Opportunistic screening for osteoporosis using artificial intelligence-based morphometric analysis of chest computed tomography images: a retrospective multi-center study in Russia leveraging the COVID-19 pandemic

A. V. Petraikin, Perry J. Pickhardt, Mikhail Belyaev, Zhanna Belaya, Maxim Pisov, Alim Bukharaev, Aleksei Zakharov, Nikita D. Kudryavtsev, Tatiana Bobrovskaya, Д. С. Семенов, Е. С. Ахмад, Z. R. Artyukova, L. R. Abuladze, Л. А. Низовцова, Ivan A. Blokhin, Anton V. Vladzymyrskyy, Yuriy А. VasilevAsian 2025 · DOI

TL;DR

Researchers leveraged a massive influx of chest CT scans from the COVID-19 pandemic to demonstrate that AI can "opportunistically" screen for osteoporosis with high accuracy. By applying a deep-learning algorithm to routine scans intended for lung analysis, the study identified vertebral fractures and low bone mineral density in over a quarter of the high-risk patient cohort—pathologies that radiologists frequently overlook during standard clinical reviews.

What Problem Does This Paper Address?

Osteoporosis is often called a "silent" disease because it remains asymptomatic until a debilitating fracture occurs. While Dual-energy X-ray absorptiometry (DXA) is the gold standard for diagnosis, it isn't always feasible for population-wide screening due to cost and access.

Meanwhile, millions of chest Computed Tomography (CT) scans are performed annually for other reasons (like pneumonia or lung cancer screening). These scans capture the thoracic and upper lumbar spine, yet radiologists—focused on the lungs or heart—rarely report on bone health unless a fracture is glaringly obvious. This paper explores whether an AI-driven "opportunistic" approach can turn these existing images into a powerful, no-cost screening tool for bone health.

Methods

The researchers conducted a retrospective multi-center study in Russia, analyzing 1,888 patients over the age of 50. They employed a specialized AI algorithm called Genant-IRA, which uses a two-stage pipeline:

  1. Localization and Straightening: A 3D UNet-like architecture identifies the spinal column and "straightens" it into a 2D curved reconstruction.
  2. Morphometric Analysis: Using a YOLOv3-based architecture, the AI detects individual vertebral bodies and calculates a compression deformity index based on height loss.
  3. Asynchronous Quantitative CT (QCT): To measure Bone Mineral Density (BMD), the team used a phantom-calibrated method to convert Hounsfield Units (HU) from the CT scans into actual mineral density (mg/mL). This allowed them to diagnose osteoporosis (<80 mg/mL) without needing a calibration phantom present during every single patient scan.

Datasets

The AI model was developed and validated using several distinct datasets:

  • Training: 100 images from the Moscow Radiology CTLungCa-500 dataset, a private dataset of 402 annotated chest/abdominal CTs, and 191 additional scans for spine localization.
  • Testing/Benchmarking: The publicly available VerSe-2020 dataset (300+ scans) and the Mosmed.ai test dataset (100 studies).
  • Study Cohort: 1,888 routine chest CTs from seven medical facilities in Moscow, collected during June 2021.

Key Insights and Ideas

  • The "Invisible" Fracture: One of the most striking findings was that radiologists only correctly identified vertebral fractures (VF) in 19% of the cases during routine clinical reading. The AI acts as a "second pair of eyes" that doesn't get distracted by lung pathology.
  • The T11-L1 Sweet Spot: While the lumbar spine is the traditional site for BMD measurement, chest CTs often cut off at L1 or L2. The study found that the T11 vertebra is a highly reliable alternative for opportunistic screening, as it was present in 96% of the scans.
  • Morphometry vs. Probability: Unlike many AI tools that simply give a "probability" of a fracture, this model performs automated morphometry—measuring the actual height of the vertebrae. This mimics the rigorous Genant classification used by human experts, making the results more interpretable for clinicians.

Key Results

  • Prevalence: The AI detected vertebral compression deformities (>25% height loss) in 17.8% of the patients.
  • Bone Health: Low BMD (<80 mg/mL), indicating osteoporosis, was found in 26.5% of the total cohort (31.0% of females and 18.6% of males).
  • AI Performance: The algorithm demonstrated a sensitivity of 0.982 and an overall accuracy of 0.874 for detecting vertebral fractures when compared against a ground truth established by expert radiologists with decades of experience.
  • Clinical Impact: The study suggests that implementing this AI could identify roughly 21,000 patients with fractures and 68,000 with osteoporosis annually in Moscow alone, based on current CT volumes.

Limitations and Future Work

  • Anatomical Coverage: Chest CTs do not include the lower lumbar spine (L3-L5), which are the areas under the greatest mechanical load and the preferred sites for standard BMD measurement.
  • False Positives: The authors noted that the algorithm occasionally struggles to differentiate between true osteoporotic fractures and other deformities like Schmorl’s nodes or congenital anomalies.
  • Technical Variability: HU values can be affected by the "sharp" kernels used for lung imaging or low-dose CT protocols, which may slightly skew BMD calculations.
  • Future Direction: The researchers aim to integrate additional data—such as age, gender, and clinical history—into the AI model to further improve its specificity and help clinicians better stratify high-risk cohorts for follow-up DXA scans.

Reference: Petraikin, A. V., et al. (2025). Opportunistic screening for osteoporosis using artificial intelligence-based morphometric analysis of chest computed tomography images. Asian Spine Journal, 19(3), 355-371.

Summary updated

Abstract

We observed high diagnostic accuracy of opportunistic osteoporosis screening using AI algorithms for detecting VF and low BMD. AI-based opportunistic screening of osteoporosis and VF in chest CT scans can facilitate the identification of high-risk cohorts.
BibTeX
@inproceedings{petraikin2025opportunistic,
  author = {A. V. Petraikin and Perry J. Pickhardt and Mikhail Belyaev and Zhanna Belaya and Maxim Pisov and Alim Bukharaev and Aleksei Zakharov and Nikita D. Kudryavtsev and Tatiana Bobrovskaya and Д. С. Семенов and Е. С. Ахмад and Z. R. Artyukova and L. R. Abuladze and Л. А. Низовцова and Ivan A. Blokhin and Anton V. Vladzymyrskyy and Yuriy А. Vasilev},
  title = {{Opportunistic screening for osteoporosis using artificial intelligence-based morphometric analysis of chest computed tomography images: a retrospective multi-center study in Russia leveraging the COVID-19 pandemic}},
  booktitle = {Asian},
  year = {2025},
  doi = {10.31616/asj.2024.0314},
}
Figure