Noise-Generating Mechanism-Driven Implicit Diffusion Prior for Low-Dose CT Sinogram Recovery

Abstract

Low-dose computed tomography (CT) images often suffer from noise and artifacts from photon starvation and electronic noise. Recent advancements in deep learning (DL) techniques have significantly improved outcomes in low-dose CT (LDCT) imaging. However, many existing methods require costly low-dose/high-dose CT image pairs for supervised training, which is difficult to obtain in clinical. In this article, we propose a novel unsupervised approach for LDCT sinogram recovery based on the noise generation mechanism within the Bayes framework. Specifically, we introduce a novel formulation of sinogram recovery model based on the noise-generating mechanism without additional regularization terms. Then, we design an efficient algorithm that utilizes Bayes rules to solve the sinogram recovery model, offering approximate and analytical solutions for all decomposed score functions. Instead of relying on deep network priors, we adopt an implicit diffusion model to characterize the common latent prior of sinogram data and enable the iterative algorithm more efficient and interpretable. Extensive experiments conducted on two datasets demonstrate the superiority of our proposed method over competing techniques in both denoising and generalization performance.