Coupling-Randomized Simulation for Separating Bone from Vessels in Perfused Micro-CT

Bone grows where blood vessels are. A specialized class of capillaries in the growing ends of long bones, called Type H vessels, feeds signals to the cells that lay down bone. When this pairing breaks down, as under steroid exposure, radiation, or loss of specific growth factors, bone is lost. Biologists measure the pairing by filling the vasculature with a radiopaque casting agent and scanning the bone with micro-computed tomography (micro-CT, a laboratory CT scanner with micrometre voxels). The casting agent and mineralized bone look almost identical in the scan. So vessels and bone are usually separated by a destructive second scan after dissolving the mineral, or by hand. We propose training a 3D segmentation network entirely on simulated volumes to do this separation from a single scan. The simulated volumes come from a procedural generator of vessel trees and trabecular bone (the thin struts of spongy bone), passed through a physical model of the CT scanner.

The main point is a hazard that is specific to this biology. In real healthy bone, vessels sit right next to newly forming trabeculae. A network trained on data that has this pairing can learn to predict "vessel" from "nearby bone" instead of from vessel shape. That shortcut is harmless on healthy controls. It fails exactly on the animals scientists care about, those in which the pairing has been broken, and it pushes their measured vessel loss toward zero. We therefore make the strength of the vessel–bone coupling a randomized simulator parameter, including uncoupled and anti-coupled regimes. We then test whether this is what lets a synthetic-trained model recover real uncoupling phenotypes.

Core Claim

For separating vessels from bone in contrast-perfused micro-CT, synthetic-only training transfers to real scans only when two conditions hold. First, the simulator models the scanner's blur, partial-volume mixing and noise. Second, the simulator randomizes the spatial coupling between vessels and bone. A simulator that faithfully reproduces healthy coupling produces a model that looks accurate on control animals but underestimates vessel loss in coupling-disrupted animals.

The shortcut argument follows directly from sim-to-real work. Adversarial Spatial Masking showed that networks seize on whatever cue separates domains or classes most cheaply. In their case, domain discriminators latched onto the presence or absence of LiDAR point dropout. Here the cheap cue is biological rather than sensory.

The claim is also narrower than "synthetic data for bone imaging." Ferreira et al. already showed that a 3D nnU-Net trained only on diffusion-generated clinical CT segments macroscopic head-and-neck bone well, with a Dice score of 0.841. The same approach collapses to 0.064 on low-contrast, texture-defined tumours, even though the images passed a radiologists' visual Turing test. Our target, vessel against bone at trabecular scale, is neither high-contrast nor texture-defined. The two materials have similar intensity, and they differ mainly in topology and shape: branching tubes versus plates and struts. That is why we generate geometry procedurally and render it physically, rather than learning an image generator.

Research Program

The simulator has three parts. The first grows a vessel tree inside a bone-shaped domain, using a branching process with randomized radii, branching angles and density. It includes a dense metaphyseal capillary bed that mimics the Type H region described by Tang et al. and reviewed in Cell communication and relevant signaling pathways in osteogenesis–angiogenesis coupling.

The second part deposits trabecular bone. It uses a randomized plate-and-rod generator whose local deposition probability depends on distance to the nearest vessel:

$$p_{\text{bone}}(x) = \sigma\big(a + \lambda\, e^{-d_v(x)/\ell}\big),$$

where $x$ is a voxel, $d_v(x)$ is its distance to the nearest vessel surface, $\ell$ is a length scale of tens of micrometres, $a$ sets baseline bone density, $\sigma$ is the logistic function, and $\lambda$ is the coupling strength. A "faithful" simulator fixes $\lambda$ to a positive value fitted to control animals. The coupling-randomized simulator draws $\lambda$ from a range that includes zero and negative values. Separately, it can thin the vessel tree to mimic the rarefaction produced by the endothelial-to-mesenchymal transition described in steroid-induced osteonecrosis.

The third part is the scanner model. It assigns attenuation values to bone, casting agent and soft tissue, with randomized overlap between the bone and agent intensity ranges. It then applies a point-spread blur, integrates to the target voxel size, and adds photon-count noise and mild beam hardening, which is the tendency of dense material to shift the X-ray spectrum. Following the logic of DODA's virtual scan simulation, the aim is to model the acquisition process that produces real artifacts rather than to imitate their appearance.

A 3D segmentation network (nnU-Net configuration) is trained only on these volumes to output vessel, bone and background.

We also run a lower-resolution variant, down to clinical-CT-like voxels. Bissinger et al. showed that at such voxels bone volume is overestimated and mineral density underestimated, while bone mineral content, the product of the two, is conserved. We use their result as a physical sanity check: the model's bone mineral content estimate should stay stable as resolution degrades, even though volume and density individually drift.

The real evaluation data must be collected. The file contains no public perfused-bone benchmark. The cohort includes control mice and two coupling-disruption groups: mice with megakaryocyte-specific TGF-β1 depletion, which Tang et al. report reduces vessel volume by about 32.5%, and glucocorticoid-treated mice. Every specimen is perfused and scanned intact. It is then decalcified and rescanned, and the two scans are registered. The decalcified scan contains only the casting agent and gives a vessel label. The intact scan minus the registered vessels gives a bone label. This provides dense ground truth without manual slice annotation.

Baselines

The naive reference is the standard laboratory pipeline: a global intensity threshold on the intact scan, followed by morphological opening to split tubes from struts. It sets the floor, and it shows how badly vessel and bone intensities overlap in practice.

The strongest real-data competitor is a supervised network trained on a small number of real registered pairs. We run it both as a plain nnU-Net and as a conditioned transformer in the style of ConUNETR, which adds learnable developmental-stage tokens and spatial encodings. On a 5% slice budget, ConUNETR improved Dice on an unseen mutation by 2.9% at embryonic day 14.5 and 5.2% at day 16.5 over unconditioned baselines. Conditioning of this kind is the most direct real-data answer to biological distribution shift.

The strongest synthetic competitor is Ferreira-style training on scans generated by a conditional diffusion model fit to real intact scans and labels. It reaches 0.841 Dice on clinical CT bone (Ferreira et al.). No reported result was found on Lacuna for vessel–bone separation in perfused micro-CT under any method, so all numbers on our task will be new.

We also run feature-level adversarial alignment (a DANN-style domain discriminator) from our synthetic data to unlabeled real scans. This tests whether alignment can rescue a faithful simulator.

Three ablations isolate the proposal's ingredients. The first uses the faithful simulator with $\lambda$ fixed. The second uses coupling randomization but no scanner model, rendering clean label volumes with only additive noise. The third is a "random geometry" simulator that places random tubes and random struts with no bone-shaped layout, which checks whether any biological structure is needed at all.

All learned methods share the same nnU-Net backbone, except the ConUNETR variant. They also share the same compute budget, the same hyperparameter search budget, and the same held-out test animals. Real-supervised methods train on control animals only when we test transfer to held-out disrupted phenotypes.

Beating a baseline means one of two things. Either the method has higher vessel Dice on disrupted animals together with a smaller absolute error in the group difference in vessel volume fraction between knockout and control, or it matches the baseline while using no real labels.

Why not just…?

Why not just train on a few real decalcified pairs? This is the real threat, because our own labelling protocol makes supervised training possible. Two things limit it. The pairs are expensive, and they come only from the phenotypes one happened to breed. A model trained on controls meets the same coupling shortcut as the faithful simulator, and nothing in the real data breaks the correlation. We test this directly by training on controls and testing on knockouts, and by sweeping the number of real pairs, both with and without synthetic pretraining. If real-only training with a handful of pairs matches the coupling-randomized model on disrupted animals, we will report that the simulator's value is limited to label-free settings.

Why not just generate training data with a diffusion model? Ferreira et al. show that diffusion samples can look realistic while losing fine texture: the concordance of radiomic texture features was 0.546 for CT tumours. More importantly, a generator fit to real scans reproduces the real coupling. So it should inherit the shortcut. The phenotype-transfer experiment tests this prediction directly.

Why not just align features between synthetic and real scans? PixelDA reported that feature-alignment methods left large errors on real-object pose estimation (56.6° for DANN versus 23.5° for pixel-level adaptation). ASM showed discriminators exploiting sensor cues. Alignment also cannot remove a shortcut that exists in both domains. We include the adversarial-alignment baseline to confirm this empirically.

Why not just randomize everything with no biology? If random tubes and struts work as well as the full simulator, then the bone-shaped structure adds nothing, and we would say so. We expect the biological layout to help mainly through realistic vessel-to-strut proximity statistics, which the random-geometry ablation lacks.

Decisive Experiments

The central test compares the faithful and coupling-randomized simulators on real knockout animals. We measure three things. The first is vessel Dice against the decalcification ground truth. The second is the recovered knockout-minus-control difference in vessel volume fraction, compared with the ground-truth difference in the same cohort, which we expect to be near the 32.5% reported by Tang et al.. The third is a "coupling bias" score: the correlation between per-region vessel error and local bone density. The core claim is supported if the faithful model shrinks the group difference and shows positive coupling bias while the randomized model does not. It is refuted if the two simulators behave the same on disrupted animals.

Secondary checks cover three questions. We test whether removing the scanner model reproduces the collapse Ferreira saw on low-contrast targets. We test whether bone mineral content from the low-resolution variant tracks micro-CT values the way Bissinger et al. found for real scanners. And we measure how many real pairs are needed before real-only supervision catches up with synthetic pretraining.