Faint Ends Look Shorter: Auditing Length Bias in Medical Segmentation

Two old vision-science papers predict a specific, directional error. One predicts that faint boundaries make a structure look shorter. The other predicts that uniform thinning from all sides clips the ends of wide shapes. We argue that the same error is hiding inside modern medical image segmentation, and that current benchmarks cannot see it. Consider a thin anatomical tube, such as a nerve canal or a blood vessel, whose wall fades into surrounding tissue near its ends. A segmentation model will probably stop the tube early, and the human annotator who drew the reference mask will probably stop early too. Overlap scores barely register a few missing millimetres at the tip. Yet those millimetres are what a surgeon measures, and they shift whenever a scanner changes the local contrast. That makes the error a reproducibility problem that looks like a reporting detail.

The proposal is a controlled audit with three parts. We fade the contrast of real structures while holding their true shape fixed, so the reference geometry no longer depends on anyone's perception. We then measure the signed change in centerline length, meaning whether the structure got shorter or longer, as a function of local contrast and local width. We apply this to both trained models and the skeleton-based metrics used to score them. The result is a single number per model, a distinctness–length slope, that a benchmark can report next to Dice. We then use this number to test which fixes actually remove the bias.

Core Claim

The mechanisms come from two sources. Conjectures on the Perception of Elongation models perceived length as the length of a skeleton left after eroding a shape uniformly from all sides. Because the ends erode at the same rate as the flanks, wider shapes lose more length. Human Vision Paradox Implicates Relaxation Model explains an adaptation illusion by coupling edge signal strength to perceived extent, so that fainter contours read as shorter. Both are hypotheses about human vision. Montalvo herself flags the distinctness–length relation as not yet systematically mapped. We do not assume either is true of the brain. We turn each into a falsifiable prediction about segmentation pipelines.

Our first claim is about the metrics themselves. Centerline metrics compute skeletons by repeated erosion, which is exactly Deutsch's isotropic thinning. One example is the soft-skeleton centerline Dice (clDice) used in Automatic segmentation of mandibular canal using transformer based neural networks. Such metrics should therefore report an end-length deficit that grows with local radius, even for a perfect segmentation. Geometry suggests this: for a tube with rounded caps, the medial axis stops roughly one radius short of each tip.

Our second claim is about trained models. Their predicted centerlines should end earlier as local wall contrast drops, with a positive slope between contrast and signed length error. Unsigned metrics such as Dice and the 95th-percentile Hausdorff distance (HD95) will largely hide this slope.

Our third claim is about annotation. Human annotators are subject to the same coupling, so single-rater reference masks may carry it too. A benchmark scored against such masks can report a biased model as accurate. Evidence on real data would take the form of annotated canal length correlating with end contrast.

What would be new is the combination: signed, contrast-conditioned and width-conditioned length measurement, under geometry-preserving contrast manipulation, applied to models, metrics and annotations alike. The closest work does not do this. Deep Distance Transform for Tubular Structure Segmentation in CT Scans thins a probability map into a pseudo-skeleton and regrows the tube from predicted radii. clDice-based training penalizes broken centerlines. Neither measures whether the ends recede with contrast, and both use erosion-style thinning that our first claim says is itself biased.

Research Program

The core tool is a geometry-preserving contrast attenuator. We take a scan with a reference mask and extract the mask's centerline by geodesic tracing between endpoints rather than by erosion. Along a chosen terminal segment, we blend the intensities inside the structure and in a thin wall band toward the statistics of the surrounding tissue. The mask is left untouched. An attenuation factor $a \in [0,1]$ controls the blend, where $a=0$ leaves the scan unchanged and $a=1$ makes the wall indistinguishable from background.

Synthetic fading could look unlike real fading, which would undermine the audit. To guard against this, we measure the local wall contrast at the naturally faded canal ends in real scans. We restrict $a$ to the range that reproduces those contrast values and their noise spectrum. We also run the whole analysis a second time without synthetic manipulation, on the natural variation in contrast along real structures.

For each prediction, we trace its centerline with the same geodesic procedure. Our main quantity is the signed endpoint deficit $\Delta = L_{\text{pred}} - L_{\text{ref}}$, where $L$ is centerline length in millimetres. We then fit

$$\Delta_i = \beta\, c_i + \gamma\, r_i + \epsilon_i,$$

where $c_i$ is the measured local wall contrast at terminal segment $i$, $r_i$ is its local radius, and $\epsilon_i$ is residual error. $\beta$ is the distinctness–length slope, which Montalvo's coupling predicts is positive. $\gamma$ is the width slope, which Deutsch's account predicts is negative for erosion-based length.

We estimate the same regression for the metric acting on the reference mask itself (claim one), for trained models (claim two), and for existing annotations on real scans (claim three). Confidence intervals come from bootstrapping over patients.

The rotation arm connects to Deutsch's argument that erosion-based thinning is used because it does not depend on orientation. On the xrayto3D benchmark, we repeat the length analysis under the view tilts that Benchmarking Encoder-Decoder Architectures for Biplanar X-ray to 3D Bone Shape Reconstruction showed to collapse rib accuracy. We ask whether the tilt damage appears as tip truncation or as general overlap loss. That paper also defined a clinical length measure, vertebra canal length, which serves as a clinically anchored check of $\Delta$.

Finally, we compare remedies against the slope rather than against Dice. The deliverable is released code for the attenuator and the slope estimator, packaged as a benchmark add-on in the spirit of the standardization arguments in The Medical Segmentation Decathlon.

Baselines

Do-nothing reference. A standard nnU-Net trained and scored with Dice and HD95 only. This is what current benchmarks report, and it defines the bias we expect to find. Because nnU-Net configures itself, our tuning is limited to its own automatic configuration. On the mandibular canal, the file reports nnU-Net at a Dice of 0.796 (Automatic segmentation of mandibular canal using transformer based neural networks). "Beating" this baseline does not mean a higher Dice. It means reducing $\beta$ toward zero while keeping Dice within its bootstrap interval.

Strongest existing method on thin, fading structures. Swin-UNetR with a clDice loss plus that paper's feature-fusion module and contrast equalization, which reports a Dice of 0.844 and an HD95 of 2.947 mm on the mandibular canal data (Automatic segmentation of mandibular canal using transformer based neural networks). It is the strongest competitor because it already enforces centerline continuity, and its authors name boundary fading at the canal ends as its residual failure. We retrain it on identical ToothFairy splits with the same GPU budget and epoch cap as every other arm.

Strongest vessel baseline. Deep Distance Transform, reported at a Dice of 63.43% on the Decathlon hepatic vessel task (Deep Distance Transform for Tubular Structure Segmentation in CT Scans), run with its published configuration.

Boundary-loss comparators. Boundary DoU loss (Boundary Difference Over Union Loss For Medical Image Segmentation) and active contour loss (Learning Active Contour Models for Medical Image Segmentation), each on the nnU-Net backbone. Active contour loss contains a length penalty that we predict will make $\beta$ worse on faint tips, which gives a directional test.

For none of these methods does the file report a signed or contrast-conditioned length result. There is no reported result found on Lacuna for endpoint length bias on any of these datasets.

Ablation. The key ingredient isolated by the ablation is geometry-preserving attenuation versus natural variation. If the slope from synthetic fading disagrees with the slope estimated on real, naturally faded ends, the synthetic protocol is not measuring the real phenomenon, and we report that.

Why not just…?

Why not just train with contrast augmentation? This is the most serious threat to any claim that the bias needs a specialized fix. Randomly lowering local contrast during training may well teach models to continue faint tubes, which would flatten $\beta$. We therefore make this a first-class remedy arm rather than a strawman: nnU-Net plus random terminal-contrast attenuation, with the same budget. If it drives $\beta$ to zero, that is a useful, cheap result, and our contribution reduces to the audit itself, which is still needed to show that the fix worked. A distinctness-weighted centerline loss enters only if augmentation leaves a residual slope, particularly for real rather than synthetic fading. Two of our predictions also survive augmentation regardless. It cannot remove the width bias in erosion-based metrics (claim one), because that bias lives in the scoring. And it cannot correct references whose ends were drawn too short (claim three).

Why not just report clDice or HD95, which already target thin structures? HD95 is unsigned, and it discards the worst 5% of boundary distances, which is where short tip truncations live. clDice is computed by erosion, so under claim one it shares the bias it is meant to detect. The metric-on-reference experiment settles this directly. If clDice computed between a perfect mask and itself-after-width-change shows no radius-dependent deficit, claim one is false and we drop it.

Why not just collect multiple raters? Multiple raters measure disagreement, not shared bias. The GlaS contest (Gland Segmentation in Colon Histology Images: The GlaS Challenge Contest) and the Decathlon both flag single-rater references on indistinct boundaries as a limitation. But if the distinctness–length coupling is perceptual, every rater stops early, and averaging them preserves the error. Geometry-preserving attenuation sidesteps annotators entirely, because the true mask is fixed by construction.

Why not just apply stronger contrast equalization at test time? The mandibular canal paper already applies adaptive histogram equalization and still reports fading at the ends. This suggests that global or regional contrast stretching does not restore the local wall signal. We still include test-time equalization as an arm on the attenuated scans.

Proposed Experiments

We use three datasets. The ToothFairy mandibular canal CBCT set (153 densely annotated volumes) is the primary testbed, since it has long tubes with documented natural end fading. The Decathlon hepatic vessel task tests a second modality and branching topology. The xrayto3D benchmark ribs and vertebrae supply the tilt arm and the clinical vertebra canal length.

The decisive checks, in order, are as follows.

First, synthetic tubes of known length and radius are scored by erosion-based versus geodesic centerlines. Claim one predicts $\gamma<0$ for erosion only.

Second, the attenuation sweep is run on held-out canals for every baseline and remedy. Claim two predicts $\beta>0$ with a confidence interval excluding zero for the do-nothing and clDice baselines, alongside Dice changes smaller than their bootstrap spread. That pattern is the "invisible to Dice" result.

Third, we look for the natural-variation replication: the same sign of $\beta$ estimated on unmodified scans.

Fourth, we test the annotation effect. Annotated canal length should correlate with end contrast after controlling for patient anatomy. As a secondary test, a small blinded re-annotation of attenuated versus unattenuated copies of the same scans should show shorter traced ends on the attenuated copies.

Fifth, in the tilt arm, the xrayto3D rib degradation reported by Shakya and Khanal is decomposed into tip loss versus body loss.

Any one of these can fail independently. A null $\beta$ across all models would itself be informative: it would mean the psychophysical coupling does not transfer to learned segmenters, and that the observed canal-end failures come from something other than contrast.