HarDNet-based deep learning model for osteoporosis screening and bone mineral density inference from hand radiographs

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Bone Health and Osteoporosis Research

Background:

• Osteoporosis affects over 200 million people globally, often undetected, leading to increased fracture risk in older adults.
• Diagnosis typically relies on bone mineral density (BMD) measured by dual-energy X-ray absorptiometry (DXA).
• There is a need for accessible and opportunistic screening methods for osteoporosis.

Purpose of the Study:

• To develop and validate DeepDXA-Hand, a deep learning model for opportunistic osteoporosis screening using hand radiographs.
• To assess the model's ability to predict bone mineral density (BMD) non-invasively.
• To evaluate the clinical potential of the model for early osteoporosis detection.

Main Methods:

• Utilized a Convolutional Neural Network (CNN)-based HarDNet architecture for BMD prediction.
• Trained and validated the model on 10,351 hand radiographs and corresponding DXA scan pairs.
• Employed GradCAM for hotspot analysis to enhance model interpretability and identify key areas of focus.

Main Results:

• Achieved a significant correlation (0.745) between predicted and ground truth BMD.
• Demonstrated high performance in binary classification of osteoporosis: 0.73 sensitivity, 0.83 specificity, and 0.80 accuracy.
• Identified carpal bones (capitate, trapezoid, hamate, triquetrum) and the second metacarpal head as key areas for BMD inference.

Conclusions:

• DeepDXA-Hand shows significant potential as an opportunistic screening tool for early osteoporosis detection.
• The model exhibits high sensitivity and specificity, suggesting clinical utility.
• Further research is warranted to explore its efficacy in predicting fracture risk.