Advancing Opportunistic Screening of Stratified Bone Mineral Density Disorders Using Deep Learning on Chest X-ray

Objective: 

Bone mineral density (BMD) disorders, including osteoporosis and its osteopenia subtypes (mild, moderate, advanced), are highly prevalent conditions that significantly increase the risk of fractures if left undiagnosed or untreated. While Dual-Energy X-ray Absorptiometry (DXA) remains the gold standard for BMD assessment, its high cost and limited accessibility hinder its use in opportunistic screening. This study aims to develop a deep learning model leveraging chest X-rays (CXRs) to classify BMD into normal BMD or osteoporosis categories and further stratify osteopenia into specific subtypes. By enabling precise and early detection, this approach seeks to enhance bone health management through cost-effective and scalable solutions. 

Materials and Methods: 

We retrospectively collected 69,201 CXR images and corresponding DXA scans from a single tertiary care hospital in South Korea. BMD was categorized using DXA T-scores into normal BMD (≥ -1.0), mild osteopenia (-1.0 to -1.49), moderate osteopenia (-1.5 to -1.99), advanced osteopenia (-2.0 to -2.49), and osteoporosis (≤ -2.5). The dataset was split into 80% for training the deep learning model and 20% for internal validation. For external validation, an additional 3,338 chest X-ray images were acquired from an external institution to evaluate the model’s generalizability. We developed a cascaded deep learning architecture that first classifies each chest X-ray as normal BMD, osteopenia, or osteoporosis, and then further stratifies osteopenia cases into mild, moderate, and advanced subtypes. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) to ensure accurate classification and stratification of BMD disorders.

Results: 

For internal validation, the model achieved AUC scores of 0.934, 0.947, 0.738, 0.804, and 0.867 for normal BMD, osteoporosis, and osteopenia subtypes (mild, moderate, advanced), respectively. Similarly, during external validation, the model demonstrated consistent performance, achieving AUC scores of 0.868 for normal BMD, 0.899 for osteoporosis, and 0.652, 0.754, and 0.819 for osteopenia subtypes.

Conclusion: 

The deep learning model effectively classifies BMD into normal BMD, osteoporosis, and stratifies  osteopenia into mild, moderate, and advanced subtypes using CXRs. This detailed stratification enables targeted interventions and early prevention of osteoporosis progression, improving patient outcomes and optimizing healthcare resources through scalable and efficient bone health monitoring.