Spending Annotator Disagreement on Printability: Support-Minimal Anatomical Models Within the Inter-Observer Band

When a hospital 3D-prints a patient-specific model of a kidney, a tumor, or a fractured pelvis, it starts from a segmentation. A segmentation is a voxel-wise outline drawn on a CT or MRI scan, and experts routinely disagree about where it should go by several millimetres. The printer then treats that outline as exact. Wherever the surface overhangs by more than about 45 degrees, it builds sacrificial support scaffolding, meaning temporary material that holds up the overhang and is later snapped, cut, or dissolved away. That removal costs material and time, and it can damage thin features. Our proposal is to stop treating these two facts separately. If qualified experts would accept any surface inside the band between "everyone agrees this is organ" and "someone thinks this is organ", then that band is design freedom. Within it we can choose the surface that needs the least support, or that becomes fully self-supporting under a multi-directional print plan. Every point of the printed surface stays inside the range of clinical opinion.

The work is new in how it treats disagreement. Medical imaging research models disagreement in order to report or calibrate it. Multi-axis printing research assumes the input surface is fixed and exact. We use disagreement as a budget that the fabrication planner may spend. We test this on multi-rater 3D benchmarks and on physical prints, and we measure both how much support is eliminated and whether the printed surface would pass as "one more expert's contour".

Core Claim

A meaningful share of the support volume in anatomical prints comes from small overhangs, protrusions, and undercuts that lie entirely inside the inter-observer band. Removing that support requires bending the surface only where experts already disagree. We claim that optimizing the printed shape under a hard constraint to stay within a calibrated agreement band will cut support volume substantially. We also claim it will make more parts decomposable into self-supporting pieces for multi-directional printing, beyond what two simpler routes achieve: printing the consensus mesh with a strong multi-directional planner, and picking the most printable among whole-rater or sampled contours. We claim this holds only if the resulting surfaces are statistically indistinguishable from held-out raters on standard agreement metrics. We expect the gain to be concentrated on irregular shapes such as pelvic fragments and brain tumors, and to be negligible on near-convex organs. We treat that contrast as a prediction, not a caveat.

Background and Positioning

On the manufacturing side, three lines of work remove supports by changing how material is deposited rather than what shape is deposited. The first tilts the part or nozzle mid-print, as in RoboFDM, where a robot arm holds the build plate under a fixed extruder. The second splits the shape into self-supporting sub-volumes printed along different directions. General Support-Effective Decomposition for Multi-Directional 3-D Printing does this with a beam search over clipping planes. Learning to Accelerate Decomposition for Multi-Directional 3D Printing replaces most of that search with a learned ranker over six geometric features of each candidate plane. The third uses curved, non-planar layers. Open5x provides accessible 5-axis hardware with conformal slicing. Geodesic Distance Field-based Curved Layer Volume Decomposition builds layers from geodesic distance fields. Support-free volume printing by multi-axis motion grows voxel fields along an advancing convex front.

All of these supply our downstream planners, and all of them take the target surface as given. Support-free volume printing by multi-axis motion also concedes that thin fragile features are unhandled. That concession motivates us, because thin ambiguous margins are exactly where anatomical models suffer. None of these works considers that the boundary itself might legitimately move. That is the degree of freedom we add.

On the imaging side, several methods produce the band we need. Pionono models each rater's style as a Gaussian in a small latent space and samples coherent alternative segmentations. Variational Inference for Quantifying Inter-observer Variability does something similar with a VAE (variational autoencoder). The Multi-decoder U-Net gives each annotator a dedicated decoder on a shared encoder. QUBIQ benchmarks all of these with thresholded soft Dice. QUBIQ also reports that simple ensembles are competitive, and that 3D uncertainty modeling is memory-limited. We take that as a constraint and favor cheap band estimators.

Dice Semimetric Losses matters more directly. It shows that training on majority-vote labels leaves models badly calibrated, and it reports expected calibration error on QUBIQ Prostate T1 dropping from 0.51 under majority voting to 0.20 with weighted-average soft labels and its loss. Because our band is only as trustworthy as its calibration, this result tells us which band estimator to use. It also defines an ablation. QUBIQ further notes that no work carries uncertainty representations into physical fabrication. We read that as the gap, and we are not aware of anything in the file that fills it.

The adjacent orthopedic work is PENGWIN 2024. It documents disagreement over whether hairline fractures separate fragments, but it stops at digital meshes. A Meta-Review about Medical 3D Printing shows that orthopedics and maxillofacial surgery dominate clinical printing, while treating the path from segmentation to printed part as a black box.

Research Program

We first build the band. For each case we obtain a per-voxel agreement probability $p(x)$. Where raters exist, we take the fraction of raters labeling voxel $x$ as foreground. On unseen scans, $p(x)$ is predicted by a calibrated model, trained with the semimetric Dice loss of Dice Semimetric Losses on soft multi-rater targets. For a tolerance level $\alpha$, the inner set $L_\alpha = \{p \ge 1-\alpha\}$ is what nearly everyone includes. The outer set $U_\alpha = \{p \ge \alpha\}$ is what at least some experts include.

Next we optimize the printed shape. We represent it by a signed distance field $\phi$ whose negative region is the solid $S$, and solve

$$\min_{\phi}\; V_{\text{sup}}(S; \mathcal{P}) + \lambda \int_{\partial S} \kappa^2 \, dA \quad \text{s.t.} \quad L_\alpha \subseteq S \subseteq U_\alpha .$$

Here $V_{\text{sup}}(S;\mathcal{P})$ is the support volume required under print plan $\mathcal{P}$. For a single build direction, it is the volume swept down to the platform from surface patches whose outward normal points more than the critical angle below horizontal. $\kappa$ is surface curvature and $\lambda$ weights a smoothness penalty that prevents jagged surfaces. We solve this by projected level-set descent on a relaxed overhang penalty. After each step the field is clamped into the band, so the constraint is never violated.

For multi-directional printing we alternate two steps. We run the decomposition planner of General Support-Effective Decomposition, accelerated by the learned ranker of Learning to Accelerate Decomposition, to get $\mathcal{P}$. We then reshape the surface within the band to shrink the "risky" unsupported area that remains under that plan. The planner's own risky-area feature serves as the objective.

Clinically functional regions pose a separate problem. Examples are fracture reduction interfaces or surfaces a guide must seat against. There we set $\alpha$ locally to zero, so the planner may not move them. This keeps the method honest where millimetres matter mechanically.

Baselines

The naive reference is the current practice described in the meta-review. We take the majority-vote mesh (or STAPLE mesh), choose the single best build orientation for it by exhaustive search, and print on a 3-axis printer with standard tree supports. This fixes the support volume, print time, and post-processing time that everything else must beat.

The strongest existing method is support-effective multi-directional decomposition of the consensus mesh, using General Support-Effective Decomposition with the learned ranker of Learning to Accelerate Decomposition. That paper reports, on its Thingi10k subset, 17% less risky area at beam width 2, roughly a 3× overall speedup over beam search, and a test case falling from 1,424 s to 120 s. No reported result was found on Lacuna for this or any support-free planner on anatomical or multi-rater geometry. We will therefore run it ourselves on every case, with the same beam width and time budget given to our alternating loop. Beating it means lower residual risky area, and a higher fraction of cases fully decomposable without support at equal planning time.

The most dangerous baseline is best-of-K selection. We generate K plausible segmentations, either the real raters' contours or K samples from Pionono or the Multi-decoder U-Net. We run the same planner on each and keep the one needing least support. This baseline is clinically valid by construction and requires no new optimizer. Our method must beat it at matched K and planning compute. The argument for why it can is that our optimizer can combine the locally most printable parts of different hypotheses, which no single hypothesis offers.

Two ablations isolate the key ingredient. First, a uniform offset baseline dilates or erodes the consensus mesh to the band edges. It tests whether shape-aware movement matters or whether moving the boundary is enough. Second, a band-source ablation builds the band from a majority-vote-trained model instead of the calibrated soft-label model. It tests whether calibration is necessary for staying within real rater opinion.

Why not just…?

Why not just print the consensus mesh with water-soluble supports? Soluble supports remove the risk of breaking the part during removal. They do not remove the extra material, the print time, the soaking time, or the surface marks, all of which the meta-review and Support-free volume printing by multi-axis motion identify as costs. Collapsing to a consensus boundary also throws away calibration: the 0.51 versus 0.20 calibration error in Dice Semimetric Losses. We include soluble-support prints as a physical arm and report the full wall-clock and material cost, so a reader can see whether our savings matter against this option.

Why not just run a good multi-directional planner on the consensus mesh? This is a real threat. For near-convex organs such as kidney and prostate, the consensus mesh may already decompose into few self-supporting pieces, and then the band adds nothing. We state this in advance. The experiment that settles it is the stratified comparison against the strongest baseline, broken down by shape irregularity. If band optimization does not beat consensus decomposition on pelvic fragments and irregular tumors, the idea fails.

Why not just pick the most printable rater or sampled hypothesis? This baseline may capture most of the gain, especially when raters differ globally rather than locally. We run it at several values of K. If best-of-K matches our method within noise, we will report that the practical result is "sample, then select", and the continuous optimizer is unnecessary. Our prediction that it will not match rests on disagreement being local: different raters are generous at different places on the same boundary.

Proposed Experiments

The data are the 3D multi-rater tasks of QUBIQ, including brain tumor, kidney, prostate, and the pancreas-related tasks, plus pelvic CT from PENGWIN 2024. PENGWIN does not, as far as the file shows, release multiple raters. For it we build bands from multi-decoder or ensemble hypotheses, and we collect a second and third annotation on a small subset to check those bands against people.

The first decisive check is a cheap test done before any optimization. For each case we compute the best-orientation support volume of the inner set, the consensus mesh, and the outer set. If the spread between them is small relative to consensus support, the premise is false and we stop.

The main digital experiment compares all baselines on the following measures:

  • support volume and overhang area under a single build direction;
  • residual risky area and number of sub-volumes under multi-directional planning;
  • the fraction of cases that become fully support-free;
  • planning time.

These are reported across $\alpha \in \{0.1, 0.2, 0.3\}$ and stratified by shape irregularity.

Clinical validity is judged by a leave-one-rater-out test. The optimized surface counts as acceptable if its Dice score and 95th-percentile Hausdorff distance to the remaining raters fall within the distribution of held-out-rater scores against the same raters. We also run a blinded reading in which clinicians try to pick the optimized contour out of a lineup of real ones.

Physically, we print a subset of cases in two conditions. The first is a 3-axis printer, comparing consensus-with-supports against band-optimized geometry. The second is Open5x-style multi-axis hardware, comparing consensus-decomposed against band-optimized-decomposed geometry. For each print we record material mass, print and post-processing time, and surface damage after support removal. Damage is measured by 3D scanning each part and computing the fraction of the scanned surface lying outside $U_\alpha$ or inside $L_\alpha$. That fraction is the physical analogue of staying within expert opinion.

The idea succeeds if two things hold together. On irregular anatomy, band optimization must cut support volume or risky area clearly more than both consensus decomposition and best-of-K, at equal compute. And its surfaces must remain statistically indistinguishable from held-out raters.