AI Reads Routine Chest CTs to Flag Osteoporosis in Early Study

Artificial intelligence software can estimate bone mineral density from ordinary chest CT scans and, in a small preliminary study, matched the performance of a dedicated, higher-dose quantitative CT exam for detecting osteoporosis and low bone mass. The finding, published in Frontiers in Medicine, points toward opportunistic screening: measuring bone health from scans patients already receive, without extra radiation, extra scanning time, or a separate referral.

The study, led by Runhai Zhang and Caijun Huang at Guiqian International Hospital in Guiyang, China, with a co-author from GE HealthCare’s CT research center, set out to address a practical gap. Quantitative computed tomography, or QCT, is an established volumetric method for measuring bone mineral density, but it involves higher radiation exposure than a routine chest scan and depends on specialized equipment and post-processing. Current diagnostic criteria for individual osteoporosis are largely based on the lumbar spine, so patients with thoracic conditions often need an additional lumbar scan just to obtain a QCT bone measurement.

The researchers analyzed 98 patients, 48 women and 50 men with a mean age of 49.48 years, who underwent both chest and abdominal CT within a seven-day period between August 2024 and March 2025. All had presented with abdominal pain and were drawn from a registered prospective study of first-episode acute pancreatitis. Forty-seven candidates were excluded, including those with malignant tumors or spinal surgery, significant image artifacts, compression fractures, severe scoliosis, or other vertebral abnormalities that would prevent reliable density measurement by any method. Of the final cohort, 61 had normal bone mass, 22 had osteopenia, and 15 had osteoporosis according to American College of Radiology QCT criteria, which classify normal bone as above 120 mg/cm³, osteopenia between 80 and 120 mg/cm³, and osteoporosis below 80 mg/cm³.

The AI component was a pre-trained, commercially available, regulatory-cleared phantom-less bone mineral density system, uAI-Discover BMD from United Imaging Intelligence. The software automatically segmented the twelve thoracic vertebrae, T1 through T12, on non-contrast chest CT images and converted CT values into density values in milligrams per cubic centimeter, using stable internal references to correct for scanner differences and factors such as beam hardening. It mimicked the anatomical logic of lumbar QCT, placing cylindrical regions of interest 9 mm high in the center of each vertebral body while avoiding cortical bone and basivertebral veins. No model training or fine-tuning was performed on the study data, so there was no risk of data leakage into the model’s development.

To mirror standard QCT practice and reduce the variability of single-vertebra measurements, the team averaged every three consecutive vertebrae, creating four subgroups: T1–T3, T4–T6, T7–T9, and T10–T12. These AI-derived thoracic values were compared against lumbar QCT measurements of L1 to L3 performed on a Mindways workstation with daily phantom calibration, taken as the reference standard. Across all patients, AI-measured bone density fell progressively from T1 toward L3, and thoracic values were consistently higher than lumbar values. Density in both thoracic and lumbar segments declined significantly with age, with correlation coefficients between −0.546 and −0.722.

The lower thoracic spine proved the strongest proxy for the lumbar reference. The T10–T12 average correlated with QCT at r = 0.98, the highest of any group, and Bland-Altman analysis showed good agreement, with a mean deviation of 7.1 mg/cm³ and 95% limits of agreement spanning −13.9 to 28.1 mg/cm³. The researchers attribute this caudal gradient to biomechanics: mechanical loading increases from the cervical region toward the lower lumbar spine, and the more comparable loading conditions of the lower thoracic and lumbar segments plausibly explain why T10–T12 values track lumbar density most closely.

Diagnostic performance was striking. For osteoporosis, the T10–T12 measurement achieved an area under the receiver operating characteristic curve of 0.998 (95% CI 0.960–1.000), with an optimal threshold of 90.86 mg/cm³ yielding 100% sensitivity (95% CI 78.2–100.0), 97.6% specificity, a negative predictive value of 100%, and a positive predictive value of 88.24%. Bootstrap internal validation produced an optimism-corrected AUC of 0.998. For low bone mass, defined as osteopenia plus osteoporosis, the T10–T12 AUC was 0.990 (95% CI 0.945–1.000) at a threshold of 126.0 mg/cm³, with 97.3% sensitivity and 96.7% specificity. The authors note the more conservative low-bone-mass figure likely reflects true discriminative ability more accurately, given the larger number of positive cases, and that misclassifications clustered near the decision threshold.

The practical advantages were quantified as well. The volumetric CT dose index was 4.13 ± 1.01 mGy for chest CT versus 14.63 ± 2.30 mGy for the QCT scans, roughly a quarter of the exposure, and the AI post-processing took 7.50 ± 0.78 seconds per case compared with 55.76 ± 9.05 seconds for manual QCT processing. Because AI-derived density requires no additional acquisition, the authors argue the approach could enable large-scale opportunistic case-finding, with the 100% negative predictive value particularly valuable for reliably ruling out osteoporosis; the few false positives would be resolved by confirmatory QCT in the proposed workflow.

Important caveats temper the enthusiasm. The cohort consisted of pancreatitis patients rather than a screening population, raising selection and spectrum bias concerns; all osteoporosis cases were aged 50 or older, so performance in younger people could not be assessed; a single scanner model, fixed 120 kVp tube voltage, and one reconstruction protocol were used; only non-contrast images were included, leaving contrast effects untested; and confounders such as menopausal status, corticosteroid use, and diabetes were unavailable for analysis. No formal failure analysis in challenging cases, such as motion artifacts or degenerative changes, was performed. DXA, the clinical diagnostic standard, was not systematically available, and areal and volumetric density measures are not interchangeable, so validation against DXA-defined outcomes including fracture prediction remains future work.

The authors themselves are measured in their conclusion: thoracic BMD from routine chest CT with deep learning may serve as a potential opportunistic screening-support tool, with T10–T12 values comparable to lumbar measurements, but given the small sample, medically selected population, lack of external validation, and limited assessment of scanner variability, the method should not yet replace conventional osteoporosis assessment. Larger, multicenter, prospective, externally validated studies are required before clinical adoption, and any positive AI finding should trigger confirmatory QCT or DXA before treatment decisions.

Subject of Research: Using artificial intelligence to measure bone mineral density from routine chest CT scans for opportunistic osteoporosis screening

Article Title: Can routine chest CT with artificial intelligence tool be used for opportunistically osteoporosis screening? A preliminary study

Article References: Zhang, R., Li, H., Lu, M., Chen, S., & Huang, C. (2026). Can routine chest CT with artificial intelligence tool be used for opportunistically osteoporosis screening? A preliminary study. Frontiers in Medicine, 13, Article 1936400. https://doi.org/10.3389/fmed.2026.1936400

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1936400

Keywords: osteoporosis, bone mineral density, chest CT, artificial intelligence, QCT, opportunistic screening, deep learning, thoracic vertebrae, medical imaging, radiation dose, low bone mass, diagnostic accuracy