Introduction
Lung cancer screening (LCS) using low-dose computed tomography (LDCT) can reduce lung cancer-specific mortality and is being implemented globally.1 LCS targets individuals with significant smoking exposure, who are at risk of over 40 other smoking-related diseases. LCS provides an opportunity to systematically assess for clinically significant non-lung cancer conditions that could reduce all-cause mortality in the screened population.2
Osteoporosis is one such condition. Despite its high prevalence among older adults who have smoked, osteoporosis is often overlooked.3 Smoking increases osteoporotic fracture risk by 30-50%,4 and these fractures are associated with all-cause mortality.3 Spinal vertebrae imaged on LDCT can be assessed for bone mineral density (BMD) and vertebral fracture (VF), facilitating early detection of low BMD (osteopenia and osteoporosis) and timely intervention.
The WHO reference-standard test for osteoporosis is dual-energy X-ray absorptiometry (DXA).5 DXA T-scores can be incorporated into fracture risk prediction tools, such as FRAX and Garvan.6, 7 DXA is widely available and has a low radiation dose, but factors such as degenerative spinal disease, obesity, and aortic calcification may erroneously inflate BMD measurements when measured using a standard anteroposterior approach, thereby underdiagnosing osteoporosis.8 Another validated test, unaffected by these confounders, is volumetric trabecular BMD measured using quantitative CT (QCT). However, standard QCT requires dedicated CT imaging with in-scan phantom calibration, comparatively higher radiation exposure, and specialized analysis software.
An alternative CT-based measurement is vertebral trabecular bone attenuation measured in Hounsfield units (VHU), which correlates with BMD.9 VHU can be measured from existing CT scans, avoiding the limitations of DXA and the additional radiation required for QCT. Tested retrospectively, VHU has comparable accuracy against DXA and QCT and may be superior to DXA for fracture prediction.10–12 Threshold values for categorizing osteoporosis have been retrospectively determined, ranging from 99 to 136 HU, depending on skeletal site and apparatus.13 The International Early Lung Cancer Action Program screening protocol recommends “case-finding” LCS participants with low BMD using VHU, or an alternative visual scoring method, and referring them for appropriate diagnostic work-up.14 While visual methods are convenient, quantitative VHU measurements can be paralleled against standard BMD tests in osteoporosis and fracture risk assessment.15
Prospective validation of VHU against standard DXA in an LCS cohort has not yet been undertaken. Furthermore, no within-population comparison of VHU, DXA, QCT, and clinical fracture risk prediction tools has been performed. We aimed to address these gaps by prospectively assessing BMD and VF in LCS participants, using VHU, QCT, and DXA methods and clinical fracture risk tools (FRAX and Garvan) to discriminate low BMD (using DXA as the reference standard), prevalent VF, and 2-year incident VF.
Methods
Study design
This was a cross-sectional study, performed at the 2-year follow-up time point of participants enrolled in the Osteoporosis substudy of the International Lung Cancer Screening Trial (ILST) (ACTRN12621001367831).16, 17 Participants were healthy, asymptomatic men and women, aged 55-80 years, at risk of lung cancer, who met USPSTF eligibility criteria (current or former smoking; quit smoking ≤15 years ago; ≥30 pack-year history) and/or had an estimated PLCOM2012 lung cancer risk ≥1.51%.17, 18 Participants were recruited consecutively via telephone between July and December 2021. Participants underwent screening LDCT scans at baseline and at 2 years, DXA at 2 years, and annual collection of demographic and osteoporotic fracture risk variables. Participants provided written informed consent.
Ethical approval was obtained through the Hospital Human Research Ethics Committee (HREC/2021/QPCH/76621).
DXA acquisition
DXA scans (Lunar Prodigy, GE Healthcare) were performed within 6 months of participants’ 2-year follow-up screening LDCT scans using a standard protocol (anteroposterior spine and left femur, 0.018-0.17 mSv effective dose) and assessed by a nuclear medicine physician for areal BMD, T-score, and Z-score from the lumbar spine (average of L1-L4 vertebrae), L1 vertebra, left femoral neck, and total hip. DXA assessment, using WHO criteria, was blinded to CT results (T-score> −1 (normal BMD); T-score −2.5 to −1 (osteopenia); T-score< −2.5 (osteoporosis).5
CT acquisition
Non-contrast LDCT scans were obtained with a GE Healthcare Revolution 256-detector row CT scanner using a standardized protocol (120 kV, 40-50 mA, ≤1.5 mSv effective dose) at baseline and repeated after 2 years.17
Vertebral Hounsfield unit density assessment (VHU)
Reconstructed images, with a standard soft-tissue kernel and ≤1 mm thickness and spacing, were assessed for blinded BMD and VF measurements by a single observer. VHU was measured from reconstructed axial images (bone window, W: 2500, L: 480, compression 4:1) using InteleViewer software (Intelerad Medical Systems Inc., Montreal, QC, Canada).9, 16 VHU was derived from the mean Hounsfield unit attenuation of a region-of-interest in the anterosuperior portion of the most caudally imaged vertebral body (Figure 1).9 The first lumbar vertebra (L1) was preferentially measured, but if it was incompletely visualized or pathological, the next superior, visually normal vertebra was measured. VHU < 110 HU defined low BMD (osteoporosis and osteopenia).19
Quantitative CT density assessment
Asynchronous QCT was performed on LDCTs using Mindways QCT Pro software (Austin, TX, USA). Mean and individual volumetric BMD of imaged vertebrae T11-L1 were measured from the upper half of each vertebral body. Mean QCT-BMD thresholds were defined using standard criteria: ≥120 mg/cm3 (normal); 80-120 mg/cm3 (osteopenia); < 80 mg/cm3 (osteoporosis).20
Vertebral fracture assessment
Prevalent radiological VF were detected from baseline midsagittal CT images across all imaged vertebrae using the semiquantitative Genant method; 20%-25% vertebral height loss (Grade 1); 25-40% height loss (Grade 2); >40% height loss (Grade 3).21 Grades 2-3 were considered clinically significant, but all grades were included in sensitivity analyses. Incident radiological VF were assessed from CT scans repeated after 2 years, blinded to baseline results. We defined incident VF as a Genant grade increase ≥2 from baseline.
We performed sensitivity analyses using a less conservative definition (Genant grade increase ≥1). Incident osteoporotic (low-trauma) fractures at other skeletal sites were self-reported in annual participant questionnaires and verified by the study team using radiological and clinical reports. Unverified self-reported fractures were excluded from the analysis.
Clinical osteoporotic fracture risk assessment
Risk variables were obtained from participant questionnaires. Missing responses were sought through follow-up phone calls, but those whose responses remained incomplete were excluded. Ten-year hip and major osteoporotic fracture risks (for hip, spine, distal forearm, or proximal humerus fracture) were estimated using FRAX (Australia-specific version) and Garvan calculators, with and without inclusion of the DXA T-score at the 2-year follow-up time point.7, 22 10-year FRAX or Garvan hip fracture risk >3% or major osteoporotic fracture risk >20% defined “high-risk”.23
Analysis
Statistical analysis was performed using R (RStudio, V1.3.1073, Boston, MA, USA). BMD methods were compared pair-wise using correlation (Pearson’s coefficient, R), and inter-method reliability (kappa coefficient, ĸ). We used area under the receiver operating characteristic curves (AUC) to measure method discrimination of DXA-defined low BMD, prevalent VF and incident fracture, with Delong method to test AUC differences, and bootstrapping used for non-inferiority testing.24 Optimal VHU threshold for discriminating low DXA-BMD was determined using Youden index.
P value < .05 distinguished statistical significance. We performed decision curve analysis to assess the net benefit of VHU compared with other methods of fracture prediction.25
Sample size
A sample of ≥94 participants was projected to obtain a power of 80% with Type 1 error of 0.05. This was determined from our previously published AUC values to allow non-inferiority testing between VHU and QCT (using DXA as the reference standard), with a non-inferiority margin of 0.1.16 This margin was considered a clinically acceptable difference in AUC, lying within the range of other non-inferiority diagnostic imaging studies.26, 27
Results
Participants
Study flow is shown in Figure 2. Ninety-seven participants were recruited. Ninety-one participants had complete imaging data for DXA, QCT, and VHU (median age, 67 [IQR = 13] years; 43 [47%] male; 37 (53%) actively smoking).
Six participants failed QCT analysis (2 scans were performed externally and were incompatible for software analysis; 4 scans had encountered calibration module errors).
Vertebral BMD assessments
VHU was measured at vertebral levels L1 (n = 63), T12 (n = 25), T11 (n = 1), and T10 (n = 2). QCT was measured individually at L1 (n = 64), T12 (n = 88), T11 (n = 90), and at all 3 levels (n = 63). Examples of paired DXA and VHU measurements are shown in Figure 3.
Prevalence of low BMD
Of 91 participants with complete imaging data, 63 (69%) participants had low DXA-defined BMD (M:F ratio, 0.75) of whom 15 (16%) had DXA-defined osteoporosis (Table 1). When restricted to spine T-score only, 44 (48%) participants had low DXA-BMD. In comparison, 58 (64%) participants had low QCT-BMD and 44 (48%) had low VHU-BMD.
Prevalent vertebral fractures
Grade 2-3 VF were identified in 14 (15%) of participants (M:F ratio 1.3) (Table 1). In a sensitivity analysis, 25 (27%) participants had prevalent Grade 1-3 VF (M:F ratio 1.5).
Incident radiological and low trauma fractures
Two (2%) participants had incident Grade 2-3 VF radiologically. In sensitivity analysis, 13 (13%) participants had incident Grade 2-3 VF. Five of 71 (7%, M:F ratio 0.67) participants who completed questionnaires reported verified low-trauma fractures (wrist (n = 1), rib (n = 2), metatarsal (n = 1), and humerus (n = 1)).
Combining radiological VF and self-reported fractures, 6 (8%) participants had incident fractures, increasing to 15 (21%) in the sensitivity analysis.
Prevalence of high osteoporotic fracture risk
FRAX scores were measurable in 79 participants, and Garvan scores in 72 participants. Of 71 participants with paired scores, 34 (47%) were identified as “high-risk.”
Discrimination of low BMD
Discrimination of low DXA-BMD by VHU was moderate (AUC, 0.74 [95% CI, 0.63-0.85]) (Figure 4) and not significantly different when VHU was measured from T12 or L1 (AUC, 0.79 vs 0.75, P > .05). Discrimination was similar compared to DXA spine, L1, femoral neck and total hip T-score (Table 2). VHU was non-inferior to QCT (AUCQCT, 0.73 [95% CI, 0.61-0.86], non-inferiority P < .05) and was not significantly different to FRAX+T-score or Garvan with/without T-score inclusion (AUCFRAX+T, 0.82 [95% CI, 0.71-0.92], [Figure 4B], AUCGarvan, 0.71 [95% CI, 0.57-0.85]) [Figure 4A] (P > .05)).
Exclusion of T-score from FRAX significantly reduced discriminative accuracy (AUCFRAX, 0.52 [95% CI, 0.36-0.68], P < .05).
VHU threshold of 110HU had 57% (95% CI, 45-69%) sensitivity, 71% (95% CI, 53-85%) specificity, 43% (95% CI, 30-57%) negative predictive value (NPV) and 82% (95% CI, 68-91%) positive predictive value (PPV) in identifying participants with low DXA-BMD. A higher threshold (126 HU), determined in our previous study, improved sensitivity to 83% (95% CI, 71-90%) and reduced specificity to 46% (95% CI, 30-64%) (NPV 78% [95% CI, 66-86%], PPV 54% [95% CI, 35-72%]) for low DXA-BMD.16 The optimal VHU threshold for low DXA-BMD categorization in this dataset was 117HU (specificity 68%, sensitivity 71%).
Discrimination of DXA-defined osteoporosis
Discrimination of DXA-defined osteoporosis was better using FRAX and Garvan risk scores inclusive of T-score; however, it was not statistically different from that the VHU and QCT AUCs: VHU, 0.63 (95% CI, 0.46-0.80); QCT, 0.67 (95% CI, 0.51-0.83); FRAX+T-score, 0.80 (95% CI, 0.69-0.90); and Garvan+T-score, 0.77 (95% CI, 0.65-0.89), respectively (P > .05).
A VHU threshold of 110 HU had a 53% (95% CI, 30-75%) sensitivity, a 53% (95% CI, 42-64%) specificity, an 85% (95% CI, 72-93%) NPV and 19% (95% CI, 10-33%) PPV in correctly classifying osteoporosis. The optimal threshold for osteoporosis categorization in this dataset was 88 HU (specificity, 85%; sensitivity, 47%).
Discrimination of prevalent fracture
VHU showed moderate discrimination for individuals with prevalent VF (AUC, 0.71 [95% CI, 0.58-0.84]) and was not statistically different from any other method (Table 3). A sensitivity analysis using any grade VF (Genant score ≥ 1) found similar results.
Discrimination of incident fracture
VHU moderately discriminated individuals with incident vertebral or other osteoporotic fractures (AUC, 0.71) and was not significantly different to other methods: QCT (0.72), DXA spine (0.69), DXA femoral neck (0.75), DXA L1 (0.68) or the lowest DXA T-score (0.66) (Table 4). Excluding participants with baseline VF or osteoporotic fracture history, VHU maintained moderate discrimination for participants with incident fracture (AUC, 0.75). VHU showed no statistical difference compared to other methods; however, non-inferiority of VHU against DXA T-scores in incident fracture discrimination could not be concluded (P > .05).
Correlation between methods and inter-method reliability
VHU correlated strongly with average QCT-BMD (R = 0.89 [95% CI, 0.84-0.93, P < .05) but poorly with spine DXA-BMD (R = 0.42 [95% CI, 0.23-0.58, P < .05). Inter-method reliability between VHU and DXA and between QCT and DXA was fair (ĸ = 0.24, P < .05, and ĸ = 0.39, P < .05, respectively).
Decision curve analyses
VHU showed greater net benefit compared to DXA T-score, FRAX+T-score and Garvan+Tscore between threshold probabilities of 15% to 28% for incident VF (Figure 5).
Discussion
This study confirmed the high prevalence (69%) of low BMD measured by DXA (spine or hip) among LCS participants. When restricted to spine DXA T-score, low BMD prevalence was similar when using VHU (48% vs 48%, respectively). Although VHU correlation with standard DXA lay in the lower range of published values (r = 0.399-0.891),13 VHU had moderate discrimination for low DXA-BMD (AUC, 0.74) and DXA-defined osteoporosis (AUC, 0.63).
Most importantly, VHU identified screenees with prevalent VF (27% of participants) and with recent incident osteoporotic fractures as comparably as DXA, QCT, and clinical risk scores (AUC, 0.71 and 0.71 respectively).
Osteoporosis prevalence in LCS has been previously reported. One study (n = 457) reported a 73% prevalence of low spine QCT-BMD in LCS participants.11 This higher estimate was likely due to the inclusion of lower lumbar vertebrae (L2 and L3), which have lower BMD than thoracic vertebrae, thus lowering the average BMD. Similar to our findings, another LCS study (n = 302) reported 71% prevalence of low hip or spine DXA-BMD and a 43% prevalence of low spine VHU among screenees.28
While those with low BMD were predominantly female, a higher proportion of men had prevalent low-trauma VF. In addition to the small sample size, selection bias may explain this finding. LCS participants have a different baseline fracture risk compared with the general population, owing to significant cigarette smoke exposure and potentially higher corticosteroid exposure for the treatment of smoking-related airways disease.
Additionally, women may be more likely to have had prior bone screening and initiation of bisphosphonate therapy for low BMD, which may reduce their baseline fracture risk.
Previous studies in LCS and other populations reported similar AUCs for L1-VHU discrimination of DXA-defined osteoporosis (between 0.6728 and 0.839). VHU showed similar discrimination for low spine DXA-BMD (AUC, 0.78) in spinal surgery patients, compared to our estimate (AUC, 0.72).29 Smoke exposure and aging in LCS participants contribute to degenerative bone disease and vascular calcification, affecting DXA estimates and may explain the low correlation and discrimination estimates. VHU was, unsurprisingly, non-inferior to QCT, a CT-based method unaffected by these confounders, in low DXA-BMD discrimination.
Fracture prevention is the goal for patients and clinicians. Low BMD is an important proxy for fracture risk, although it is but one of several risk factors. Prevalent VF is independently associated with subsequent fracture risk.30 We found that VHU identified screenees with low BMD, existing fracture and recent osteoporotic fracture, on par with other standard methods.
Additionally, VHU has further advantages. It is comparatively time-saving, does not expose participants to additional radiation and does not require complex training or software, a limitation we noted as 6% of QCT analyses failed in our study. Therefore, in a high-throughput LCS setting, the moderate accuracy of VHU to identify low BMD and fractures may be an acceptable tradeoff, considering these other benefits.
Indeed, the greater net benefit of VHU compared to DXA and clinical risk scores appears to support this.
Few studies have compared VHU to other imaging and clinical risk tools in incident fracture prediction. A large retrospective study (n = 507) reported that higher L1-VHU was associated with decreased fracture risk (HR, 0.63 for 10-unit change in VHU), but did not compare with other methods.31 Two meta-analyses compared different imaging modalities in discriminating vertebral fragility fracture status and osteoporotic fracture prediction. One reported similar ability of 4 methods to classify patients with vertebral fragility fractures, although the pooled QCT AUC was higher than that for HU values (0.88 and 0.76, respectively).32 The second reported no significant difference in predicting osteoporotic fractures between HU and DXA (pooled AUCs, 0.73 and 0.73, respectively).33 This is similar to our findings regarding VHU ability to discriminate low BMD, prevalent, and incident VF.
Our results extend this literature to include clinical risk prediction calculators and show similar statistical outcomes.
Although the literature reports encouraging statistical metrics for VHU compared with the reference-standard DXA, the use of VHU as a diagnostic or treatment decision tool to improve fracture outcomes has yet to be studied. Treating DXA-defined osteoporosis with bisphosphonate therapy reduces fracture incidence in general populations and improves BMD.34 Theoretically, VHU might show increases in response to bone-strengthening therapies and lead to improved fracture outcomes. Further work is needed to define the role of VHU in the diagnostic process, either to guide therapy or to prompt further evaluation using standard methods.
Population-based screening for osteoporosis is not recommended due to paucity of evidence for fracture reduction.23 However, in LCS, case-finding of individuals, who may benefit from further investigation or treatment is recommended, and VHU could be an effective tool.14, 35
A limitation of this study is the cohort size. The sample size was selected primarily to assess non-inferiority between VHU and DXA in low BMD assessment rather than fracture prediction. A selected non-inferiority margin of 0.1 lies within the range of other diagnostic imaging studies (0.05 to 0.1).
However, due to missing participant responses (21.9%) and the low fracture incidence (n = 13), conclusions regarding non-inferiority between VHU and other methods in fracture prediction are limited. Our findings support the potential inclusion of VHU in fracture prediction tools, although this requires larger-scale validation.
The absence of agreed upon VHU cutoff values to classify low BMD limits its clinical application. Appropriate thresholds can be influenced by both the target population and CT parameters, which vary across different VHU studies. However, LCS programs internationally use convergent CT acquisition parameters, potentially mitigating this source of variability and bias.
Using the Youden index to equally maximize sensitivity and specificity, we determined optimal thresholds of 117 HU to detect low BMD and 88 HU to detect osteoporosis in this cohort. As LCS participants are at high risk of low BMD and osteoporotic fractures,36 we considered a higher threshold (126 HU) proposed in a prior study, to have better sensitivity for low BMD detection in screenees.16 With the adoption of standardized LCS CT protocols globally, and the future potential to implement phantom use to calibrate scanners for lung nodule assessment, VHU thresholds could potentially be standardized.14, 37 Further automation of VHU may minimize inter-scanner and inter-reader variability, in addition to radiologists’ workload.38, 39
Though L1 is the desired site for VHU measurement based on previous studies, it was not consistently imaged on CT in this study. VHU was measured from T11, T12 and L1. Spinal BMD reduces caudally; however, we found that VHU from T12 and L1 had comparable discrimination for DXA-defined osteoporosis and osteopenia.12 As mentioned, vertebral-specific optimal thresholds could be determined from larger cohorts to accommodate this.
In conclusion, VHU as an opportunistic BMD assessment tool in the LCS setting may improve the detection of DXA-defined osteoporotic disease in screenees. The simplicity of performing VHU and its similar discriminative ability for both prevalent VF and incident osteoporotic fracture, compared to standard DXA and fracture risk assessment methods, renders it a favorable parameter for use in fracture prediction tools. Compared to DXA and QCT, the VHU method is accessible to non-radiologists, including physicians or general practitioners who are commonly involved in osteoporosis management.
Large prospective studies, involving assessment of VHU in treatment decisions and associated fracture outcomes in individuals who smoke, are warranted to support its clinical implementation in this cohort.