An Improved Phase-Contrast X-Ray Tomography Based on Dual-Domain Combined Deep Learning

To address the low contrast of X-ray weakly absorbing materials, phase retrieval (PR) is employed for contrast enhancement and material identification but inevitably leads to structural blurring and boundary smoothing due to its equivalent low-pass filtering. A space–frequency physics-guided neural network (SF-PGNet) is proposed to improve the reconstruction quality of phase-sensitive hard X-ray tomography without additional hardware assistance for weakly absorbing materials. It consists of three sub-networks: (i) the spatial-domain network (SD-Net) predicts a projection-conditioned effective deconvolution kernel (PSF_N) from the measured projection, using the experimentally calibrated effective kernel PSFtot as physical supervision, and applies PSF_N to obtain a sharpness-enhanced sub-image; (ii) the frequency-domain network (FD-Net) predicts the inverse transfer function HN−1 from the measured projection, using analytical H−1(γ) functions generated from physically selected phase-retrieval parameters as supervision, and generates a phase-contrast-enhanced sub-image; and (iii) the dual-domain fusion network integrates the complementary sub-images to produce the final reconstruction. Validation on a polymer phantom and testing on a biological specimen demonstrate that SF-PGNet enhances material contrast and boundary representation, while the learned spatial- and frequency-domain branches remain highly consistent with their corresponding physics-based correction references.