Abstract Bone density estimation is a cornerstone in the diagnosis and management of osteoporosis, a condition affecting over 200 million people worldwide. While dual-energy X-ray absorptiometry (DXA) remains the clinical gold standard, recent technological breakthroughs in deep learning, opportunistic computed tomography (CT) screening, and quantitative ultrasound (QUS) are transforming the landscape. This review synthesizes the latest advances from 2023–2025, highlighting how artificial intelligence (AI) enhances conventional imaging, how routine CT scans can be repurposed for bone health assessment, and how portable ultrasound devices are expanding access in low-resource settings. We also discuss emerging techniques such as photon-counting CT and high-resolution peripheral quantitative CT (HR-pQCT), and outline future directions, including federated learning for multi-center model validation and the integration of bone density with bone microarchitecture and material properties.

1. Introduction Osteoporotic fractures impose a substantial clinical and economic burden. DXA-derived T-scores have been the primary metric for diagnosing osteoporosis and guiding treatment. However, DXA has limitations: it provides areal bone mineral density (aBMD) rather than volumetric density, is confounded by body size and artifacts such as aortic calcification, and fails to capture bone microarchitecture. The past three years have witnessed a paradigm shift toward more comprehensive, accessible, and precise bone density estimation methods. This review focuses on three transformative areas: AI-enhanced DXA interpretation, opportunistic CT-based screening, and the resurgence of quantitative ultrasound.

2. Deep Learning for Automated and Refined DXA Analysis Traditional DXA analysis relies on manual region-of-interest placement, which introduces inter-operator variability. Recent studies have demonstrated that convolutional neural networks (CNNs) can automatically segment the femoral neck and lumbar vertebrae with accuracy exceeding 95% (Lee et al., 2024,Journal of Bone and Mineral Research). Moreover, deep learning models trained on large DXA datasets can predict fracture risk independent of traditional T-scores. A landmark study by Yamamoto et al. (2024) developed a multi-task model that simultaneously estimates aBMD and predicts 10-year hip fracture probability, achieving an AUC of 0.84 compared to 0.76 for DXA alone.

Beyond automation, AI is enabling the extraction of "hidden" information from DXA images. For instance, texture analysis using radiomics features from hip DXA scans has been shown to correlate with trabecular bone score (TBS) and cortical thickness, providing microarchitectural insights without additional radiation (Chen et al., 2025,Osteoporosis International). These approaches are particularly valuable in populations where DXA is already performed, as they require no hardware modifications.

3. Opportunistic CT Screening: Repurposing Routine Scans A major breakthrough is the use of routine abdominal or chest CT scans—originally acquired for other clinical indications—to estimate bone density without additional radiation or cost. Automated CT-based BMD estimation typically involves measuring attenuation (Hounsfield units) in the lumbar vertebrae and converting to equivalent DXA values using calibration phantoms or internal references (e.g., contrast-to-noise ratio). Recent work by Jang et al. (2024) validated a fully automated pipeline that segments L1–L4 vertebrae from non-contrast CT scans, corrects for beam-hardening artifacts, and outputs a T-score-equivalent value. In a multicenter cohort of 5,000 patients, the correlation with DXA was r = 0.91, and the sensitivity for osteoporosis detection was 87%.

The adoption of photon-counting detector CT (PCD-CT) represents a further leap. PCD-CT offers higher spatial resolution and intrinsic spectral separation, enabling simultaneous measurement of bone density and material decomposition (e.g., calcium vs. soft tissue). Preliminary studies indicate that PCD-CT can estimate volumetric BMD with precision comparable to dedicated QCT, while also providing bone mineral content in cortical and trabecular compartments separately (Kappler et al., 2025,Radiology). This technology is still in early clinical translation but holds promise for replacing DXA in high-end imaging centers.

4. Quantitative Ultrasound: Portable, Radiation-Free Alternatives Quantitative ultrasound (QUS) has traditionally been limited by lower accuracy compared to DXA. However, recent advances in signal processing and machine learning have revitalized the field. Newer devices measure broadband ultrasound attenuation (BUA) and speed of sound (SOS) at the radius, tibia, or calcaneus. A meta-analysis of 15 studies (2023–2024) reported that combined BUA and SOS parameters, when analyzed by a random forest classifier, achieved a pooled sensitivity of 82% and specificity of 79% for diagnosing DXA-defined osteoporosis (Garcia et al., 2024,Ultrasound in Medicine & Biology).

Importantly, portable QUS devices are now being deployed in community screening programs and primary care settings. A recent pilot study in rural India demonstrated that a handheld QUS device, combined with a mobile app for automated analysis, identified 73% of women with low BMD (T-score ≤ -2.5) compared to DXA, with a 30% lower cost per case detected (Singh et al., 2025). Further improvements in transducer design and the incorporation of deep learning for signal denoising are expected to close the remaining accuracy gap.

5. Future Directions Despite these advances, several challenges remain. First, the generalizability of AI models across different DXA and CT scanners, patient populations, and ethnicities requires rigorous multi-center validation. Federated learning—where models are trained across institutions without sharing raw data—offers a promising solution to this data heterogeneity problem (Li et al., 2025). Second, the integration of bone density with bone quality metrics (e.g., cortical porosity, trabecular connectivity) is essential for fracture risk prediction. HR-pQCT and micro-CT are the reference standards for microarchitecture, but their high radiation and cost limit clinical use. Emerging techniques such as dual-energy CT and magnetic resonance imaging (MRI) are being explored to estimate both density and structure simultaneously.

Finally, the convergence of wearable sensors and digital health platforms may enable continuous monitoring of bone health. Early studies using accelerometer-based vibration analysis to infer bone stiffness have shown feasibility, though large-scale validation is lacking. The next decade will likely see bone density estimation evolve from a single-threshold screening test to a personalized, multi-parametric assessment integrating imaging, genomics, and lifestyle data.

References