Osteoporosis is a significant health concern worldwide, particularly among the aging population. It is characterized by low bone mineral density (BMD) and deteriorated bone microarchitecture, leading to an increased susceptibility to fractures [1]. Postmenopausal women and men aged over 50 years are at a high risk of osteoporosis [2]. The prevalence of osteoporosis is expected to rise due to increasing life expectancy and the aging population.

Dual-energy X-ray absorptiometry (DXA) is widely used for assessing BMD and diagnosing osteoporosis based on T-scores, as recommended by the World Health Organization [1]. However, some studies have suggested that routine DXA screening for at-risk individuals may not be economically feasible [2]. Vertebral fractures (VF) are another indicator of osteoporosis, detectable through various imaging modalities [3]. VFs are typically underreported [4] due to patients not exhibiting specific symptoms and radiologists overlooking vertebral abnormalities on computed tomography (CT) scans [5]. Some studies suggest an opportunistic screening approach for osteoporosis, leveraging CT scans performed for other clinical indications [6].

Opportunistic CT screening for osteoporosis utilizes various approaches. These include qualitative models for VF detection [7], morphometric models for vertebral height measurement [8], and prognostic models for determining repeated fractures [9]. Artificial intelligence (AI) algorithms are also employed for determining regions of interest (ROI) inside the trabecular bone and assessing BMD or Hounsfield units (HU) [10]. Additionally, indirect BMD assessment using HU values [6,11], fully automated bone mineral densitometry following an asynchronous calibration [12], and automated BMD assessment with texture analysis [13] are used. A recent review analyzed different AI solutions for osteoporosis, focusing on VF detection and osteoporosis screening [14].

In individuals at risk (men and women over 50 years old), a decline in BMD is more frequently observed than VF. While VF occur in only 7% of osteoporotic cases [15], low BMD is found in 28.99% of women and 13.53% of men [16]. Therefore, opportunistic screening for osteoporosis using chest CT data, combined with the application of AI for VF detection [6,8,17] and ROI selection for BMD measurement [16], may maximize the effectiveness.

A previous paper proposed a morphometric algorithm for VF detection using convolutional neural networks (CNN), allowing quantification of vertebral morphometry [8]. This algorithm was updated to allow for the selection of the ROIs within the vertebrae and measurement of HU values, which are correlated with BMD [18]. As part of the project “Experiment on the use of innovative computer vision technologies for medical image analysis and subsequent applicability in the healthcare system of Moscow” (Experiment), this algorithm processed 59,482 CT chest studies within the Moscow Unified Radiological Information System (URIS) system over a period of 7 months (01.10.2021–29.04.2022).

This study aimed to confirm and evaluate the performance of opportunistic osteoporosis screening using AI in chest CT scans. The AI algorithm assessed both VF and low trabecular BMD. The objective was to determine the efficiency of this AI-enabled opportunistic screening approach, which combines VF detection and BMD measurement.

In this study, opportunistic screening of osteoporosis using the AI algorithm for VF detection and BMD quantification showed satisfactory performance. The overall prevalence of VF, as detected by the AI algorithm, was 5.5% (4.9% in men and 5.9% in women) (Table 1). Our findings are consistent with the incidence of VF reported in a previous study using abdominal CT scans for opportunistic screening in a different patient sample, where VFs were detected in 6.3% of patients from both genders [12]. In our study, the prevalence of VCD and VF in patients aged ≥50 years was 17.8% (Table 1). These findings are consistent with data from a European study spanning 19 countries [25], which reported an average VCD prevalence of 12% in both men and women aged 50–79 years using the morphometric method of McCloskey et al. [26] and Eastell et al. [27]. In previous Russian studies, VF was detected in 6.6%–6.9% of women and in 6.4%–10.2% of men aged over 50 years [28]. Additionally, Lesnyak et al. [15] reported VF detection rates of 7% in women and 7.2% in men [15]. The slightly lower relative number of detected VF in our study is probably related to the analysis being limited to the thoracic spine and the first lumbar vertebrae. The Th11 vertebra was included in the CT scanning length for 96% of cases but the L3 level was included in only 4%. The distribution of VF among vertebrae corresponds with data acquired by Fujiwara et al. [29], who reported the prevalence of VF cases for T11–T12 and L1 vertebrae. De Smet et al. [30] found a similar distribution of VCD among vertebrae, with the highest fractions for T7–T8 and T11–T12. Minor discrepancies have been observed in the accuracy metrics of the AI algorithm for detecting vertebrae in the thoracic and lumbar spine. The sensitivity was 0.95 (95% CI, 0.947–0.953) for the thoracic spine and 0.91 (95% CI, 0.905–0.916) for the lumbar spine. The specificity was 0.89 (95% CI, 0.887–0.895) for the thoracic spine and 0.94 (95% CI, 0.933–0.939) for the lumbar spine [31].

The importance of VF diagnosis and the differential diagnosis of VCD are highlighted in related work [24,28]. Osteoporosis diagnosis poses a challenge due to its asymptomatic nature, resulting in a large pool of patients with evidence of osteoporosis going undetected during routine CT studies. To address this, various approaches have been developed to perform opportunistic screening for VF and reduced BMD using existing CT data from other clinical indications [17,32,33], including CT coronary angiography for the T6–T9 thoracic spine [34]. The validity of this approach could be explained by comparing the potential number of VF in patients over 50 years old with cases of decreased BMD. While incidents of VF were observed in 7% of cases [15], a decrease in BMD to the level of osteoporosis was recorded in 28.99% of women and 13.53% of men [16]. Notably, our study identified 23 patients (1.2%) with VF and BMD >80 mg/mL, highlighting the importance of BMD measurement in avoiding misdiagnosis. Therefore, analyzing both VF and reduced BMD in chest CT imaging enhances the efficiency of opportunistic screening.

It is also possible to determine a decrease in BMD from HU data without CT scanner calibration. Alacreu et al. [11] showed that the criterion for diagnosing osteoporosis could be an L1 vertebra density <73 HU (90% probability), while “suspected osteoporosis” corresponded to 74< HU <116 and “norm” to >160 HU. Graffy et al. [19] substantiated the possibility of opportunistic osteoporosis screening using only HU values. Another study noted that densities below 100 HU indicate severe osteoporosis, with a VF probability of up to 31.9%. Others have suggested an L1 vertebra HU value of 110 as the threshold level for osteoporosis, with >90% specificity [35].

This study was performed during the COVID-19 pandemic using chest CT scans performed for various clinical indications worldwide. While radiologists primarily focus on lung pathology, leveraging opportunistic data on VF and BMD has enabled the detection of relevant, yet often unsuspected pathology [36]. In 2020, approximately 439,000 chest CT studies were performed in Moscow during the COVID-19 pandemic in patients aged over 50, corresponding to 7% of this age group. Considering the incidence of osteoporosis across the age groups of interest (7% with VF) [15], osteoporotic BMD in 13% of men and 29% of women [22], the introduction of AI-powered opportunistic screening for osteoporosis is expected to yield 21,000 patients with VF and 68,000 patients diagnosed with osteoporotic BMD annually. Moreover, radiologists accurately detect and report VF on CT images in only a minority of cases [4], although increasing awareness may lead to improvements in this area. Our data reveal that radiologists correctly interpreted VF during routine CT image readings in only 20 out of 105 patients (19.0%). These data underscore the importance of incorporating opportunistic screening capabilities into routine clinical practice.

In the future, the AI algorithm can serve as a double-reading tool for radiologists, similar to its suggested application in mammography [37]. After uploading a CT study into URIS, the AI algorithm automatically processes the CT images to detect any vertebral VCD and VF, while also estimating BMD values, generating a report in the DICOM Structured Reporting format. By leveraging the AI algorithm, radiologists can identify vertebrae with suspected VCD or reduced BMD, thereby minimizing false-negative results.

It is important to note that most qualitative AI algorithms only detect the presence of VF by indicating the probability of pathology, without specifying the compression deformity index. VF typically alters the vertebral shape, but not all deformities are caused by fractures [25]. Recent studies, such as Wang et al. [38], have demonstrated the potential of AI algorithms in diagnosing various types of vertebral diseases, including osteoporotic VF, old fractures, Schmorl’s node, Kummell’s disease, and previous surgeries. Although the results are promising, the algorithm struggled with accurately identifying Schmorl’s node [38]. In this context, morphometric AI algorithms offer greater accuracy and promise for opportunistic osteoporotic VF diagnosis. However, these models may lead to false-positive VF diagnoses and require further development, including improved identification of VF signs or incorporation of additional data (BMD, gender, age, and others) to enhance specificity.

Opportunistic screening of osteoporosis by chest CT has certain limitations. The most significant one is that chest CT scans do not include most of the vertebrae with the greatest mechanical load, specifically the lumbar spine. Only 59.1% of studies include the L1 and L2 vertebrae, which are recommended by the ISCD 2023 position for measuring BMD by 3D CT scans [19]. Additionally, not all studies can be correctly processed by AI algorithms for automatic BMD determination. Some scans may be performed on non-calibrated scanners or asynchronous scanners, which can affect the accuracy of BMD measurements. Furthermore, according to the ISCD 2023 position, an HU value of ≤100 is considered indicative of osteoporosis; however, using this threshold may reduce the accuracy of the methods [19]. Another limitation is the potential distortion of HU estimates when using the “sharp” kernel for visualization of the lung parenchyma. Additionally, HU values may be inaccurate when using low-dose or ultra-low-dose CT technologies. Moreover, the patient population referred for chest CT scans may differ from the general population, as they often have suspected chest pathology or undergo repeat studies. Therefore, opportunistic osteoporosis screening based on chest CT may not provide full population coverage for detecting osteoporosis. More effective approaches, such as the Fracture Risk Assessment Tool (FRAX) assessment followed by DXA, may be necessary for comprehensive osteoporosis screening. However, special epidemiological conditions, such as the COVID-19 pandemic, can significantly increase the number of chest CT scans performed, thereby enhancing the detection of osteoporosis when opportunistic patient screening methodologies are in place.

This study has several limitations. The data on the prevalence of VCD/VF are a byproduct of AI validation and, although performed opportunistically, only indirectly reflect real-world values. To quantify the population prevalence of these conditions more accurately, further investigations considering age group sizes and lumbar spine CT studies are necessary. Additionally, the L1 and L2 vertebrae, typically included in BMD assessments, were only present in a small percentage of chest CT scans. Therefore, T11 vertebrae were used as a more reliable alternative to determine BMD.

Our study demonstrates the efficiency of opportunistic osteoporosis screening using chest CT scans performed for diagnosing COVID-19-associated pneumonia, in conjunction with a morphometric AI algorithm. This approach enabled the detection of VF and reduced BMD in patients over 50 years of age. Opportunistic screening facilitates the identification of high-risk patients who require further evaluation to confirm the diagnosis. The AI algorithm can serve as a valuable second-reader tool, enhancing osteoporosis diagnostics by detecting both VF and reduced BMD.