U-Net-Based Deep Learning Hybrid Model: Research and Evaluation for Precise Prediction of Spinal Bone Density on Abdominal Radiographs

Abstract

Osteoporosis is a metabolic bone disorder characterized by the progressive loss of bone mass, which significantly increases the risk of fractures. While dual-energy X-ray absorptiometry is the standard technique for assessing bone mineral density, its use is limited in high-risk female populations. Additionally, quantitative computed tomography offers three-dimensional evaluations of bone mineral density but is costly and prone to motion artifacts. To overcome these limitations, this study proposes a hybrid model integrating U-Net and artificial neural networks, specifically focusing on abdominal X-ray images in the anteroposterior view for detailed skeletal analysis and improved accuracy in L2 vertebra mineral density measurement. The model targets female patients, who are at a higher risk for spinal disorders and osteoporosis. The U-Net model is employed for image preprocessing to reduce background noise and enhance bone tissue features, followed by analysis with the artificial neural network model to predict bone mineral density through nonlinear regression. The performance of the model, demonstrated by a high correlation coefficient of 0.77 and a low mean absolute error of 0.08 g per square centimeter, highlights its significance and effectiveness, particularly in comparison to dual-energy X-ray absorptiometry.