Introduction
Various pulmonary diseases, such as chronic obstructive pulmonary disease and cystic fibrosis, along with glucocorticoid therapy for these conditions, elevate the risk of osteoporosis. Osteoporosis is a skeletal disorder characterized by diminished bone mass and a deterioration of bone microarchitecture (1, 2). Its most common and serious complication is fracture, which represents a major cause of disability and reduced mobility (3–5).
Consequently, the simultaneous screening and evaluation for osteoporosis in patients with pulmonary diseases is a critical component of clinical prevention and intervention (6, 7).
Although dual-energy x-ray absorptiometry (DXA) has been widely adopted for bone mineral density (BMD) measurement, its application requires considerable time; furthermore, its diagnostic utility is limited to assessing two-dimensional planar bone mineral density, and its results can be adversely affected by degenerative changes, osteophytes, aortic calcification, and other factors. Quantitative Computed Tomography (QCT) is increasingly being adopted as a screening tool for osteoporosis, because it enables volumetric analysis, provides precise anatomical localization, and is relatively insensitive to confounding skeletal pathologies (8, 9). Current diagnostic criteria for individual osteoporosis, however, are largely based on the lumbar spine rather than the thoracic spine.
Standard QCT protocols involve higher radiation exposure, and the analysis often depends on specialized CT equipment and complex post-processing. As a result, diagnosing osteoporosis in patients with thoracic conditions via QCT frequently requires an additional lumbar scan. Concerns regarding radiation dose and the frequently asymptomatic nature of osteoporosis contribute to the common oversight of BMD measurement in pulmonary disease patients.
In clinical practice, patients with pulmonary diseases routinely undergo chest CT scans for diagnostic purposes (10–12). The radiation dose for a chest CT, however, is typically lower than that for a lumbar QCT due to anatomical differences, which complicates opportunistic osteoporosis screening through QCT analysis. Advances in artificial intelligence have promoted research on using machine learning and deep convolutional neural networks to measure BMD and diagnose osteoporosis from low-dose chest CT images (13–16), offering a potentially simpler screening approach.
However, most of these studies have focused on the lumbar spine or abdominal CT examinations, with relatively few studies exploring the value of thoracic vertebra BMD measured from chest CT for individual osteoporosis diagnosis (17, 18). This study aims to technically validate AI-based thoracic vertebral BMD measurement from routine chest CT as an opportunistic screening-support tool for osteoporosis and low bone mass.
Methods
Study population
Patient data were retrospectively derived from a prospective study of patients admitted for first-episode acute pancreatitis, registered with the Chinese Clinical Trial Registry (ChiCTR 2400082188), conducted in accordance with the Declaration of Helsinki (2013 revision), and approved by the Ethics Committee of Guiqian International Hospital (Approval Number: Guiqian 2024 Ethics Review No. 04). All methods were carried out in accordance with relevant guidelines and regulations.
From this retrospective cohort, 145 participants who had undergone both chest and abdominal CT examinations within a 7-day period between August 2024 and March 2025 were consecutively included. We excluded 47 participants with a history of malignant tumors or spinal surgery, images with significant artifacts, and other conditions precluding reliable BMD measurement by any method, such as compression fractures, severe scoliosis, or other vertebral abnormalities (Figure 1). A total of 98 participants were included in the final study cohort (48 female and 50 male; mean age 49.48 ± 14.64 years; range, 18–87 years).
CT data acquisition
All patients presented with abdominal pain and underwent combined chest and abdominal CT scans within a seven-day period. All scans were acquired on a single 256-slice CT scanner (Revolution CT, GE HealthCare, USA) using 120 kVp tube voltage, Smart mA tube current, 5 mm slice thickness and 5 mm slice interval for the original acquisition, and a pitch of 0.992:1. Only non-contrast chest and abdominal CT images were included in this study.
Images were reconstructed with 1.25 mm thickness, standard (Stnd) reconstruction kernel, helical plus mode, window width 400, window level 40, and 1.25 mm interval. AI-BMD analysis was performed using this standard reconstruction. The volumetric CT dose index (CTDIvol) of each scan was recorded.
The chest scan extended from the lung apices to the 12th thoracic vertebra (T12), while the abdominal scan covered the entire pancreas.
QCT bone mineral density measurement
BMD measurements were performed using QCT (American Mindways Company, Austin), with daily phantom calibration performed prior to measurement and a standard reconstruction algorithm applied. CT data were analyzed on a Mindways 4th Generation QCT Pro workstation (version 6.1). BMD was measured within the cylindrical regions of interest (ROI), which were 9 mm high and placed in the center of L1 to L3 vertebral bodies avoiding cortical bone and basivertebral veins.
The BMD values (mg/cm3) of L1 to L3 were recorded and analyzed, respectively, and the average was calculated. The total processing time was recorded. To ensure consistency, all BMD measurements were performed by one radiologist (R.Z.).
A randomly selected subset of 30 samples was independently re-measured at a two-week interval by a second radiologist (H.L.) with more than 5 years of experience in musculoskeletal imaging, blinded to the initial results.
American college of radiology (ACR) QCT classification
Each participant bone status was categorized into normal (BMD >120 mg/cm3), osteopenia (80 mg/cm3 ≤ BMD ≤120 mg/cm3), and osteoporosis (BMD<80 mg/cm3) in accordance with the American College of Radiology QCT diagnostic criteria (19). Low BMD was defined as BMD ≤120 mg/cm3, including osteopenia and osteoporosis. Accordingly, patients were classified as normal (n = 61), osteopenia (n = 22), and osteoporosis (n = 15).
Artificial intelligence bone mineral density measurement
This study employed an artificial intelligence-based CT bone mineral density detection system utilizing deep learning and machine learning algorithms (Phantom-less Bone Mineral Density Calculation Software, uAI-Discover BMD, version R001, United Imaging Intelligence Co., Ltd., Shanghai, China). It corrects for scanner differences and factors affecting CT values (e.g., beam hardening) using stable internal references and other measurement parameters, thereby converting CT values into BMD values. After importing chest CT data, the software automatically segmented thoracic vertebrae (T1-T12) and measured BMD—mimicking the anatomical localization logic of lumbar QCT to ensure comparability between thoracic and lumbar BMD results.
BMD was measured within the cylindrical ROI, which were 9 mm high and placed in the center of vertebral bodies avoiding cortical bone and basivertebral. The BMD values (mg/cm3) of T1 to L3 were recorded and analyzed, respectively, and the average was calculated. We recorded the time from software launch to the successful acquisition of thoracic BMD values.
Subsequently, to reduce the variability in single vertebral measurements, and following the lumbar spine QCT measurement protocol, we exploratively calculated the mean values of the obtained bone mineral density data across every three consecutive vertebrae, thereby establishing four subgroups: T1–T3, T4–T6, T7–T9, and T10–T12. Notably, this software is a pre-trained, commercially available, regulatory-cleared system; no model training, fine-tuning, or optimization was performed on the present study data, and therefore there was no risk of data leakage into model development.
Statistical analysis
Statistical analyses were conducted with SPSS 27.0 and MedCalc® V22.001. Continuous variables were assessed for normality using the Shapiro–Wilk test. Normally distributed variables are presented as mean ± standard deviation (SD), and non-normally distributed variables are presented as median (interquartile range, IQR), while categorical variables are expressed as counts (percentages).
Normally distributed paired data were compared using paired t-tests with Bonferroni correction for multiple comparisons. Agreement between AI-BMD and QCT-BMD was assessed via Bland-Altman analysis. Pearson correlation coefficients were used to assess the correlations between AI-BMD and both lumbar QCT-BMD and age.
The diagnostic performance of AI-BMD for osteoporosis and low bone mass was evaluated using receiver operating characteristic (ROC) curve analysis. The Area Under the Curve (AUC) was calculated, and the cutoff threshold corresponding to the maximum Youden index was derived from the ROC curve to determine sensitivity and specificity. The 95% confidence intervals (CI) for the AUC were calculated using the non-parametric method of DeLong method.
Internal validation was performed using bootstrap resampling (1000 iterations) with the rms R package validate() function (Harrell's method), and the optimism-corrected C-index was calculated, optimal thresholds, and Youden indices. Sensitivity and specificity at the optimal thresholds were reported with exact 95% CI, and positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1-score were derived from the corresponding confusion matrices.
Results
Demographic information and BMD measurements of the vertebrae
Among the 98 patients, 61 had normal bone mass, 22 had osteopenia, and 15 had osteoporosis. In all patients, AI-BMD values showed a progressive decrease from T1 to L3, and thoracic BMD values were consistently higher than lumbar BMD values (all P < 0.001) (Table 1). Bone mineral density in both the thoracolumbar segments showed a significant negative correlation with age (r = −0.546 to −0.722, all P < 0.001).
(Figure 2).
Osteoporosis prevalence was higher in women than in men (11/48, 22.9% vs. 4/50, 8.0%; P = 0.046), and all patients with osteoporosis were aged ≥50 years (P < 0.001). BMI was significantly higher in men than in women (24.70 ± 2.97 vs.
22.73 ± 3.01 kg/m2, P = 0.007), but BMI did not differ significantly among the normal, osteopenia, and osteoporosis groups (23.60 ± 3.28 vs 23.88 ± 2.96 vs. 25.24 ± 1.31 kg/m2, P = 0.599), BMD values at all vertebral levels did not differ significantly by sex (all P > 0.05). Both AI-BMD and QCT-BMD values were significantly higher in the younger subgroup (P < 0.001) (Table 1).
These subgroup results should be considered exploratory given the small sample size. Agreement was assessed using the intraclass correlation coefficient (ICC). The single-measure ICCs for L1, L2, and L3 were 0.998 (95% CI: 0.990–0.999), 0.994 (95% CI: 0.978–0.999), and 0.998 (95% CI: 0.995–1.000), respectively, and the average-measure ICC for L1–L3 was 0.998 (95% CI: 0.992–1.000), indicating excellent inter-observer agreement (Supplementary Table S3).
Differences and correlation between thoracic and lumbar BMD
The mean deviation of thoracic vertebral BMD (from T1 to T12) relative to lumbar vertebral BMD gradually decreased. Among the four thoracic vertebral groups, the T10-T12 group showed the best consistency with QCT-BMD(r = 0.98, P < 0.001); Bland-Altman analysis confirmed good agreement between T10-T12 BMD and QCT-BMD (mean deviation: 7.1 mg/cm3; 95% LOA: -13.9 to 28.1 mg/cm3), further supporting its potential utility as a screening tool for osteoporosis (Figure 3). All thoracic vertebral BMD values were strongly positively correlated with lumbar vertebral BMD (r = 0.88 to 0.98).
The strongest correlations was T10–T12 group (r = 0.98). The full correlation matrix is presented as a heatmap (Figure 4).
The diagnostic efficacy of AI-BMD for osteoporosis and low bone mass
Among the 98 included subjects, 15 were diagnosed with osteoporosis. ROC analysis demonstrated that the AUC values for the T10–T12 segments were superior to those of the other three segment groups(AUC = 0.998, 95% CI 0.960-1.000), with a sensitivity, specificity, NPV, PPV and accuracy of 100.0% (95% CI 78.2–100.0), 97.6% (95% CI 91.6–99.7), 100.00%, 88.24% and 98.00%, respectively (Figure 5). To further evaluate the predictive ability of T10-T12 without overfitting, the corrected AUC obtained by internal validation using the bootstrap method was 0.998 (95% CI 0.984-1.000), the calibration slope was 0.781; the calibration intercept was -0.455; and the Brier score was 0.016.
In the low-bone-mass analysis, ROC analysis showed that T10–T12 also provided better AUC values than other three-vertebra groups (AUC = 0.990, 95% CI 0.945-1.000) with a sensitivity, specificity, NPV, PPV and accuracy of 97.30% (95% CI 85.8–99.9), 96.72% (95% CI 88.7–99.6), 98.33%, 94.74% and 96.94%, respectively. To further evaluate the predictive ability of T10-T12 without overfitting, the corrected AUC obtained by internal validation using the bootstrap method was 0.989(95% CI 0.961-0.998), the calibration slope was 0.870; the calibration intercept was -0.012; and the Brier score was 0.037 (Table 2). The more conservative AUC in the low-bone-mass analysis likely reflects the true discriminative ability more accurately, given the larger number of positive cases.
Borderline cases (BMD within ±5 mg/cm3 of the threshold) accounted for both false positives in the osteoporosis analysis, while all true osteoporosis cases had BMD values clearly below the threshold. These findings indicate that misclassification occurs mainly near the threshold and that the method has a high NPV. Vertebral-level-specific analyses for both osteoporosis and low bone mass are provided in Supplementary Tables S1, S2.
Diagnostic performance increased caudally, with T12 showing the closest correlation with lumbar QCT (r = 0.98) and excellent diagnostic accuracy (osteoporosis: AUC 0.998; low bone mass: AUC 0.987). For osteoporosis, no pairwise differences reached statistical significance (all P > 0.05), although T7–T9 (P = 0.057) and T10–T12 (P = 0.063) showed a trend toward higher AUC than T1–T3. For low bone mass, T7–T9 and T10–T12 were both significantly superior to T1–T3 (P = 0.030 and P = 0.036, respectively), whereas no significant differences were observed among T4–T6, T7–T9, and T10–T12.
These findings indicate that diagnostic performance improved caudally and plateaued at T7–T12.
In addition, the CT dose index (CTDIvol) of QCT scans was significantly higher than that of chest CT scans (14.63 ± 2.30 vs. 4.13 ± 1.01 mGy, P < 0.001); the post-processing time of QCT was significantly longer than the AI-based BMD measurement time of chest CT (55.76 ± 9.05 vs. 7.50 ± 0.78 s, P < 0.001).
Discussion
This study is a technical validation of an AI-based phantom-less BMD measurement tool in a cohort of patients with acute pancreatitis who underwent both chest and abdominal CT. This cohort was selected because patients with pulmonary diseases typically undergo chest CT only and rarely have paired abdominal CT for QCT reference. Our results suggest that AI-based thoracic vertebral BMD measurement on routine chest CT may serve as a promising tool for opportunistic osteoporosis screening, addressing limitations of traditional methods such as radiation exposure and low efficiency.
However, these findings are preliminary and require validation in pulmonary-disease and general screening populations before broader clinical application. In our cohort, thoracic vertebral BMD was consistently higher than lumbar BMD measured. A gradual decrease in BMD was observed from the T1 to T12 vertebral bodies, aligning with findings reported by George Simion et al.
(9, 20). This trend likely relates to differences in vertebral stress. The sagittal curvature of the spine is known to directly influence load distribution (21), and variations in the load-strength ratio (LSR) can further alter stress distribution (22), ultimately leading to regional differences in BMD.
From the cervical to the lower lumbar segments, increasing mechanical stress correlates with decreasing BMD, which helps explain the higher fracture incidence observed in the lower thoracic and lumbar regions (22). Given the more comparable mechanical loading conditions between the lower thoracic and lumbar spine, the AI-BMD values for the T10-T12 group are numerically closer to and show a stronger correlation with those of the L1-L3 group. The AI-BMD derived from T10-T12 demonstrates excellent diagnostic efficacy for lumbar osteoporosis (AUC = 0.998, P < 0.001).
This performance is superior to that of other thoracic segment groups, particularly compared to the T1-T3 and T4-T6 group. Three consecutive vertebrae were averaged for two reasons: (1) to reduce the impact of individual vertebral abnormalities (e.g., compression fractures, implants, or degenerative changes) that may preclude reliable BMD measurement at a single level, consistent with standard QCT practice; and (2) Three-vertebra averaging improved correlation with QCT compared with individual vertebrae (T10–T12: r = 0.98 vs. T10: r = 0.96, T11: r = 0.97, T12: r = 0.98).
These findings are consistent with previously reported by Ya Ling Pan et al. (23). Moreover, the vertebral-level-specific analysis demonstrated a caudal gradient of diagnostic performance for osteoporosis(AUC = 0.954-0.998), with the individual T12 vertebra showing the closest correlation with lumbar QCT (r = 0.98), further supporting the biomechanical similarity between the lower thoracic spine and the lumbar region.
The 100% NPV of the T10-T12 measurement is a particularly favorable property for an opportunistic case-finding tool, for which reliably excluding osteoporosis is the key requirement; the few false positives would be clarified by confirmatory QCT in the proposed clinical workflow.
Although the AI-BMD of the T10-T12 group demonstrated high diagnostic efficacy for osteoporosis, its diagnostic threshold requires adjustment. In this study, a BMD threshold of ≤ 90.86 mg/cm3 yielded a sensitivity of 100% (95% CI 78.2–100.0) and a specificity of 97.6% (95% CI:91.6–99.7) for diagnosing osteoporosis in the T10-T12 group (Table 2); the corresponding confusion matrix yielded a PPV of 88.24%, an NPV of 100.00%, an accuracy of 98.00%, and an F1-score of 0.938. In 2021, Maximilian T Löffler et al.
(24) proposed an artificial intelligence-based method, which achieved a slightly lower diagnostic efficacy for osteoporosis (AUC=0.86) than the present study. Furthermore, Sebastian Rühling et al. (25) established thoracic vertebral BMD thresholds for osteoporosis diagnosis from the cervical to lower thoracic vertebral bodies, proposing a general thoracic threshold of 100 mg/cm3 and a specific threshold of 83.8 mg/cm3 for the T12 vertebral body.
These published values are broadly consistent with our findings, though slightly lower than the threshold identified here. The discrepancy likely arises from two methodological differences. First, our analysis employed the average bone mineral density (BMD) across three vertebrae (T10-T12), which is inherently higher than the BMD of a single T12 vertebra.
Second, the threshold established here was validated against the QCT gold standard, whereas the prior threshold was derived solely from correlation analysis without such validation.
Both AI-BMD and QCT-BMD exhibited a strong negative correlation with age, a finding consistent with prior reports (25–27). Age-related declines in overall bone mineral density are well-documented, with osteoporosis being more prevalent in women. This increased prevalence is linked to factors including mineral loss and hormonal changes.
The patient cohort in this study reflected this trend, with a higher proportion of female osteoporosis cases (11/15, 73.3%), but all osteoporosis cases were diagnosed in individuals over 50 years old, the diagnostic performance of the AI bone mineral density threshold in younger populations should be interpreted with caution.
In this study, the radiation dose for chest CT scanning was significantly lower, measuring only 28% of that for abdominal CT. This difference enables patients with pulmonary diseases to undergo both pulmonary diagnosis and osteoporosis assessment in a single CT examination, thereby avoiding the additional radiation exposure of a dedicated lumbar QCT scan. Furthermore, the AI-assisted diagnostic system automatically segments vertebral bodies and rapidly calculates BMD.
Relative to manual QCT post-processing, this system reduces processing time by 87%, and its automated segmentation significantly improves inter-observer consistency by eliminating examiner subjectivity. Consequently, deep learning-based BMD measurement from chest CT scans may optimize osteoporosis prevention and management in this patient population. In addition, 19 patients were excluded because reliable BMD measurement was not feasible by any method, including QCT and AI, due to multiple compression fractures, severe scoliosis, or other vertebral abnormalities.
In clinical practice, it is recommended to use the average value of adjacent vertebrae as a surrogate indicator, a approach also consistently adopted in this study. From a practical perspective, AI-BMD can be derived from already-acquired chest CT without additional scanning, radiation, or acquisition time, with an automated post-processing time of only 7.50 ± 0.78 s per case, making this approach feasible for large-scale opportunistic case-finding. The vertebral-level results further suggest that single-vertebra measurement at T12 may suffice, simplifying clinical implementation.
Positive AI-BMD findings should nevertheless trigger confirmatory QCT or DXA before any treatment decision. DXA remains the clinical standard for osteoporosis diagnosis; however, DXA (areal BMD) and QCT (volumetric BMD) are not interchangeable, and DXA was not systematically available in our cohort. Validation of AI-BMD against DXA-defined outcomes, including fracture prediction, represents an important direction for future research.
This study has several limitations. First, The study population consisted of patients with first-episode acute pancreatitis rather than a population-based screening cohort, which may introduce selection and spectrum bias and limit the generalizability of our findings to routine opportunistic osteoporosis screening, therefore, the generalizability of our findings to these populations requires further validation. Second, This study used a single scanner model, a fixed tube voltage (120 kVp), and a single reconstruction protocol; therefore, the robustness of the AI system across different scanners, tube voltages, slice thicknesses, and reconstruction algorithms remains to be determined, and further multicenter validation is warranted.
Third, Only non-contrast chest and abdominal CT images were included; therefore, the potential effect of intravenous contrast on AI-derived BMD estimation and osteoporosis classification could not be evaluated, and further studies are needed to determine whether contrast-enhanced CT affects measurement accuracy. Fourth, Other potential confounding factors, including menopausal status, corticosteroid use, anti-osteoporotic treatment, diabetes, renal disease, and vertebral fractures, were not consistently available and could not be incorporated into subgroup or multivariable analyses. In addition, because all osteoporosis cases were aged ≥50 years, the performance of AI-BMD in younger populations could not be assessed.
Fifth, A formal robustness or failure-analysis study of the AI system in challenging cases (e.g., motion or metal artifacts, vertebral compression fractures, degenerative changes, osteophytes, vascular calcification, or incomplete vertebral visualization) was not performed, and future work should evaluate the AI system in these cases.
Conclusions
The measurement of thoracic vertebral BMD using chest CT combined with deep learning algorithms may serve as a potential opportunistic case-finding or screening-support tool for osteoporosis and low bone mass. T10–T12 BMD measurements are comparable to lumbar BMD. However, given the small sample size, medically selected population, lack of external validation, and limited assessment of scanner/protocol variability, this approach should not yet be considered a replacement for conventional osteoporosis assessment.
Larger multicenter, prospective, and externally validated studies are warranted before clinical adoption.