Medical segmentation masks are least reliable exactly where the image itself is least informative: along the faint edge of a flat colon polyp, a diffuse tumor, or an optic cup. The standard tool for noisy labels, small-loss selection, trains only on the examples the network currently fits well. It relies on the observation that networks learn clean patterns before they memorize wrong labels. At a low-contrast boundary this rule breaks. A correctly labeled but genuinely ambiguous edge pixel and a pixel mislabeled by a sloppy outline both have high loss, so both get thrown away. The model then learns smooth interiors and never learns where the edge is.
We propose to change what small-loss selection ranks. It should rank short stretches of contour, scored by how well the label fits after the contour is allowed to shift. Most boundary annotation error is not random pixel flipping. An annotator traces a little outside or inside the true edge along part of the outline. Such an error has a clear signature: the disagreement between label and model all sits on one side of the contour, and it disappears once the drawn contour is displaced by a few pixels along its normal. Genuine ambiguity looks different. Disagreement scatters on both sides and no shift removes it. Separating these two cases gives two numbers per contour segment, a displacement and a residual loss. The method then corrects displaced segments, keeps ambiguous ones with soft supervision, and discards nothing merely because it is hard.
Core Claim
A single network with an averaged copy of itself, using displacement-then-residual scoring of contour segments, will match or beat peer-network noisy-label methods on overall overlap. It will clearly beat them on boundary metrics for low-contrast targets. It does this without knowing the noise rate and without Monte Carlo sampling. The mechanism also makes a falsifiable prediction. On multi-rater datasets, the displacement it estimates for one rater's contour should correlate with that rater's measured offset from the other raters' consensus.
Where This Sits Relative to Prior Work
The noisy-label side supplies the selection principle and its known failure modes. Co-teaching trains two networks that each pass their small-loss examples to the other. Co-teaching+ restricts this exchange to examples where the two networks disagree. JoCoR instead adds an agreement regularizer and selects on the joint loss. All three operate on whole examples in classification, and all three need a schedule tied to the noise rate. We keep their premise that early-training fits are trustworthy. We do not adopt their two-network cost or their example-level unit.
Segmentation-specific work motivates the proposal by documenting the gap. Co-Seg ports Co-teaching to lung masks. It selects at the image level, corrects labels by peer consensus, and retrains from scratch, assuming a known corruption level. Its chest X-ray targets also have high contrast. PINT names the "boundary paradox": pixel-level filters flag true edges as noise. It recovers edges only for slices it judges globally clean, using entropy from several noisy forward passes per step. CLCS adds class-dependent curriculum thresholds and reverse cross-entropy for flagged pixels, but its thresholds ignore where a pixel lies relative to the boundary. NA-CLD shows that confidence-based cleaning collapses toward background, and it names hard-pixel mining at blurry boundaries as open. None of these models the geometry of boundary error. They decide pixel by pixel or image by image whether a label is trustworthy. We instead ask how far a stretch of contour is off.
Two further papers constrain the design. Beyond Boundary Noise shows that aggregated output entropy rings every boundary yet scores 0.48 AUROC at detecting whether annotators disagree about a nodule's existence. Entropy is therefore a poor selection signal, and our criterion deliberately avoids it. The same paper also tells us our method addresses boundary placement, not presence ambiguity, and we will report that limit rather than hide it. JCAS corrects losses through pixel-pair affinities and noise transition matrices. It is strong on rigid surgical tools, but its authors note that transition matrices struggle with spatially varying noise, which is exactly what per-segment displacement is.
The boundary-architecture side supplies the intuition and the backbone. Polyper resolves ambiguous polyp edges by attending from the confident interior. It defines the boundary band with fixed-kernel erosion and dilation, which the research file flags as failing for very small objects. We borrow the interior-first logic for selection rather than for architecture, and we replace fixed kernels with bands scaled to object size.
Research Program
Training uses one segmentation network (the student) and an exponential-moving-average copy of it (the teacher), as in Mean Teacher. This requires no second backbone and no stochastic passes. After a short warm-up that exploits early learning, each annotated mask is processed as follows.
We compute the signed distance $\phi(x)$ of every pixel $x$ to the annotated contour, negative inside and positive outside. A shifted label is then just a re-threshold, $y^{(d)}(x) = \mathbb{1}[\phi(x) < d]$, where $d$ is a displacement in pixels. We split the contour into arcs of roughly equal length. Each arc owns the pixels within a band whose half-width is proportional to the object's equivalent radius, so tiny polyps get thin bands and large lesions get wide ones. For arc $s$ with band pixels $B_s$ and teacher probabilities $p(x)$, we compute
$$ L_s(d) = \frac{1}{|B_s|}\sum_{x \in B_s} \mathrm{CE}\big(p(x),\, y^{(d)}(x)\big) + \lambda |d|, $$
where $\mathrm{CE}$ is cross-entropy and $\lambda$ penalizes large shifts. We search $d$ over a small integer range scaled with the band, so the search costs a few extra loss evaluations and no extra forward passes. The minimizer $d^*_s$ is the estimated displacement of that arc. The minimum $r_s = L_s(d^*_s)$ is its residual loss.
Small-loss selection is applied to $r_s$. We do not use a preset noise rate. Each epoch we fit a two-component mixture to the residuals and treat the low-residual component as reliable. Reliable arcs with nonzero $d^*_s$ are supervised by the shifted label $y^{(d^*_s)}$: these are fixable outline errors. Reliable arcs with $d^*_s = 0$ keep their original label. High-residual arcs are judged genuinely ambiguous or randomly corrupted. They are not dropped. They are trained toward a soft target that averages the annotation and the teacher, with reduced weight, so the edge branch still receives gradient. Interior pixels, outside all bands, go through ordinary per-class small-loss selection. Class-dependent thresholds follow CLCS so that rare foreground classes are not filtered out. Corrections are recomputed each epoch rather than written into the dataset, so there is no offline relabel-and-retrain stage as in Co-Seg.
We will plug the procedure into two backbones: a plain U-Net, and a boundary-attention polyp network in the style of Polyper or MISNet. The second tests whether an edge branch trained on corrected arcs behaves better than one trained on raw edges. Gap 5 in the research file notes that no prior work gives explicit edge branches noise-tolerant supervision.
The main risk is self-confirmation. A teacher that over-smooths low-contrast objects will systematically pull arcs inward and "correct" true edges into its own bias. We address this in three ways. The shrinkage penalty $\lambda$ opposes large shifts. We impose a symmetry check: if the $d^*$ values within an image are all the same sign while the image's labels are known-clean, that image's corrections are vetoed. And we report the distribution of $d^*$ on clean labels as a standing diagnostic, which should be centered at zero. If this diagnostic fails, the method fails, and we will say so.
Baselines
The naive reference is the same backbone trained with Dice plus cross-entropy on the noisy masks, at the same epochs, augmentations and learning-rate schedule. It shows how much damage the noise does.
The first boundary-specific naive baseline is a fixed ignore band. It simply removes supervision within $k$ pixels of every annotated contour, with $k$ tuned on the validation set. It is concrete, cheap, and plausibly competitive on Dice, because it removes most displacement noise without any modeling.
The stratified small-loss ablation is the "obvious" Gap 1 idea. It ranks pixel losses separately in the interior and in the size-scaled band, with a mixture-fitted threshold per stratum, but performs no displacement search. It isolates whether the shift-then-residual step adds anything beyond spatially separated selection.
A second ablation keeps the displacement search but discards high-residual arcs instead of soft-supervising them. It isolates the "never drop hard edges" ingredient.
Among peer and teacher methods, we run Co-teaching adapted to pixels via its official code bhanML/Co-teaching. We also reimplement Co-teaching+ and JoCoR at the pixel level, and include Co-Seg. The strongest competitors are PINT and CLCS, because both were built specifically to protect boundary pixels and rare classes under noisy segmentation labels. PINT reports 76.25% Dice on Left Atrial MRI under 75% erosion/dilation noise, against 72.76% for a plain V-net, and 75.31% Dice on Cervical Cancer CTV against a 77.26% noise-free bound. CLCS reports more than 3% Dice and mIoU over SCE and Co-teaching+ on EndoVis18; the file gives no absolute value, so we will reproduce it. We also run JCAS from its official code CityU-AIM-Group/JCAS as the affinity-based alternative.
Fairness means the same backbone, input resolution and epoch budget for every method. Every method receives the same number of hyperparameter trials, about twenty, on a held-out validation split with noisy labels; clean validation labels are never used for tuning. Methods that need a noise rate receive the true rate, which advantages them. Training wall-clock and peak memory are reported alongside accuracy, since two-network and Monte Carlo methods spend more compute.
Beating a baseline means a gain on 95th-percentile Hausdorff distance and boundary F-score that exceeds seed-to-seed variation across five seeds, with no loss on Dice. On Left Atrial specifically, it means exceeding PINT's 76.25% Dice under its 75% noise protocol at lower training cost.
Why not just…?
Why not just ignore a fixed band around every annotated boundary? This is the most serious threat. On mild, uniform erosion/dilation it may match us on Dice, because it removes most of the noise. It cannot recover the information inside the band, so it should lose on boundary metrics. It should also fail on small polyps, where any fixed band swallows a large fraction of the object; the fixed-kernel failure for small structures noted around Polyper predicts this. The ETIS small-polyp split and a band-width sweep will settle it. If the ignore band matches us on HD95 as well, our displacement modeling is not earning its cost, and we will report that.
Why not just run small-loss selection separately inside the boundary band and in the interior? Stratification fixes the bias of comparing edge losses against interior losses. Within the band, though, a correctly labeled faint edge and a displaced outline still both have high loss, which is PINT's boundary paradox restated inside the stratum. The stratified ablation measures this directly. Our method should differ most on arcs whose true label is known to be correct but low-contrast; we can identify these in the synthetic-noise setting.
Why not just use a strong boundary architecture with a noise-robust loss? Edge-attention networks such as MISNet and Polyper are trained to reproduce whatever edges they are shown. Given the memorization capacity documented by Co-teaching, their edge branches will fit displaced outlines. Reverse cross-entropy of the kind CLCS uses bounds the damage but does not move the contour back. Our backbone experiment runs Polyper-style networks with and without our selection, and with CLCS's loss, on identical noisy masks.
Why not just use affinity-based loss correction like JCAS? Pairwise affinity is genuinely more noise-tolerant. However, JCAS's transition matrices assume noise that is uniform across space, while outline displacement varies arc by arc. JCAS was also shown only on rigid instruments. We run it on our per-arc displacement noise and on real multi-rater data, where we expect the gap to be largest. If JCAS wins there, the affinity route is the better bet.
Proposed Experiments
The synthetic study uses Kvasir-SEG and CVC-ClinicDB for training, with ETIS and CVC-ColonDB as low-contrast generalization sets, and the Left Atrial MRI volumes under PINT's 25/50/75% protocol. Beyond the global erosion/dilation used by Co-Seg and PINT, we add per-arc displacement noise, in which each arc is shifted independently by a random signed amount, and a mixture of displacement and random flips. Because the true contour is known here, we can score the estimated displacements directly against the injected ones.
The real-noise study uses RIGA for optic disc and cup, which has multiple raters. It also uses LIDC-IDRI, training on a single radiologist and testing against the consensus of the remaining radiologists, and the Cervical CTV setting if access permits. Metrics are Dice, mIoU, 95th-percentile Hausdorff distance, boundary F-score at a tolerance scaled to object size, and results stratified by object area so that polyps under 6% of the image are reported separately.
Three checks decide the claim. First, on the synthetic per-arc noise, $d^*_s$ should recover the injected shift, measured by correlation and mean absolute error in pixels. On clean data its distribution should center at zero; an inward bias would confirm the over-smoothing risk. Second, on LIDC-IDRI and RIGA, the displacement estimated for one rater's contours should correlate with that rater's signed offset from the other raters' consensus. Training never sees this quantity, so it is an honest external test of the mechanism. Third, the method should beat the fixed ignore band and the stratified ablation on HD95 and boundary F-score for small, low-contrast objects, at equal Dice.
As a scope check, we will apply the Ambiguity Head probe from Beyond Boundary Noise to LIDC-IDRI cases where raters disagree about whether a nodule exists. We expect our method not to help on these cases, and we will report it as bounded to boundary placement.
References
- Co-teaching: Robust training of deep neural networks with extremely noisy labels
- How does Disagreement Help Generalization against Label Corruption?
- Combating Noisy Labels by Agreement: A Joint Training Method with Co-Regularization
- Co-Seg: An Image Segmentation Framework Against Label Corruption
- Distilling Effective Supervision for Robust Medical Image Segmentation with Noisy Labels
- Imbalanced Medical Image Segmentation With Pixel-Dependent Noisy Labels
- Clean Label Disentangling for Medical Image Segmentation with Noisy Labels
- Beyond Boundary Noise: Aggregated Aleatoric Uncertainty Fails to Capture Presence Ambiguity in 3D Lung Nodule Segmentation
- Joint Class-Affinity Loss Correction for Robust Medical Image Segmentation with Noisy Labels
- Polyper: Boundary Sensitive Polyp Segmentation
- Multi-scale information sharing and selection network with boundary attention for polyp segmentation