Opportunistic Osteoporosis Screening on Chest CT Using Vertebral Normal-Density Voxel Distribution: A Voxel Density Graph–Derived Bone Mineral Content Proxy

1 Promedius Inc., Seoul, Republic of Korea 

² Department of Radiology, Guro Hospital, Korea University Medical Center, Seoul, Republic  of Korea 

Purpose 

To evaluate a voxel density graph (VDG) that constructs graph representation of the  vertebral-wide distribution and connectivity of normal-density voxels to derive a bone  mineral content proxy (BMCP) for opportunistic osteoporosis screening. 

Methods 

This single-center retrospective study included 301 patients (50.8±6.9 years) from a tertiary hospital with non-contrast chest CT and lumbar spine DXA within 12 months; cases with  prior spinal surgery, vertebral fracture, or osteolytic lesions were excluded. Vertebral bodies  

were segmented and normal density voxels (≥160 HU) within the vertebral body were  selected as nodes. Starting from the vertebral centroid, edges were iteratively added to link  each node to at least one neighboring node within a 10-pixel radius; the radius was  empirically selected. BMCP was defined as average graph density 2E/N (E; edges, N; nodes)  and computed on the L1 mid-vertebral slice ±1. For comparison, a conventional ROI-based  volumetric BMD measure was obtained as the median HU from an elliptical L1 trabecular  ROI centered at the centroid while excluding basivertebral vein. Osteoporosis was defined as  lumbar spine DXA T-score ≤ −2.5. ROC analysis compared AUCs between the ROI-based  measure and BMCP. 

Results 

Osteoporosis was present in 38/301 patients (12.6%). BMCP discriminated osteoporosis with  an AUC of 0.83 (95% CI, 0.75–0.91), outperforming the ROI-based approach (AUC 0.76;  95% CI, 0.67–0.85). VDGs demonstrated distinct vertebral-wide voxel-organization patterns  between osteoporotic and non-osteoporotic subjects, providing interpretable information  beyond summary HU. 

Conclusion 

BMCP improved CT-based osteoporosis detection compared with focal ROI metrics and may  better approximate DXA-referenced assessment for opportunistic screening on chest CT. 

Limitations 

Normal density HU threshold can vary across CT protocols. Also, BMCP was derived from  three L1 slices rather than full vertebral volumes; full-volume/multi-level extension,  resampling-based uncertainty estimates, and external/prospective validation are warranted.

Figure 1. Workflow for voxel density graph (VDG)–based bone mineral content proxy (BMCP)  computation. Vertebral bodies were automatically segmented from chest CT using segmentation AI  model (TotalSegmentator), and normal-density voxels were selected using an attenuation threshold of  ≥160 HU (Pickhardt et al., 2013). The selected voxels were represented as nodes and organized into a  voxel density graph based on spatial connectivity. BMCP was defined as the average graph density  (2E/N; E = number of edges, N = number of nodes). Representative cases illustrate the normal-density  voxel distributions, corresponding VDGs, and resulting BMCP values. 

Figure 2. Representative examples of VDG–based BMCP across DXA categories. Cases are  grouped by lumbar spine DXA T-score as normal (T-score ≥ −1.0), osteopenia (−1.0 > T-score >  −2.5), and osteoporosis (T-score ≤ −2.5). For each subject, an axial CT slice at the L1 level

is shown  with the segmented vertebral body mask and selected normal-density voxels (≥160 HU), followed by  the resulting normal-density voxel distribution, the corresponding VDG, and the computed BMCP  value. Lower BMCP values were observed in osteoporosis compared with normal/osteopenia in these  examples.

Figure 3. Receiver operating characteristic (ROC) curves comparing the proposed VDG–based  BMCP with a conventional ROI-based volumetric BMD (vBMD). ROC

curves were generated by  varying the decision threshold for each metric, with lumbar spine DXA T-score ≤ −2.5 used as the  reference standard. The VDG-based BMCP achieved a higher discriminative performance (AUC =  0.83) than the ROI-based vBMD approach (AUC = 0.76). Markers indic

ate the optimal operating  points selected by the Youden index for each method. 

Table 1. Patient chara

cteristics and CT acquisition parameters. Summary demographics (age and  sex) and CT tube potential distribution are reported for the study cohort (N = 301). Age is presented as  mean ± standard deviation and as median (range), and sex and tube potential are presented as n (%).