Detector nonlinearity in measuring the bone mineral density based on neural networks

Objective.A dual-layer flat-panel detector (DFD) allows for the acquisition of dual-energy images with a single x-ray exposure without scanning. In this paper, we aim to perform noise sensitivity analyses on a single-shot approach using DFD in measuring bone mineral density (BMD) and propose a method for calculating BMD using features acquired from a nonlinear response detector without a nonlinearity correction scheme.Approach.We evaluate the condition numbers of the BMD estimate functions and observe the mean squared error (MSE) for ranges of additive and multiplicative errors to compare with a dual-shot FD approach. To accurately describe the BMD surface, we use a fully connected neural network (FCNN) model instead of the conventional multiple regression model. We conduct experiments to observe and model the nonlinear response of DFD, followed by simulations based on a BMD model with nonlinear intensities.Main results.Direct BMD estimation from uncorrected nonlinear intensities using the FCNN produced lower estimation errors than the conventional second-order polynomial estimator. For Case S-sub at a multiplicative noise variance of, the FCNN reduced the BMD estimation MSE fromto.Significance.The results support the feasibility of direct FCNN-based BMD estimation from nonlinear DFD signals without applying an explicit nonlinearity correction. The proposed approach may support single-exposure BMD pre-screening, subject to further validation using physical phantoms and clinical data.