Copy Only What the Scan Confirms: Reference-Guided Super-Resolution for Long-Term Bone Implant Monitoring

Judging whether a bone scaffold or biodegradable implant actually works means watching the same animal or patient for months. The fine structure that decides mechanical success is the lattice of thin bone struts called trabeculae, together with the pores of the scaffold itself, and both are smaller than what safe, repeatable scanners can see. Laboratory micro-CT resolves 8 µm detail but delivers 272–1088 mGy per scan. Clinical CT delivers about 25 mGy but blurs everything below roughly half a millimetre (Bissinger et al.). Deep super-resolution can sharpen clinical scans, but it does this by inventing plausible fine texture. The one bone-specific system in the literature reports false structures near enamel and metal edges (Clinical super-resolution CT of bone microstructure). In an efficacy study, invented trabeculae are worse than blur.

We propose a different source for the fine detail. In a longitudinal study one high-resolution image is usually affordable, for example a single baseline micro-CT or the exact print geometry of a 3D-printed scaffold. Every later low-dose scan is then reconstructed by copying fine structure from that reference through a learned 3D correspondence field. Any copied structure must, when blurred by the scanner's measured point-spread function, reproduce the low-dose scan actually acquired. Where the reference no longer agrees with what the scanner sees, because new bone has grown or a strut has resorbed, the method refuses to copy. It marks the region as "changed, unresolved" instead of inventing detail. The output is a sharp volume plus a map of where sharpness is earned, which is what a long-term efficacy readout needs.

Core Claim

Coordinate-sampling generators from person-image synthesis keep high frequencies sharp because they fetch real pixels from a source image instead of predicting them. Grigorev et al. predict per-pixel source coordinates, and Leffa turns attention maps into flow fields supervised by a warping loss. Neither method is safe in a remodeling tissue, for the same reason it works: it will faithfully copy structure that no longer exists. Our claim is that one physical constraint makes reference copying usable for longitudinal bone assessment. That constraint is consistency with the low-dose measurement under a known scanner model, combined with a per-voxel gate that falls back to blur-honest output where the constraint fails. With it, the method should match single-image super-resolution on morphometry where tissue is stable. In regions that changed since the reference, it should do clearly better than both single-image super-resolution and plain registration.

Research Program

The model receives a low-dose scan $x_t$ at time $t$ and a high-resolution reference $r$. The reference is a baseline micro-CT of the defect site, or a voxelized scaffold print file plus a baseline scan of the surrounding host bone. A 3D network predicts a dense correspondence field $\phi_t$, in the spirit of Leffa's flow-in-attention but over volumes, and produces a candidate high-resolution volume $\hat y_t = r \circ \phi_t$ by sampling the reference. Large spatial stretches cause aliasing, so we sample through a multi-scale band decomposition chosen by the local stretch of $\phi_t$, borrowing the Laplacian pyramid warping of LookingGlass. The reason for doing this in voxel space rather than a compressed latent is that latent warps smear boundaries, a failure that paper documents.

The scanner is modeled by an operator $A$ that blurs with the measured point-spread function and downsamples to clinical voxel size, 210×210×500 µm for the MDCT in Bissinger et al. We then form a per-voxel residual between the warped reference viewed through the scanner, $A(\hat y_t)$, and what the scanner recorded, $x_t$. A gate $g_t \in [0,1]$ is driven by that residual relative to the scanner noise level $\sigma$:

$$g_t(v) = \exp\!\left(-\frac{\big[(A\hat y_t - x_t)(v)\big]^2}{2\sigma^2}\right), \qquad y_t = g_t\,\hat y_t + (1-g_t)\,u_t .$$

Here $v$ indexes voxels. The fallback $u_t$ is a conservative upsampling that adds no high-frequency content. Training penalizes $\|A y_t - x_t\|^2$ everywhere, plus a structural similarity loss against micro-CT where paired ground truth exists. We choose structural similarity because both MVTec AD and the clinical bone super-resolution paper found it preserves fine structure better than squared error. Adversarial and perceptual losses are deliberately excluded, since the clinical bone paper ties them to hallucinated mineral features. The ungated regions, $1-g_t$, are themselves an output: their volume over time is a direct readout of how much tissue has turned over.

Downstream, we compute the standard morphometrics, which are bone volume fraction, trabecular thickness and bone mineral content. We also compute grayscale texture descriptors (local binary patterns and pattern entropy) following Thevenot et al., who showed these track architectural shifts that binary thresholding misses. Every metric is reported separately for gated and ungated regions.

Two roles in the literature motivate this design. Bissinger et al. show that clinical CT reproduces mineral content (BMC 6.48 vs 6.15 mg, $p=0.40$) while overestimating bone volume and missing architecture. Thevenot et al. and Gao et al. argue that architecture, not mass, governs long-term stability. The scaffold studies explain why the reference must be allowed to go stale. The PTMC/PLA/HA composites of Kang et al. and the magnesium anchors of Hsu et al. resorb while bone forms, and these studies stop at 8–12 weeks precisely because longer imaging is costly.

Baselines

The naive reference point is trilinear upsampling of the clinical scan followed by standard morphometry. The clinical bone super-resolution paper reports a bone volume fraction correlation of $r_{BV/TV}=0.67$ for interpolation on its ex vivo test set. We run it on identical inputs.

A second naive point is to report only bone mineral content from clinical CT. This needs no fine detail and agrees with micro-CT per Bissinger et al. Our method must match it on mineral content while adding architecture metrics that this approach cannot produce at all.

The strongest existing method is the "structure model" of the clinical super-resolution CT paper. It is a resize-convolution network trained with a structural similarity loss for 4× magnification. On its ex vivo human osteochondral and dental test set it reports $r_{BV/TV}=0.817\pm0.005$, against 0.64 for conventional segmentation. We retrain it in 3D on exactly our training pairs, with the same scanner operator, compute and hyperparameter search budget. It therefore differs from ours only in lacking the reference. On that paper's own test protocol, beating it means exceeding 0.817 in stable regions or matching it. In changed regions, it means a significantly lower rate of false structure at equal or better correlation.

The strongest no-learning competitor is deformable registration of the high-resolution reference onto each low-dose scan, with morphometry read off the warped reference. This isolates whether learning adds anything beyond alignment.

Two ablations isolate the key ingredient. The first removes the gate, which gives an ungated Leffa-style volumetric correspondence model. The second removes the scanner-consistency loss but keeps a learned confidence head. If either matches the full model in changed regions, the physical constraint is not what matters.

Why not just…?

Why not just use single-image super-resolution, which already reaches $r_{BV/TV}=0.817$? That figure comes from 4× magnification of CBCT on ex vivo blocks. Rat trabeculae are thinner than the 210 µm in-plane MDCT voxel, so in longitudinal rodent work the information is simply not in the scan, and any network must invent it. The reference-free baseline in our experiments measures how often it does. If its false-structure rate in change regions is already as low as ours, the reference adds nothing and we would say so.

Why not just register the baseline micro-CT to each follow-up and skip learning? This is the real threat in stable host bone, where we expect registration to perform nearly as well as our method. It fails exactly where efficacy is decided. A warped baseline shows intact scaffold struts after they have resorbed, and empty space where new bone formed. The synthetic-remodeling experiment below measures this directly. If registration plus a simple residual check matches the full model, we would adopt that simpler pipeline.

Why not just track mineral content, which clinical CT already gets right? Bissinger et al. show mineral content agrees across modalities only because volume overestimation and density underestimation cancel. Thevenot et al. show that architectural reorganization correlates with disease grade ($R^2$ up to 0.67) independently of mass. Mineral content stays in our outputs as a sanity metric, not a substitute.

Why not just scan repeatedly with in vivo micro-CT? At 272–1088 mGy per scan, repeated imaging over months risks altering the healing response under study (Bissinger et al.). Our design spends that dose once.

Proposed Experiments

The first stage uses existing paired data with controlled change. We take high-resolution scans from ex vivo osteochondral samples and from harvested rat mandibles, of the kinds in the clinical super-resolution and Bissinger cohorts. We create a "past" reference by computationally editing each volume: eroding scaffold-like struts, deleting or adding trabeculae in marked regions, and applying smooth deformations. The true current volume is degraded through the measured scanner model to give the low-dose input. Because the edited regions are known, we can measure four things: morphometric correlation and error (bone volume fraction, trabecular thickness, texture entropy) in stable and changed regions separately; a false-structure rate, meaning bone voxels predicted where truth has none, inside changed regions; how well the gate detects changed regions, as area under the curve; and agreement in mineral content.

The decisive check is the changed-region false-structure rate. To support the claim, the full model must show a lower rate than both registration and the ungated ablation, and must not lose to the single-image baseline on stable-region correlation. We also confirm that real clinical scans, rather than only simulated ones, behave the same way. For this we use the Bissinger cohort, where each harvested jaw has both MDCT and micro-CT at timepoints from 7 to 112 days. A different animal's earlier-timepoint micro-CT, registered to the current jaw, serves as a deliberately mismatched reference, a stress test of the gate.

The second stage is a prospective longitudinal rat defect study using a 3D-printed resorbable scaffold of the Kang et al. type. Each animal gets one baseline in vivo micro-CT and its print file, then monthly low-dose clinical CT well beyond the usual 12 weeks. Staggered sacrifice gives ex vivo micro-CT and histology at several timepoints, which serve as ground truth. Here the question is the efficacy readout itself. Does the method recover the trajectory of strut resorption and trabecular bridging seen in terminal micro-CT closely enough that group differences between scaffold formulations reach the same statistical conclusions as micro-CT? Does the ungated fraction track histologically measured tissue turnover? If the gate turns out to be ungated almost everywhere by late timepoints, the method has degraded into honest blur. That is itself an informative negative result: it would bound how long a single reference stays useful.