Surgeons and engineers increasingly judge implants with finite element analysis, a computer simulation that splits a 3D bone model into small elements and computes how it deforms and where stress concentrates under load. Before any simulation can run, someone has to trace the bone on the CT scan. Published studies of screw and plate fixation treat that tracing as exact. Each builds one model from one operator's segmentation of one scan and then reports sub-millimetre differences between implant options. In fracture imaging, though, the tracing is where experts disagree most. A key disagreement is whether a faint hairline crack splits a bone into two fragments or leaves it intact. We propose to measure, on real fractured pelvises traced independently by several annotators, how often that disagreement reverses a simulation's verdict. A verdict here means which screw path is better, or whether fragment displacement stays under a clinical threshold.
The second half of the proposal tests a specific explanation for any such reversals. Our hypothesis is that the disagreement that matters is structural rather than a uniform fuzz along the boundary. It comes from discrete fragment split-or-merge decisions and from local cortical thickness where screw threads grip. If that is right, a cheap procedure can stand in for collecting multi-rater data on every case. The procedure enumerates the plausible fragment configurations from one segmentation and re-runs the simulation on each. Our main comparison pits this procedure against the obvious cheaper alternative, which is to randomly grow and shrink a single mask by a millimetre. That alternative may turn out to be sufficient, and we treat that as a real possibility.
Core Claim
Fixed-angle fixation (screws that lock rigidly into a plate, or long screws whose path is fixed once inserted) gives a construct its stability through screw-thread purchase along a chosen corridor, not through friction against the bone surface. That makes simulated stability sensitive to two things in the segmentation: whether the corridor crosses one fragment or two, and how much cortical bone surrounds the threads.
The fixation studies show how small the reported margins are. Lü et al. simulated superior pubic ramus screws under a 500 N standing load. They found Zone II retrograde screws limited displacement to 0.59 mm with 5.51 MPa peak stress, against 0.88 mm and 15.53 MPa for antegrade screws. Wang et al. found that only a 3-screw tibial tuberosity construct stayed below a 2.0 mm displacement threshold, at 1.97 mm. Both studies built geometry from a single interactive segmentation and treated it as ground truth.
The segmentation literature shows why that assumption is risky. The PENGWIN 2024 challenge benchmarked pelvic fragment segmentation and reports that expert annotators disagreed sharply over whether hairline fractures separate fragments. It also states that failing to separate two fragments is worse than slight overlap errors. Taken together, these two literatures motivate the study but do not answer it. Nobody has measured whether real observer disagreement is large enough, relative to margins like 0.29 mm, to change which implant option wins.
Our falsifiable claim has two parts. First, across independent annotators of the same fractured pelvis, the preferred screw path or threshold verdict will flip in a non-trivial fraction of cases. Second, most of those flips will be explained by fragment-topology and local cortical differences, not by global boundary offset. As a result, a topology-enumeration procedure applied to one segmentation will predict the real multi-rater spread of simulation outputs better than uniform boundary perturbation of equal average size.
Research Program
The first component is a small multi-rater fracture corpus. PENGWIN provides 150 pelvic CT scans from six centres with expert fragment annotations. As far as the file indicates, these are single-annotator labels, so we will collect new independent tracings. We will select about 30 scans with superior pubic ramus involvement, chosen to span clean fractures and comminuted or hairline cases. Four orthopaedic or radiology annotators will each segment every fragment independently, with no access to the others' work.
To capture the "interactive prompt" variability that FI-GAC introduces but never measured, half of the annotation sessions will use an interactive contour tool. In those sessions, annotators place virtual edges across gaps where image intensity gives no boundary, which is how FI-GAC closes fracture gaps. The other half will be manual. Each annotator will repeat a subset of cases after a washout period so that same-person inconsistency can be separated from between-person disagreement. This mirrors the split that Pionono models in histopathology.
The second component is a fixed simulation pipeline applied identically to every segmentation, so that segmentation is the only thing that varies. For each segmentation we will generate a tetrahedral mesh, assign stiffness from CT intensity as in the statistical finite element model of Belenguer Querol et al., and place the Lü et al. configurations: 6.5 mm retrograde and antegrade ramus screws under a 500 N load on the S1 endplate. We will add one anterior locking-plate configuration so that a true fixed-angle plate is also tested. Screw entry and exit points will be defined once per case in a shared anatomical coordinate frame, then transferred to every annotator's geometry so that trajectories do not drift with the segmentation. The outputs are maximum fragment displacement, peak implant stress, peak bone stress, the fraction of screw length lying in cortical bone, and two binary verdicts: which path is preferred, and whether displacement crosses a pre-registered threshold.
The third component is the estimator we are testing. Starting from one segmentation, which can be any single annotator's or an automatic model's, we will detect candidate fracture lines where annotators could plausibly disagree. Candidates are thin low-intensity bands inside a fragment, and narrow bridges between fragments. We will then enumerate the configurations that split or merge at those sites. Within a narrow band around each screw corridor we will also perturb cortical thickness, with the size of that perturbation set from same-person repeat tracings rather than tuned to the cross-annotator spread we are trying to predict. Every configuration is simulated. The estimator's prediction is the range and flip rate of the outputs across configurations. We then ask how well this predicted spread matches the spread observed across the real annotators.
Baselines
The do-nothing reference is the practice of the existing fixation papers: one segmentation, one simulation, one verdict, as in Lü et al. and Wang et al.. We run it by simulating each annotator's segmentation alone. If segmentation were exact, these single-observer verdicts would agree. Their disagreement rate is therefore the primary quantity of the study rather than a score to beat.
The strongest consensus baseline is a voxel-wise majority vote across the four annotators, which is the standard way to "cancel out" human error, meshed and simulated once. Wu et al. report that majority-vote-style fusion underperforms task-calibrated fusion in downstream diagnosis, with AUC gains of up to 2.73% on glaucoma. That evidence comes from diagnostic classification, not mechanics, so it motivates testing consensus rather than settling the question. Consensus counts as beaten if its single verdict disagrees with the majority of individual-annotator verdicts more often than a randomly chosen single annotator does. That would show consensus is not a good stand-in for any annotator.
The strongest automatic segmenter is the PENGWIN winning approach, a two-stage pipeline that predicts fragment cores and boundary shells. On CT it reports a fragment-wise intersection over union of 0.930 and a 95th-percentile Hausdorff distance of 5.866 mm (PENGWIN 2024). We will train the publicly described configuration on the PENGWIN scans outside our 30 cases and treat its output as a fifth "observer." This tests whether automation simply moves the verdict to one point inside the human spread.
The strongest existing uncertainty method is sampling from a probabilistic segmentation model trained on multi-rater labels. We use Pionono, which reports an unweighted Cohen's kappa of 0.758 on Gleason 2019. We also use the variational model of Liu et al., which reports a continuous Dice of 0.5372 on QUBIQ. We found no result on Lacuna for either method on 3D bone fracture segmentation, and none for any method's effect on finite element outputs.
The key ablation isolates topology. We apply uniform dilation and erosion to the same starting mask, matched to the observed mean surface distance between annotators, and give it the same simulation budget as topology enumeration: the same number of meshes per case. All uncertainty methods receive the same starting segmentations, the same meshing settings, and the same per-case simulation count.
Beating the alternatives means one thing: a smaller error between predicted and observed spread, with spread measured by the range of displacement and stress and by verdict flip rate. We will also check calibration, meaning whether cases the estimator flags as unstable are in fact the cases where annotators' verdicts flip.
Why not just…?
Why not just perturb one mask uniformly by a millimetre? This is the real threat. If observer disagreement is mostly smooth boundary offset, uniform perturbation is cheaper and good enough. Belenguer Querol et al. found that smooth shape and density variation alone produced a nearly six-fold stiffness range, from 0.36 to 2.14 N/mm² between −2 SD and +2 SD instances. That shows smooth variation can matter a great deal. On the other side, a split-or-merge change alters the load path in a way no surface offset can reproduce, and PENGWIN singles out split decisions as the dominant disagreement. The size-matched ablation decides the question. If uniform perturbation matches topology enumeration in predicting the real spread, we will report that and recommend the simpler method.
Why not just use consensus or an automatic segmenter so there is one answer? A single answer hides the variability; it does not remove it. The PENGWIN authors themselves say fully autonomous segmentation is not yet clinically practical, and FI-GAC shows a pelvis-trained U-Net falling to 64.69% Dice on ankle scans. Our design measures directly whether the consensus or automatic verdict sits at a stable point or on a knife-edge.
Why not just sample a probabilistic segmentation network? These models need multi-rater training data, which barely exists for 3D trauma, and they have only been validated on 2D slices and histology. If, with our small corpus, Pionono samples reproduce the real spread as well as enumeration does, that is a legitimate outcome. Enumeration's advantage would then be only that it needs no multi-rater training data, and we would state the claim in those terms.
Proposed Experiments and Decisive Checks
The first decisive check is whether the problem exists. Across the 30 cases and four annotators, we compute the fraction of annotator pairs whose simulated retrograde-versus-antegrade preference or threshold verdict differs. We compare the spread of displacement against the 0.29 mm retrograde-antegrade gap reported by Lü et al. If between-annotator spread is consistently far below that gap and verdicts essentially never flip, the honest conclusion is that single-observer simulation is safe for this application. That is a useful negative result, and the estimator work would stop there.
The second check decomposes the cause. We categorize each annotator pair's difference as a topology difference (different fragment count or connectivity near the corridor), a cortical-thickness difference within the screw band, or a diffuse offset. We then fit a regression predicting verdict flips from these categories. The hypothesis fails if diffuse offset explains most flips.
The third check is head-to-head prediction. For each case, starting from each single annotator's mask, we run topology enumeration, uniform perturbation, and probabilistic sampling under equal simulation budgets. We score the error between predicted and observed output ranges and the calibration of flip predictions, with confidence intervals obtained by bootstrapping over cases. A secondary analysis compares interactive and manual annotation sessions to quantify how much variability FI-GAC-style prompts add or remove.
The main limitation is shared with the source simulations. Bone is modelled as isotropic, loads are static, and soft tissue is excluded. Absolute stresses are therefore not clinically predictive, and we interpret only differences across segmentations, with the simulation held fixed.
References
- Finite Element Analysis of Percutaneous Anterograde and Retrograde Screws in the Treatment of Superior Rami Pubis Fracture
- Comparative finite element analysis of fixation strategies for sports-related tibial tuberosity avulsion fractures in adolescents: biomechanical evidence for optimized clinical decision-making
- Benchmark of Segmentation Techniques for Pelvic Fracture in CT and X-Ray: Summary of the PENGWIN 2024 Challenge
- Fracture interactive geodesic active contours for bone segmentation
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Statistical Finite Element Model for Bone Shape and Biomechanical Properties
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Variational inference for quantifying inter-observer variability in segmentation of anatomical structures