Ask several doctors to trace the same small, faint lesion, and their outlines will differ by a band of roughly fixed width. For a large organ that band is a thin rind around a large, agreed-upon interior. For a tiny or ragged target, the band can be as large as the interior itself. At that point the target is mostly boundary, and there is no longer a single correct answer for a segmentation model to learn. We propose to measure this crossover directly and to state it as a law. The fraction of an object's area that lies inside the rater-disagreement band should scale as $(w/R)^{2-D}$. Here $w$ is the width of the disagreement band, $R$ is the object's size, and $D$ is the fractal dimension of its boundary, a number between 1 and 2 that says how ragged the edge is. The law predicts where standard overlap scores stop being meaningful. It also predicts where methods built to sharpen edges should stop helping and start hurting.
The framing comes from metallurgy. In the inverse Hall-Petch effect, shrinking the grains of a metal makes it stronger, until the thin disordered layers between grains take up as much volume as the grain interiors. Below that point the relationship reverses and the metal softens, because deformation now happens by grains sliding along those layers (Atomic-Scale Modeling of the Deformation of Nanocrystalline Metals). The lesson we borrow is narrow. When the boundary's share of the volume crosses a threshold, a method's gains can reverse sign, not merely saturate. The segmentation claim does not depend on the physics: the law above follows from plain geometry, and the experiments test it on multi-rater medical data. The metallurgy only tells us to look for a reversal as a function of one dimensionless boundary fraction. Reporting average Dice per dataset would hide it.
Core Claim
For each annotated object, let $w$ be the local half-width of the band where rater masks disagree, $R$ a characteristic radius of the consensus region, and $D$ the box-counting dimension of the band's centerline. Box-counting estimates $D$ by counting how many grid boxes of shrinking side the curve touches. The area of a width-$w$ neighborhood of a curve with dimension $D$ scales as $R^{D} w^{2-D}$; this is the standard Minkowski-content estimate. The core area scales as $R^2$. The boundary fraction is therefore
$$f = \frac{A_{\text{band}}}{A_{\text{core}}} \approx c\left(\frac{w}{R}\right)^{2-D},$$
where $c$ is a shape constant. For a smooth edge ($D=1$) this reduces to the familiar $f \approx 2w/R$ for a disc. Two predictions follow.
The first concerns the ceiling set by the raters themselves. The shortfall of inter-rater Dice below 1 should scale with $f$. On a log-log plot against $R/w$, its slope should therefore be $-(2-D)$, where $D$ is measured independently. Ragged boundaries should flatten the curve, so the ambiguous regime extends to larger objects than size alone would suggest.
The second concerns method rankings. Methods that sharpen boundaries using crisp labels should beat plain training when $f \ll 1$ and lose to it when $f \gtrsim 1$. Methods that model the distribution of rater outlines should show the reverse pattern. We call this sign change the inversion point $f^*$, and we claim it is shared across tasks once they are indexed by $f$ rather than by size or by dataset.
No prior work makes either claim. The QUBIQ challenge (QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge) collected multi-rater data on six tasks and found that ensembles win. It scores boundary errors uniformly across object sizes and does not characterize margin geometry. Fractal analysis in medicine has been applied to single-pipeline masks, never to the geometry of disagreement itself. For example, Robust and efficient computation of retinal fractal dimension through deep approximation regresses one deterministic pipeline's fractal-dimension output from fundus images.
Research Program
Measurement. For every object in the multi-rater datasets, we compute three quantities. The first is the local band width $w$: half the distance between the 0.1 and 0.9 rater-agreement contours, taken along normals to the 0.5 contour, with its median and spread recorded. The second is $R$, the square root of the consensus area divided by $\pi$. The third is $D$, the box-counting dimension of the 0.5-agreement contour, fitted only over the scale range where the log-log plot is actually linear. Fractal scaling in real structures holds over one to two orders of magnitude at most, and we report that range for every object. We do not trust $D$ for objects whose contours span too few pixels for a stable fit. These objects are flagged and analyzed separately rather than dropped, because the file documents how dropping hard cases biases conclusions (More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel).
Law fitting. We regress the log of each object's inter-rater Dice shortfall on $\log(R/w)$. Then we test whether the per-task slope matches $-(2-D)$, using the independently measured median $D$ as the prediction.
Model sweep. A fixed set of segmentation methods is trained on each task. Per-object gains relative to a plain baseline are then plotted against $f$ to locate $f^*$.
Method test. We then build one method that follows from the law, a regime-switching loss. For each object, a small head predicts $\hat f$ from the image, trained against measured $f$. The loss then blends a crisp contour term with a soft distribution-matching term, with a weight that is a sigmoid in $\log(\hat f/f^*)$. This is the only new architectural ingredient. It exists to test whether the law is useful. It is not the headline.
The prior work plays three distinct roles.
- Work that motivates the proposal. QUBIQ, Pionono, and the Liu et al. variational model all show that disagreement concentrates at margins, and all model it as either a global per-rater latent code or a per-pixel average. None asks how the magnitude of the problem scales with object geometry. Pionono is Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation, which embeds each rater as an 8-dimensional Gaussian applied uniformly across the image. The variational model is Variational Inference for Quantifying Inter-observer Variability in Segmentation of Anatomical Structures.
- Work that supplies the contrasting method family. Reverse attention removes confident foreground predictions so that later layers focus on the transition zone, and it trains on crisp masks (Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation). Contour-loss branches do the same with an auxiliary edge loss (A Medical Image Segmentation Method based on Multi-scale Features and Contour Loss Constrain (S)). These are exactly the methods our second prediction says should invert at high $f$.
- Work that constrains the proposal. The diagnosis-first method Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle deliberately smooths away high-frequency parts of multi-rater "expertness" maps and gains diagnostic AUC by doing so. That appears to cut against treating boundary roughness as signal. We address it directly with an experiment described below, which asks whether measured $D$ carries information beyond size.
Baselines
The naive reference point is a ResNet34 U-Net trained with cross-entropy plus Dice loss on majority-vote masks. Its results are reported both as an overall average and stratified by $f$. Every gain in the proposal is measured against this model object by object, so it defines the zero line for the inversion analysis.
The strongest existing competitor is a deep ensemble of five deterministic U-Nets trained on soft averaged rater labels. This is the approach that top QUBIQ teams used; they reached average Dice between 0.80 and 0.85 according to the challenge summary (QUBIQ). It is the strongest competitor because QUBIQ found such ensembles beat variational and dropout models across all six tasks. It is also the natural high-$f$ method.
Two rater-modeling baselines are included. The first is Pionono, reported at Cohen's κ = 0.758 on Gleason 2019 and 0.711 on the 25-rater breast cancer dataset (Pionono). The second is the variational model, reported at a Continuous Dice of 0.5372 on QUBIQ brain growth, against 0.5083 for Bayes dropout (Variational Inference). The boundary-sharpening baseline is reverse attention retrained on majority-vote masks with the same U-Net backbone, so that the effect of backbone choice is removed.
All methods share the same splits, backbone, augmentation, and epoch budget. Each receives the same hyperparameter search budget of twenty trials per task, and the ensemble's five-fold compute is reported alongside its scores. We found no reported result on Lacuna for any of these methods stratified by object size or boundary fraction. The stratified numbers will therefore all be produced by this study. Published aggregate numbers serve only as a check that each reimplementation is faithful.
The key ablations isolate the geometric ingredient. The first stratifies by size alone, which is $f$ with $w$ and $D$ ignored. The second uses $f$ with $D$ fixed at 1. The third uses $f$ with measured $D$. For the method test, we compare the regime-switching loss against its two fixed endpoints, pure crisp and pure soft, and against a version switched by size alone.
The proposal beats the baselines in two separate senses. Scientifically, the full $f$ must explain significantly more held-out variance in per-object Dice shortfall than size alone. Practically, the regime-switching single model must match the ensemble's Q-Dice on high-$f$ objects, and the reverse-attention model's Q-Dice on low-$f$ objects, while using one-fifth of the ensemble's inference compute.
Why not just…?
Why not just stratify results by object size? This is the most serious threat to the scientific claim. If size alone predicts the Dice shortfall and the location of the inversion as well as $f$ does, then the fractal dimension and band-width terms add nothing and the law reduces to "small things are hard." The first ablation settles this. We require a significant held-out gain from including $D$, which is only possible if tasks differ in boundary roughness at matched size. The QUBIQ pancreatic-lesion and Gleason tasks should supply such differences. If they do not, we will report that the simpler description suffices.
Why not just train on soft labels everywhere? This is the most serious threat to the method. Soft averaged labels already encode the disagreement band, and QUBIQ's winners used essentially this recipe. Our second prediction says soft targets should lose to crisp boundary training only for low-$f$ objects. If the pure-soft endpoint ties the regime-switching loss at every $f$, the switch is unnecessary and we will say so. The measurement contribution would still stand.
Why not just use an ensemble or Pionono? Both capture rater variation, but neither tells you when crisp training suffices. The ensemble costs five times the compute, and Pionono's single global rater code cannot express disagreement that varies locally along a boundary. We still include both as baselines. If either one dominates at every $f$, the inversion prediction is falsified.
Why not drop the metallurgy analogy? The analogy makes no quantitative claim; the equation comes from geometry alone. What the analogy contributes is the hypothesis of a sign reversal at a boundary-fraction threshold, as opposed to a gain that simply shrinks toward zero. A reviewer can discard the analogy and keep every experiment unchanged.
Proposed Experiments
Data. The primary data are the six QUBIQ tasks and Gleason 2019. The breast cancer dataset with 25 raters serves for the most precise estimates of $w$, and REFUGE-2 multi-rater optic cups serve as a smooth-boundary, low-$D$ control.
Decisive check 1: the slope. The fitted slope of $\log$(inter-rater Dice shortfall) against $\log(R/w)$ must lie within its confidence interval of $-(2-D)$ in most tasks, where $D$ is measured independently. If slopes cluster at $-1$ regardless of $D$, fractal roughness is irrelevant to disagreement and the core claim fails.
Decisive check 2: the inversion. The per-object gain of reverse attention over the naive U-Net, plotted against $f$, must cross zero at an $f^*$ whose confidence intervals overlap across at least four tasks. A gain that shrinks toward zero without turning negative would support a weaker saturation story instead.
Decisive check 3: the diagnosis-first tension. On REFUGE-2, where diagnostic labels exist, we compare two versions of the diagnosis-first pipeline. One uses smoothed expertness maps as in the original paper; the other preserves the multi-rater band at the measured $D$. This asks whether the high-frequency content that the method removes carries the roughness signal our law depends on.
Decisive check 4: the method. The regime-switching loss is compared with the ensemble and with its own two endpoints on Q-Dice and calibration error, stratified by $f$, with inference compute reported next to each score.
References
- Atomic-Scale Modeling of the Deformation of Nanocrystalline Metals
- QUBIQ: Uncertainty Quantification for Biomedical Image Segmentation Challenge
- Robust and Efficient Computation of Retinal Fractal Dimension Through Deep Approximation
- More is more: leveraging multi-rater information for whole slide images grading via virtual expert panel
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Variational inference for quantifying inter-observer variability in segmentation of anatomical structures
- Parallel Res2Net-based Network with ReverseAttention for Polyp Segmentation
- A Medical Image Segmentation Method based on Multi-scale Features and Contour Loss Constrain (S)
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle