numbers come from full training data and the hidden test set. Our reduced-tumor protocol will not reproduce them, so they serve as an upper reference. We will rerun nnU-Net's default configuration on our splits, both with and without the same synthetic lesions, and that rerun is the baseline to beat. Beating it means equal or higher tumor Dice and recall, together with a clearly better uncertainty–error area under the ROC curve and lower calibration error in background.

For calibration, the strongest competitor is DSC++. On KiTS19 tumor in its authors' setup, it reports Dice 0.429 against 0.232 for standard Dice loss Calibrating the Dice Loss. We run it with synthesis under our protocol. For evidential imbalance handling, the strongest competitors are the Evidential Critical loss, which reports sensitivity 0.870 versus 0.801 for nnU-Net at three false positives per patient on its internal prostate set Cross-Slice Attention and Evidential Critical Loss, and class-balanced EDL. For class-balanced EDL, no reported result was found on Lacuna for KiTS19 or for any segmentation benchmark; its reported numbers, such as ECG5000 macro-recall rising from 0.388 to 0.778 Class-Balanced Evidential Deep Learning, are for classification. Both evidential baselines are trained on the same synthetic-augmented data with labels treated as real. That makes them the cleanest ablation of the provenance flag.

Three further ablations isolate the key ingredient. The first is a softmax model whose synthetic voxels get a fixed down-weight. The second uses label smoothing on synthetic voxels only. The third is our method with the Dice term applied to all voxels. All baselines share the architecture, patch size, augmentation, epochs, and a fixed budget of twenty tuning trials each on a validation split, with three seeds.

Why not just…?

Why not just skip synthesis and use focal or class-weighted compound loss? This is a real threat, and the diagnostic study tests it directly. On SVHN, naive augmentation lowered average class-specific accuracy from 0.74 to 0.69 while boundary-targeted synthesis raised it to 0.76 Generative Adversarial Minority Oversampling. So whether synthesis helps depends on how it is done, and nothing in the file establishes this for 3D tumors. If the no-synthesis evidential model with class-balanced priors matches the synthetic-augmented one on recall, the recommendation becomes "do not synthesize." That would be a useful negative result, and we would report it as such.

Why not just down-weight synthetic voxels, or label-smooth them, in an ordinary softmax network? This is the closest simple competitor, since a capped evidence target behaves somewhat like label smoothing. The difference shows up at test time. A smoothed softmax network has no quantity that grows only with real evidence, so it cannot report "this tumor-like region resembles only my synthetic training data." The ablation decides this. If label smoothing on synthetic voxels matches our method on uncertainty–error AUROC and on detecting corrupted synthetic lesions, the evidential machinery is not earning its keep, and we would say so.

Why not just treat synthetic lesions as unlabeled and use mean-teacher pseudo-labels, as the S3VM approach does? That approach discards the label entirely. In 3D segmentation the location and extent of a pasted lesion are known exactly, so discarding them wastes the signal that makes synthesis useful. We include a mean-teacher variant to measure how much that costs.

Why not just apply temperature scaling afterwards? Hybrid-EDL reports that temperature scaling fitted on clean validation data fails under shift. It also cannot undo confident evidence that was learned from fabricated labels; it can only rescale it uniformly. We still report it on top of every softmax baseline.

Proposed Experiments

The main benchmark is KiTS19, with the tumor-containing training cases subsampled to 25% and 10% to simulate scarcity. Synthetic tumors are made by pasting real tumor crops into healthy kidney regions of other patients. This keeps synthesis simple and its noise controllable, so the study does not depend on a particular generative model. Pancreas-CT provides a second, smaller-target setting, with synthetic pancreatic lesions made the same way. We measure tumor Dice, recall, precision and 95% Hausdorff distance for overlap. For uncertainty, we measure voxel-level expected calibration error and Brier score, the AUROC of uncertainty for predicting voxel errors, and the mean confidence of false-positive background voxels. In the controlled-noise study, we also measure the AUROC of per-lesion uncertainty for flagging corrupted synthetic lesions, at corruption rates of 10%, 30% and 50%.

The decisive checks are three. First, synthesis must measurably worsen uncertainty–error AUROC for standard models; without this, the premise fails. Second, our method must recover that AUROC while keeping at least most of the recall gain, and it must beat label smoothing on synthetic voxels. Third, under the Hybrid-EDL overconfidence paradox, uncertainty on corrupted synthetic lesions must stay high rather than collapse; if it collapses, evidential budgeting is not enough and the per-lesion agreement-based budget becomes necessary. A sweep over $E$ should show a clear trade-off between recall and calibration, with an interior optimum. If the curve is flat, the budget is not doing the work attributed to it.