When a pancreas fades into surrounding fat on CT, or a stroke lesion blurs into healthy tissue on MRI, there is no single correct boundary to trace. Experts still spend most of their annotation time pushing a cursor along an edge they cannot actually see, and the line they produce disagrees with the next expert's. We propose to stop asking for that line. Instead, the annotator draws two loose outlines on a few slices: an inner one around tissue they are sure is lesion, and an outer one beyond which they are sure it is not. A standard segmentation network learns to predict both outlines across the whole volume. A small calibration step then widens or tightens the predicted outlines so that, at a rate the user chooses, the outer outline contains everything an expert would call lesion and the inner outline contains nothing an expert would call background. This calibration is checked using only the same cheap two-outline labels, so no expert ever traces a full boundary.
The bet is that, for a fixed amount of expert time on low-contrast targets, these "band" labels produce better point segmentations than scribbles or a few densely traced volumes. They also produce something those approaches cannot: a map of doubt, drawn as two contours, with a stated error rate. This form is what clinicians said they wanted when they rejected continuous heatmaps (Confidence Contours).
Core Claim
Three bodies of work each have half of this. Confidence Contours introduced the min/max two-polygon protocol and trained a two-channel network on it. Experts preferred it, and it cut representation error by 19.9% on LIDC. However, it requires both polygons on every 2D training image, makes no use of unlabeled slices, and gives no guarantee on the predicted band.
Conformal methods such as Conformal Lesion Segmentation and Conformal Semantic Image Segmentation supply the guarantee: they pick a threshold on a held-out calibration set so that an error rate stays below a target. On KiTS21, for example, the false negative rate drops from about 0.53 at the default 0.5 threshold to about 0.10. These methods calibrate a single mask, though, and they assume a fully traced calibration set.
The scribble literature, culminating in ScribbleBench, shows that a plain nnU-Net trained only on annotated voxels (a "partial loss") beats elaborate propagation pipelines. That baseline reaches 0.813 mean Dice across seven 3D tasks, while ShapePU collapses to 0.369. This is a constraint on our design: whatever we build should stay that simple.
The claim we will test has three parts. First, sparse band annotations on a few slices per volume are enough to train a nested pair of 3D predictions with nothing more than a partial loss. Second, both predictions can be conformally calibrated using band labels alone, which gives a finite-sample guarantee with respect to the annotators' own certainty statements. Third, under matched annotation budgets on low-contrast targets, this beats scribbles and dense-few labels on point accuracy and dominates them on certified uncertainty.
What is new is not a network component. It is the pairing of a cheaper annotation form with a calibration target that this annotation form can verify.
Research Program
Each annotated slice gives three regions: the "certain-inside" region within the inner outline, the "certain-outside" region beyond the outer outline, and the ambiguous band between them. The network has two output channels, which we call inner and outer.
On annotated slices, the inner channel is trained toward foreground on certain-inside voxels and toward background on band and certain-outside voxels. The outer channel is trained toward foreground on certain-inside and band voxels and toward background on certain-outside voxels. Unannotated slices and volumes receive no loss, exactly as in nnU-Net+pL. We keep nnU-Net's architecture, augmentation and schedule untouched, so that any gain comes from the labels rather than engineering. A point mask is read off as the outer channel's 0.5 level set, or as the midpoint of the two channels, chosen on validation.
Calibration follows conformal risk control as used by Mossina et al.. Let $\hat O_\lambda$ be the voxels where the outer channel exceeds $1-\lambda$ and $\hat I_\mu$ the voxels where the inner channel exceeds $\mu$. For a calibration case with annotated certain-inside region $M^-$ and outer region $M^+$ (everything within the outer outline), we define two losses, each computed only on that case's annotated slices:
$$ L_{\text{out}}(\lambda) = \frac{|M^+ \setminus \hat O_\lambda|}{|M^+|}, \qquad L_{\text{in}}(\mu) = \frac{|\hat I_\mu \setminus M^-|}{\max(|\hat I_\mu|,1)}. $$
$L_{\text{out}}$ is the fraction of possibly-lesion voxels missed by the predicted outer contour. $L_{\text{in}}$ is the fraction of the predicted inner contour that leaks outside certain lesion. $L_{\text{out}}$ is monotone in $\lambda$, so standard conformal risk control picks the smallest $\hat\lambda$ whose adjusted empirical risk on $n$ calibration cases is at most $\alpha$. This guarantees that the expected risk on a new exchangeable case is at most $\alpha$; the $1/(n+1)$ correction is where the finite-sample bound comes from. $L_{\text{in}}$ is a ratio and need not be monotone in $\mu$. For it we will either use the monotone surrogate $|\hat I_\mu \setminus M^-|/|M^-|$ or run a learn-then-test style threshold search; which one is used will be stated. The calibration set consists of band-labeled cases only, so the guarantee costs no dense tracing.
There is an honest caveat. The guarantee is relative to the annotators' band, not to some unobservable truth. Whether that band actually brackets individual experts' tracings is an empirical question, and it is the central check below. The guarantee is also marginal over cases and assumes calibration and test cases are exchangeable, the limitation that CLS itself flags. We will report per-site breakdowns on ISLES-2022, whose test set comes from an unseen center, to see how it degrades.
Bands must be obtained for existing datasets. On LIDC-IDRI we use real multi-reader data: the inner region is the intersection of readers' masks and the outer region is their union, and we hold out readers to test coverage. On single-annotator 3D datasets we simulate bands. The inner outline is the ground-truth mask eroded and the outer outline is the mask dilated, using two schemes: a uniform margin, and a margin scaled by inverse local intensity gradient, so bands widen where contrast is low.
The second scheme risks circularity, because the model could learn the simulator rather than anatomy. We therefore report both schemes and treat the uniform-margin result as the conservative one. Annotation cost is modeled as polygon vertex count plus slice count, calibrated against scribble and dense-tracing cost models used for ScribbleBench-style simulation. A small timing study with two readers on a subset is a stretch goal. We do not claim real-time savings without it.
Baselines
Naive reference. An nnU-Net trained on the dense-few budget-matched labeled volumes, thresholded at 0.5, with a fixed-width band formed by eroding and dilating the prediction by a constant margin tuned on validation. This shows what a band costs without learning.
Scribble supervision. nnU-Net+pL trained on simulated scribbles at the same annotation budget, following Gotkowski et al., followed by post-hoc CLS-style conformal thresholding for the outer contour. It is the strongest competitor for two reasons: it is the best reported weak-supervision method across diverse 3D tasks (0.813 mean Dice on ScribbleBench), and it shares our exact training machinery, so any gap isolates the label type.
Semi-supervised learning from dense masks. DyCON trained on a budget-matched fraction of densely traced volumes plus all unlabeled volumes, followed by the same post-hoc conformal step. It is the strongest dense-label method on the low-contrast targets in the file, reporting 61.48% Dice on ISLES-2022 with 5% labels and 84.81% Dice on Pancreas-NIH with 20% labels. We will rerun it with its published hyperparameters, including its dataset-specific focal setting, rather than copy numbers, because our budgets differ.
Multi-annotator ambiguity. On LIDC only, we compare against AmbiSSL, which reports 89.86% soft Dice and 0.1620 generalized energy distance with 10% labels. This baseline needs multi-reader masks, which our method does not.
Dense band reference. Confidence Contours applied densely on every slice, extended to 3D nnU-Net. This is an upper reference showing what sparsity costs.
Ablations isolating the key ingredient. First, the same sparse band labels but a single-channel partial-loss model trained on certain-inside versus certain-outside, with the band unlabeled, and post-hoc thresholds producing the two contours. This tests whether the nested two-channel training matters. Second, the full method with calibration done on dense masks instead of bands. This tests whether band-only calibration loses tightness.
Fairness conditions. All methods use the same nnU-Net preprocessing, the same compute budget, the same validation split and a matched hyperparameter search of at most ten configurations.
What counts as winning. The method beats a baseline if, at equal annotation cost, it achieves equal or higher point Dice, and its certified band at $\alpha=0.1$ is narrower, measured as mean surface distance between the inner and outer contours, while meeting the target risk.
Why not just…?
Why not just use nnU-Net with scribbles and then calibrate the threshold post hoc? This is the real threat. If scribble-trained nnU-Net with CLS calibration matches our point Dice and band width at equal cost, the band protocol adds nothing and we will say so. Our reason to expect otherwise is structural. Scribbles say nothing about where certainty ends, so the network's probabilities near the boundary are shaped only by its own extrapolation. Band labels supervise exactly that transition. Gap 3 in the file notes that scribble methods have not been validated on low-contrast, scattered pathologies like ISLES-2022 or Pancreas-NIH, which is where we expect the difference to appear. The head-to-head comparison at matched budget on those datasets settles it.
Why not just trace a few volumes densely and run a semi-supervised method like DyCON? Dense tracing of an ambiguous edge spends expert time on information that is partly noise, and it forces a single boundary that AmbiSSL and Confidence Contours show experts do not agree on. DyCON also needs per-dataset tuning of its focal parameter (0.5 versus 0.8). Still, dense-few labels might win on point Dice if our cost model overstates the speed of band drawing. We will therefore report the full cost-versus-Dice curve and the crossover budget rather than a single point.
Why not just widen a point prediction by a calibrated fixed margin? A constant margin, conformally chosen, already gives a guaranteed outer contour. Its weakness is that it is equally wide at a crisp kidney edge and at a fuzzy tumor edge, so it should be much wider on average for the same coverage. The naive-reference baseline with conformal margin selection tests this directly. If its band width is close to ours, learned bands are not worth it.
Why not just use Monte Carlo dropout or an ensemble to get the band? Sampling-based uncertainty reflects model disagreement, not annotator doubt. It multiplies inference cost. Its continuous heatmaps are what clinicians found hard to act on (Confidence Contours). We include a five-member nnU-Net ensemble trained on scribbles, with its variance thresholded and conformally calibrated, as an extra comparison on band width and per-reader coverage.
Proposed Experiments
The main study runs on Pancreas-NIH, ISLES-2022 and KiTS21 with simulated bands, and on LIDC-IDRI with real reader-derived bands. LiTS and BraTS from ScribbleBench serve as checks that we do not lose on higher-contrast tasks. Budgets are set at three levels of simulated annotation cost. Metrics are point Dice and 95th-percentile Hausdorff distance, empirical risk against target $\alpha \in \{0.05, 0.1, 0.2\}$, band width, and on LIDC generalized energy distance plus per-reader coverage.
The first decisive check is on LIDC. Calibrate the bands on one set of readers, then measure the fraction of each held-out reader's tracing that falls inside the predicted outer contour and outside the complement of the inner contour. If coverage of held-out readers' tracings falls well below $1-\alpha$, the band-relative guarantee does not transfer to expert tracings and the clinical claim fails.
The second decisive check is the matched-budget comparison against nnU-Net+pL with scribbles on Pancreas-NIH and ISLES-2022. Both use identical training code, so the label type is the only variable. We need non-overlapping confidence intervals over three seeds on band width at fixed risk, and no loss in Dice.
The third check repeats the main results with uniform-margin simulated bands only, to rule out the model exploiting the gradient-scaled simulator.
References
- Confidence Contours: Uncertainty-Aware Annotation for Medical Semantic Segmentation
- Conformal Lesion Segmentation for 3D Medical Images
- Conformal Semantic Image Segmentation: Post-hoc Quantification of Predictive Uncertainty
- Revisiting 3D Medical Scribble Supervision: Benchmarking Beyond Cardiac Segmentation
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation
- Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation