Deep Learning-Based Classification of Osteoporosis and Osteopenia from Chest Radiographs: A Multi-Center Validation Study

Purpose or Learning Objective

Osteoporosis significantly affects the elderly, characterized by decreased bone density and increased fracture risk. Osteopenia, a precursor to osteoporosis, involves lower-than-normal bone density and can lead to similar health issues if untreated. Diagnosis typically relies on DXA scans, which are often inaccessible and costly. To address this, we aim to develop a deep learning model that utilizes readily available chest X-ray (CXR) images to classify osteoporosis and osteopenia effectively, providing a practical solution for early screening and intervention.

Methods or Background

We conducted training using a self-distillation method with 69,201 CXR images paired with corresponding DXA scores obtained from a specific hospital. The CXR images were pre-processed and used as input data for the model. Subsequently, the model's performance was validated using datasets from various external institutions.

Results or Findings

In the internal validation at hospital A, the AUC scores for the classes normal, osteopenia, and osteoporosis were 0.92, 0.76, and 0.91. In an external validation conducted at the same institution using differently collected data, the AUC scores were 0.94, 0.75, and 0.87. Additionally, in external institutions, Hospital B and Hospital C reported AUC scores of 0.85, 0.70, 0.89 and 0.82, 0.63, 0.78, respectively.

Conclusion

This study introduces a deep learning model capable of classifying osteopenia and osteoporosis based on CXR images. The model was validated using data from multiple external institutions, confirming that CXR images can also be effectively utilized for the classification of osteopenia and osteoporosis.

Limitations

One challenge we encountered was the difficulty in obtaining paired DXA and CXR image data, which hinders the detection of underrepresented classes like osteoporosis and osteopenia. We believe that increased collection of this paired data would enhance model training and improve performance outcomes.

Funding for this study

No funding was received for this study