One-Step Material Decomposition for Photon-Counting CT Using Implicit Neural Representation and Physics-Guided Model

Photon-counting computed tomography (PCCT) demonstrates significant potential for advancing clinical applications through its ability to precisely measure individual X-ray photon energies, enabling enhanced material and tissue differentiation. The technique employs multi-energy bin projections for material decomposition-a complex nonlinear, nonconvex inverse problem that is often compromised by physical non-idealities, including charge-sharing and pulse pile-up effects in photon-counting detector (PCD) and application-specific integrated circuit (ASIC) responses. Material decomposition algorithms are categorized into three types: image-domain, projection-domain, and one-step inversion decomposition. One-step inversion decomposition, though computationally intensive, provides a more comprehensive solution by integrating decomposition and reconstruction within a unified optimization framework. Recent advances in Implicit Neural Representation (INR) have introduced novel approaches to address the reconstruction inverse problem, offering a promising framework for one-step material decomposition. This study implements INR methodology for one-step inversion decomposition by developing a direct mapping between position distributions and basic material length fractions, while incorporating a physics-guided detector model to enhance system response characterization and decomposition accuracy. Experimental results from water and Gammex phantom studies validate the method's capability for one-step material decomposition. However, further optimization is necessary to enhance decomposition accuracy and mitigate artifacts visible in narrow clinical display windows to meet diagnostic quality standards.