PrideDiff: Physics-Regularized Generalized Diffusion Model for CT Reconstruction

Abstract

Achieving a lower radiation dose and a faster imaging speed is a pivotal objective of computed tomography (CT) reconstruction. However, these often come at the cost of compromised reconstruction quality. With the advent of deep learning, numerous CT reconstruction methods rooted in this field have significantly improved the reconstruction performance. Recently, diffusion models have further enhanced training stability and imaging quality for CT. However, many of these methods only focus on CT image domain features, ignoring the intrinsic physical information of the imaging process. Although compressive sensing-based iterative reconstruction algorithms utilize physical prior information, their intricate iterative process poses challenges in training, subsequently influencing their efficiency. Motivated by these observations, we introduce a novel physics-regularized generalized diffusion model for CT reconstruction (PrideDiff). On the one hand, our method further improves the quality of reconstructed images by fusing physics-regularized iterative reconstruction methods with diffusion models. On the other hand, we propose a prior extraction module embedded with temporal features, which effectively improves the performance of the iteration process. Extensive experimental results demonstrate that PrideDiff outperforms several state-of-the-art methods in low-dose and sparse-view CT reconstruction tasks on different datasets, with faster reconstruction speed. We further discuss the effectiveness of relevant components in PrideDiff and validate the stability of the iterative reconstruction process, followed by detailed analysis of computational cost and inference time.