Generative Image Priors for Artifact Suppression in Low-Dose CT Scans: Simulation Study

Abstract

The widespread utilization of computed tomography in clinical diagnostics has brought immense benefits to patient care, yet the associated ionizing radiation poses non-negligible long-term health risks. Consequently, the transition toward low dose computed tomography has become an urgent clinical imperative, guided by the principle of keeping radiation doses as low as reasonably achievable. However, reducing the radiation dose intrinsically leads to severe degradation in image quality, primarily manifested as excessive quantum mottle, electronic noise, and complex streak artifacts resulting from photon starvation. Traditional analytical and iterative reconstruction methods often struggle to strike an optimal balance between aggressive noise reduction and the preservation of crucial anatomical details. In recent years, data-driven approaches, particularly those leveraging deep learning, have demonstrated remarkable potential in addressing these inverse problems. Among these, generative image priors derived from unsupervised generative models have emerged as highly promising tools, offering powerful regularization capabilities without relying on strictly paired training data. This paper presents a comprehensive simulation study investigating the efficacy of generative image priors in suppressing artifacts and restoring image fidelity in low dose computed tomography scans. By simulating a rigorous forward projection model incorporating realistic noise characteristics, we evaluate an iterative reconstruction framework guided by a score-based generative prior. Extensive quantitative and qualitative analyses demonstrate that the proposed method significantly outperforms conventional techniques in mitigating streak artifacts, preserving high-frequency structural edges, and maintaining natural image textures.