Physics-Constrained Deep Learning for Effective Atomic Number and Density Calculation of Biological Tissues in Photon-Counting Spectral CT.
Authors
Abstract
Photon-Counting Detector Computed Tomography (PCD-CT) boasts excellent spectral utilization capability. Combined with material decomposition methods, it enables the calculation of the effective atomic number ($Z_{eff}$) and density ($\rho$) of scanned materials. Traditional material decomposition methods derive $Z_{eff}$ and $\rho$ from physical models using either the basis material model or the dual-effect model. However, these methods generally fail to meet high-precision decomposition requirements due to model approximations and suffer from the limitation of severe noise amplification in $Z_{eff}$ images. This study proposes a physics-constrained deep learning network that achieves high-precision joint estimation of $Z_{eff}$ and $\rho$ by dynamically modeling nonlinear X-ray interactions. A Multilayer Perceptron (MLP) is employed to construct dynamic compensation functions dependent on energy ($E$) and $Z_{eff}$ for the photoelectric effect exponent and Compton scattering model in the X-ray interaction model. The decomposition network adopts Swin-Unet as its backbone and utilizes a hybrid loss function, which consists of L1 loss and SSIM loss for $Z_{eff}$/$\rho$, as well as a physics-informed loss derived from the L1 loss between the predicted $Z_{eff}$/$\rho$ and the monoenergetic linear attenuation coefficient images generated by the optimized X-ray model. This design allows the network to simultaneously learn data-driven features and physical principles. Comparative experiments were conducted on a PCD-CT system between the proposed method and four methods (Lan, U-Net, Butterfly-net, and Swin-Unet without physical constraints). The results demonstrate that: for standard materials, the proposed method achieves a Mean Absolute Percentage Error (MAPE) below 5% for both $Z_{eff}$ and $\rho$ decomposition, with superior Noise Power Spectrum (NPS) performance; for biological samples including freshwater crayfish and mouse, the $Z_{eff}$ images generated by the proposed method exhibit higher Multi-Exposure Fusion Structural Similarity Index (MEF-SSIM), reaching 0.9559 for crayfish and 0.8950 for mouse. The method also demonstrates superior detail recovery capability in the restoration of the speckled tissue structure of crayfish and the tissue regions of mouse. By incorporating constraints from monoenergetic images generated based on the dynamic X-ray interaction model, the network effectively learns the nonlinear decomposition process of $Z_{eff}$ within a data-driven framework. The proposed method improves decomposition accuracy while reducing decomposition noise and enhancing the quality of decomposed images.