Medical segmentation models are expensive to train because a specialist has to trace every organ or tumor border by hand. Two cheaper sources of signal exist. The first is scribbles: quick strokes drawn through the middle of each structure. The second is the much larger pile of scans nobody has annotated at all. Neither tells the model where a low-contrast border lies, such as the pancreas fading into fat or a tumor infiltrating brain tissue. Scribbles avoid borders on purpose, and the pseudo-labels a model invents for unlabeled scans are least reliable exactly at borders. We propose a simple way to find those borders without any edge labels. Run the model on the same scan at several zoom levels and compare the answers. Where the zoom levels agree, the region is almost always object interior or clear background. Where they disagree, the model is guessing about a border.
The proposal turns that disagreement into a training rule. Voxels where the zoom levels agree get confident hard labels. Voxels where they disagree get only a soft averaged target and are never hardened into a guess. Everything is built on top of the plain nnU-Net-with-scribble-loss recipe, which a recent benchmark found to be the only thing that reliably works outside cardiac MRI. The claim is modest and testable. Multi-scale agreement locates the voxels a semi-supervised model gets wrong better than entropy or dropout-based uncertainty does. Using it to decide where hard pseudo-labels are allowed is what lets unlabeled scans improve boundary accuracy, rather than degrade it as they currently tend to.
Core Claim
Three findings in the literature motivate the work.
First, Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation shows that nnU-Net trained with a partial loss, meaning loss computed only on scribbled voxels, reaches 0.813 mean Dice across seven 3D tasks. Specialized scribble methods collapse off cardiac data, to 0.05 on brain tumors. This constrains the design: anything we add must sit on that baseline and must not introduce fragile machinery. The same paper also leaves the gap we target. Partial loss cannot use scans that have no scribbles, and it gives no gradient at borders.
Second, ScaleMatch shows that semi-supervised networks lose up to 12% mIoU when inputs are rescaled by 0.5x or 2x. It treats this sensitivity as a defect and fuses scales into one pseudo-label everywhere, which requires a 10-epoch warm-up to avoid corrupting training. We read the same sensitivity as information. Scale instability concentrates where local evidence is ambiguous, and in low-contrast medical images that means borders.
Third, On the pitfalls of entropy-based uncertainty for multi-class semi-supervised segmentation shows that Shannon entropy can stay low where two classes compete at an interface, so entropy-gated methods confidently propagate boundary errors. That paper's fix is to replace entropy with a Bhattacharyya divergence, a measure of overlap between two probability distributions. We reuse this, but compute it between predictions at different scales rather than within one prediction.
The new element is using cross-scale disagreement as a region selector that decides where hard pseudo-labels are permitted. This is distinct from ScaleMatch's fusion-everywhere approach. It is also distinct from CoraNet, which finds uncertain regions by training two extra models with fixed 5:1 and 1:5 error costs. Our selector needs no extra networks and no cost ratios, and it is applied in a scribble-plus-unlabeled regime that none of these papers evaluate.
Research Program
A Mean Teacher pair is used: a student network trained by gradient descent and a teacher whose weights are a slow moving average of the student's. Both use the nnU-Net 3D configuration.
On each training scan, the teacher predicts at scales $s \in \{0.5, 1, 2\}$, and each output is resampled to the native grid, giving class-probability maps $p_s(v)$ at voxel $v$. We define disagreement as
$$d(v) = \max_{s \neq s'} \; D_B\big(p_s(v), p_{s'}(v)\big),$$
where $D_B$ is the Bhattacharyya distance restricted to the union of each prediction's two most probable classes. Restricting to the top two classes keeps the measure sensitive to the two-class competition that entropy misses.
Voxels are split by a threshold chosen by a target band fraction rather than a raw value. The 90th percentile of $d$ within the predicted foreground's dilated envelope is a candidate default, chosen to avoid the per-dataset threshold retuning that DyCON needed. The student's loss is
$$\mathcal{L} = \mathcal{L}_{pL} + \lambda \Big[\sum_{d(v) < \tau} \mathrm{CE}\big(f(v), \hat y(v)\big) + \sum_{d(v) \geq \tau} \mathrm{KL}\big(\bar p(v) \,\|\, f(v)\big)\Big].$$
The terms are defined as follows. $\mathcal{L}_{pL}$ is the scribble partial loss. $f(v)$ is the student's native-scale prediction. $\hat y(v)$ is the argmax of the scale-averaged teacher prediction $\bar p(v)$. $\tau$ is the band threshold. $\lambda$ is ramped up over training as in standard Mean Teacher.
Interior voxels thus receive hard labels, and band voxels receive only a soft target that encodes the scale ensemble's uncertainty. Fully unlabeled scans contribute through the bracketed term alone. The teacher's extra cost is two additional forward passes with no backpropagation. Inference uses the student at native scale, so there is no added test-time cost.
The analysis component asks whether the disagreement band actually captures errors. On held-out scans with dense ground truth, we measure the fraction of the student's misclassified voxels that fall inside the band, at a fixed band volume. The same measurement is made for four competing region selectors: entropy, Monte Carlo dropout variance, CoraNet-style conservative/radical disagreement, and geodesic distance from the nearest scribble. If scale disagreement does not capture errors better than these at equal volume, the method has no reason to exist, and we stop there.
Baselines
The naive reference is nnU-Net with partial loss trained only on scribbled scans, as in Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation. That paper reports a mean Dice of 0.813 over its seven tasks, against 0.856 for full dense supervision. We re-run it in our own splits, where only a fraction of scans carry scribbles, so the comparison uses identical data and preprocessing.
The strongest competitor is the same baseline extended with plain Mean Teacher on unlabeled scans, using hard pseudo-labels at a confidence threshold, with identical architecture, epochs, and augmentation. It is the obvious way to use unlabeled data, and it isolates whether our gain comes from unlabeled data or from the scale-disagreement gating.
We also run Bidirectional Copy-Paste and DyCON from DUCore, adapted so that their labeled loss is the scribble partial loss. These are the strongest dense-label semi-supervised methods in the file, with DyCON reporting 84.81% Dice on Pancreas-NIH at 20% dense labels. ScribbleVS, which reports 90.6% Dice on ACDC with scribbles, represents scribble-specific pseudo-labeling.
Every method gets the same total training iterations and a hyperparameter search of the same size, eight configurations, on a validation split. Teacher multi-scale passes are counted in wall-clock reporting.
Three ablations isolate the key ingredient. The first swaps the selector, replacing $d(v)$ with entropy, dropout variance, or scribble distance while holding band volume fixed. The second uses the scale-averaged pseudo-label everywhere with no band, which approximates ScaleMatch's fusion. The third uses the band but applies hard labels inside it.
Beating the field means two things on the low-contrast tasks (pancreas, BraTS, KiTS, AMOS): higher mean Dice than Mean Teacher on top of partial loss, and a lower 95th-percentile Hausdorff distance, with improvements outside seed variance over three seeds. On ACDC we require no regression.
Why not just…?
Why not just add standard Mean Teacher to nnU-Net plus partial loss? This is the real threat, and it is our main baseline. Our reason to expect it to fall short is that confidence-thresholded hard pseudo-labels are least trustworthy at borders, and the entropy pitfalls paper found entropy-gated semi-supervised training sometimes did worse than training on 3 labeled images alone. If plain Mean Teacher matches us on Hausdorff distance in the low-contrast tasks, the proposal's core claim fails, and we will report it that way.
Why not just average predictions over scales at test time? Multi-scale test-time averaging is cheap and uses the same signal. It smooths predictions, though, and cannot fix a model whose interior was trained on corrupted boundary pseudo-labels. The inference-only variant is run as an ablation on the partial-loss baseline. If it recovers most of our gain, we will conclude that the training-time gating is unnecessary.
Why not just use entropy or dropout uncertainty to choose the band? Uncertainty-gated consistency already exists in DyCON and CoraNet. The pitfalls paper's evidence that entropy misses two-class interfaces is the reason to expect scale disagreement to do better, and the error-capture analysis tests this directly before any full training. If scale disagreement ties with these selectors, the contribution reduces to a reuse of the Bhattacharyya measure.
Why not just refine borders with a CRF? Learning to Segment Medical Images with Scribble-Supervision Alone found its CRF loop unstable and hyperparameter-sensitive. CRFs also depend on intensity gradients that are nearly flat in pancreas CT. We include CRF post-processing on the partial-loss baseline as a comparison point.
Proposed Experiments
The main benchmark is the seven-task ScribbleBench suite (ACDC, MSCMR, AMOS2022, WORD, LiTS, BraTS2020, KiTS2023). It is run with scribbles on 10% and 25% of training scans, with the rest left unannotated, and with full-scribble training reported for comparability to the published 0.813. For per-task scribble results we rely on our re-runs, since the file quotes only the aggregate.
Pancreas-NIH and the Left Atrium dataset are added with simulated scribbles. These allow comparison against published dense-label semi-supervised numbers: DyCON reports 84.81% on pancreas at 20% labels and BCP reports 88.02% on Left Atrium at 5%. This shows how far scribbles plus unlabeled scans close the gap to dense labels.
To test the multi-class interface claim specifically, we run on dHCP with 3 labeled scans. The pitfalls paper reports only relative gains there (+2.40% Dice for UA-MT with Bhattacharyya), and no absolute baseline number was found on Lacuna for dHCP, so we re-run UA-MT.
Metrics are Dice, 95th-percentile Hausdorff distance, and surface Dice within 2 mm, the last because border quality is the point of the method.
The decisive checks come in order. The error-capture comparison at fixed band volume runs first and gates everything else. Next is the hard-versus-soft-in-band ablation. Last is the Hausdorff comparison against Mean Teacher on top of partial loss on pancreas and BraTS, the two settings where the benchmark found specialized methods collapse.
References
- Embarrassingly Simple Scribble Supervision for 3D Medical Segmentation
- ScaleMatch: Multi-scale Consistency Enhancement for Semi-supervised Semantic Segmentation
- On the Pitfalls of Entropy-Based Uncertainty for Multi-class Semi-supervised Segmentation
- Inconsistency-Aware Uncertainty Estimation for Semi-Supervised Medical Image Segmentation
- DUCore: Dual Uncertainty-Guided Consistency and Regional Contrastive Learning for Semi-supervised Medical Image Segmentation
- Bidirectional Copy-Paste for Semi-Supervised Medical Image Segmentation
- ScribbleVS: Scribble-Supervised Medical Image Segmentation via Dynamic Competitive Pseudo Label Selection
- Learning to Segment Medical Images with Scribble-Supervision Alone