Letting tissue mechanics arbitrate between radiologists' prostate contours

When a surgeon or physicist builds a patient-specific model of how the prostate deforms under an ultrasound probe, the model starts from one radiologist's outline of the gland and of its internal zones. Radiologists disagree substantially about those outlines, and the disagreement is worst at the soft internal boundary between the transition zone and the peripheral zone. That boundary is also where the model places a jump in stiffness. Recent work shows that a nonlinear rubber-like tissue model can recover the stiffness ratio between the two zones from surface motion alone, while a linear spring-like model cannot. That work assumed the boundary was known exactly. We do not know whether rater disagreement destroys this ability. We also do not know whether the reverse holds: the stiffness contrast leaves a mark on how the organ moves, and that mark might reveal which of the disagreeing contours is mechanically consistent with the observed motion.

The proposal is to treat the zonal boundary as uncertain and to infer it jointly with tissue stiffness. A learned model of how different raters draw the boundary supplies the prior. A nonlinear hyperelastic simulation, meaning one whose stiffness grows with strain the way real tissue stiffens, supplies the likelihood of the observed surface displacements. The output is a stiffness estimate whose error bars honestly include contouring disagreement, plus a mechanics-reweighted distribution over plausible boundaries. The work stands or falls on one testable question: does the observed motion carry enough information about where the internal boundary lies to beat simply averaging over raters?

Core Claim

We claim two things. First, propagating inter-observer variability through a compressible neo-Hookean model (a standard nonlinear hyperelastic law) inflates the error of inverse stiffness estimation well beyond what the single-segmentation benchmark suggests. Second, joint inference over the boundary and the stiffness ratio recovers most of that loss, and it does better than consensus contours (STAPLE, majority vote) and better than unweighted averaging over raters. We expect nonlinearity to matter for the second claim as well as the first. A linear model is a poor inverse estimator even with a perfect boundary, so it should also be a poor judge of which boundary is right.

Prior work frames this from both sides without closing it. Biomechanics-informed Non-rigid Medical Image Registration and its Inverse Material Property Estimation with Linear and Nonlinear Elasticity embeds linear and neo-Hookean mechanics in a physics-informed neural network (a network trained to satisfy the governing equations as part of its loss). It reports a mean absolute percentage error of 14.20% for the nonlinear model on the zone stiffness ratio versus 76.28% for the linear one. Its boundaries are deterministic, so it supplies both our method and our clean-boundary reference point. Patient-specific Deformation Modelling via Elastography builds a neo-Hookean prostate model with inner and outer zones, each assigned a stiffness from elastography. It finds that registration degrades once assigned stiffness is off by more than about 15 kPa. Its zones came from single manual segmentations. We use its 15 kPa figure as a clinically meaningful error tolerance.

On the segmentation side, Pionono maps each rater to a Gaussian in a low-dimensional latent space, so that sampling yields coherent rater-specific segmentations. Disentangling Human Error from Ground Truth models per-rater confusion to recover a latent truth. Both work purely on image appearance. We reuse Pionono-style latents as the prior over boundaries and add a physical likelihood that neither has.

The closest conceptual neighbour is From Registration Uncertainty to Segmentation Uncertainty. It links deformation uncertainty to label error in cardiac MRI. It uses generic smoothness regularization and single-expert labels, and it reports that raw registration uncertainty barely correlates with segmentation error (r < 0.10). That negative result constrains us. Deformation variability does not automatically say anything about boundaries, so identifiability has to be tested, not assumed.

Research Program

Let $z$ be a rater latent code that, through a trained multi-rater segmentation model, produces a gland surface and a zonal boundary. Let $\theta$ be the transition-to-peripheral stiffness ratio, and let $d$ be the observed surface displacements under probe loading. We target the joint posterior

$$ p(z,\theta \mid d) \;\propto\; \mathcal{N}\big(d \,;\, u(z,\theta),\, \sigma^2 I\big)\; q(z)\; p(\theta), $$

where $u(z,\theta)$ is the surface displacement predicted by a compressible neo-Hookean solve on the geometry decoded from $z$, $\sigma^2$ is the measurement noise variance, $q(z)$ is the rater-mixture prior learned from multi-rater masks, and $p(\theta)$ is a broad prior on the stiffness ratio.

Running a full finite-element solve for every posterior sample is too slow. We therefore train a neural emulator of $u(z,\theta)$ on a few thousand finite-element runs spanning the prior. The emulator is either the physics-informed network of Min et al. conditioned on $z$, or a direct regressor. We check its accuracy against held-out solves before using it for inference, and we sample with gradient-based Markov chain Monte Carlo.

Three things get built. The first is the rater-latent boundary prior. The second is the emulator, trained for both constitutive laws so that linear and nonlinear versions differ only in the physics. The third is the inference loop. The analysis then reports how much of the posterior over $z$ moves away from the prior, which measures boundary identifiability. It also reports whether that movement is toward the held-out true boundary.

Baselines

The naive reference is the current practice: one rater's segmentation, a neo-Hookean inverse fit, and a point estimate of $\theta$. We run it once per rater and report the spread. This number shows how badly current pipelines are exposed to contouring choice.

The strongest existing method is the neo-Hookean physics-informed inverse estimator of Min et al.. The file reports for it a ratio mAPE of 14.20% and a surface RMSE of 1.24 mm, against 1.83 mm for their weaker baseline, all on clean boundaries. No reported result was found on Lacuna for this estimator on rater-perturbed boundaries. We will produce that number ourselves by feeding it STAPLE consensus contours and majority-vote contours. These consensus-fed runs are the fair version of the strongest competitor. Every run uses the same simulated displacement data, the same emulator architecture, and the same optimisation budget.

The key ablation is unweighted Monte Carlo propagation. We sample $z$ from the prior, fit $\theta$ for each sample, and pool the estimates, with no mechanical reweighting of boundaries. A second ablation keeps the full joint inference but swaps in linear elasticity, to test whether nonlinearity is what makes the boundary identifiable.

We count the proposal as beating the baselines if three things hold. Its ratio mAPE must be significantly lower than both the consensus-fed estimator and unweighted propagation, using paired tests across cases. Its 90% credible intervals must cover the true ratio at close to nominal frequency. Its implied absolute stiffness errors must fall within 15 kPa more often than the competitors' errors do.

Why not just…?

Why not just fuse the raters with STAPLE and run the existing estimator? Consensus fusion averages away exactly the interface the mechanics depends on. Tanno et al. found that STAPLE-style fusion trails explicit rater modelling, including by 14.4% in confusion estimation on BraTS. That evidence concerns segmentation quality, not stiffness recovery, so it does not settle the question here. Our consensus-fed baseline settles it directly. If consensus matches joint inference on mAPE and coverage, we will report that the cheaper route suffices.

Why not just sample rater contours and average the stiffness estimates? This is the real threat. It quantifies uncertainty honestly and needs no new machinery. It wins if surface motion carries little information about the internal boundary, in which case the posterior over $z$ equals the prior and our method collapses into it. The unweighted-propagation ablation is designed to detect this outcome. Before running full comparisons, we will measure the divergence between posterior and prior over $z$. If that divergence is negligible, we will narrow the contribution to uncertainty-aware estimation and drop the arbitration claim.

Why not just use linear elasticity inside the joint inference, since it is cheaper? Min et al. show that linear models already fail at inverse estimation with perfect boundaries, at 76.28% mAPE. A model that cannot recover stiffness is unlikely to tell good boundaries from bad ones. The linear ablation tests this rather than assuming it.

Why not just train a better segmentation network and remove the disagreement? The QUBIQ organisers (QUBIQ) treat multi-rater ambiguity as a property of the data, and the transition-to-peripheral boundary is genuinely graded rather than sharp. A better network still has to choose one contour. Our method instead asks which contours the physics supports.

Proposed Experiments

The first stage is a controlled study with known ground truth. Gland geometries come from the multi-rater prostate MRI task of the QUBIQ challenge. For each case, one rater is held out and that rater's contour defines the "true" geometry. The QUBIQ prostate masks are whole-gland, so internal-zone variability must come from elsewhere. We will either collect a small set of independent zonal delineations from several readers or, failing that, perturb zone boundaries using the learned rater latents. We will state clearly which option was used.

On each true geometry we simulate probe-induced deformation with a neo-Hookean finite-element model, following the probe loading setup of Wang et al.. We sweep the stiffness ratio and add measurement noise to the surface. All methods then see only the remaining raters' contours and the noisy surface displacements.

Results are reported on several metrics: ratio mAPE, credible-interval coverage, the fraction of cases within the 15 kPa tolerance, surface RMSE, and the Dice score of the posterior-mean zonal boundary against the held-out truth compared with the prior mean. We also measure how much a fixed rater spread inflates stiffness error under linear versus nonlinear laws, which tests whether nonlinearity amplifies rater variance.

The decisive checks are three. Joint inference must move the boundary posterior measurably toward the truth, as shown by higher Dice than the prior. It must cut ratio error relative to both the consensus-fed and unweighted-propagation baselines. And these gains must shrink or vanish under the linear ablation. Failing the first check redirects the project to honest uncertainty propagation. Passing all three would justify a follow-up on a phantom or a clinical cohort with elastography and multiple zonal readers.