Gradient-enhanced neural network and extreme gradient boosting modeling for the prediction of the 3D bone mineral density distribution from 2D-DXA scans.

Objectives: This study aims to predict the volumetric bone mineral density (BMD) distribution from a dual-energy X-ray absorptiometry (DXA) scan. By employing machine learning, this study bridges the gap between DXA and computed tomography (CT) in terms of volumetric bone assessment, suggesting an approach for a cost-effective and low-radiation alternative for bone health evaluation in a three-dimensional (3D) fashion from a two-dimensional (2D) scan. Methods: Data from 34 participants included aligned DXA and CT scans for the proximal femur. Intensity values were extracted in Hounsfield units, with 3D information mapped as target variables and 2D information as features. Two machine learning models, Extreme Gradient Boosting (XGB) and Gradient-Enhanced Neural Network (GENN), were trained using 5-fold cross-validation strategy and show an average registration accuracy of 0.89 ± 0.04, assessed thorough structural similarity index measure. Results: Both models were built to predict the statistics of the 3D structure of the bone from a 2D image. The GENN model outperformed XGB, achieving mean absolute percentage errors (MAPE) of 12.98 ± 1.70%, 13.28 ± 2.01%, and 9.63 ± 1.66% for minimum, maximum, and the number of nonzero pixel intensities, respectively. In contrast, XGB's errors exceeded 16% across these metrics. The loss stabilized within 100 epochs, indicating model robustness and reliability across diverse test sets. Conclusions: The proposed GENN framework offers a method for predicting 3D BMD distributions from a 2D-DXA scan, rivaling CT-based assessments. This approach reduces costs and radiation exposure, presenting a viable solution for personalized bone health evaluation and early osteoporosis diagnosis.