Letting the Image Decide Where One Annotator's Boundary Can Be Trusted

Medical segmentation models are usually trained on masks drawn by a single expert, and they learn that expert's habits along with the anatomy. One radiologist may draw consistently loose margins and another tight ones. Semi-supervised learning trains on a few labeled images plus many unlabeled ones by having the model label the unlabeled images for itself. This makes the problem worse: the model's own guesses, called pseudo-labels, carry the annotator's habits, and training on them spreads those habits across the whole unlabeled set. Multi-rater methods that model disagreement between experts need several masks per image, which most clinical datasets do not have.

We propose using the image itself as an independent referee. A hybrid active contour energy is a classical score that says how well a curve fits the image, combining edge strength with how uniform the intensities are on each side. For each point on a predicted boundary, we slide the curve inward and outward along the boundary normal and ask how sharply this energy changes. Where the energy has a sharp minimum, the image fixes the boundary. There, a lone annotator's systematic offset can be measured and corrected, and pseudo-labels can be pulled toward the image-supported boundary. Where the energy is flat, as in low-contrast tissue, the image is silent, and the boundary is a matter of convention. There, the model should predict a band of plausible boundaries rather than one confident line. The hypothesis is that this image-derived split lets a model trained on one annotator's masks recover much of the between-expert spread it never saw.

Core Claim

Boundary uncertainty in medical images has two sources that current semi-supervised methods lump together as generic entropy. The first is image-limited ambiguity: contrast is low and no annotator can do better. The second is convention: raters disagree even where the image is clear. We claim a hybrid active contour energy separates the first from the second without any multi-rater labels. This gives a semi-supervised objective in which one annotator's labels are treated as a single sample from a boundary distribution. The spread of that distribution is set at each location by image evidence. We expect such a model, trained on one rater, to beat standard semi-supervised training when scored against other raters and the consensus. We also expect its predicted spread to track real inter-rater disagreement on benchmarks where several raters are available for evaluation.

This is not the obvious combination of adding an active contour loss to a semi-supervised method. Learning Active Contour Models for Medical Image Segmentation turned the Chan-Vese energy into a differentiable length-plus-region loss. It trained fully supervised and pushed every boundary toward one sharp answer. We use the energy's local sharpness as a measurement, not as a uniform smoothing penalty. Shape Transformation Driven by Active Contour for Class-Imbalanced Semi-Supervised Medical Image Segmentation (STAC) evolves pseudo-label boundaries outward along normals only as data augmentation. It assumes the pseudo-labels are correct and ignores who drew the seed labels. We evolve along the same normals to test whether the image supports a boundary, and our pseudo-label targets depend on that test.

The closest competitor is Annotation Ambiguity Aware Semi-Supervised Medical Image Segmentation (AmbiSSL). It is the only semi-supervised method in the file that models annotator ambiguity. It makes pseudo-labels diverse by randomly zeroing decoder weights, and it needs multi-rater labels on the labeled subset. Our proposal differs on both counts. It works when the labeled set has one rater. Its diversity is placed where the image is uninformative, rather than produced by random weight perturbation spread uniformly over the output.

The Mechanism

Represent the predicted boundary through its signed distance function, which gives each boundary point $s$ an outward normal $n(s)$. For a displacement $t$ along that normal, let $E_s(t)$ be a local hybrid energy computed on the image profile around $s$. It sums three terms. The first is a Chan-Vese region term: how well the inside and outside intensity means fit the pixels on either side. The second is a local-variance term in the spirit of A Novel Hybrid Active Contour Model for Intracranial Tuberculosis MRI Segmentation Applications, which combined global statistics with local intensity variance to remain sensitive under brightness shifts. The third is an edge-magnitude term, weighted adaptively as in the signed-pressure-force balancing of Land use Detection in Nusajaya using Higher-Order Modified Geodesic Active Contour Model. These papers supply the energy; we reuse their ingredients rather than invent new ones.

From each profile we extract two quantities. The first is the minimizing displacement $t^*_s$. The second is the image-evidence score

$$\kappa_s = \frac{\partial^2 E_s}{\partial t^2}\Big|_{t=t^*_s} \Big/ \bar{E}_s,$$

where $\bar{E}_s$ is the mean energy on the profile, so $\kappa_s$ is scale-free. A large $\kappa_s$ means a sharp, image-determined minimum. A small $\kappa_s$ means a flat landscape.

The network predicts a boundary center $\mu$ and a normal-direction spread $\sigma$, both as signed-distance maps. The unified objective is

$$\mathcal{L} = \mathcal{L}_{\text{sup}}(\mu, \sigma; y_A) + \lambda \sum_{s} \Big[ w(\kappa_s)\, \ell\big(\mu_s, t^*_s\big) + \big(1 - w(\kappa_s)\big)\, \ell_{\text{band}}\big(\mu_s, \sigma_s; \hat{\mu}^{T}_s, \sigma_0\big) \Big].$$

Each symbol is as follows. $y_A$ is rater A's mask on labeled images. $\mathcal{L}_{\text{sup}}$ is the negative log-likelihood of A's boundary under a Laplace distribution along the normal with center $\mu$ and scale $\sigma$. $w(\kappa)$ is a sigmoid gate on $\kappa$. $\ell$ is a robust penalty pulling the student's center toward the image-supported minimum. $\ell_{\text{band}}$ asks the student to match the teacher's center $\hat{\mu}^T_s$ only up to a spread of at least $\sigma_0$, so ambiguous regions are not forced into a false sharp answer. Following Mean Teacher, the teacher is a slowly updated average of the student's weights.

On labeled images, the average offset between A's boundary and $t^*_s$ at high-$\kappa$ points gives an estimate $\delta_A$ of A's systematic margin style. The model can then output either the style-free center $\mu$ or an "A-style" boundary $\mu + \delta_A$. This mirrors the consensus-versus-personal split of Modeling annotator preference and stochastic annotation error for medical image segmentation, but it is estimated from one rater plus image physics.

The file supports DyCON's point that uncertain regions should be down-weighted rather than discarded (DUCore). That paper decays the weight on uncertain voxels using the model's own entropy. We make the same kind of move, but gate on the image energy rather than on network entropy. Network entropy inherits the seed annotator's bias; image energy does not.

One risk is built into the design. With a single rater, $\delta_A$ mixes rater style with any anatomical convention in which boundaries are defined away from intensity edges. The optic cup, for example, is partly defined by vessel bending rather than intensity. Correcting $\delta_A$ could therefore remove a legitimate convention along with the bias. For this reason the style correction is reported separately, and the spread prediction stands as the primary claim.

Research Program

The work has three stages. The first is a measurement study, run before any model training, that tests the premise directly. On RIGA (six raters), LIDC-IDRI (up to four), and the QUBIQ tasks (three to seven raters, QUBIQ), we compute $\kappa_s$ along the consensus boundary. We also compute the per-point standard deviation of rater boundaries along the same normals, and correlate the two. We record, per anatomy, whether raters at high-$\kappa$ points cluster around $t^*_s$ or around a shared offset from it. This decides where the style correction is valid.

The second stage builds the method on a 2D U-Net with a Mean Teacher backbone. The normal profiles are one-dimensional samples taken on the teacher boundary. The added cost is small next to iterative contour solvers, whose slowness Unsupervised Deep Learning Meets Chan-Vese Model documents (about 2 seconds per image versus 0.02 seconds feed-forward).

The third stage trains in the single-annotator regime and evaluates against raters and consensus the model never saw.

Baselines

The naive reference is a supervised U-Net trained only on rater A's labeled images. Beside it is plain Mean Teacher trained on A's labels plus the unlabeled pool. This second model is exactly the pipeline the file warns amplifies single-rater bias. Both get the same backbone, augmentations, and a fixed tuning budget of equal trials per method on a validation split scored against rater A only. This mirrors what a practitioner with one rater could actually do.

The strongest same-information competitor is Mean Teacher plus the uniform Chan-Vese loss of Learning Active Contour Models for Medical Image Segmentation, applied to unlabeled images. That paper reports LV Dice 0.986 and a Hausdorff drop from 5.43 mm to 4.73 mm on ACDC, so it is a credible boundary regularizer. Beating it means matching its mean Dice against consensus and improving GED by a clear margin across seeds. GED is generalized energy distance, which scores how well a set of predicted masks matches the set of rater masks.

We also run DyCON's entropy-scaled consistency (DUCore) as the strongest uncertainty-aware semi-supervised method. It reports 61.48% Dice on ISLES-2022 with 5% labels and 84.81% on Pancreas-NIH with 20%. We also run AmbiSSL restricted to single-rater labels.

As an upper reference we report AmbiSSL with full multi-rater labels. On LIDC-IDRI with 10% labels it reaches 89.86% soft Dice and 0.1620 GED (AmbiSSL). It sees roughly four times our annotation, so closing much of the GED gap to 0.1620 counts as success, not surpassing it. On RIGA, TAB's fully supervised multi-rater soft Dice of 96.70% (disc) and 85.52% (cup) (Transformer-based Annotation Bias-aware Medical Image Segmentation) serves as an oracle ceiling. No reported result was found on Lacuna for single-rater semi-supervised segmentation on RIGA, LIDC-IDRI, or QUBIQ, so our single-rater baselines are run fresh.

Three ablations isolate the key ingredient, the image-evidence gate. The first replaces $w(\kappa)$ with a constant. The second replaces $\kappa$ with raw gradient magnitude. The third replaces it with teacher entropy. A fourth removes the $\delta_A$ correction.

Why not just…?

Why not just add the Chan-Vese loss to Mean Teacher? This is the most serious threat to mean Dice, and we expect it to be competitive on overlap. Its length and region terms push every boundary toward one sharp contour, so it cannot express that a pancreas margin is uncertain while a disc rim is not. The decisive comparison is GED and the correlation of predicted spread with inter-rater spread. If uniform active contour loss matches us there too, the gating contribution is refuted.

Why not just use raw image gradient magnitude as the contrast signal? It is cheaper and might track disagreement nearly as well. The energy's region and local-variance terms should matter where edges are weak but intensity distributions still differ, and in noisy CT where gradients spike. This is a real possibility, and the ablation settles it. If gradient gating comes within noise of $\kappa$ gating, we adopt it and the contribution reduces to image-evidence gating of pseudo-label spread.

Why not just use the model's own entropy, ensembles, or dropout for spread? These capture the model's own uncertainty, which is trained on rater A and so inherits A's view of where the boundary is uncertain. On the Calibration of Image Semi-Supervised Learning Models shows semi-supervised pseudo-labels are overconfident. QUBIQ notes ensembles are costly on volumes. The entropy-gate ablation measures directly whether network uncertainty tracks real rater disagreement worse than image energy does.

Why not just widen A's mask into a uniform band, or smooth its labels? A uniform band cannot be both tight at clear edges and loose at faint ones. This baseline is included as the constant-gate ablation.

Proposed Experiments

The first decisive check is the stage-one correlation. If $\kappa$ explains little of the per-point rater spread on RIGA, LIDC-IDRI, and QUBIQ, the idea fails before any training, and we report that as a finding.

The main experiment trains each method once per rater on RIGA, LIDC-IDRI, and the QUBIQ prostate and kidney tasks, using that rater's masks on 10% and 20% labeled subsets and all remaining images unlabeled. We score against the held-out raters' masks and the consensus using soft Dice, GED, and QUBIQ's Q-Dice, which binarizes predictions at thresholds 0.1 to 0.9. We also report the rank correlation between predicted $\sigma$ and observed rater spread. Rotating over raters shows whether gains hold for both loose and tight annotators.

A bias-transfer test checks whether the $\delta_A$ correction helps. It measures whether the model's style-free center sits closer to raters other than A than Mean Teacher's does, and whether its A-style output still matches A. We report this per anatomy and expect it to hurt on the optic cup if the stage-one study shows convention-defined boundaries there.

Finally, on the single-annotator low-contrast volumes of Pancreas-NIH and ISLES-2022, extended to 3D normals, we test only whether the method keeps DyCON-level Dice. Bias cannot be measured there.