Teaching Evidential Segmenters to Say "Ambiguous" Instead of "Unknown"

Semi-supervised segmentation trains on a handful of expert-traced scans plus many unlabeled ones. It usually works like this: a teacher network labels the unlabeled scans, and a student learns from those labels, skipping voxels where the teacher looks unsure. In low-contrast organs such as the pancreas or the thin left-atrial wall, the voxels the teacher is unsure about are mostly the boundary itself. Filtering them out throws away the part of the image that decides the boundary metrics. Evidential deep learning is meant to help here. It makes the network output a Dirichlet distribution, a "distribution over class probabilities" whose total mass says how much evidence the network has seen. In principle this lets a network separate two situations. In the first, it has plenty of evidence and that evidence is split between classes, which is real ambiguity at a blurred edge. In the second, it has almost no evidence, which is ignorance from too few labels. Recent evidential semi-supervised methods still reduce this to one uncertainty score and filter or re-weight by it.

We argue the separation fails before the semi-supervised stage begins. Expert masks are one-hot, meaning every boundary voxel is labeled as fully one class. The standard evidential loss then pushes evidence for all non-target classes toward zero. A voxel that sits on a blurred edge, and gets opposite labels from nearly identical image patches, cannot build balanced evidence for two classes. It can only lose evidence. So the network learns to express boundary ambiguity as ignorance. Our proposal has two parts. First, train the labeled scans with soft boundary targets under an evidence-preserving loss, so that real boundary ambiguity appears as conflicting evidence. Second, weight the unlabeled consistency loss by lack of evidence only, not by total uncertainty. Conflicting boundary voxels are then kept and supervised with the teacher's soft class distribution rather than discarded. The two parts only make sense together, and the experiments are built to test that interaction.

Core Claim

Write a voxel's Dirichlet parameters as $\alpha_k = e_k + 1$ over $K$ classes, with strength $S=\sum_k \alpha_k$. Following subjective logic, each class gets a belief mass $b_k = e_k/S$, and vacuity (lack of evidence) is $u = K/S$. We define conflict as the belief mass that does not go to the top class:

$$c = 1 - u - \max_k b_k .$$

This gives $c \approx 0$ for a confident voxel ($\alpha=[1000,1,1]$). It also gives $c = 0$ for an ignorant voxel ($\alpha=[1,1,1]$, where $u=1$). For a well-informed but split voxel ($\alpha=[50,50,1]$), $c \approx 0.49$ while $u \approx 0.03$.

Our first claim is diagnostic. Under standard evidential training, where the loss is digamma cross-entropy plus a KL penalty toward a flat Dirichlet on non-target evidence, the uncertainty in the boundary band of LA and Pancreas-CT models will be dominated by $u$, not $c$. Our second claim is that fixing this on labeled data allows a simple unlabeled loss to beat both entropy-weighted and threshold-filtered consistency on boundary metrics without losing Dice. That loss weights each voxel by $(1-u_{\text{teacher}})$ and targets the teacher's Dirichlet mean $\alpha/S$.

Relation to Prior Work

Several evidential semi-supervised methods supply the machinery and motivate the gap. EVIL pairs an evidential network with a softmax network and filters pseudo-labels by a fixed uncertainty threshold, so high-uncertainty boundary voxels are dropped. EPL is the closest in spirit. It fuses student and teacher evidence with Dempster's rule and computes two uncertainties, evidential mass and belief entropy. However, it uses both only to mask voxels out of prototype extraction. It never changes how labeled boundaries are supervised, and it reports that its evidential loss becomes less favorable to boundary metrics as the labeled fraction grows. We read that report as a symptom of our diagnostic claim. MEDL and Co-evidential fusion add curricula that bring uncertain voxels in late in training. Both rank voxels by a single entropy-like score. MEDL's ablations show a Dice-versus-95HD trade-off when its loss weights are tuned. All four methods treat "uncertain" as one quantity. We change which uncertainty gates supervision, and we change the labeled-data training so that the distinction actually exists.

A second group constrains the design. DyCON argues that discarding uncertain voxels loses exactly the lesion margins. It keeps them with entropy-decayed weights and a focal contrastive loss, but it uses softmax entropy, which cannot tell ambiguity from ignorance, and it needs per-dataset focal settings. UA-MT filters by Monte Carlo dropout variance, which requires eight stochastic passes. EvidenceCap notes that evidential models fail on very small targets and ambiguous boundaries. That warns against rigid vacuity cutoffs, so we use continuous weights rather than thresholds.

Our labeled-data change borrows from the boundary literature. Bounding Maps replaces binary lesion labels with continuous targets that decay toward the border. Combating Uncertainty with Novel Losses shows that penalizing foreground/background overlap sharpens left-atrium margins under full supervision. We reuse the soft-border idea, but as a target for Dirichlet evidence rather than for a softmax.

Finally, AmbiSSL models multi-annotator disagreement under limited labels using randomized decoder pruning. We use its benchmark, LIDC-IDRI, as ground truth for what conflict should look like.

Research Program

The first stage is measurement. We train a standard evidential V-Net at 10% and 20% labels on LA and Pancreas-CT. We then decompose uncertainty in a band of a few voxels around the ground-truth surface and compare it with the interior and with distant background. On LIDC-IDRI and ISIC 2018 we compute pixel-wise annotator disagreement and ask which of $u$ or $c$ predicts it better, measured by area under the ROC curve. For the claim to hold, disagreement should be predicted mainly by $u$ under standard training, which would show the conflation. If $c$ already carries it, the premise is wrong, and the project reduces to the routing rule alone.

The second stage is labeled training. Each hard mask is replaced by a soft target $y_i$ in a band around the surface, with class probabilities blending linearly toward 0.5 at the traced contour, as in bounding maps. We test two band widths. One is fixed. The other adapts to contrast: the width shrinks where the local intensity gradient across the traced contour is steep and widens where it is flat. The loss is the expected cross-entropy under the Dirichlet with soft $y_i$, written $\sum_k y_{ik}\,(\psi(S_i)-\psi(\alpha_{ik}))$, where $\psi$ is the digamma function. The KL shrinkage applies only to classes with $y_{ik}=0$, so it leaves evidence for both classes intact inside the band. A small term also discourages high vacuity on labeled voxels, since these voxels are by definition not unknown.

The third stage is the unlabeled loss in a mean-teacher setup, where the teacher is a running average of the student's weights. For each unlabeled voxel, the student's Dirichlet mean is pulled toward the teacher's Dirichlet mean under a strongly perturbed view. The pull is weighted by $(1-u_{\text{teacher}})$, with no threshold and no argmax. Confident voxels therefore receive near-hard targets. Conflicting boundary voxels receive soft targets at nearly full weight. Voxels the teacher knows nothing about are down-weighted but not zeroed, which avoids the small-target failure EvidenceCap describes. Everything runs in a single forward pass.

Baselines

The naive reference is a V-Net trained only on the labeled subset. UA-MT reports it at 84.13% Dice on LA with 20% labels. Any semi-supervised result must clear this by a wide margin on the same splits.

The classical uncertainty-filtered baseline is UA-MT, which reports 88.88% Dice and 95HD of 7.32 on LA at 20% labels. It represents the filter-out approach that our method rejects.

The strongest non-evidential competitor is DyCON. It reports 91.18% Dice on LA at 5% labels and 84.81% Dice on Pancreas-CT at 20%. It is the most direct test of whether Dirichlet machinery is needed at all, because it already keeps boundary voxels using softmax entropy.

The strongest evidential competitors are Co-evidential fusion and MEDL. Co-evidential fusion reports 82.93% Dice and 71.22% Jaccard on Pancreas-CT at 5%, 90.98% Dice on LA at 10%, and 79.46% Dice on TBAD at 5%. MEDL reports 77.67% Dice on TBAD at 5%. EPL is included because it is the only prior method that computes two uncertainties. The file gives no exact LA or Pancreas Dice for EPL, so no reported result was found on Lacuna for EPL's numbers.

On LIDC-IDRI, the reference is AmbiSSL, which reports soft Dice of 89.86% and generalized energy distance of 0.1620 at 10% labels.

For fairness, all methods are re-run in one codebase. They share the same V-Net backbone (DyCON reports with V-Net), the same published splits, the same number of training iterations, the same augmentations, and an equal hyperparameter search of about twenty configurations each, tuned on a held-out labeled validation fold. Published numbers are shown alongside for reference but are not the basis of claims. Beating a baseline means a better 95HD and ASD with Dice no worse by more than seed-to-seed standard deviation, over three seeds, on at least LA and Pancreas-CT at 5% and 10%.

The key ablation is a two-by-two design. One factor is the labeled targets: hard, or soft band. The other is the unlabeled weighting: teacher entropy of the Dirichlet mean, or teacher vacuity. Our prediction is an interaction. Vacuity weighting should help only when labels are soft, because only then does vacuity exclude conflict. A further ablation replaces the evidential conflict map with a purely geometric band, taken as the distance from the teacher's predicted surface.

Why not just…?

Why not just use softmax entropy with soft teacher targets, as DyCON does? This is the most serious threat. Entropy weighting with soft targets could recover most of the gain. Entropy, however, is high at conflicting boundary voxels, so any entropy-decreasing weight down-weights precisely the band we want supervised. DyCON works around this with a decay schedule and dataset-specific focal settings (0.5 on ISLES versus 0.8 on LA and Pancreas). The two-by-two ablation settles the question directly. If entropy weighting with soft labels matches vacuity weighting with soft labels, the Dirichlet adds nothing for accuracy, and we will report that plainly.

Why not separate aleatoric and epistemic uncertainty with Monte Carlo dropout, using the usual split into expected entropy and mutual information? Dropout can make the same split. UA-MT, however, uses eight passes, and EVIL reports that UA-MT adds about 233% training overhead against about 46% for its single-pass evidential design. More importantly, dropout sampling would also turn boundary ambiguity into epistemic spread when trained on one-hot masks. We will run a dropout variant with the same soft labels and the same routing rule. If it matches our separation quality on LIDC-IDRI, the contribution is the soft-label training and the routing, not the Dirichlet itself. We will report the cost difference.

Why not just mark the band around the teacher's predicted contour as ambiguous? A geometric band captures "near the edge" but not "genuinely low-contrast here." It treats a sharp blood-pool edge the same as the pancreas–duodenum interface. The geometric ablation tests this directly. It also has a second weakness: it cannot flag teacher ignorance away from the predicted surface, such as a missed small structure. Our method should win mainly on TBAD thrombus and small pancreatic regions. If it does not, the band is enough.

Why not add Overlap and Focal Positive losses to an existing evidential method such as MEDL? Those losses lifted fully supervised LA Dice from 82.01% to 91.54% in Combating Uncertainty with Novel Losses. They need trusted contours, though. On unlabeled scans, the only available contours come from the teacher, and the losses would sharpen the teacher's errors. We include MEDL with these losses added on the labeled scans as a baseline.

Decisive Checks

Three results would confirm the proposal. First, under standard evidential training, boundary-band uncertainty is mostly vacuity. After soft-band training it becomes mostly conflict, and conflict predicts LIDC-IDRI and ISIC annotator disagreement better than vacuity does. Second, the two-by-two ablation shows the predicted interaction on LA and Pancreas-CT. Third, the full method improves 95HD and ASD over DyCON and Co-evidential fusion at 5% and 10% labels while staying within noise on Dice, which breaks the trade-off MEDL documents.

Two results would weaken or overturn it. If the first stage shows standard evidential models already express boundary ambiguity as conflict, the premise fails. If entropy weighting with soft labels ties vacuity weighting, the evidential part of the method is unnecessary. We also report calibration error restricted to the boundary band and generalized energy distance on LIDC-IDRI against AmbiSSL's 0.1620. These test whether the conflict map is a trustworthy picture of where experts would disagree, which is what the method claims to provide.