When several clinicians outline the same thin structure in a 3D scan, such as a coronary artery on a CT not taken for cardiac imaging, their outlines can overlap poorly. They may still agree on what matters downstream: where the vessel starts, where it branches, how far it runs, and that it is one continuous tube. Standard training fits a 3D segmentation network to a single outline per scan, or to an average of outlines. That spends most of the learning signal on boundary voxels that humans cannot agree on. It spends almost none on the vessel's connectivity, which humans do agree on and which models routinely break. We propose to split expert disagreement into two levels. The voxel level is where the boundary falls. The graph level is which anatomical landmarks the vessel connects, and how directly. The model learns a distribution over each rater's boundary style, but is held to the consensus at the graph level.
The graph-level quantity comes from network science. Global efficiency is the average inverse shortest-path length between pairs of nodes; in brain connectomics it measures how directly a network routes information. We apply it to a sparse set of rater-agreed vessel landmarks. Distance is measured through the network's predicted vessel probability, so a single gap in the vessel sharply reduces efficiency, and a gap near the root reduces it more than one near a tip. The claim is that this landmark-efficiency term is a better training target than voxel overlap with any one rater. It should also make a fairer evaluation target: a prediction is judged by whether its graph falls inside the spread of graphs drawn by human raters, not by its Dice score against one of them.
Where this sits relative to existing work
Three lines of work each cover one piece and leave the joint problem open.
Multi-rater modeling motivates the rater side but is confined to 2D and to pixel likelihoods. Pionono represents each rater as a Gaussian in an 8-dimensional latent space and reaches κ = 0.758 on Gleason 2019, but only on 2D histology patches, and nothing stops its sampled outlines from being structurally invalid. DiFF argues that the useful label is whichever one best supports a downstream task, and weights annotators by how well their masks help a diagnosis classifier. It is also 2D, and its downstream anchor is a class label. We borrow DiFF's principle that the target should be defined by downstream use, but our downstream object is the vessel graph itself.
Topology-aware 3D segmentation shows that overlap losses miss structural errors, but it assumes a single ground truth. Topology-Aware Focal Loss matches persistence diagrams, which count connected components, loops and voids, between prediction and one reference mask on BraTS 2019. When raters disagree about the topology itself, the target is ill-defined. FeTA 2024 showed that teams with nearly identical Dice can differ by about 50% in Euler-characteristic error. That is evidence that overlap and structure diverge, but it again uses a single reference. Counting components says that a vessel is broken. It does not say where, or how much the break matters for reaching the distal territory. Path-based efficiency does.
Volumetric vessel segmentation supplies the application and the backbone. NA-UNETR segments the left anterior descending (LAD) artery on non-contrast planning CT using 3D neighborhood attention, with pretraining on ImageCAS. It reports 45.64% Dice on the 20-scan LAD-SEG set, against 42.54% for nnU-Net. It treats one physician's contour as fixed truth and does not measure connectivity. Its low absolute Dice in a regime the authors describe as highly ambiguous is exactly where we expect voxel overlap to be the wrong yardstick.
The brain-network papers constrain the evaluation design rather than supply the method. MSN vs. MIND shows that graph metrics swing with construction choices such as threshold density. Connectome harmonization shows they carry large acquisition-site biases. Together these warn that any graph metric must be reported against an explicit noise floor, which here is the inter-rater envelope, before differences can be read as signal.
Core Claim
Experts disagree far more about voxel boundaries than about landmark-to-landmark connectivity. A 3D model trained this way should match human connectivity much more often while staying within the inter-rater range on overlap. The model is held to rater consensus on a path-based efficiency between agreed landmarks, and models rater style only at the boundary. The first half of this claim is an empirical premise that we test before building anything.
Research Program
Annotation. We need real multi-rater 3D data, and the file lists no public multi-rater vascular set. We will have three or four clinicians independently re-contour the 20 LAD-SEG scans. Each rater also places a small set of landmarks: the LAD origin, diagonal branch points, and the most distal point they consider visible. Landmark placement is cheap relative to full contouring. Landmarks marked by at least $k$ of the raters, matched within a tolerance radius, become consensus nodes.
Pairs of consensus nodes are labelled "connected" if every rater's mask joins them. Pairs are labelled "separated" if a nearby non-target structure lies between them and no rater joins them. Scans or segments where raters disagree about connectivity get no graph-level target and are reported as their own category.
Model. The backbone is NA-UNETR, pretrained on ImageCAS as in the original paper. Following Pionono's construction, a rater-conditioned latent is injected into the decoder in 3D form. A voxel-level Dice-focal loss is applied per rater, using that rater's own mask and latent.
Graph term. Let $p_v \in (0,1)$ be the predicted vessel probability at voxel $v$. For landmarks $i$ and $j$, define the soft path cost
$$ d_{ij} = \min_{\pi:\, i \to j} \sum_{v \in \pi} -\log p_v , $$
the cheapest 26-connected voxel path through the predicted probability field. This cost is near zero along a confidently segmented vessel and grows sharply across even one low-probability gap. Landmark efficiency is $e_{ij} = 1/(1+d_{ij})$. Writing $C$ for the set of consensus-connected pairs and $S$ for the set of consensus-separated pairs, the graph loss is
$$ \mathcal{L}_{\text{graph}} = \frac{1}{|C|}\sum_{(i,j)\in C} (1 - e_{ij}) + \frac{1}{|S|}\sum_{(i,j)\in S} e_{ij}. $$
The first term rewards a continuous vessel between agreed landmarks; the second penalizes leaks into neighboring structures. The minimum is subdifferentiable in $p$: gradient flows to the voxels on the current cheapest path. This is the familiar minimal-path trick. Our contribution is not this operator but what it is attached to: rater-consensus graph targets, kept separate from rater-specific boundary targets.
The graph term is computed from the rater-marginal prediction, meaning the average over sampled rater latents. Connectivity therefore cannot be reduced by choosing a particular rater style.
Controlled study. Before the clinical data are complete, we run a controlled study on ImageCAS's 1,000 contrast volumes. We synthesize "raters" from the single reference by applying style-specific perturbations: systematic dilation or erosion, boundary jitter, and rater-specific truncation of the distal vessel. Because the true connectivity is known here, we can check whether the method recovers structure that no single synthetic rater reproduces faithfully. These results only support the clinical study; they cannot replace it.
Baselines
The naive reference is nnU-Net trained on the per-voxel majority vote of the raters. This is what most groups would do, and it shows how far plain consensus gets. It is trained with nnU-Net's own self-configuration on the same folds.
The strongest existing method is NA-UNETR. On the single-physician LAD-SEG labels it reports 45.64% Dice and 38.16 mm HD95, and it reports 79.49% Dice on ImageCAS (NA-UNETR). We rerun it under three conditions: the original single contour, the majority-vote target, and soft labels averaged over raters. It uses the same ImageCAS pretraining and low-rank fine-tuning, the same five-fold split of the 20 scans, and the same hyperparameter-search budget as our model.
The strongest structural competitor is NA-UNETR plus the persistent-homology topology loss of TAFL, applied to the majority-vote mask. It represents "add a topology prior to consensus training."
The strongest rater-modeling competitor is our 3D port of Pionono on the same backbone without the graph term. It represents "model raters, ignore structure."
No graph-efficiency or rater-envelope results exist for any of these: no reported result found on Lacuna for landmark efficiency or within-envelope rates on LAD-SEG or ImageCAS. We will compute them for every method.
Two ablations isolate the key ingredient. In the first, the graph term is replaced with a component-count (Betti-0) term computed on the same consensus graph. This tests whether path-based efficiency adds anything beyond "be one piece." In the second, graph targets are built from all raters' graphs, including the disputed ones, rather than consensus only. This tests whether restricting supervision to agreement is what matters.
Beating the field means two things on held-out folds. First, a clearly higher fraction of consensus pairs connected, with fewer separated-pair leaks than the TAFL and Pionono variants. Second, Dice against each individual rater that is not significantly lower than NA-UNETR's and lies within the raters' own pairwise Dice range.
Why not just…?
Why not just average the raters into a soft or STAPLE label and add a topology loss? This is the real threat, and it is run as a baseline. Averaging does preserve some disagreement information. Still, DiFF found that majority vote and STAPLE erase clinically relevant borders and lose 2–3 AUC points downstream, and Pionono found consensus training suppresses legitimate rater variation. The bigger structural problem is that a topology loss on an averaged mask inherits whatever topology the average happens to have. Where one rater truncates the distal LAD and others continue it, the averaged target has a faint, possibly disconnected tail. If NA-UNETR with soft labels and TAFL matches our connectivity numbers, we will report that the simpler recipe suffices.
Why not just count components with persistent homology instead of computing path efficiency? The Betti-0 ablation settles this directly. Our reason to expect a difference is that component counts give the same penalty to a one-voxel gap near a vessel tip and one near the origin. They also give no signal about leaks that merge the LAD with an adjacent vein into a single component, which the separated-pair term targets. If the Betti-0 ablation performs within noise of the efficiency term, the efficiency machinery is unnecessary and we will say so.
Why not just bridge gaps after the fact with morphology or a conditional random field? 3D DSN found that a fully connected conditional random field took 85 of 90 seconds per volume and still left shape outliers, because it optimizes local intensity smoothness rather than structure. Morphological closing can join fragments, but it cannot tell a true gap from a nearby non-target vessel, which inflates exactly the leak errors we measure. We will apply closing plus largest-path reconnection to the NA-UNETR output as an additional post-hoc baseline, under the same leak metric.
Proposed Experiments
Premise test. This test can stop the project. On the re-annotated LAD-SEG scans, we compute rater agreement at two levels. At the voxel level we use pairwise Dice between raters. At the graph level we compute pairwise agreement on landmark connectivity and on landmark efficiency computed from each rater's binary mask. If graph-level agreement is not substantially higher than voxel-level agreement, the decomposition has no basis. In that case we would stop and publish the measurement, which would itself show that LAD connectivity is as subjective as the boundary.
Main comparison. All baselines and ablations are evaluated on held-out folds with four metrics:
- Fraction of consensus pairs connected.
- Separated-pair leak rate.
- Within-envelope rate: the fraction of test scans whose predicted graph efficiency lies between the minimum and maximum rater value.
- Dice and HD95 against each individual rater, alongside the rater-to-rater Dice range.
We report Euler-characteristic difference as in FeTA 2024 for comparability.
Disputed segments. Separately, on the disputed segments, where raters disagree on connectivity, we check whether sampling different rater latents reproduces the raters' different connectivity choices. This is where rater modeling should matter, and where consensus methods have nothing to say.
Controlled transfer. On ImageCAS with synthetic raters, we test whether the graph term recovers known true connectivity as the synthetic raters' distal-truncation rate grows. We expect the margin over consensus training to widen as raters disagree more at the boundary while agreeing on the graph.
Rater noise floor. Following the warning from connectome work that graph metrics are construction-sensitive (MSN vs. MIND), every efficiency difference is reported relative to the inter-rater spread. A model improvement smaller than that spread is not counted as an improvement.
References
- Probabilistic Modeling of Inter- and Intra-observer Variability in Medical Image Segmentation
- Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle
- Topology-Aware Focal Loss for 3D Image Segmentation
- Advances in automated fetal brain MRI segmentation and biometry: Insights from the FeTA 2024 challenge
- A Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery
- Revealing Structural Brain-Cognition Relationships in Children: A Comparison of Morphometric Similarity and INverse Divergence Networks
- Learning disentangled representations to harmonize connectome network measures
- 3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes