Can AI Detect Osteoporosis from Routine Chest Radiographs?

A probability score, not a bone density reading

AI-OsRM does not measure bone mineral density (BMD). It takes a standard chest radiograph (CXR) and outputs probabilities that a person falls into each of three DXA-defined categories: normal BMD, osteopenia, or osteoporosis. The reference standard came from dual-energy X-ray absorptiometry (DXA), which assigns T-scores from the lumbar spine (L1–4), femoral neck, and total hip; osteoporosis was defined as a T-score of −2.5 or lower, osteopenia as between −2.5 and −1.0, and normal as −1.0 or higher.

The paper also reports a second target, "reduced BMD," which groups osteopenia and osteoporosis together (T-score below −1.0). This distinction matters because the two outputs come from different model values: the osteoporosis output is the predicted probability of osteoporosis, while the reduced BMD output is 1 minus the predicted probability of normal BMD. Their optimal cutoffs, chosen on internal validation by maximizing the Youden J index, were 0.19 and 0.74 respectively. A reader comparing sensitivity numbers across the paper must check which of these two targets and which cutoff produced them.

The authors applied T-score categories to all patients regardless of age or menopausal status, even though guidelines recommend T-score classification only for men aged 50 or older and postmenopausal women. They state this choice was made to reflect real-world screening practice and improve generalizability.

Training on paired X-rays and DXA scans

The model was developed retrospectively at a tertiary referral hospital in Seoul, using chest radiographs matched to DXA examinations performed within one month of each other between January 2016 and December 2022. After excluding cases with inconclusive reports needed to compute a final T-score and radiographs from patients later used in external validation, 51,076 radiographs from 26,757 patients remained. These were split by patient in a 3:1:1 ratio into training (30,918 radiographs from 16,053 patients), tuning, and internal validation sets, so no patient appeared in more than one set.

The backbone was ResNet-18, selected after comparing convolutional networks and vision transformers on internal validation discrimination. Before training, images were intensity-windowed to emphasize bone, contrast-limited adaptive histogram equalization was applied to harmonize intensity across scanners, and images were resized to 224×224 pixels. The authors used masked autoencoder pretraining to capture subtle intensity and texture features relevant to BMD estimation.

A multi-task design fed the final ResNet-18 features into two parallel linear branches: one for three-class classification and one for continuous T-score regression, optimized jointly with cross-entropy and mean squared error losses, with the regression weight set to 0.5. The stated aim was to preserve T-score ordering in the learned feature space. In internal validation, the three-class ResNet-18 model reached an AUROC of 0.90 (95% CI 0.88–0.90) for osteoporosis and was chosen as the final model.

External cohorts differ sharply in disease prevalence

External validation used three independent Korean cohorts. The SNUH cohort was temporally separated, drawn from patients who had CXR and DXA within one month in 2015: 6,391 patients, 80% women, median age 61, with 15% osteoporosis and 48% osteopenia. The two health checkup cohorts were larger and younger. HPC included 24,132 individuals (64% women, median age 58) with 5% osteoporosis and 35% osteopenia, and YSMC included 122,535 individuals (91% women, median age 46) with 6% osteoporosis and 30% osteopenia.

These differences are not incidental. The health checkup cohorts had same-day CXR and DXA, while the SNUH cohort allowed up to a month between examinations. Prevalence of osteoporosis ranged from 5% to 15%, so identical sensitivity and specificity values would produce very different positive predictive values across cohorts. The paper reports discrimination and threshold-based operating characteristics separately for each cohort rather than pooling them.

Discrimination holds up, but fixed cutoffs behave differently

AUROC, which measures discrimination across all thresholds, was consistent. For osteoporosis it was 0.88 (95% CI 0.87–0.89) in SNUH, 0.93 (0.92–0.93) in HPC, and 0.92 (0.92–0.92) in YSMC. For reduced BMD it was 0.88 in SNUH, 0.89 in HPC, and 0.88 in YSMC.

Performance at the fixed internal-validation cutoffs was less uniform. For osteoporosis at the 0.19 cutoff, sensitivity was 84% (818/979) in SNUH, 83% (1,044/1,251) in HPC, but only 53% (3,905/7,342) in YSMC, where specificity was 96% (110,945/115,193). For reduced BMD at the 0.74 cutoff, sensitivity was 77% (3,139/4,065) in SNUH, 71% (6,910/9,730) in HPC, and 55% (24,015/43,624) in YSMC. The high AUROCs therefore do not mean the same cutoff yields the same sensitivity everywhere; the YSMC cohort's younger, largely female population and low prevalence shifted the operating point.

Subgroup analysis in the SNUH cohort found discrimination varied significantly by age (P for interaction < .001), with AUROC falling from 0.91 in patients under 50 to 0.84 in those aged 65 or older. Sensitivity was generally preserved across age groups, while specificity was lower in older patients. No significant interaction was found for sex, site of the lowest T-score, metal instrumentation, vertebroplasty, or CXR vendor. The authors suggest degenerative spinal changes and aortic calcifications can artifactually raise DXA-derived T-scores in older patients, and similar findings on radiographs may confound the model.

The prescreening trade-off in health checkups

The clinical utility analysis simulated AI-OsRM as a filter before DXA in the two health checkup cohorts, sweeping cutoffs from 0.20 to 0.60 on the reduced BMD output. In HPC, DXA examinations fell by 33%–56% while osteoporosis sensitivity stayed near 99% and reduced BMD sensitivity ranged from 81% to 95%. In YSMC, DXA use fell by 42%–70%, with osteoporosis sensitivity at 96%–99% and reduced BMD sensitivity at 68%–91%.

A cutoff of 0.40 was described as a balanced operating point in both cohorts: it detected 98–99% of osteoporosis cases and 80%–90% of reduced BMD cases while reducing DXA examinations by roughly 50%. Negative likelihood ratios for osteoporosis stayed below 0.10 across cutoffs, which the authors interpret as support for a rule-out role: a low-risk AI-OsRM result makes osteoporosis unlikely. Positive likelihood ratios rose with higher cutoffs. Decision curve analysis showed positive net benefit for both targets in HPC and YSMC, with and without cohort-specific probability calibration.

These are simulation results, not outcomes of a deployed screening program. Every included individual had already undergone DXA, so the study cannot show how many people would actually complete a DXA referral, nor whether earlier detection would change fracture outcomes.

Where the evidence stops

The study is retrospective and multicenter but confined to three Korean institutions, with DXA reference standards calibrated to Asian reference databases. The authors note that calibration to local reference standards remains to be established and may require cohort-specific tuning and recalibration. Grad-CAM maps were relatively diffuse, limiting interpretability, although representative cases highlighted the lower thoracic vertebrae, shoulder girdle, and thoracic vertebrae.

The authors also position their work against prior CXR-based osteoporosis models, noting that earlier external validations were often under 1,000 cases, and that a randomized trial of AI-based CXR screening was conducted at the model's own development institution with roughly 10% DXA adherence in the intervention group. The contribution claimed here is breadth of external validation across three independent consecutive cohorts totaling more than 150,000 individuals, plus an explicit prescreening simulation in health checkup populations.

The cutoff decides the use case

The most decision-relevant result is not a single AUROC but the pairing of a threshold with a cohort. At 0.40 in the health checkup setting, AI-OsRM is presented as a rule-out filter that halves DXA volume while missing about 1–2% of osteoporosis cases; at the internal-validation cutoff of 0.19, sensitivity in the younger YSMC cohort dropped to 53%, which would be unsuitable for screening. The paper's own framing therefore favors a deliberately high-sensitivity cutoff in low-prevalence checkup populations, with the caveat that the 0.40 operating point was derived from simulation in two Korean cohorts and has not been tested prospectively.

For readers weighing opportunistic screening, the practical question is which population and which threshold, not whether the model discriminates. The SNUH cohort's age gradient—AUROC 0.91 under 50 versus 0.84 at 65 or older—suggests the model's advantage is clearest in younger health checkup populations, precisely where baseline osteoporosis prevalence is lowest and where a rule-out filter has the most DXA volume to save.

Sources used in this answer