X-ray Fluorescence Computed Tomography (XFCT) enables element-specific and high-sensitivity imaging, offering distinct advantages in nanoparticle tracking and cancer diagnosis. However, current benchtop XFCT systems using polychromatic Xray sources suffer from prolonged acquisition times. While sparse-view acquisition strategy can accelerate scanning, they inevitably lead to image quality degradation due to data insufficiency. To address this challenge, we propose Diff-PCGS, a novel reconstruction framework that synergizes generative diffusion models with physics-constrained 3D Gaussian Splatting.First, we introduce a diffusion model in the projection domain to suppress noise while preserving structural information.Subsequently, an iterative 3D Gaussian Splatting reconstruction is performed on the denoised projections, incorporating constraints based on photon attenuation and propagation physics to enhance physical consistency and accuracy. Validation on both simulation data and phantom experiments demonstrates that Diff-PCGS maintains superior reconstruction quality when reducing the number of projection angles from 45 to 5, outperforming conventional iterative methods in image fidelity, noise robustness, and quantitative accuracy. The framework's practical utility is further confirmed through in vivo sparse-view imaging of orthotopic liver tumors in mice. In summary, this study presents an effective solution for high-fidelity fast XFCT imaging, enabling accurate element-specific imaging in practical scenarios.
Pu et al. (Wed,) studied this question.