Blur-Consistent Cortical Shell Segmentation for Opportunistic Spine CT

The outer shell of a vertebra, the cortex, carries roughly half to three quarters of the spine's mechanical load. On routine CT it is often thinner than one voxel. Scanner blur, together with averaging of different tissues inside a single voxel (the partial volume effect), smears this 1 mm shell into the soft tissue outside and the spongy trabecular bone inside. The best reported deep network for this task reaches only a Dice overlap of 0.71. Opportunistic screening tools therefore measure trabecular density and ignore the cortex. We argue that the bottleneck is mainly physical, not architectural. The fine detail of the shell is removed by the scanner before any network sees the image, so a network that preserves high frequencies internally cannot bring back what the scan never recorded. What can be recovered is a sub-voxel estimate of shell position and thickness that is consistent with the blurred measurement.

We propose a segmentation model with two parts. It predicts the inner and outer cortical surfaces as continuous, sub-voxel fields. During training, it is also required to explain the observed image through a differentiable model of the scanner's blur: the predicted shell is rendered at high resolution, blurred with the scan's own point spread function (the scanner's blur pattern, set by slice thickness and reconstruction kernel), downsampled, and compared with the measured Hounsfield values. This borrows a principle from generative image synthesis, where fine detail is kept faithful by tying it to a reference signal rather than letting a decoder invent it. Here the reference is the physics of the acquisition itself. The consistency term needs no cortex labels. That lets the model adapt to elderly, osteoporotic and thick-slice scans, where no annotations exist and where the cortex matters most clinically.

Core Claim

A cortical segmentation trained with a blur-consistency loss will estimate cortical thickness more accurately and more stably across scan protocols than voxel-wise networks, including frequency-aware ones, at equal training data. The resulting cortical measure will add fracture-prediction value on top of automated trabecular density. The claim is falsifiable at each step. If a plain nnU-Net trained on upsampled labels matches thickness error and protocol stability, the physics term is unnecessary. If cortical thickness adds nothing to fracture prediction, the clinical motivation fails.

Relation to Prior Work

Three papers sit closest to this idea, and each leaves a different piece open.

Automated segmentation of vertebral cortex with 3D U-Net-based deep convolutional neural network segments the T11–L2 cortex with a 3D U-Net and Dice loss after 100–800 HU windowing. It reports a Dice of 0.71, against 0.62 for Res U-Net and 0.59 for SegNet, on 105 young, healthy patients. It motivates this proposal by showing that the cortex task is hard and that failures cluster where cortex and trabecular bone fuse. It treats the problem as voxel classification at 1 mm, however, and never tests older spines.

Frequency-enhanced Multi-granularity Context Network for Efficient Vertebrae Segmentation (FMC-Net) is the obvious "high-frequency plus vertebrae" combination, and it already exists. It replaces pooling with wavelet downsampling so that edge information survives inside the network, and it reaches 79.92% Dice on VerSe 2019. It segments whole vertebrae only. Its high-frequency path preserves detail that is present in the input; it has no model of detail the scanner removed. Our proposal does not compete on network internals. It adds an output representation and a training loss grounded in acquisition physics.

Modeling the acquisition shift between axial and sagittal MRI for diffusion superresolution to enable axial spine segmentation supplies the key methodological idea we reuse: an explicit forward degradation model (point spread function, slice gaps, noise) makes learned detail recovery realistic. It also supplies the key constraint. Its authors note that beyond roughly 6× upscaling the diffusion model produces a generative "best guess" that can invent bone structure. That is why we do not super-resolve the image and then segment. We estimate a low-dimensional shell geometry whose re-blurred version must match the data, which leaves far less room for invention than generating full-resolution texture.

From generative vision, Learning Flow Fields in Attention for Controllable Person Image Generation (Leffa) shows that a training-only loss can keep fine structure faithful at zero inference cost: it tethers attention to a reference image through a warping loss. Our blur-consistency term has the same shape, a training-time tether with no change at inference, but the reference is the measured scan passed through the scanner model. TryOn-Adapter motivates separating coarse structure from fine detail. We apply that separation by taking the outer surface largely from existing whole-vertebra masks and learning the inner surface, which is where the difficulty lies.

The clinical anchor comes from two further studies. Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography and Automated Opportunistic Trabecular Volumetric Bone Mineral Density Extraction Outperforms Manual Measurements for the Prediction of Vertebral Fractures in Routine CT show that automated trabecular density is robust across protocols (ICC 0.992) and predicts fractures (β = −0.32). Neither measures the cortex. The cadaver study also finds that a spondylophyte, a bony spur on the vertebral edge, raises density error from 8.93 to 12.65 mg/cm³. That is a boundary failure our method should reduce.

Method

For each vertebra, the network outputs two signed distance fields: continuous functions giving each point's distance to the outer or inner cortical surface. It also outputs a trabecular density field. These are rendered on a grid finer than the scan (0.25 mm), producing a density volume $\rho_\phi$. The cortex is assigned a density $c$ that is either fixed from physiological priors or constrained to a narrow range. The training loss adds a consistency term to the usual supervised terms:

$$\mathcal{L}_{\text{blur}} = \big\| \, D_s\!\left(h_{\psi} * \rho_\phi\right) - I \, \big\|_1 ,$$

Here $I$ is the observed CT volume in Hounsfield units. $h_\psi$ is the scan's point spread function, a parametric blur whose in-plane width is set by the reconstruction kernel and whose through-plane profile is set by the slice thickness in the scan metadata; its parameters $\psi$ can be refined per scan. $*$ is convolution, and $D_s$ resamples onto the scan grid with spacing $s$. The loss is computed only in a band around the predicted cortex.

The main technical risk is identifiability. When the shell is thinner than the blur, a thin dense shell and a thick faint one produce nearly the same blurred profile. What the image preserves is the integral of density across the shell. Fixing or tightly constraining the cortical density $c$ turns that integral into a thickness estimate. This is the reason for the density prior, and the reason we report thickness uncertainty whenever the blur width far exceeds the estimated thickness.

Training uses three data sources. Whole-vertebra masks from CTSpine1K and VerSe supervise the outer surface. A small set of cortex-labelled scans supervises the inner surface; we would request the Li et al. data or annotate about 30 VerSe lumbar vertebrae ourselves. Blur consistency on all scans, including unlabelled older and thick-slice ones, supplies the remaining signal.

Research Program

The work proceeds in three stages. The first builds the differentiable renderer and scanner model and validates it on synthetic shells of known thickness blurred with realistic point spread functions. That measures how thin a shell can be recovered at each slice thickness before the thickness estimate becomes uninformative. The second trains the model and baselines on shared data. The third tests the clinical payoff: whether cortical thickness, measured alongside trabecular density, improves fracture discrimination. That final question decides whether this line of work matters to opportunistic screening.

Baselines

The naive reference point is HU-window thresholding (100–800 HU) followed by morphological erosion of the whole-vertebra mask to form a fixed-width shell. It shows how much a learned model adds beyond "cortex is the bright rim."

The strongest existing cortex-specific method is the 3D U-Net of Li et al., which reports a cortex Dice of 0.71 and mean pixel accuracy of 0.83. We will reimplement it with its 256×256×16 patches and Dice loss.

The strongest general-purpose competitor is nnU-Net, the self-configuring segmentation pipeline benchmarked in CTSpine1K. We run it twice: once at native resolution, and once with images and labels resampled to 0.25 mm so it has every chance to represent sub-voxel geometry.

The strongest frequency-aware competitor is FMC-Net, retrained for three classes (cortex, trabecular, background). No reported cortex result exists on Lacuna for FMC-Net; its reported number is whole-vertebra Dice of 79.92% on VerSe 2019, which is not comparable to cortex Dice.

All methods get the same labelled scans and the same unlabelled pool. Baselines use the unlabelled pool through standard self-training with pseudo-labels, so our method's access to unlabelled data is not an unfair advantage. Every method receives the same GPU-hour budget and hyperparameter search budget.

Two ablations isolate the key ingredient. The first removes the blur-consistency loss while keeping the sub-voxel surface output. The second keeps the loss but uses a deliberately wrong point spread function, blurring with 1 mm when the true slices are 3 mm. If the wrong-blur variant performs as well as the correct one, the gain comes from regularization and not from physics.

Beating the baselines means three things. Cortical thickness error against the reference must be at least 25% lower than nnU-Net at 0.25 mm. Cortex Dice on the Li-style evaluation must exceed 0.71. Thickness estimates must vary less across the six cadaver scan protocols than every baseline.

Why not just…?

Why not just run nnU-Net on upsampled scans? This is the most serious threat. The answer depends on what the labels contain. If the cortex labels are drawn at 1 mm, upsampling images and labels cannot teach sub-voxel thickness, because the targets never contained it. The prediction is that nnU-Net at 0.25 mm will still quantize thickness to roughly the label resolution and will drift as slice thickness changes. The synthetic-shell experiment and the cadaver protocol comparison test this directly. If nnU-Net matches us on both, we would report that and drop the physics claim.

Why not just use a frequency-preserving architecture like FMC-Net? FMC-Net protects high-frequency features the input already contains. The cortex's fine detail is lost at acquisition, not inside the network. Retraining FMC-Net for three classes tests this. We expect it to beat the plain U-Net on Dice, where sharper edges help, but not on thickness accuracy under thicker slices. The frequency-aware design might combine well with our loss, and we will also report FMC-Net trained with blur consistency.

Why not super-resolve the scan first, then segment? The diffusion super-resolution work Modeling the acquisition shift... did improve axial Dice, from 0.684 with cubic upsampling to 0.718. Its authors also warn that large upscaling produces invented bone detail. Synthesizing trabecular texture that looks right is exactly the failure a density biomarker cannot tolerate. We will include a super-resolve-then-segment pipeline, applying an RCAN-style model to synthetically thickened VerSe scans, and measure thickness bias against the thin-slice originals.

Why not just threshold? Li et al. needed a deep network on top of HU filtering because thresholding leaks into soft tissue. Manual ROI methods also predict fractures worse than automated ones, with β = −0.15 against −0.32 in the fracture study. Thresholding is kept as the floor baseline.

Proposed Experiments

The first experiment uses controlled degradation. We take thin-slice (≤1 mm) scans from VerSe and CTSpine1K, simulate 2, 3 and 5 mm slices and smoother reconstruction kernels with the same forward model, and compare each method's thickness map on the degraded scan with its map on the original. The model that produced the degradation must not be reused as the evaluation reference, so we also degrade with a mismatched blur family, to avoid rewarding our own assumptions.

The second experiment is the decisive real-world check. It uses the cadaver protocol set from the SpineQ cadaver study: 14 scans of the same specimens under six protocols (80–140 kVp, 0.5–1.0 mm). The 0.5 mm acquisitions serve as reference. The metrics are thickness error, and variance of per-vertebra thickness across protocols. The spondylophyte vertebra is analysed separately. Access to these scans must be arranged with the authors.

The third experiment evaluates cortex overlap on Li-style labels, using Dice and average surface distance, and adds a new split of older patients drawn from CTSpine1K's colonography cohort with our own annotations. This tests the untested generalization to older spines.

The fourth experiment tests clinical value on a fracture cohort like the 157-patient one in the fracture study, subject to data access. We ask whether adding cortical thickness to automated trabecular density improves the fracture association and discrimination (AUC) beyond the reported β = −0.32 for density alone. No reported result was found on Lacuna for cortical thickness as a fracture predictor from routine CT, so this is the question the proposal ultimately exists to answer.