Development of an automated landmarking tool for the femur in dual-energy X-ray absorptiometry scans using...

Osteoporosis is a skeletal disease that significantly increases fracture risk and imposes a growing public health and economic burden. Notably, hip fractures are associated with high mortality, long-term disability, and loss of independence. When evaluating osteoporosis and estimating fracture risk, DXA is commonly used to determine the BMD. To capture geometrical morphology from these DXA scans as factors in fracture risk prediction, landmarking around the femur is conducted manually around the ROI. However, this process is labor-intensive and prone to variability. This study presents the development of a fully automated femoral landmarking tool that integrates a U-Net convolutional neural network for femur contour segmentation with a geometric algorithm for consistent landmark placement. A heterogeneous dataset of 555 DXA scans was used to train and evaluate the U-Net model, achieving a pixel-wise accuracy of 97.55%, an Intersection over Union of 91.55%, and a Dice coefficient of 95.56% on the test set. When incorporated into a statistical shape and appearance modeling (SSAM) framework for fracture risk prediction, predictions using the automatically generated landmarks achieved a test AUC of 0.831 (95% CI: 0.698-0.965), compared with 0.780 (95% CI: 0.635-0.925) for manual landmarks; a paired DeLong test showed no significant difference (p = .53), indicating a comparable performance. The proposed pipeline produces anatomically relevant, reproducible landmarks, and supports fracture prediction performance similar to manual methods. It presents a scalable and objective solution for morphometric analysis in DXA imaging.