Learning-Based Material Decomposition for Spectral X-Ray Imaging

Spectral computed tomography (CT) presents significant advantages by furnishing precise material information. However, the practical utility of this technique faces challenges arising from the instability or over-determination of the material decomposition model, thereby compromising the accuracy of material decomposition. Traditional methodologies encompass projection-based, image-based, and one-step inversion approaches, all adhering to the foundational solution principles of variational methods designed to address the intricacies of the nonlinear convex problem. Despite their robustness, the performance of these methods relies on a priori knowledge of the scanner’s energy response and the judicious selection of the regularization function. Moreover, their iterative nature may result in computational sluggishness. Recent advancements in deep-learning-based methodologies have demonstrated substantial efficacy in tackling multiple nonlinear challenges, showcasing their potential as effective tools for material decomposition in the realm of spectral computed tomography.