The association between bone density evaluated on CT or MRI versus the gold standard dual-energy X-ray absorptiometry (DEXA) in the lumbar spine and hip: a meta-analysis Authors: Luc W.F. van Haaster, Erem Tutal, Rania A. Mekary, Carmen L.A. Vleggeert-Lankamp Published: 2026 ABSTRACT IntroductionCT-based Hounsfield units (HU) and MRI-based vertebral bone quality (VBQ)-scores are opportunistic markers of bone mineral density and potential alternatives for dual-energy X-ray absorptiometry (DEXA).Research questionWhat is the correlation between HU values and VBQ-scores in relation to DEXA, and what is the corresponding area-under-the-curve (AUC)-based diagnostic discrimination to detect osteopenia and osteoporosis?Material and methodsPubMed, Embase, Web of Science, Cochrane Library, and CINAHL were searched through August 2025 for studies reporting correlations between DEXA and CT and/or MRI or the AUC-based diagnostic discrimination of CT and/or MRI to detect osteopenia and osteoporosis. Random effects models were used to calculate pooled correlation coefficients and AUC values along with 95% confidence intervals.ResultsForty-three studies, including 8222 patients, were analyzed. HU demonstrated stronger correlations with DEXA T-scores than the VBQ-score (maximum correlation coefficient: 0.741 (95% CI: 0.578-0.847) vs. −0.476 (95% CI: −0.548 to −0.398) and greater AUC-based diagnostic discrimination for osteopenia and osteoporosis (maximum AUC: 0.875 (95% CI: 0.841-0.909) vs. 0.806 (95% CI: 0.684-0.929) across all vertebrae and DEXA sites. Correlation and AUC-based diagnostic discrimination of HU were consistent across vertebrae, with slightly better results at L2 and the mean of L1–L4. VBQ-score correlations were highest at L1 and gradually decreased from L1 to L4.Discussion and conclusionWhen CT is available, HU appears to be the better-supported opportunistic marker for low BMD than VBQ. All vertebrae demonstrated comparable correlation and AUC-based diagnostic discrimination, with L2 HU and the mean of L1-L4 HU showing slightly superior results. Introduction Osteoporosis is a systemic bone disorder characterized by low bone mass and deteriorated microarchitecture of the trabecular bone (Akkawi and Zmerly, 2018; Föger-Samwald et al., 2020). The prevalence of osteoporosis and the accompanying financial burden are expected to further increase in the future, due to an ageing population (Hernlund et al., 2013; Rachner et al., 2011). Preoperative detection and treatment of osteoporosis are important in spinal surgery patients, as low bone mineral density (BMD) is associated with complications such as pedicle screw loosening and cage subsidence, which can compromise surgical outcomes and necessitate revision surgery (Galbusera et al., 2015; Cho et al., 2018; Khalid et al., 2020). Despite the clinical relevance and rising prevalence, a recent Dutch study suggested that a significant number of patients with osteoporosis remains undiagnosed, estimating a treatment gap of 60% to 72% (Dunnewind et al., 2017). The current gold standard for BMD assessment is dual-energy X-ray absorptiometry (DEXA) measured at the hip or femur (No authors listed. World Health Organization, 2003). Additionally, DEXA scans derived from the vertebra of the lumbar spine are more commonly performed (Akkawi and Zmerly, 2018; Föger-Samwald et al., 2020; Hernlund et al., 2013; Rachner et al., 2011; Pennington et al., 2021). According to the World Health Organization (WHO), a T-score ≥ −1 indicates normal bone density, a T-score < −1 and > −2.5 indicates the presence of osteopenia, and a T-score ≤ −2.5 indicates the presence of osteoporosis (No authors listed. World Health Organization, 2003). However, increasing evidence suggests that in vivo assessment of bone density using DEXA is inaccurate. Factors such as osteophytes, vascular calcifications, scoliosis, and excessive fat deposits can lead to overestimation of bone density, while variability in the absorptiometry measurement site further limits the generalizability of the results (Liu et al., 1997; Yu et al., 2012; Rand et al., 1997; Pappou et al., 2006). Furthermore, although DEXA is considered to be the gold standard for the diagnosis of osteoporosis, DEXA remains underutilized among eligible patients (Amarnath et al., 2015; Gillespie and Morin, 2017). Alternative methods to assess the BMD of the lumbar spine have been proposed, such as computed tomography (CT)-based Hounsfield units (HU) and magnetic resonance imaging (MRI)-based vertebral bone quality (VBQ) score (Ahmad et al., 2023; Deshpande et al., 2023; Zhu et al., 2021a; Pinto et al., 2022; Zaidi et al., 2019). One major advantage of these imaging modalities is that CT- and MRI-scans are commonly obtained during the preoperative assessment and can therefore be used opportunistically, as compared to DEXA. HU represents a standardized scale used in CT imaging to quantify the relative radiodensity of tissues and can therefore be used as a measure of bone density (Schreiber et al., 2011; Hendrickson et al., 2018). The VBQ-score assesses the amount of fat infiltration in the vertebral body through T1-weighted MR images of the lumbar spine (Salzmann et al., 2022; Ehresman et al., 2020). The attenuation (i.e., BMD-indicative) values derived from these opportunistic imaging modalities have been shown to correlate with DEXA, but varying sites of measurement and variability in measurement methods are likely to be responsible for the spread in reported correlation coefficients and diagnostic accuracy (ROC-analysis) (Ahmad et al., 2023; Deshpande et al., 2023; Zhu et al., 2021a; Pinto et al., 2022; Zaidi et al., 2019; Chen et al., 2023). Existing meta-analyses comparing BMD measured with DEXA scans and opportunistic methods do not differentiate between specific DEXA-measurement sites or methods and measurement sites (varying lumbar vertebra) of CT/MRI assessments (Zhu et al., 2021b; Ahern et al., 2021). Given that bone density varies across the body and along the vertebral column, it is reasonable to expect that results may differ for each vertebral body and measurement site. This meta-analysis aimed to provide a better understanding of the correlation and area-under-the-curve (AUC)-based diagnostic discrimination of CT-based Hounsfield units and MRI-based VBQ-scores in assessing bone mineral density relative to the current gold standard, DEXA. In addition, the optimal vertebral level for opportunistic assessment of BMD was evaluated. Ultimately, this could support improved screening and diagnosis of osteoporosis, especially in patients undergoing preoperative imaging, enhancing the overall quality of spinal care. Material and methods Search methods for the identification of studies The PRISMA 2020 27-item checklist was used as the reporting guideline for this review (Page et al., 2021). A systematic search was conducted in five databases of medical literature: PubMed, Embase, Web of Science, Cochrane Library, and CINAHL. The search spanned from inception to August 2025 (Appendix A). Measurement methods DEXA In clinical practice, DEXA measurements are most commonly performed at the femoral neck or total hip and are increasingly obtained at the lumbar spine. The resulting individual T-scores can be used for BMD assessment, but often the lowest T-score is used among the assessed regions. CT-based HU measurement Several methods exist for measuring Hounsfield units. Measurements are typically obtained at L1 to L4, and can be performed in the midaxial, midsagittal, and midcoronal plane, with the midaxial plane being most commonly used. With regard to measurement methods, some approaches measure the entire midvertebral body, whereas others use a standardized circular ROI. As HU represents the quantified relative radiodensity of tissues, a higher HU value corresponds to higher bone density. MRI-based VBQ measurement The VBQ-score is measured exclusively in the midsagittal plane and most commonly from L1 to L4. As the VBQ-score represents the amount of fat infiltration in the vertebral body, a higher VBQ-score corresponds to lower bone density. Criteria for considering studies for this review This systematic review and meta-analysis included retrospective and prospective studies aiming to compare and/or correlate lumbar bone density assessment using conventional CT and/or MRI with DEXA measurements of the lumbar spine, hip, and/or femur. The following inclusion criteria were used: 1) patients aged 18 years or older, 2) sample size larger than 20 patients, 3) conventional CT/MRI protocols used for measuring respectively lumbar HU values and VBQ-scores, 4) DEXA T-score measurement in the lumbar spine, femur or hip, scored according to the WHO-definition of osteoporosis and 5) reporting either the correlation coefficient between CT/MRI attenuation values and DEXA T-scores and/or the receiver operating characteristic (ROC) analysis, including the AUC and corresponding 95% confidence interval (95% CI) or p-value (World Health Organization, 1994). Exclusion criteria included: 1) time between CT/MRI and DEXA measurements longer than 1 year, 2) CT/MRI measurements at non-lumbar spinal sites, such as cervical or thoracic spine, mandibula, tibia, fibula, radius or ulna, 3) study cohort containing patients with previous lumbar fusion, due to the inability to accurately measure CT/MRI attenuation values after lumbar fusion, 4) study cohort containing patients with systemic pathology possibly affecting bone quality (i.e. skeletal hyperostosis, severe adult spinal deformity, and malignancies), except for osteopenia and osteoporosis, 5) study cohort containing patients receiving medication specifically aiming to enhance bone quality, 6) use of contrast on radiological imaging, 7) biomechanical, animal or cadaver studies, 8) meeting abstracts, 9) case reports and 10) duplicate articles. Two researchers (LWFH and ET) independently screened the titles and abstracts of all retrieved records for eligibility, and disagreements were discussed with a third reviewer (CLAVL). Full text screening of remaining records was performed independently by both researchers, and discrepancies were discussed with the third reviewer. Assessment of risk-of-bias for the included studies Two researchers (LWFH and ET) independently assessed the risk-of-bias using the Quality Assessment Tool for Diagnostic Accuracy Studies-2 (QUADAS-2), and discrepancies were discussed with a third reviewer (CLAVL) (Whiting et al., 2011). This tool was used to evaluate the risk-of-bias and applicability of diagnostic accuracy studies based on four domains: patient selection, index test (i.e., CT or MRI), reference standard (i.e., DEXA scan) and flow and timing, resulting in low, high or unclear risk-of-bias for each domain and low, high, or unclear applicability concerns for the first three domains. In the patient selection domain, factors such as consecutive patient enrollment, avoidance of a case-control design, and the risk-of-bias associated with inappropriate exclusions were considered. In the flow and timing domain, various factors were evaluated, including whether all patients received the same reference standard and whether all patients were included in the analysis (Appendix B). While the original QUADAS-2 guidance recommends that studies should be at low risk in all domains to be considered high quality, we applied a slightly less stringent criterion, allowing studies with a maximum of one domain at high or unclear risk-of-bias to be classified as high quality, as adapted from Brunelli et al. (2023). Consequently, records identified as having a ‘’high’’ or ‘’unclear’’ risk-of-bias in at least three of the four domains were considered to be of low quality. Studies were deemed to be of high quality with a ‘’low’’ risk-of-bias assessment in at least three of the four domains. Other classification patterns in studies were regarded as moderate quality. The high/moderate/low quality classification is an adapted overall categorization and should not be interpreted as an original QUADAS-2 summary score. Data collection and analysis Two researchers (LWFH and ET) extracted the data from all included studies and entered the data into an Excel spreadsheet. The subsequent data were extracted: publication year, study design, country, number of patients, attenuation value protocol, location of CT/MRI- and DEXA-measurements, time between CT/MRI and DEXA, patient characteristics, attenuation values (HU value or VBQ-score), correlation parameters, ROC-analysis outcomes, and other inclusion and exclusion criteria. ROC-analysis is a statistical method used to assess the AUC-based diagnostic discrimination of a test by evaluating its ability to discriminate between different outcomes, such as normal BMD and low BMD, and is typically expressed through a curve that plots sensitivity against 1-specificity. When an article reported multiple measurements at the same vertebral level obtained in different imaging planes (e.g., axial, sagittal, and coronal), measurements from the axial plane were selected for analysis to avoid double-counting of patients. Furthermore, when multiple measurement methods were reported (e.g., region-of-interest and volume-of-interest approaches), the most conventional method (region-of-interest) was selected to enhance the generalizability of the findings and prevent double-counting patients. All extracted data were transferred to the Comprehensive Meta-Analysis version 4 software for data analysis. The pooled correlation coefficients and AUCs were determined using a random-effects model based on the DerSimonian and Laird method and were presented with the 95% CI (DerSimonian and Laird, 1986). The Cochrane Q statistic was used to evaluate the presence of heterogeneity with a significance level of p < 0.10 (Huedo-Medina et al., 2006). Heterogeneity was quantified using the I-squared (I2) index, reported as low (0-25%), moderate (26-50%), substantial (51-75%), or high (76-100%). To address heterogeneity, all analyses were stratified by the location of DEXA and the location of the CT/MRI attenuation value. Additionally, correlation analyses were stratified by correlation type (Pearson for normally distributed data vs. Spearman for not normally distributed data), and AUC-analyses were stratified by outcome type (osteopenia and/or osteoporosis). Egger's regression intercept was used to assess potential asymmetry in the funnel plot due to an alleged small-study effect. In case of any asymmetry in the funnel plot proven visually or statistically, the trim-and-fill method was applied to impute potentially missing studies and calculate the adjusted correlation coefficient, contingent on publication bias being the source of asymmetry. Univariate meta-regression on 1) time interval between CT and DEXA imaging, 2) percentage of osteoporotic patients, and 3) study quality was performed when at least ten studies were available for a single outcome to evaluate whether the correlation coefficient varied across the levels of these trial-level covariates. A two-sided p-value <0.05 was considered statistically significant, unless specified otherwise. Results Search and selection results The initial search yielded a total of 3035 records, of which 1676 duplicates were removed, resulting in 1359 articles eligible for title and abstract screening. A total of 1205 articles were excluded due to irrelevance to the research topic or failure to meet the predefined in- and exclusion criteria based on the abstract, after which full text screening was performed for the remaining 154 articles. Based on the full-text review, 111 articles were excluded due to failure to meet the predefined inclusion and exclusion criteria, resulting in the inclusion of 43 articles in the analysis (Hendrickson et al., 2018; Hiyama et al., 2024, 2025; Zhang et al., 2021, 2023; Wang et al., 2023, 2024; Kim et al., 2019, 2024; Saad et al., 2024; Li et al., 2024a, 2024b, 2025; Yang et al., 2022; Özmen et al., 2023; Viswanathan et al., 2023; Yao et al., 2022; Chang et al., 2021, 2025; Elarjani et al., 2021; Booz et al., 2020; Berger-Groch et al., 2020; Zou et al., 2019; Fujimoto et al., 2024; Zheng et al., 2024; Pu et al., 2023a, 2023b; Courtois et al., 2023; Choi et al., 2016; Xu et al., 2023; Lee et al., 2013, 2015; Vadera et al., 2023; Jameshouran et al., 2020; Nishi et al., 2025; Yurac-Barrientos et al., 2025; Pabbisetti et al., 2025; Su et al., 2025; Yan et al., 2025; Langer et al., 2025; Najafi et al., 2024; Cetik et al., 2025; Khoroushi et al., 2024). An overview of the identification and selection process is provided in Fig. 1.Fig. 1Flow diagram of the identification and selection of studies. Abbreviations: CT; computed tomography, MRI; magnetic resonance imaging. Flow diagram of the identification and selection of studies. Abbreviations: CT; computed tomography, MRI; magnetic resonance imaging. The 43 analyzed articles included 8222 patients over five prospective cohort studies (n = 417) and 38 retrospective cohort studies (n = 7805). The median of the reported mean ages was 65.8 years (range 48.8-73.2 years). For articles evaluating CT, the mean time between CT and DEXA was 95.2 days, whereas this was 93.4 days for articles examining MRI. An overview of all included articles is provided in Appendix C and Appendix D. Risk-of-bias assessment All 43 included articles were evaluated for risk-of-bias using the QUADAS-2 tool. Twenty articles were classified as high quality. Eighteen articles were of moderate quality. Five studies were categorized as low quality. For both the risk of bias assessment as the applicability concerns, the domain that most frequently raised concerns was patient selection. The results are illustrated in Table 1 and Fig. 2.Table 1Overview of risk-of-bias and applicability concern assessment, according to the domains of the QUADAS-2 tool, presented as review authors’ judgments about each domain for each included study. QUADAS-2 results were primarily interpreted at the domain level for risk of bias and applicability. The high/moderate/low quality classification was an adapted overall categorization and should not be interpreted as an original QUADAS-2 summary score. *Identical author-publication year combination, but separate articles.Table 1Fig. 2Overview of risk-of-bias and applicability concern assessment, according to the domains of the QUADAS-2 tool, presented as total percentage. Table 1Overview of risk-of-bias and applicability concern assessment, according to the domains of the QUADAS-2 tool, presented as review authors’ judgments about each domain for each included study. QUADAS-2 results were primarily interpreted at the domain level for risk of bias and applicability. The high/moderate/low quality classification was an adapted overall categorization and should not be interpreted as an original QUADAS-2 summary score. *Identical author-publication year combination, but separate articles.Table 1 Overview of risk-of-bias and applicability concern assessment, according to the domains of the QUADAS-2 tool, presented as review authors’ judgments about each domain for each included study. QUADAS-2 results were primarily interpreted at the domain level for risk of bias and applicability. The high/moderate/low quality classification was an adapted overall categorization and should not be interpreted as an original QUADAS-2 summary score. *Identical author-publication year combination, but separate articles. Overview of risk-of-bias and applicability concern assessment, according to the domains of the QUADAS-2 tool, presented as total percentage. Correlation outcomes Correlation between HU and lumbar T-score Sixteen studies (n = 3288) reported correlations between vertebral HU values and lumbar T-scores (Hiyama et al., 2024; Kim et al., 2024; Yang et al., 2022; Zhang et al., 2021; Elarjani et al., 2021; Kim et al., 2019; Zou et al., 2019; Choi et al., 2016; Xu et al., 2023; Pu et al., 2023b; Lee et al., 2015; Lee et al., 2013; Vadera et al., 2023; Nishi et al., 2025; Yurac-Barrientos et al., 2025; Cetik et al., 2025). Pooled Pearson's correlation coefficients (r) ranged from 0.551 (95% CI: 0.406-0.669) at L1-L4 to 0.617 (95% CI: 0.536-0.686) at L2, whereas this ranged from 0.629 (95% CI: 0.497-0.732) at L1 to 0.741 (95% CI: 0.578-0.847) at L1-L4 for pooled Spearman's rank correlation coefficients (ρ). Heterogeneity was high across all analyses (Table 2).Table 2Pooled correlation coefficients for different CT-based Hounsfield units measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman).Table 2Reference standardCT-based HU-measurementPearson or SpearmanNumber of studies |nPooled correlation coefficient (95% CI)I2p-heterogeneityLumbar T-scoreL1 HUPearson12 | 20710.592 (0.528-0.649)73.6%<0.001L2 HUPearson9 | 12910.617 (0.536-0.686)73.8%<0.001L3 HUPearson9 | 12910.593 (0.516-0.660)69.8%<0.001L4 HUPearson9 | 13290.572 (0.480-0.651)77.5%<0.001L1-L4 HUPearson4 | 4500.551 (0.406-0.669)70.9%0.01L1 HUSpearman3 | 9990.629 (0.497-0.732)77.5%0.01L2 HUSpearman2 | 1600.725 (0.557-0.835)86.5%0.01L3 HUSpearman2 | 1600.723 (0.569-0.827)80.7%0.02L4 HUSpearman2 | 1600.653 (0.445-0.794)85.7%0.01L1-L4 HUSpearman2 | 1600.741 (0.578-0.847)68.3%0.08Femoral neck T-scoreL1 HUPearson2 | 2250.371 (0.165-0.545)0.0%0.67L2 HUPearson2 | 2250.482 (0.369-0.581)0.0%0.64L3 HUPearson2 | 2250.402 (0.166-0.595)0.0%0.61L4 HUPearson2 | 2630.395 (0.128-0.609)0.0%0.32L1-L4 HUPearson1 | 1800.393 (0.143-0.596)NANAL1 HUSpearman2 | 1600.638 (0.482-0.755)73.3%0.05L2 HUSpearman2 | 1600.646 (0.541-0.730)46.2%0.17L3 HUSpearman2 | 1600.631 (0.447-0.764)79.5%0.03L4 HUSpearman2 | 1600.624 (0.411-0.772)81.7%0.02L1-L4 HUSpearman2 | 1600.675 (0.532-0.781)49.2%0.16Total hip T-scoreL1 HUPearson1 | 450.466 (0.200-0.668)NANAL2 HUPearson1 | 450.584 (0.351-0.749)NANAL3 HUPearson1 | 450.524 (0.272-0.709)NANAL4 HUPearson2 | 2630.491 (0.181-0.712)78.6%0.03Lowest T-scoreL1 HUPearson1 | 1910.532 (0.330-0.687)NANAL2 HUPearson1 | 1910.570 (0.327-0.742)NANAL3 HUPearson1 | 1910.563 (0.136-0.814)NANAL4 HUPearson1 | 1910.573 (0.139-0.822)NANAL1-L4 HUPearson1 | 2120.630 (0.292-0.828)NANAL1 HUSpearman3 | 6500.613 (0.507-0.701)57.8%0.09L2 HUSpearman2 | 1600.701 (0.550-0.808)58.2%0.12L3 HUSpearman2 | 1600.684 (0.430-0.837)81.1%0.02L4 HUSpearman2 | 1600.655 (0.380-0.824)81.9%0.02L1-L4 HUSpearman3 | 3220.647 (0.463-0.778)81.4%0.01 Table 2Pooled correlation coefficients for different CT-based Hounsfield units measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman).Table 2Reference standardCT-based HU-measurementPearson or SpearmanNumber of studies |nPooled correlation coefficient (95% CI)I2p-heterogeneityLumbar T-scoreL1 HUPearson12 | 20710.592 (0.528-0.649)73.6%<0.001L2 HUPearson9 | 12910.617 (0.536-0.686)73.8%<0.001L3 HUPearson9 | 12910.593 (0.516-0.660)69.8%<0.001L4 HUPearson9 | 13290.572 (0.480-0.651)77.5%<0.001L1-L4 HUPearson4 | 4500.551 (0.406-0.669)70.9%0.01L1 HUSpearman3 | 9990.629 (0.497-0.732)77.5%0.01L2 HUSpearman2 | 1600.725 (0.557-0.835)86.5%0.01L3 HUSpearman2 | 1600.723 (0.569-0.827)80.7%0.02L4 HUSpearman2 | 1600.653 (0.445-0.794)85.7%0.01L1-L4 HUSpearman2 | 1600.741 (0.578-0.847)68.3%0.08Femoral neck T-scoreL1 HUPearson2 | 2250.371 (0.165-0.545)0.0%0.67L2 HUPearson2 | 2250.482 (0.369-0.581)0.0%0.64L3 HUPearson2 | 2250.402 (0.166-0.595)0.0%0.61L4 HUPearson2 | 2630.395 (0.128-0.609)0.0%0.32L1-L4 HUPearson1 | 1800.393 (0.143-0.596)NANAL1 HUSpearman2 | 1600.638 (0.482-0.755)73.3%0.05L2 HUSpearman2 | 1600.646 (0.541-0.730)46.2%0.17L3 HUSpearman2 | 1600.631 (0.447-0.764)79.5%0.03L4 HUSpearman2 | 1600.624 (0.411-0.772)81.7%0.02L1-L4 HUSpearman2 | 1600.675 (0.532-0.781)49.2%0.16Total hip T-scoreL1 HUPearson1 | 450.466 (0.200-0.668)NANAL2 HUPearson1 | 450.584 (0.351-0.749)NANAL3 HUPearson1 | 450.524 (0.272-0.709)NANAL4 HUPearson2 | 2630.491 (0.181-0.712)78.6%0.03Lowest T-scoreL1 HUPearson1 | 1910.532 (0.330-0.687)NANAL2 HUPearson1 | 1910.570 (0.327-0.742)NANAL3 HUPearson1 | 1910.563 (0.136-0.814)NANAL4 HUPearson1 | 1910.573 (0.139-0.822)NANAL1-L4 HUPearson1 | 2120.630 (0.292-0.828)NANAL1 HUSpearman3 | 6500.613 (0.507-0.701)57.8%0.09L2 HUSpearman2 | 1600.701 (0.550-0.808)58.2%0.12L3 HUSpearman2 | 1600.684 (0.430-0.837)81.1%0.02L4 HUSpearman2 | 1600.655 (0.380-0.824)81.9%0.02L1-L4 HUSpearman3 | 3220.647 (0.463-0.778)81.4%0.01 Pooled correlation coefficients for different CT-based Hounsfield units measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman). Correlation HU and femoral neck T-score Five studies (n = 603) reported correlations between vertebral HU values and femoral neck T-scores (Hiyama et al., 2024; Kim et al., 2019; Pu et al., 2023b; Nishi et al., 2025; Cetik et al., 2025). Pooled Pearson's r ranged from 0.371 (95% CI: 0.165-0.545) at L1 to 0.482 (95% CI: 0.369-0.581) at L2, while overall Spearman's ρ ranged from 0.624 (95% CI: 0.411-0.772) at L4 to 0.675 (95% CI: 0.532-0.781) at L1-L4 (Table 2). Correlation HU and total hip T-score Two studies (n = 263) reported correlations between vertebral HU values and total hip T-scores (Nishi et al., 2025; Cetik et al., 2025). Overall Pearson's r ranged from 0.466 (95% CI: 0.200-0.668) at L1 to 0.584 (95% CI: 0.351-0.749) at L2. Spearman's ρ and the correlation between the mean lumbar HU value and total hip T-score were not reported (Table 2). Correlation HU and lowest T-score Six studies (n = 1216) reported correlations between vertebral HU values and lowest T-scores (Hiyama et al., 2024; Pu et al., 2023b; Chang et al., 2025; Langer et al., 2025; Li et al., 2024b; Hiyama et al., 2025). Pooled Pearson's r ranged from 0.532 (95% CI: 0.330-0.687) at L1 to 0.630 (95% CI: 0.292-0.828) at L1-L4, whereas this ranged from 0.613 (95% CI: 0.507-0.701) at L1 to 0.701 (95% CI: 0.550-0.808) at L2 for pooled Spearman's ρ (Table 2). Correlation between VBQ-score and lumbar T-score Eight studies (n = 865) reported correlations between VBQ-scores and lumbar T-scores (Hiyama et al., 2024; Li et al., 2024a; Özmen et al., 2023; Chang et al., 2021; Pu et al., 2023a; Xu et al., 2023; Pabbisetti et al., 2025; Najafi et al., 2024). Pooled Pearson's r ranged from −0.340 (95% CI: −0.453 to −0.216) at L4 to −0.460 (95% CI: −0.594 to −0.302) at L1, whereas the pooled Spearman's ρ was −0.166 (95% CI: −0.440 to 0.136) at L1-L4. Spearman's correlations were not evaluated for individual vertebrae (Table 3).Table 3Pooled correlation coefficients for different vertebral bone quality score measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman).Table 3Reference standardMRI-based VBQ-score measurementPearson or SpearmanNumber of studies |nPooled correlation coefficient (95% CI)I2p-heterogeneityLumbar T-scoreL1 VBQPearson2 | 218−0.460 (−0.594 to −0.302)47.4%0.17L2 VBQPearson2 | 218−0.442 (−0.575 to −0.285)44.3%0.18L3 VBQPearson2 | 218−0.341 (−0.515 to −0.139)60.6%0.11L4 VBQPearson2 | 218−0.340 (−0.453 to −0.216)0.0%0.84L1-L4 VBQPearson7 | 765−0.408 (−0.496 to −0.312)55.8%0.03L1-L4 VBQSpearman1 | 100−0.166 (−0.440 to 0.136)NANAFemoral neck T-scoreL1 VBQPearson3 | 658−0.465 (−0.523 to −0.402)0.0%0.80L2 VBQPearson3 | 658−0.441 (−0.501 to −0.377)0.0%0.79L3 VBQPearson3 | 658−0.422 (−0.483 to −0.357)0.0%0.99L4 VBQPearson3 | 658−0.411 (−0.473 to −0.346)0.0%0.91L1-L4 VBQPearson4 | 388−0.361 (−0.506 to −0.197)48.0%0.12L1-L4 VBQSpearman2 | 244−0.333 (−0.507 to −0.132)76.6%0.01Total hip T-scoreL1-L4 VBQPearson2 | 231−0.265 (−0.386 to −0.136)0.0%0.53L1-L4 VBQSpearman1 | 144−0.408 (−0.539 to −0.257)NANALowest T-scoreL1-L4 VBQPearson7 | 1148−0.476 (−0.548 to −0.398)45.9%0.07L1-L4 VBQSpearman2 | 244−0.315 (−0.467 to −0.146)78.3%0.01 Table 3Pooled correlation coefficients for different vertebral bone quality score measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman).Table 3Reference standardMRI-based VBQ-score measurementPearson or SpearmanNumber of studies |nPooled correlation coefficient (95% CI)I2p-heterogeneityLumbar T-scoreL1 VBQPearson2 | 218−0.460 (−0.594 to −0.302)47.4%0.17L2 VBQPearson2 | 218−0.442 (−0.575 to −0.285)44.3%0.18L3 VBQPearson2 | 218−0.341 (−0.515 to −0.139)60.6%0.11L4 VBQPearson2 | 218−0.340 (−0.453 to −0.216)0.0%0.84L1-L4 VBQPearson7 | 765−0.408 (−0.496 to −0.312)55.8%0.03L1-L4 VBQSpearman1 | 100−0.166 (−0.440 to 0.136)NANAFemoral neck T-scoreL1 VBQPearson3 | 658−0.465 (−0.523 to −0.402)0.0%0.80L2 VBQPearson3 | 658−0.441 (−0.501 to −0.377)0.0%0.79L3 VBQPearson3 | 658−0.422 (−0.483 to −0.357)0.0%0.99L4 VBQPearson3 | 658−0.411 (−0.473 to −0.346)0.0%0.91L1-L4 VBQPearson4 | 388−0.361 (−0.506 to −0.197)48.0%0.12L1-L4 VBQSpearman2 | 244−0.333 (−0.507 to −0.132)76.6%0.01Total hip T-scoreL1-L4 VBQPearson2 | 231−0.265 (−0.386 to −0.136)0.0%0.53L1-L4 VBQSpearman1 | 144−0.408 (−0.539 to −0.257)NANALowest T-scoreL1-L4 VBQPearson7 | 1148−0.476 (−0.548 to −0.398)45.9%0.07L1-L4 VBQSpearman2 | 244−0.315 (−0.467 to −0.146)78.3%0.01 Pooled correlation coefficients for different vertebral bone quality score measurements and DEXA T-scores, stratified by correlation type (Pearson vs. Spearman). Correlation of the VBQ-score and femoral neck T-score Seven studies (n = 1133) reported correlations between VBQ-scores and femoral neck T-scores (Hiyama et al., 2024; Li et al., 2024a; Özmen et al., 2023; Zheng et al., 2024; Pu et al., 2023a; Su et al., 2025; Wang et al., 2024). Pooled Pearson's r ranged from −0.361 (95% CI: −0.506 to −0.197) at L1-L4 to −0.465 (95% CI: −0.523 to −0.402) at L1, whereas the pooled Spearman's ρ was −0.333 (95% CI: −0.507 to −0.132) at L1-L4. Spearman's correlations were not reported for individual vertebrae (Table 3). Correlation between VBQ-score and total hip T-score Three studies (n = 375) reported correlations between VBQ-scores and total hip T-scores (Li et al., 2024a; Özmen et al., 2023; Wang et al., 2024). The pooled Pearson's r was −0.265 (95% CI: −0.386 to −0.136) at L1-L4, whereas the pooled Spearman's ρ was −0.408 (95% CI: −0.539 to −0.257) at L1-L4. Correlation was not evaluated for individual vertebral levels (Table 3). Correlation between VBQ-score and lowest T-score Nine studies (n = 1392) reported correlations between VBQ-scores and lowest T-scores (Hiyama et al., 2024; Wang et al., 2023; Saad et al., 2024; Li et al., 2024a; Özmen et al., 2023; Pu et al., 2023a; Li et al., 2025; Wang et al., 2024; Li et al., 2024b). The pooled Pearson's r was −0.476 (95% CI: −0.548 to −0.398) at L1-L4, whereas the pooled Spearman's ρ was −0.315 (95% CI: −0.467 to −0.146) at L1-L4. Correlation was not evaluated for individual vertebral levels (Table 3). ROC-analyses outcomes There were no studies that reported the AUC for vertebral HU values or VBQ-scores to diagnose osteopenia and osteoporosis based on the total hip T-score as the reference standard. The range of reported HU cut-off values for diagnosing osteoporosis was more restricted when based on the lowest T-score, compared to the lumbar T-score (Fig. 3). Across studies, the optimal HU cut-off values, defined as the threshold with the highest combined sensitivity and specificity, varied according to the reference standard and vertebral level assessed. Using the lumbar T-score as the reference standard, the optimal cut-offs were 180 HU at L1 (sensitivity 100.0%, specificity 91.5%), 101 HU at L2 (75.0%, 77.8%), 199 HU at L3 (92.3%, 79.6%), 193 HU at L4 (81.8%, 83.2%), and 88 HU for the average L1–L4 measurement (87.5%, 85.5%). Using the femoral neck T-score as the reference standard, the optimal cut-off for L1–L4 was 99 HU (81.2%, 71.0%). When the lowest T-score was used as the reference standard, the optimal cut-offs were 160 HU at L1 (93.6%, 86.9%), 92 HU at L2 (94.1%, 74.4%), 84 HU at L3 (94.1%, 76.7%), 79 HU at L4 (82.4%, 88.4%), and 92 HU for L1–L4 (94.1%, 74.4%). These findings demonstrate substantial variability in proposed HU thresholds across studies and reference standards. An overview of HU cut-off values to diagnose osteopenia and osteoporosis is provided in Appendix E.Fig. 3HU cut-off ranges for the diagnosis of osteoporosis, separated by vertebral level and based on lowest T-score and Lumbar T-score as DEXA reference sites. HU cut-off ranges for the diagnosis of osteoporosis, separated by vertebral level and based on lowest T-score and Lumbar T-score as DEXA reference sites. ROC-analysis HU and lumbar T-score Twelve studies (n = 3023) reported the AUC-based diagnostic discrimination of vertebral HU values using the lumbar T-score as the reference standard (Hendrickson et al., 2018; Hiyama et al., 2024; Zhang et al., 2023; Yang et al., 2022; Booz et al., 2020; Kim et al., 2019; Zou et al., 2019; Fujimoto et al., 2024; Xu et al., 2023; Lee et al., 2015; Vadera et al., 2023; Yurac-Barrientos et al., 2025). For the combined diagnosis of osteopenia and osteoporosis, pooled AUC-values ranged from 0.694 (95% CI: 0.547-0.841) to 0.802 (95% CI: 0.672-0.931), whereas for the diagnosis of osteoporosis alone, pooled AUC-values ranged from 0.807 (95% CI: 0.754-0.860) to 0.853 (95% CI: 0.815-0.891) (Table 4).Table 4Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using Hounsfield units, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard.Table 4Reference standardCT-based HU-measurementOutcomeNumber of studies |nPooled AUC (95% CI)I2p-heterogeneityLumbar T-scoreL1 HUOsteopenia + osteoporosis2 | 9830.802 (0.672-0.931)0.0%0.56L1-L4 HUOsteopenia + osteoporosis1 | 530.694 (0.547-0.841)NANAL1 HUOsteoporosis7 | 24160.834 (0.761-0.907)94.2%<0.001L2 HUOsteoporosis3 | 6450.853 (0.815-0.891)0.0%0.68L3 HUOsteoporosis3 | 6450.821 (0.735-0.908)70.2%0.04L4 HUOsteoporosis3 | 6450.807 (0.754-0.860)0.0%0.62L1-L4 HUOsteoporosis5 | 6070.845 (0.787-0.902)58.6%0.05Femoral neck T-scoreL1-L4 HUOsteoporosis2 | 2800.773 (0.736-0.810)0.0%0.51Lowest T-scoreL1 HUOsteopenia + osteoporosis3 | 6720.819 (0.683-0.954)35.8%0.21L2 HUOsteopenia + osteoporosis1 | 870.750 (0.592-0.908)NANAL3 HUOsteopenia + osteoporosis1 | 870.670 (0.512-0.828)NANAL4 HUOsteopenia + osteoporosis1 | 500.730 (0.599-0.861)NANAL1-L4 HUOsteopenia + osteoporosis1 | 2120.870 (0.790-0.950)NANAL1 HUOsteoporosis6 | 10210.831 (0.735-0.926)94.4%<0.001L2 HUOsteoporosis3 | 3380.822 (0.764-0.881)0.0%0.68L3 HUOsteoporosis3 | 3380.829 (0.780-0.877)0.0%0.75L4 HUOsteoporosis3 | 3380.859 (0.798-0.920)27.1%0.25L1-L4 HUOsteoporosis3 | 3510.875 (0.841-0.909)0.0%0.37 Table 4Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using Hounsfield units, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard.Table 4Reference standardCT-based HU-measurementOutcomeNumber of studies |nPooled AUC (95% CI)I2p-heterogeneityLumbar T-scoreL1 HUOsteopenia + osteoporosis2 | 9830.802 (0.672-0.931)0.0%0.56L1-L4 HUOsteopenia + osteoporosis1 | 530.694 (0.547-0.841)NANAL1 HUOsteoporosis7 | 24160.834 (0.761-0.907)94.2%<0.001L2 HUOsteoporosis3 | 6450.853 (0.815-0.891)0.0%0.68L3 HUOsteoporosis3 | 6450.821 (0.735-0.908)70.2%0.04L4 HUOsteoporosis3 | 6450.807 (0.754-0.860)0.0%0.62L1-L4 HUOsteoporosis5 | 6070.845 (0.787-0.902)58.6%0.05Femoral neck T-scoreL1-L4 HUOsteoporosis2 | 2800.773 (0.736-0.810)0.0%0.51Lowest T-scoreL1 HUOsteopenia + osteoporosis3 | 6720.819 (0.683-0.954)35.8%0.21L2 HUOsteopenia + osteoporosis1 | 870.750 (0.592-0.908)NANAL3 HUOsteopenia + osteoporosis1 | 870.670 (0.512-0.828)NANAL4 HUOsteopenia + osteoporosis1 | 500.730 (0.599-0.861)NANAL1-L4 HUOsteopenia + osteoporosis1 | 2120.870 (0.790-0.950)NANAL1 HUOsteoporosis6 | 10210.831 (0.735-0.926)94.4%<0.001L2 HUOsteoporosis3 | 3380.822 (0.764-0.881)0.0%0.68L3 HUOsteoporosis3 | 3380.829 (0.780-0.877)0.0%0.75L4 HUOsteoporosis3 | 3380.859 (0.798-0.920)27.1%0.25L1-L4 HUOsteoporosis3 | 3510.875 (0.841-0.909)0.0%0.37 Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using Hounsfield units, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard. ROC-analysis HU and femoral neck T-score Two studies (n = 280) reported the AUC-based diagnostic discrimination of the mean lumbar HU values for diagnosing osteoporosis based on the femoral neck T-score as the reference standard, for which the pooled AUC was 0.773 (95% CI: 0.736-0.810) (Table 4) (Hiyama et al., 2024; Kim et al., 2019). ROC-analysis HU and lowest T-score Eleven studies (n = 1626) reported the AUC-based diagnostic discrimination of vertebral HU values using the lowest T-score as the reference standard (Hiyama et al., 2024; Kim et al., 2024; Viswanathan et al., 2023; Yao et al., 2022; Berger-Groch et al., 2020; Pu et al., 2023b; Jameshouran et al., 2020; Chang et al., 2025; Langer et al., 2025; Li et al., 2024b; Khoroushi et al., 2024). For the combined diagnosis of osteopenia and osteoporosis, pooled AUC-values ranged from 0.670 (95% CI: 0.512-0.828) to 0.870 (95% CI: 0.790-0.950), whereas for the diagnosis of osteoporosis alone, pooled AUC-values ranged from 0.822 (95% CI: 0.764-0.881) to 0.875 (95% CI: 0.841-0.909) (Table 4). ROC-analysis VBQ-score and lumbar T-score Three studies (n = 314) evaluated the AUC-based diagnostic discrimination of L1-L4 VBQ-scores using the lumbar T-score as the reference standard (Xu et al., 2023; Pabbisetti et al., 2025; Langer et al., 2025). The pooled AUC was 0.612 (95% CI: 0.460-0.764) for the combined diagnosis of osteopenia and osteoporosis and was 0.806 (95% CI: 0.684-0.929) for osteoporosis alone (Table 5).Table 5Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using the vertebral bone quality score, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard.Table 5Reference standardMRI-based VBQ-score measurementOutcomeNumber of studies |nPooled AUC (95% CI)I2p-heterogeneityLumbar T-scoreL1-L4 VBQOsteopenia + osteoporosis2 | 1960.612 (0.460-0.764)91.5%<0.001L1-L4 VBQOsteoporosis3 | 3140.806 (0.684-0.929)0.0%0.47Femoral neck T-scoreL1-L4 VBQOsteopenia + osteoporosis1 | 5010.774 (0.672-0.876)NANAL1-L4 VBQOsteoporosis2 | 7540.692 (0.613-0.772)68.8%0.07Lowest T-scoreL1-L4 VBQOsteopenia + osteoporosis3 | 5830.766 (0.675-0.856)0.0%0.65L1-L4 VBQOsteoporosis2 | 2560.726 (0.621-0.831)85.7%0.01 Table 5Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using the vertebral bone quality score, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard.Table 5Reference standardMRI-based VBQ-score measurementOutcomeNumber of studies |nPooled AUC (95% CI)I2p-heterogeneityLumbar T-scoreL1-L4 VBQOsteopenia + osteoporosis2 | 1960.612 (0.460-0.764)91.5%<0.001L1-L4 VBQOsteoporosis3 | 3140.806 (0.684-0.929)0.0%0.47Femoral neck T-scoreL1-L4 VBQOsteopenia + osteoporosis1 | 5010.774 (0.672-0.876)NANAL1-L4 VBQOsteoporosis2 | 7540.692 (0.613-0.772)68.8%0.07Lowest T-scoreL1-L4 VBQOsteopenia + osteoporosis3 | 5830.766 (0.675-0.856)0.0%0.65L1-L4 VBQOsteoporosis2 | 2560.726 (0.621-0.831)85.7%0.01 Pooled AUC-values for the diagnosis of osteopenia and osteoporosis using the vertebral bone quality score, based on different DEXA reference standards, stratified by outcome type. No studies used total hip T-score as the reference standard. ROC-analysis VBQ-score and femoral neck T-score Two studies (n = 754) reported the AUC-based diagnostic discrimination of L1-L4 VBQ-scores using the femoral neck T-score as the reference standard (Su et al., 2025; Yan et al., 2025). For the combined diagnosis of osteopenia and osteoporosis, the pooled AUC value was 0.774 (95% CI: 0.672-0.876), whereas this was 0.692 (95% CI: 0.613-0.772) for osteoporosis alone (Table 5). ROC-analysis VBQ-score and lowest T-score Four studies (n = 739) assessed the AUC-based diagnostic discrimination of L1-L4 VBQ-scores using the lowest T-score as the reference standard (Zhang et al., 2023; Wang et al., 2023; Pu et al., 2023a; Li et al., 2024b). The pooled AUC was 0.766 (0.675-0.856) for the combined diagnosis of osteopenia and osteoporosis and was 0.726 (95% CI: 0.621-0.831) for osteoporosis alone (Table 5). Small study effect Small-study effect was investigated for the Pearson correlation between L1 HU and lumbar T-score, as this was the only outcome with ten or more studies included. There was no evidence of small study effects based on Egger's regression intercept (p = 0.54). Nevertheless, based on visual assessment of asymmetry in the funnel plot, the trim-and-fill method was used to correct for potential publication bias (Fig. 4). Three potentially missing studies were imputed to the right of the pooled point estimate, causing an increase in Pearson's correlation coefficient from 0.592 (95% CI: 0.529-0.648) to 0.626 (95% CI: 0.558-0.686).Fig. 4Funnel plot of the standard error against the Fisher's Z value generated for the Pearson's correlation coefficient between L1 HU and lumbar T-score (12 included studies and three imputed potentially missing records). The vertical line represents the logit of the overall pooled correlation coefficient, and the two oblique lines depict the 95% confidence interval for a specific standard error. The trim-and-fill method was conducted to further diagnose potential publication bias. The clear diamond indicates the pooled correlation coefficient of all included studies as represented by the clear dots, while the black diamond indicates the adjusted correlation coefficient after imputing potentially missing records showing a stronger correlation, as represented by the black dots. Funnel plot of the standard error against the Fisher's Z value generated for the Pearson's correlation coefficient between L1 HU and lumbar T-score (12 included studies and three imputed potentially missing records). The vertical line represents the logit of the overall pooled correlation coefficient, and the two oblique lines depict the 95% confidence interval for a specific standard error. The trim-and-fill method was conducted to further diagnose potential publication bias. The clear diamond indicates the pooled correlation coefficient of all included studies as represented by the clear dots, while the black diamond indicates the adjusted correlation coefficient after imputing potentially missing records showing a stronger correlation, as represented by the black dots. Meta-regression of the pooled correlation coefficient by the time interval between CT and DEXA Meta-regression was performed only for the correlation between L1 HU and lumbar T-score, as fewer than ten studies were available for all other comparisons. Twelve studies including 2071 patients were evaluated, and the pooled Pearson's correlation coefficient showed no significant association with the time interval between CT and DEXA: slope: 0.0000 (95% CI: −0.0009-0.0008; p = 0.96). Meta-regression of the pooled correlation coefficient by study quality The same twelve studies were evaluated, and the pooled Pearson's correlation coefficient demonstrated no statistically significant association with study quality (p = 0.10). The regression formula is as follows: Intercept = 0.6666 + (−0.0768 x Low) + (0.1306 x Moderate) The intercept represents the predicted outcome for studies of high quality, which served as the reference category. Studies of low quality exhibited slightly weaker correlation, while studies of moderate quality demonstrated stronger correlation compared to high-quality studies. However, this difference was not statistically significant (p = 0.10). Meta-regression of pooled correlation coefficient by percentage of osteoporotic patients Seven studies (n = 1335) of the total group on which meta-regression could be performed reported the percentage of osteoporotic patients. A significant inverse association was observed between the pooled Pearson's correlation coefficient between L1 HU and lumbar T-score, and the percentage of osteoporotic patients: slope: −0.0066 (95% CI: −0.0106 to −0.0025; p = 0.002). This indicates that every 10 percent increase in the proportion of osteoporotic patients was associated with a decrease of 0.066 (6.6%) in the correlation coefficient. Discussion This meta-analysis aimed to provide insights into the correlation between CT-based HU-measurement, MRI-based VBQ score, and DEXA T-scores, as well as to evaluate the AUC-based diagnostic discrimination of both attenuation measures for detecting osteopenia and osteoporosis across different vertebral levels and according to various DEXA reference sites. HU values appeared to correlate more strongly with DEXA T-scores than VBQ-scores. Moreover, the AUC-based diagnostic discrimination, as reflected by the AUC-values, tended to be higher for HU measurements than for VBQ-scores. Consequently, when available, HU appears to be the better-supported opportunistic marker for low BMD compared to the VBQ-score. However, this should be interpreted in light of indirect comparisons, heterogeneous imaging protocols, and the absence of adjusted head-to-head evidence. Although correlation coefficients were not substantially different across vertebral levels, L2 HU and the mean L1-L4 HU demonstrated the strongest correlations with DEXA T-scores. In addition, correlations between HU and lumbar- and lowest T-scores were the strongest on average, compared to femoral neck- and total hip T-scores. The correlations between the VBQ-score and lumbar- and femoral neck T-scores were highest at L1 and progressively declined from L1 to L4. This can potentially be explained by the increase in lumbar spine fat fraction from L1 to L4 (Lee et al., 2024). Because the VBQ-score estimates the signal intensity based on the vertebral fat content, higher baseline fat in the lower vertebrae reduces the range for detecting differences in bone mineral density, thereby potentially weakening the correlation with DEXA T-scores. Future research could compare the diagnostic performances of L1 VBQ versus L1-L4 VBQ to determine the optimal location to diagnose low BMD. Finally, for HU measurements, Spearman's rank correlation coefficients were consistently higher than Pearson's correlation coefficients across all vertebral levels and DEXA reference sites. In contrast, this pattern was inconclusive for the VBQ-score, likely due to the limited number of studies reporting Spearman's correlations. The higher Spearman's rank correlation coefficients can be attributed to the sensitivity of Pearson's correlation to deviations from normality, including the presence of outliers. In clinical imaging data, extreme HU values may arise from degenerative changes or vertebral fractures, which can weaken Pearson's correlation. In contrast, Spearman's correlation is based on ranked data, making it distribution-free and less affected by outliers, and therefore provides a more reliable estimate of the relationship between HU and DEXA T-scores. The average L1-L4 HU demonstrated the highest AUC-based diagnostic discrimination for identifying osteoporosis alone, as well as for the combined diagnosis of osteopenia and osteoporosis based on the lowest T-score. However, for the diagnosis of osteoporosis based on the lumbar T-score, L2 HU showed the highest AUC-based diagnostic discrimination. It was not possible to determine which vertebral level is most suitable for the combined diagnosis of osteopenia and osteoporosis based on the lumbar T-score, as HU was only measured at L1 and L1-L4. The AUC-based diagnostic discrimination of the VBQ-score was only assessed for the mean of L1-L4 and not separately for individual vertebrae and showed comparable results across different DEXA measurement sites as reference. In addition, this meta-analysis demonstrated that the proportion of osteoporotic patients significantly influences the strength of the observed correlation with DEXA T-scores. According to the results of meta-regression, it can be concluded that for every 10% increase in the percentage of osteoporotic patients, the correlation coefficient decreases by 0.066 (6.6%). This finding should be considered when utilizing HU for BMD assessment. Strengths and limitations The authors acknowledge several limitations. First, while efforts were made to standardize the included studies as much as possible, such as excluding patients with systemic pathologies, those receiving BMD-enhancing medications, and studies utilizing intravenous contrast, there was still considerable variability in imaging systems, patient characteristics, and the indications for CT and MRI. These factors may have influenced the results. In addition, we relied on visual inspection to conclude that correlation and AUC-based diagnostic discrimination outcomes were higher for HU than for the VBQ-score, as any test comparing the two point estimates would be subject to confounding bias, due to the unadjusted nature of such comparison. Therefore, the hypothesis that HU outperforms the VBQ-score should be further researched in comparative studies with appropriate adjustment for confounding. Another limitation of this meta-analysis is that technical imaging parameters, such as acquisition and reconstruction settings, were not systematically extracted and analyzed across all included studies. Therefore, their potential contribution to the heterogeneity in HU values and VBQ scores could not be fully assessed. Finally, this meta-analysis did not assess the optimal cut-off values for diagnosing osteopenia and osteoporosis due to inconsistent reporting across studies. The lack of complete contingency data prevented pooling of results and calculation of summary diagnostic accuracy measures. Further research is necessary to address this, as the significant heterogeneity in diagnostic thresholds reported in the literature complicates broader clinical implementation of these opportunistic methods. However, several strengths are worth reporting. This meta-analysis represents, to the best of the authors' knowledge, the first comprehensive comparison of correlation outcomes and AUC-based diagnostic discrimination across different vertebrae and DEXA reference sites. The findings contribute to a better understanding of the utility of CT and MRI in the opportunistic measurement of BMD, particularly focusing on the preferred method and vertebra for assessment. Conclusions In conclusion, HU appeared to be the better-supported opportunistic marker for low BMD compared to the VBQ-score. With respect to HU, there was no substantial variation in correlation outcomes and diagnostic discrimination across different vertebrae. However, on average, HU values from L2 and the mean HU of L1-L4 exhibited marginally stronger correlations with DEXA T-scores and somewhat superior diagnostic discrimination in some analyses. The correlations between the VBQ-score and DEXA T-scores were strongest at L1 and progressively declined from L1 to L4. In clinical practice, these measures may assist in identifying patients at risk of low bone mineral density when dedicated DEXA assessment in unavailable. These findings support HU as a promising opportunistic marker when CT imaging is already available. However, clinical and technical heterogeneity, differences between DXA reference sites and indirect HU-VBQ comparisons restrict clinical interpretability. Standardized acquisition and measurement protocols, threshold-specific validation, and direct comparative studies are still needed before HU can be considered a validated substitute for DXA or before a universally applicable vertebral level, measurement site, or cut-off can be recommended. Authors’ contributions CLAVL and LWFH formulated the research question. CLAVL, RAM, and LWFH contributed to the development of the study design. LWFH and ET screened abstracts, performed full-text screening, and risk-of-bias assessment. LWFH and ET extracted the data. RAM and LWFH performed the data analysis. CLAVL and LWFH interpreted the results of the data analysis. LWFH prepared the first draft of the protocol paper, with contributions and editing from CLAVL, RAM and ET. All authors have read and approved the final manuscript. Funding sources This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Competing interests CLAVL: Thromboembolic complications in Neurosurgery (Covidien), Fundamentals of radiculopathy (YM Fund/Achmea health insurance and Eurospine), 5yr FU disc prosthesis in cervical spine surgery (CSRS E). Smoking in spondylodesis surgery (Achmea health insurance). Board of CSRS Europe, Netherlands Neurosurgical Society (NVvN) and Eurospine, Advisory Board for Rijndam Rehabilitation centre, faculty for EANS, CSRS, Eurospine spine courses. RAM, ET and LWFH declare that they have no competing interests to disclose.