Development of AI model for dual detection of low bone mineral density in the femoral neck and lumbar vertebrae using chest radiographs

Abstract

Introduction: Artificial intelligence (AI) technologies have demonstrated high accuracy in detecting overall osteoporosis on chest radiographs, offering significant potential for rapid and accessible osteoporosis screening. However, as bone loss varies by lifestyle and body shape, detecting low bone mineral density (BMD) in specific parts is crucial for early treatment. This study developed and evaluated two deep learning models to detect low BMD in the femoral neck and lumbar vertebrae.

Methods: Data included chest radiographs and dual-energy X-ray absorptiometry (DXA)-measured BMD values [g/cm²] of 2,728 female examinees. Chest radiographs were categorized into low BMD or normal based on the femoral neck (low: 1,358, normal: 1,370) and lumbar vertebrae (low: 562, normal: 2,166). Deep learning models were trained using the ResNet50 architecture with fine-tuning and 10-fold cross-validation. Performance metrics included sensitivity, specificity, overall accuracy, and area under the curve (AUC). Heatmaps generated using Explainable AI visualized regions related to low BMD.

Results: The model achieved 75.3 % overall accuracy (AUC: 0.82) for femoral neck detection and 89.3 % (AUC: 0.96) for lumbar vertebrae detection. Lumbar vertebrae detection showed 14.0 % higher accuracy than the femoral neck. Patients with lumbar vertebrae low BMD exhibited more advanced bone loss compared to those with femoral neck low BMD alone. Heatmaps indicated relevant regions near the clavicle and thoracic vertebrae.

Conclusion: The proposed model accurately detected low BMD in chest radiographs and identified areas of bone loss, demonstrating particularly high performance in lumbar vertebrae detection. Early identification of low BMD enables simple, effective screening and targeted prevention or treatment based on areas of bone loss.