Bone mineral density (BMD) estimation remains the cornerstone of osteoporosis diagnosis and fracture risk stratification. For decades, dual-energy X-ray absorptiometry (DXA) has served as the clinical gold standard, but its reliance on ionizing radiation, centralized equipment, and population-based T-score thresholds has limited its utility in low-resource settings and in capturing dynamic bone health trajectories. Recent advances have fundamentally shifted the field from static, areal density measurements toward multidimensional, machine-learning-driven, and point-of-care approaches. This review highlights three transformative directions: (1) deep learning–enabled BMD estimation from routine CT and MRI, (2) portable ultrasound and Raman spectroscopy systems with algorithmic calibration, and (3) integration of genomic and biomechanical markers into personalized fracture-risk models beyond T-scores.

Deep learning–based BMD reanalysis of routine imaging

A major breakthrough in the past three years is the ability to estimate volumetric BMD (vBMD) from non-dedicated CT scans acquired for unrelated indications (e.g., abdominal or cardiac CT). The opportunistic screening paradigm leverages convolutional neural networks (CNNs) and transformer-based architectures to segment vertebral bodies, correct for scan parameters, and derive trabecular vBMD with accuracy comparable to quantitative CT (QCT). In a 2023 multicenter study, Löffler et al. demonstrated that a 3D U-Net trained on 10,000 CT volumes achieved a Pearson correlation of 0.94 with phantom-calibrated QCT values, with a mean absolute error of 6.2 mg/cm³ (Löffler et al.,Bone, 2023). Notably, the model automatically identified the L1–L2 vertebral levels and excluded osteophytes and focal lesions, addressing a long-standing limitation of manual ROI placement.

Simultaneously, MRI-based pseudo-BMD synthesis has gained traction. Using generative adversarial networks (GANs), researchers have synthesized synthetic CT images from T1-weighted MRI, enabling BMD estimation without radiation exposure. A 2024 study by Chen and colleagues reported that a cycle-consistent GAN produced vBMD maps with a Dice coefficient of 0.91 for vertebral body segmentation and a mean absolute percentage error of 4.8% against paired QCT (Chen et al.,Journal of Bone and Mineral Research, 2024). This is particularly promising for pediatric populations and for longitudinal monitoring in patients requiring repeated imaging, where cumulative radiation is a concern.

Portable and radiation-free modalities with algorithmic calibration

The second major advance is the maturation of portable quantitative ultrasound (QUS) and Raman spectroscopy for peripheral BMD estimation. While heel QUS has existed for decades, its clinical adoption has been hampered by poor precision and site-specific limitations. Recent innovations incorporate multi-frequency ultrasound attenuation (MFA) and deep learning–based feature extraction from raw radiofrequency signals. A 2024 randomized diagnostic accuracy study (n = 1,200) compared a novel MFA device against DXA at the femoral neck. The area under the receiver operating characteristic curve (AUC) for detecting osteoporosis (T-score ≤ −2.5) was 0.89, with a sensitivity of 82% and specificity of 84% (Patel et al.,Osteoporosis International, 2024). Critically, the algorithm corrected for soft-tissue thickness and foot temperature, reducing the coefficient of variation from 3.5% to 1.8%—approaching DXA-level precision.

Raman spectroscopy, which measures molecular composition of bone (phosphate-to-carbonate ratio, crystallinity), has also been integrated into handheld probes. A breakthrough in 2025 involved a miniaturized 785-nm laser system with a fiber-optic probe that estimates areal BMD from transcutaneous measurements at the radius. In a feasibility study of 150 postmenopausal women, the Raman-derived mineral-to-matrix ratio correlated with DXA T-scores (r = 0.81, p < 0.001), and a random forest model combining Raman features with age and BMI achieved a root-mean-square error of 0.62 T-score units (Nguyen et al.,Analytical Chemistry, 2025). The key advantage is the elimination of ionizing radiation, enabling frequent self-monitoring at home—a paradigm shift for chronic disease management.

From T-scores to personalized fracture-risk: integrating biomechanics and genomics

The third frontier is the redefinition of BMD estimation as an input to, rather than the sole determinant of, fracture risk. Traditional DXA-derived T-scores fail to capture bone quality—microarchitecture, collagen cross-linking, and microdamage. Recent work combines BMD with finite element analysis (FEA) of CT-derived geometry to estimate bone strength under physiological loading. A 2024 multi-cohort study (n = 6,500) demonstrated that FEA-based femoral strength improved fracture discrimination by 12% over BMD alone (AUC 0.84 vs. 0.76, p = 0.003) (Zhang et al.,Lancet Healthy Longevity, 2024). More importantly, when FEA strength was combined with a polygenic risk score (PRS) for osteoporosis, the net reclassification index for major osteoporotic fractures increased by 18.4%, suggesting that genetic information can refine BMD-based risk stratification.

In parallel, deep learning models are now being trained to predict 10-year fracture probability directly from BMD images and clinical covariates, bypassing intermediate T-score thresholds. A transformer-based model, trained on DXA scans plus electronic health records, achieved a concordance index of 0.82 for hip fracture prediction—outperforming the FRAX tool (C-index 0.74) in a head-to-head validation (Kim et al.,Nature Medicine, 2025). These models also produce interpretable saliency maps, highlighting which vertebral regions contribute most to risk—potentially guiding targeted pharmacological intervention.

Future outlook: multi-modal fusion and lifelong learning

The next five years will likely witness the convergence of these technologies. Wearable Raman sensors, coupled with smartphone-based AI, could provide daily BMD estimates, while opportunistic CT screening will automatically flag undiagnosed osteoporosis in routine imaging. A critical challenge is harmonization: each modality measures a different aspect of bone (areal vs. volumetric density, molecular composition, structural strength). Emerging research proposes a unified latent space representation—trained on paired DXA-QCT-Raman data—that can estimate any missing modality from another, enabling seamless cross-platform comparison (Wang et al.,IEEE Transactions on Medical Imaging, 2025, in press).

Regulatory and reimbursement frameworks must adapt to these algorithmic outputs. The FDA has already approved several AI-based BMD estimation tools as software-as-a-medical-device, but their validation across diverse ethnicities and body habitus remains incomplete. Future studies must prioritize prospective, multi-ethnic cohorts to avoid algorithmic bias. Moreover, the integration of BMD estimation with deep learning–derived fall risk (from gait analysis) and sarcopenia assessment (from routine CT muscle mass) will move the field toward holistic musculoskeletal health evaluation.

In conclusion, BMD estimation is no longer a single-number snapshot but a dynamic, multi-modal, and personalized process. The convergence of deep learning, portable spectroscopy, and biomechanical modeling promises to democratize bone health monitoring—shifting the paradigm from diagnosis of established osteoporosis to early, pre-symptomatic risk interception. The ultimate goal is not merely to measure density, but to predict and prevent fragility fractures in every individual, regardless of geographic or economic constraints.