Predicting skeletal age from HR-pQCT imaging.
Authors
Affiliations (2)
Affiliations (2)
- McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary AB, Canada; Department of Biomedical Engineering, Schulich School of Engineering, University of Calgary, Calgary AB, Canada.
- McCaig Institute for Bone and Joint Health, Cumming School of Medicine, University of Calgary, Calgary AB, Canada; Department of Biomedical Engineering, Schulich School of Engineering, University of Calgary, Calgary AB, Canada; Department of Radiology, Cumming School of Medicine, University of Calgary, Calgary AB, Canada. Electronic address: [email protected].
Abstract
High-resolution peripheral quantitative computed tomography (HR-pQCT) provides detailed bone microarchitecture assessments, but the interpretability of its many complex parameters remains challenging. This study aimed to develop a deep learning model to estimate skeletal age from HR-pQCT scans, offering an interpretable, quantitative summary of bone health relative to chronological age. The training dataset included 1236 adults (62.1% female) from a normative cohort, and an independent test set of 460 adults (69.3% female). HR-pQCT scans of the distal radius and tibia were acquired for all participants. Five models were trained: 2D models using a single radius (2DRad) and tibia (2DTib) slice from the middle of the scan; 3D models using full volumetric radius (3DRad) and tibia (3DTib); and a combined 2D model (2DRadTib) applying linear regression to the 2D outputs. The 2DRadTib model achieved the best performance, with a validation mean absolute error (MAE) of 5.29 ± 4.60 years (R<sup>2</sup> = 0.85) and test MAE of 5.34 ± 4.38 years (R<sup>2</sup> = 0.85). Saliency maps revealed cortical bone was most influential in younger individuals, while both cortical and trabecular features contributed in older participants. Predicted skeletal age was strongly correlated with established HR-pQCT parameters, particularly cortical and density measures (ρ = -0.51 to 0.85), indicating the model relies on key bone features. We present a novel deep learning framework for skeletal age prediction from HR-pQCT, providing a concise and interpretable summary measure of bone health. This approach may enhance the clinical utility of HR-pQCT by improving interpretability and supporting early identification of accelerated skeletal aging. High-resolution peripheral quantitative computed tomography (HR-pQCT) provides detailed images of bone structure, but the volume and complexity of data can make interpretation difficult. In this study, we developed a deep learning model to estimate skeletal age from HR-pQCT scans, offering a simplified and interpretable measure of bone health. By translating complex imaging data into an age-based summary, this approach may enhance clinical use of HR-pQCT, support early identification of individuals at risk of accelerated bone loss, and improve patient understanding by providing a relatable measure of skeletal integrity.