Slice-Sweep Topology Priors for Airway Segmentation

Airway segmentation networks trained on chest CT routinely break thin peripheral branches, or let predictions leak into the surrounding lung, even when their overall voxel overlap looks excellent. A useful way to see the problem is to read the scan as a movie in which each CT slice is one frame. A tree-shaped airway then evolves in a very constrained way from frame to frame. Cross-sections persist, a cross-section occasionally splits in two at a bifurcation, a branch tip appears or vanishes, and the airway never closes a loop. We propose to make that frame-to-frame event grammar the explicit prior of the segmenter. Each topological event is assigned a cost, and each event must happen inside a single slice step. The key design choice is that event costs are defined per millimetre of anatomy rather than per slice, so the same prior stays calibrated when scanners use different slice thicknesses.

Both halves of this idea come from lattice quantum gravity. Causal dynamical triangulations build spacetime from stacked time slices and allow the spatial universe to split only if each split is confined to one time step. They also require the splitting coupling to scale with the lattice spacing, which keeps branching dilute in the fine-lattice limit (Putting a cap on causality violations in causal dynamical triangulations, Taming the cosmological constant in 2D causal quantum gravity with topology change). Separately, spin foam amplitudes turn out to be an unresolved sum of two opposite orientations, $\cos S = \tfrac12(e^{iS}+e^{-iS})$ (Asymptotic analysis of the Ponzano–Regge model for handlebodies, Asymptotic analysis of the Engle–Pereira–Rovelli–Livine four-simplex amplitude). That ambiguity is removed only by supplying explicit orientation data. We borrow the combinatorial and scaling lessons only, not the physics.

Core Claim

A segmenter that decodes the airway as a sequence of slice-to-slice topological events should do two things better than current methods. First, it should reduce breakage and leakage beyond what voxel losses or skeleton supervision achieve. Second, it should keep that advantage under changes of slice thickness, provided three conditions hold. Events are costed per physical length. Events are confined to one slice step. The scan direction (head-to-foot versus foot-to-head) is treated as a latent variable that is selected rather than blended.

The last point is the medical analogue of the orientation superposition. Split and merge swap roles when the volume is reversed. Current bidirectional scanners such as the Tri-orientated Mamba module in SegMamba sum forward and reverse hidden states. That sum is the analogue of the unresolved cosine: two incompatible directional readings added together with nothing to choose between them.

The Model

A standard 3D backbone produces a voxel probability map. We threshold it low, so that it over-segments, and extract 2D connected blobs in every axial slice. Each blob becomes a candidate cross-section with a score from the backbone. Blobs that overlap between consecutive slices become candidate transitions of five types: continue, split, merge, birth, and death.

Decoding chooses a subset of blobs and transitions. It maximizes total blob score minus event costs, subject to two constraints. Every event must involve only slices $z$ and $z+1$. The component graph traced out by the sweep must be acyclic. That graph is the Reeb graph, and it must be a tree for a cycle-free airway. We solve decoding as an integer program and add cycle-elimination constraints only when a cycle actually appears. The candidate graph is sparse, because only overlapping blobs connect.

Event costs come from rates per millimetre. Let $\rho_e(r)$ be the rate of event type $e$ per millimetre of airway length for a cross-section of radius $r$, and let $\Delta z$ be the slice spacing in millimetres. Then the per-step probability and cost are

$$p_e = 1 - e^{-\rho_e(r)\,\Delta z}, \qquad c_e = -\log p_e .$$

The rate form treats events as a Poisson process along the airway centreline. As $\Delta z$ shrinks, the per-slice event probability shrinks proportionally, and the expected number of bifurcations per millimetre stays fixed. This plays the role of the $g \propto a^3$ scaling in the causal gravity models. We estimate $\rho_e$ from the training annotations. Split rates depend on radius and are dominated by caudal splits.

Orientation enters as a binary latent $o$. We decode under both orientations, with split and merge costs swapped between them, and keep the higher-scoring solution rather than averaging the two.

Research Program

The work proceeds in three steps. Step one is inference-only. We apply the event decoder to frozen backbones, which isolates the contribution of the prior itself. Step two is training. We backpropagate a structured hinge loss that penalizes the backbone whenever the decoded event sequence of its prediction differs from that of the ground truth, so that the network learns scores the decoder can use. Step three is analysis. We examine how the gains depend on slice thickness, pathology, and orientation.

We are explicit about what this is and is not. The causal gravity results are analytic only in one space plus one time dimension, where each slice is a single loop. We use only the structural rule (single-step events, a length-scaled coupling), not the dynamics. The decoder resembles tracking-by-assignment with division events as used for cell lineages. The proposal's specific additions are the acyclicity constraint on the Reeb graph, the physical-length scaling, and orientation selection.

Baselines

The naive reference is a 3D U-Net trained with Dice plus cross-entropy, followed by keeping the largest connected component. This is the setting in which ATM'22 documents breakage and leakage.

The strongest existing reference point on the benchmark is the top ATM'22 team, timi. It reports Tree Length Detected (TD) of 95.9% and Branches Detected (BD) of 94.7% on the ATM'22 test set, according to the ATM'22 paper. We cannot rerun that multi-stage pipeline, so we treat it as an external ceiling to approach rather than as a controlled comparison.

Our controlled strong baselines are the following. SegMamba is the representative bidirectional scanner; we found no reported airway result on Lacuna for SegMamba on AIIB2023, so we train it ourselves. LTSP is the representative slice-recurrent model; it reports BD 90.83%, TD 87.59%, and DSC 92.95% on a private cohort of 70 CTs, a number not comparable to ATM'22. Skeleton supervision with geodesic distance maps from Skeleton Supervised Airway Segmentation reports 88.79% DSC on the BAS dataset.

For fairness, every method gets the same backbone family and patch size, the same training split (the ATM'22 training set of 300 scans), and the same GPU-hour budget. Every method also gets an equal number of hyperparameter trials selected on the ATM'22 validation set. Each baseline additionally receives the same connected-component post-processing.

Two ablations isolate the key ingredients. One replaces physical-length costs with constant per-slice costs. The other replaces orientation selection with summed bidirectional scores.

Beating the baselines means two things. First, a statistically significant BD and TD gain over the best controlled baseline at matched or lower false-positive (leakage) volume. Second, a smaller drop in BD when test scans are resampled to slice thicknesses unseen in training.

Why not just…?

Why not just use skeleton or centreline losses? These losses reward covering the centreline but do not forbid cycles or count spurious births. A prediction can score well on them while leaking into parenchyma. Our ablation "backbone plus skeleton loss" versus "backbone plus skeleton loss plus event decoder" measures whether the global acyclicity and event costs add anything. If they add nothing, the idea fails.

Why not just make LTSP or SegMamba bidirectional? The LTSP authors note that their propagation runs in one direction only, and SegMamba already sums forward and reverse scans. A summed bidirectional state is exactly the blended orientation we argue against. The orientation ablation, run on scans flipped along the head-foot axis, tests whether selection beats summation. We expect the difference to be small on canonically oriented scans. It should appear under flips, in the spirit of the flip failure reported for vision-language models by Your other Left!.

Why not just augment training with resampled slice thicknesses? This is the most serious threat to the scaling claim, and it may well close most of the gap. We therefore train every baseline both with and without thickness-resampling augmentation. We then ask whether per-millimetre costs still help on thicknesses outside the augmented range. If augmentation matches our result there, we will drop the scaling claim and keep only the topology decoder.

Why not just smooth in 3D? Roth et al. show that Gaussian smoothing raised a 2D pancreas model from 56% to 68% DSC. For airways only 2–5 pixels wide, however, smoothing blurs away the very branches at issue. We include smoothing only as a sanity baseline.

Proposed Experiments

The primary benchmark is ATM'22, reporting TD, BD, DSC, and leakage volume, with the COVID-19 subset reported separately. That subset is where average BD falls from about 90% to about 64%, and it is the natural stress test for a prior that bridges faint branches. Secondary evaluations use BAS and AIIB2023.

The first decisive check is thickness transfer. We train at native spacing, then test after resampling to 1.25, 2.5, and 5 mm. We plot BD against $\Delta z$ for the per-millimetre and per-slice cost variants. The claim holds only if the per-millimetre curve stays flatter.

The second decisive check is flip consistency. We measure disagreement between predicting on a flipped scan and flipping the prediction, comparing orientation selection with summation.

The third check is topological error counting. We count cycles and disconnected fragments in the predicted trees, compared against ground truth.

If the event decoder improves BD and cycle counts on the COVID-19 subset without increasing leakage, the prior is doing real work. If gains vanish once skeleton supervision and thickness augmentation are both present, we will report that honestly as a negative result.