Prediction of Elastic Behavior of Human Trabecular Bone Using A DXA Image-Based Deep Learning Model

Inspired by the recent advancement in deep learning (DL) techniques, this study intended to confirm whether DL models could be trained to predict the elastic behavior of trabecular bone, a highly hierarchical biological material, using its dual-energy x-ray absorptiometry (DXA) images. The convolutional neural network, the most successful DL model in imaging-based predictions, was trained using simulated DXA images of trabecular bone samples as input and their apparent elastic modulus (Eapparent) determined using microCT-based finite element simulations as output (label). The results showed that the DL model achieved high fidelity in predicting Eapparent of trabecular bone samples (R2 > 0.86), and its performance appeared to be better than that of histomorphometric parameter-based regression models built using the same bone samples. The outcome of this study suggests that DXA image-based DL techniques can be used for multiscale modeling of trabecular bone to predict its elastic behavior.