Mobile low-dose CT vans already screen many thousands of rural residents for lung cancer. Every scan also images the lower thoracic and upper lumbar spine, and that image data is thrown away. The vertebrae that fracture most often from osteoporosis (T12 and L1) sit at the bottom of a standard chest scan. Their thin outer bony wall, the cortical shell, is the part that carries most of the load. We propose an on-device tool that runs inside the van. It labels and measures these vertebrae and sends a few kilobytes of structured findings over a weak rural network. The goal is to flag people at high fracture risk who would never travel to a city clinic for a bone-density scan, and whom that scan would often miss anyway.
The case for this rests on three findings. Standard two-dimensional bone densitometry (DXA, dual-energy X-ray absorptiometry) missed 56% of patients who later had vertebral fractures. Volumetric CT density predicted those fractures better Improved prediction of incident vertebral fractures using opportunistic QCT compared to DXA. Rural, low-income residents face a "double burden" of longer distances to care and fewer vehicles Income disparities in driving distance to health care infrastructure in the United States: a geographic information systems analysis. Mobile CT has already reached these populations at scale Deep Learning Empowers Lung Cancer Screening Based on Mobile Low-Dose Computed Tomography in Resource-Constrained Sites.
Core Claim
The quantity worth extracting from a mobile chest scan is not the usual reading. The usual reading is the trabecular density at mid-lumbar levels, meaning the density of the spongy bone inside the vertebra. That reading is often cut off or absent on a chest scan, and low-dose noise corrupts it. We claim the better quantity is a compartment-resolved measure at the thoracolumbar junction (T11–L1). It combines cortical shell density, trabecular density, and a cheap estimate of failure load. All of it is computed without a calibration phantom, using the patient's own muscle and fat as internal references, and it stays stable as the van's scanner drifts.
The claim is falsifiable in a specific way. If T12–L1 trabecular density alone, read by an existing automated tool and thresholded at the American College of Radiology cutoffs, predicts fractures as well as our combined measure, then the cortical and biomechanical parts are not worth building. In that case the contribution shrinks to a validated drift-robust density pipeline for mobile low-dose CT. That pipeline would still be new, but it is a smaller result.
What Prior Work Does and Leaves Open
Four results motivate the focus on the thoracolumbar junction and the cortical shell. Löffler et al. showed that calibrated L1–L4 trabecular volumetric density predicts incident vertebral fracture (hazard ratio 4.07 per standard deviation, AUC 0.76). DXA did not reach significance in the same patients (HR 1.36, AUC 0.63). They also showed that the cutoffs of 80 and 120 mg/cm³ held across five tertiary-hospital scanners Improved prediction of incident vertebral fractures using opportunistic QCT compared to DXA. Their measurement is trabecular and lumbar; ours targets T11–L1 and adds the cortical shell.
Yang et al. decomposed vertebral density into shell and core using internal tissue calibration. Cortical density was lowest at T12, which is also where fractures peak (27%, then L1 at 20%). They also found that core density becomes artificially inflated after a compression fracture, while shell density stays stable The reduced cortical bone density in vertebral bodies: risk for osteoporotic fractures? Insights from CT analysis. Their work was cross-sectional, used standard-dose hospital CT, and made no prediction; we test the shell measure prospectively on low-dose mobile data.
Rayudu et al. built patient-specific finite element models of L1–L5 from routine CT. These are simulations that divide the bone into small elements and compute the load at which it fails. Failure load in osteoporotic subjects was roughly half that of controls Patient-Specific Finite Element Modeling of the Whole Lumbar Spine Using Clinical Routine Multi-Detector Computed Tomography (MDCT) Data—A Pilot Study.
That same paper also constrains us. Failure load correlated with density at $R^2 = 0.8$, so the room for finite element modeling to add information is limited. We therefore treat it as something to test, not to assume. It also cannot run inside a van, so we only consider a reduced-order surrogate, meaning a cheap approximation of the full simulation.
The method pieces come from several sources. SpineQ-style automated, phantom-free density extraction stayed accurate across 80–140 kVp and 72–200 mAs on cadavers (ICC 0.992, RMSE 8.93 mg/cm³ after excluding an osteophytic vertebra) Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography. Intraoperative CT work shows the same idea transfers to non-diagnostic scanners Opportunistic osteoporosis screening in intraoperative CT can accurately identify patients with low volumetric bone mineral density and osteoporosis during spine surgery. Neither was tested on low-dose chest protocols, mobile hardware, or longitudinal drift.
The cloud-edge design of Ting et al. sends about 6.5 KB of structured JSON instead of raw images. It held accuracy at 500 kbps A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings, but its edge tools cover only 2D signals and radiographs. We add a volumetric spine tool to that design.
The calibration step reuses a lesson from materials processing. Kampmann–Wagner precipitation models coupled with Bayesian optimization worked only once sensitive interface parameters were fitted to sparse measurements. A two-stage schedule then beat a single-stage one Utilization of Bayesian Optimization and KWN Modeling for Increased Efficiency of Al-Sc Precipitation Strengthening. We borrow only the structure: a physics model whose few uncertain parameters are fitted by Bayesian optimization against a small expensive reference. The alloy results themselves play no role.
On marrow fat, the evidence conflicts. Chang et al. report strong cross-sectional separation of normal from osteoporotic subjects Percentage fat fraction in magnetic resonance imaging: upgrading the osteoporosis-detecting parameter. The prospective ADIMOS study found no link to incident fractures Marrow Adiposity Content and Composition Are Not Associated With Incident Fragility Fractures in Postmenopausal Women: The ADIMOS Fracture Study. Fat cannot be measured reliably on CT anyway, so we leave it out and stay with structural measures.
Research Program
The edge tool has four stages.
Vertebra labeling. A vertebra segmenter and labeler is trained on the VerSe benchmark VerSe: A Vertebrae Labelling and Segmentation Benchmark for Multi-detector CT Images and CTSpine1K CTSpine1K: A Large-Scale Dataset for Spinal Vertebrae Segmentation in Computed Tomography. It is then fine-tuned on low-dose chest crops. Chest scans truncate the lumbar spine, so labeling cannot count up from the sacrum. It anchors on ribs instead, and the twelfth-rib-bearing vertebra fixes T12. CTSpine1K reports Dice near zero on transitional anatomy, so rib-anchored labeling is a measured component, not a detail.
Compartment densities. Each labeled vertebra is split into an outer shell and an inner core. Both are converted from CT numbers to density using internal muscle and fat references. At low dose and roughly millimeter resolution, the true shell is thinner than a voxel. We therefore report an apparent shell density, defined as integrated mineral within a fixed-width band. We do not claim to measure true thickness.
Drift correction. Calibration drift is handled as a parameter-fitting problem. A small set of per-van parameters (offset and gain of the CT-number-to-density map, plus a noise-dependent bias term) is re-estimated weekly. The fit uses the internal references pooled across that week's patients, plus a periodic phantom scan. Bayesian optimization fits these parameters against the phantom.
Surrogate failure load. A one-dimensional surrogate estimates failure load from shell density, core density, and endplate area. A handful of its coefficients are fitted, again with Bayesian optimization, to full finite element runs on a reference subset processed offline in the style of Rayudu et al.
The output payload holds per-level densities, the surrogate load, segmentation-quality flags, and confidence intervals. It must stay under 10 KB and be compatible with the Ting et al. message format.
Baselines
The naive reference point is what the vans do today: no skeletal reading at all. The DXA-referral equivalent is a referral rule based on age and sex. Every method is scored against this.
The strongest competitor is automated trabecular volumetric density with the American College of Radiology cutoffs, read at whatever of T11–L2 is visible. This is the approach validated by Löffler et al. Their reported strongest number is AUC 0.76 and hazard ratio 4.07 per standard-deviation decrease for incident vertebral fracture, on hospital CT. We reimplement it on exactly the same scans, segmentations, and internal calibration as our method, so the only difference is the measured compartment. Beating it means a statistically significant AUC gain, with a pre-registered target of at least 0.04, plus better calibration of predicted risk on the same follow-up data.
The upper reference is a full patient-specific finite element model in the Rayudu style, run offline. It tells us how much of its predictive value the cheap surrogate keeps.
A further competitor is an end-to-end 3D network trained to predict fracture directly from the spine crop, given the same training data and tuning budget.
Two ablations isolate the key ingredients. The first removes the shell channel, which leaves core density plus the surrogate. The second freezes calibration at installation and removes the weekly drift refit.
For segmentation, the reference is the VerSe 2020 winner (identification rate 96.6%, Dice 91.72%) on standard-dose CT. We report how much of that performance survives at low dose and with truncated fields of view.
Why not just…?
Why not just read trabecular density with an existing tool and threshold at 80 mg/cm³? This is the real threat. The $R^2$ of 0.8 between failure load and density in Rayudu et al. suggests the extra measures may add little. Two pieces of evidence argue the gap may matter at T12. Yang et al. found shell density lowest exactly where fractures peak, and found core readings inflated after fracture. The head-to-head prospective comparison and the shell ablation settle the question. If they show no gain, we report the simpler pipeline as the deliverable.
Why not just upload the scans for central reading? A low-dose chest series is hundreds of megabytes. Ting et al. showed latency rising up to 2.5-fold on degraded links, even with 2–7 MB payloads, while structured edge payloads stayed at 25–38 seconds. We will measure our own upload times at 500 kbps and 2 Mbps. We also note that even a cropped spine volume still needs a remote reader who does not exist in these settings.
Why not just send mobile DXA vans? DXA missed 56% of incident vertebral fractures in Löffler et al. A second van also duplicates the travel-and-scheduling burden that Guo et al. document. Our marginal cost is software on scans that are already acquired.
Why not just train an end-to-end fracture classifier? It is included as a baseline. The interpretable payload has two advantages over it. Clinicians can check the per-level densities against published cutoffs, and the per-level numbers degrade visibly when segmentation quality flags fire. Its main risk is that an end-to-end model overfits to scanner-specific texture, which is exactly what the drift experiment probes.
Decisive Experiments
Calibration and drift. Following the cadaver design of Ramschütz et al., cadavers and a phantom are scanned on a mobile unit using the lung-screening low-dose protocol. Scans are repeated over months of normal driving. Agreement with reference quantitative CT (ICC, RMSE in mg/cm³) is tracked with and without the drift refit. The success bar is RMSE at or below the published 12.65 mg/cm³, with no upward trend over time.
Fracture prediction. This needs follow-up of a mobile low-dose cohort such as the West China cohort. We assume access and extended follow-up, since its published 1-year window is too short for vertebral fractures. Outcomes are prevalent fractures on the baseline scan and incident fractures on rescreening. All baselines above are compared by AUC, hazard ratio per standard deviation, and calibration.
Bandwidth. Payload size and end-to-end latency are measured on the Ting et al. network profiles.
Equity. Guo et al.'s road-network method is applied to the screened catchment. It estimates the travel distance that flagged individuals would otherwise need to reach the nearest DXA site.
If the shell-plus-surrogate measure does not beat trabecular density on fracture prediction, that negative result is itself informative and directly sets what mobile programs should deploy.
References
- Improved prediction of incident vertebral fractures using opportunistic QCT compared to DXA
- Income disparities in driving distance to health care infrastructure in the United States: a geographic information systems analysis
- Deep Learning Empowers Lung Cancer Screening Based on Mobile Low-Dose Computed Tomography in Resource-Constrained Sites
- The reduced cortical bone density in vertebral bodies: risk for osteoporotic fractures? Insights from CT analysis
- Patient-Specific Finite Element Modeling of the Whole Lumbar Spine Using Clinical Routine Multi-Detector Computed Tomography (MDCT) Data—A Pilot Study
- Automated extraction of vertebral bone mineral density from imaging with various scan parameters: a cadaver study with correlation to quantitative computed tomography
- Opportunistic osteoporosis screening in intraoperative CT can accurately identify patients with low volumetric bone mineral density and osteoporosis during spine surgery
- A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings
- Utilization of Bayesian Optimization and KWN Modeling for Increased Efficiency of Al-Sc Precipitation Strengthening
- Percentage fat fraction in magnetic resonance imaging: upgrading the osteoporosis-detecting parameter
- Marrow Adiposity Content and Composition Are Not Associated With Incident Fragility Fractures in Postmenopausal Women: The ADIMOS Fracture Study
- VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images
- CTSpine1K: A Large-Scale Dataset for Spinal Vertebrae Segmentation in Computed Tomography