Predicting skeletal age from HR-pQCT imaging.

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Bone Biology

Background:

• High-resolution peripheral quantitative computed tomography (HR-pQCT) offers detailed bone microarchitecture insights but faces interpretability challenges.
• The complexity of HR-pQCT data hinders clinical application and patient understanding of bone health status.

Purpose of the Study:

• To develop and validate a deep learning model for estimating skeletal age from HR-pQCT scans.
• To provide an interpretable and quantitative measure of bone health relative to chronological age.

Main Methods:

• Trained deep learning models on HR-pQCT scans from a normative cohort (n=1236) and tested on an independent cohort (n=460).
• Evaluated 2D, 3D, and combined models using distal radius and tibia scans.
• Utilized saliency maps to interpret model reliance on specific bone features.

Main Results:

• The combined 2D model (2DRadTib) demonstrated strong performance with a test Mean Absolute Error (MAE) of 5.34 years (R²=0.85).
• Saliency maps indicated that cortical bone features were crucial for younger individuals, while both cortical and trabecular features were important for older individuals.
• Predicted skeletal age strongly correlated with established HR-pQCT parameters, particularly cortical and density measures (ρ = -0.51 to 0.85).

Conclusions:

• A novel deep learning framework accurately predicts skeletal age from HR-pQCT.
• This approach enhances the clinical utility of HR-pQCT by providing an interpretable bone health summary.
• The model facilitates early identification of accelerated skeletal aging and may improve patient comprehension of bone status.