Photon-Counting CT Reconstruction Using Separable Attention-Based Tensor Neural Network Prior

Photon-counting computed tomography (PCCT) enables multi-energy spectral imaging from a single scan. However, low photon count in each channel introduces significant noise, reduces spatial resolution, and limits fine detail, especially in low-dose or fine-structure imaging. To tackle these issues, we propose a Separable Attention-based Tensor Neural Network (SATNN) as a Bayesian reconstruction prior for PCCT. SATNN utilizes depthwise separable convolutions for efficient spatial feature extraction and squeeze-and-excitation blocks to enhance critical channel-wise features, allowing focused attention on key features. Following reconstruction, anisotropic diffusion filtering further refines images by suppressing residual noise while preserving textural details. Through realistic PCCT simulation studies, SATNN demonstrated superior performance over the existing algorithms, achieving improved texture preservation, spectral correlation, and noise suppression, making it a robust and effective solution for both research and clinical applications.