Osteoporosis is a widespread condition that often remains undiagnosed until a fracture occurs. It significantly affects health, particularly in ageing populations, and contributes to increased morbidity, healthcare costs and reduced quality of life. Dual-energy X-ray absorptiometry (DXA) is the clinical standard for diagnosing osteoporosis, but quantitative computed tomography (QCT) offers higher sensitivity by measuring volumetric bone mineral density (BMD). However, QCT is underused due to its cost, radiation exposure and technical requirements.
With lung cancer screening programmes increasingly using low-dose chest CT, there is a potential to combine cancer detection and opportunistic osteoporosis screening using a single scan. Leveraging artificial intelligence, researchers explored whether deep learning could enable accurate BMD measurement from ultralow-voltage 80 kV chest CT images, making osteoporosis screening more accessible and efficient.
Optimising Low-Dose CT for Bone Health Assessment
Chest CT scans provide detailed anatomical data with inherently high contrast, allowing for reduced radiation dose through lower tube voltage. A key advancement in the current approach was the reduction of the voltage to 80 kV, significantly decreasing radiation exposure by over 90 percent compared to standard 120 kV QCT. Such a reduction is especially beneficial in patients with lower body mass index, where reduced X-ray attenuation facilitates imaging quality even at lower energy settings. Using a constant tube current of 500 mA helped maintain image quality despite the drop in voltage. These ultralow-voltage chest CT scans, routinely acquired for lung cancer screening, offer a valuable chance to detect osteoporosis without the need for dedicated QCT protocols.
Must read: DL for Osteoporosis Detection Using Chest CT Scans
The combination of osteoporosis and lung cancer screening has clinical and logistical appeal. Both diseases are prevalent among older adults, and using a single scan to assess lung health and bone density conserves resources and minimises patient exposure. By targeting vertebrae T12 to L2, which are typically within the scan range of chest CT, meaningful BMD estimates can be obtained without extending the scan field. This approach aligns with the principle of minimising radiation exposure while maximising diagnostic output.
Deep Learning Pipeline for Fully Automated BMD Measurement
To enable automated BMD assessment, a three-stage deep learning pipeline was developed. The first stage involved vertebral body segmentation, using two convolutional neural networks: 3D VB-Net and Spatial Configuration Net (SCN). These models worked together in a coarse-to-fine framework to accurately locate and identify individual vertebrae. By combining localisation and fine segmentation, the models overcame challenges related to structural similarities between vertebrae and surrounding anatomical noise. Once the vertebrae were isolated and aligned, they were cropped into uniform volumes to ensure consistency across scans.
The second stage focused on extracting regions of interest within the trabecular bone. Using another instance of the 3D VB-Net, elliptical volumes were generated to approximate cylindrical sampling regions, similar to those used in conventional QCT. These regions avoided cortical bone, blood vessels and pathological features, enhancing the reliability of BMD estimates. The segmented vertebrae and corresponding regions of interest were then fed into the final stage: BMD calculation using regression models. Two popular network architectures, DenseNet and ResNet, were employed in parallel. Each model received either full vertebral images or extracted regions as input, producing BMD values directly from the CT images.
This automated approach aimed to replace manual annotation and measurement, which are time-consuming and subject to operator variability. By using deep learning to streamline the process, the method increases the feasibility of applying BMD screening at scale. Data augmentation techniques, such as image rotation and noise injection, enhanced the model’s robustness to differences in scanner hardware, patient positioning and image quality.
Performance Across Devices and Diagnostic Categories
The workflow was tested on CT scans from 987 patients acquired on six scanners from different manufacturers. Patients were grouped into training, validation and test sets. The CNN models demonstrated high accuracy across all sets, with relative measurement errors remaining within ±2 percent. Agreement with reference BMD values from standard 120 kV QCT was strong, as confirmed by linear regression and Bland–Altman analysis. Compared with 80 kV QCT alone, the CNN-based estimates showed better correlation and lower error margins, particularly in patients with normal or mildly reduced bone density.
Diagnostic accuracy was also high. In identifying osteoporosis, the models achieved area under the curve (AUC) values above 0.997, with sensitivity and specificity both exceeding 95 percent. When distinguishing low BMD from normal values, the models maintained similar performance. Notably, images containing entire vertebrae produced slightly better results than those based on extracted regions. This suggests that including broader anatomical context helps the models focus on relevant structures without overfitting.
The consistency of results across different CT systems and two independent test sets highlights the adaptability of the approach. The method is especially well suited to settings where lung cancer screening is already in place, allowing healthcare providers to screen for osteoporosis without increasing scanning time or radiation exposure.
A deep learning-based pipeline can enable accurate, fully automated measurement of bone mineral density from ultralow-voltage 80 kV chest CT scans. By integrating segmentation and regression models, the method achieves diagnostic performance comparable to traditional QCT, while significantly reducing radiation dose. With the growing use of low-dose CT for lung cancer screening, there is a timely opportunity to expand its utility to include osteoporosis detection. This approach offers a scalable, cost-effective solution for early identification of patients at risk of bone loss and fragility fractures, supporting preventative care and improving outcomes. Further research may refine the models, extend their application to broader populations and incorporate fracture risk prediction.
Source: Academic Radiology
Image Credit: iStock
