PhyDiff: Physics-guided latent diffusion model for sparse-view CT projection interpolation

Sparse-view Computed Tomography (SVCT) is an important strategy for minimizing radiation exposure and accelerating industrial inspection, yet it suffers from severe streak artifacts under angular undersampling. Existing sinogram completion methods often process detector rows independently, failing to exploit cross-row correlations to maintain the structural consistency required for high-fidelity reconstruction. To address this, we propose PhyDiff, a physics-guided latent diffusion framework that reformulates sparse-view projection completion as an angle-conditioned sequence interpolation. PhyDiff incorporates the known acquisition angles to condition cross-view alignment and exploit geometric correlations between adjacent projections. Specifically, a Physical-Geometric Autoencoder (PGAE) decomposes projection synthesis into geometry-guided warping and residual refinement. The warping model captures the dominant translation and scale variation between neighboring views, while the residual branch recovers spatially varying residuals that cannot be represented by the low-order transformation. Furthermore, a Spectral-Aware Latent Diffusion (SALD) model introduces angle-conditioned spectral modulation into the generative process to preserve fine structural details. Experimental results on simulated and measured datasets demonstrate that PhyDiff consistently improves projection fidelity under different acquisition geometries.