Which Segmentation Errors Matter Mechanically? Uncertainty-Aware Bone Models for Fast Clinical Review

Automated pipelines can now turn a routine clinical CT scan into a patient-specific mechanical simulation of the spine. A neural network outlines each vertebra, and a finite element solver then computes how the bone deforms and when it fails. Finite element analysis divides the object into small pieces and solves the physics on each one. These pipelines treat the network's single outline as the truth. Two expert radiologists would not draw the same outline, though, and the disagreement is concentrated in one place: the cortical shell. This is the thin outer wall of the vertebra, often thinner than a CT voxel, so it is blurred into its neighbours. We propose to sample many plausible outlines that mimic how experts actually disagree, run each through the mechanical model, and measure how much the predicted strength and the predicted ductility (how much the bone can deform past yield before it breaks) vary across those outlines. That spread then decides which cases a clinician must check by hand.

The central bet is that strength and ductility respond very differently to boundary ambiguity. Peak load is an integral over the whole vertebral body, so small boundary shifts should average out. Post-yield deformation and energy to failure depend on where strain concentrates, which is often the thin shell that the scanner cannot resolve. If this holds, triage based on how uncertain the outline looks is the wrong signal. A case can have an obviously fuzzy outline and still give a stable strength prediction. Another case can have a crisp outline and still give an unstable ductility prediction. Triage based on the variance of the mechanical output should then let clinicians skip more cases at the same downstream error.

Where this sits

The closest pipeline is Caprara et al.. It segments lumbar vertebrae with a 3D network (Dice 93.7%, mean surface distance 0.88 mm), fits a shape model, and builds FEBio spine models in about two hours. It uses one deterministic mask and reports only range of motion.

We keep its workflow goal, but replace the single mask with a distribution of masks and ask about failure behaviour rather than stiffness alone. The Latent Doctor Model supplies the method for that distribution. It is a conditional variational autoencoder, a generator that maps an image plus a latent "annotator style" code to a plausible label. It reproduces how pathologists disagree, with KL divergence 0.461 on Gleason2019 and Spearman correlation up to 0.728 between model entropy and 20-pathologist entropy on ISUP. It also triages by predictive entropy. That work stops at the label: it never asks whether the disagreement changes a downstream quantity. We carry the samples into physics.

Querol et al. varied femur shape and density jointly through a statistical model and found a roughly six-fold stiffness range across the population. Their variation is between patients, and its sizes motivate us: geometry matters mechanically. Our variation is within one patient and between observers.

Two papers explain why the shell is ambiguous and why ambiguity matters. Kirschke et al. show that clinical multi-detector CT trabecular measures and finite element moduli correlate with cadaveric failure load (r = 0.89–0.99 in 12 vertebrae). They also note that partial-volume blurring thickens and fuses thin structures. The automated vBMD study shows that operator variability alone can erase a clinical signal. Automated trabecular density separated future fracture patients from controls (p < 0.001, β = −0.32), whereas two experts placing regions by hand did not (p = 0.30, β = −0.15). That result is the motivating evidence that observer variability is not cosmetic.

The same study also constrains us: a plain density number already predicts fracture. Any mechanical model has to earn its cost against it.

From the materials side, the strength-ductility trade-off supplies the lens rather than the method. Staggered-platelet composites (Park et al.) reach maximal toughness only where competing failure modes meet. Toughness there was experimentally noisy enough to need a heteroscedastic regression model. This is consistent with our expectation that ductility-type outputs are the fragile ones.

Within the file, we found no work that pushes a sampled distribution of observer-like segmentations through a bone failure simulation. The gap list in the file names exactly this as open.

Research Program

We will build three components.

The first is an observer-structured sampler for vertebral segmentation. It is a conditional Latent Doctor Model trained on multi-observer vertebral outlines. No multi-observer bone set appears in the file, so we will collect one. Four musculoskeletal radiologists will each independently outline 60 lumbar vertebrae drawn from the Caprara cohort style of data or from the lumbar labels in CTPelvic1K. The backbone will be pretrained on CTPelvic1K's 1,184 volumes. The decoder will be trained to output both a vertebral body mask and a separate cortical-shell mask. This adapts the boundary-core decomposition that won PENGWIN (Sang et al.), so that disagreement about the shell is modelled explicitly rather than smeared into the body.

The second is a fast per-sample mechanical evaluation. We will use voxel-based finite element models with a density-to-modulus mapping and an elastic-plastic damage law under simulated axial compression. From each model we extract three quantities: apparent stiffness, peak load (strength), and post-yield displacement and energy to failure (ductility). With 20 samples per vertebra, the per-case budget must fit the minutes-scale windows that motivate workflow work. We will report wall-clock time honestly rather than assume it.

The third is a triage rule. It flags a case for expert editing when the across-sample coefficient of variation of the clinically relevant output exceeds a threshold. That output is strength for a fracture-risk report and ductility for implant or augmentation planning.

Baselines

The naive reference is the deterministic pipeline in the style of Caprara et al.. It runs one mask through one simulation and has two policies: accept everything, or send everything to review. The first is the speed ceiling and the second is the accuracy ceiling.

The strongest existing competitor is geometric-uncertainty triage as in the Latent Doctor Model. It flags cases by segmentation predictive entropy, using the same trained sampler, so the only difference is which signal drives triage. The best reported number for that signal is a Spearman correlation of 0.728 with human entropy on ISUP histopathology. There is no reported result on Lacuna for entropy-based triage on bone CT or against mechanical outputs.

We beat this competitor if, at an equal fraction of auto-accepted cases, mechanics-keyed triage leaves fewer accepted cases whose strength or ductility falls outside the range spanned by the four experts' own models.

A deep ensemble of five nnU-Nets will serve as an alternative sampler. All samplers share the training data, the 20-sample budget, and an equal hyperparameter search budget.

The ablation that isolates our key ingredient replaces observer-structured samples with independent ±1-voxel boundary erosion and dilation. If that cheap perturbation matches the learned sampler in calibration of mechanical spread, the learned sampler is unnecessary. In that case the contribution reduces to "propagate boundary sensitivity into mechanics," and we will say so.

A second ablation drops the explicit shell mask. This tests whether shell modelling is what makes ductility uncertainty visible.

Why not just…?

Why not just triage on segmentation entropy, which already works? It works for predicting where humans disagree. Our hypothesis is that it fails to predict where disagreement changes the answer. This is directly testable: we will report the rank correlation between per-case segmentation entropy and per-case expert spread in strength and in ductility. If entropy tracks mechanical spread as well as mechanical variance does, the extra solves are wasted and the proposal collapses to the Latent Doctor Model applied to bone. We treat that as a live outcome.

Why not just perturb the mask by a voxel and rerun the simulation? This is the most serious threat, because it is cheap and needs no multi-observer data. Experts may disagree non-uniformly, for example more at the posterior shell and endplates than at the anterior wall. If so, uniform perturbation should misallocate uncertainty. The erosion/dilation ablation measures exactly this. If it wins, we adopt it and refocus on the strength-versus-ductility finding.

Why not just use automated vBMD and skip finite element models entirely? The vBMD study shows density alone predicts fracture with β = −0.32. For fracture-risk screening this may be enough, and we do not claim otherwise. The case for mechanics is ductility and implant-related loading, which density does not report. We will additionally test, on the opportunistic fracture cohort if access is granted, whether uncertainty-aware finite element strength adds discrimination beyond vBMD.

Proposed Experiments

The first experiment tests the central claim: does observer ambiguity move ductility more than strength? On the 60-vertebra four-reader set, we build one model per expert outline. For each vertebra we compute the across-expert coefficient of variation of stiffness, strength, and ductility. The prediction is that ductility's spread is consistently larger and concentrates in vertebrae whose experts disagree about the shell. A paired test across vertebrae settles this. If strength and ductility spreads are equal, the trade-off lens contributes nothing, and we fall back to a single-output version of the method.

The second experiment tests calibration. For held-out vertebrae, we check whether the sampler's 90% interval for each mechanical output covers the expert-derived values at the nominal rate. We also check whether predicted per-case spread rank-correlates with expert spread, comparing the observer-structured sampler, the ensemble, and voxel perturbation.

The third experiment tests workflow. We sweep the triage threshold and plot the fraction of cases auto-accepted against the fraction of accepted cases outside the expert range, for mechanics-keyed and entropy-keyed triage. We report the review time saved relative to reviewing everything, measured in a small reader study with the same radiologists editing flagged cases.

The decisive check is a single comparison. At the operating point where entropy triage accepts the same share of cases, mechanics-keyed triage should cut the number of out-of-range accepted cases. It should also do so specifically for ductility, which is where our hypothesis says geometric uncertainty is least informative.