Unsupervised CT Metal Artifact Reduction by Plugging Diffusion Priors in Dual Domains

TL;DR

This work addresses CT metal artifact reduction under limited paired data by introducing DuDoDp, a dual-domain MAR framework that leverages diffusion priors trained on artifact-free CT images. At each diffusion timestep, priors are injected in both the sinogram and image domains, with a sinogram inpainting module complemented by image-domain fusion guided by temporally dynamic weight masks. The approach achieves state-of-the-art performance among unsupervised MAR methods and approaches supervised performance on synthetic data, while delivering visually superior results on clinical data. Although computationally intensive, DuDoDp demonstrates strong potential for clinical deployment where paired training data are scarce, and suggests directions for faster inference and content-aware fusion to further close the gap with supervised methods.

Abstract

During the process of computed tomography (CT), metallic implants often cause disruptive artifacts in the reconstructed images, impeding accurate diagnosis. Several supervised deep learning-based approaches have been proposed for reducing metal artifacts (MAR). However, these methods heavily rely on training with simulated data, as obtaining paired metal artifact CT and clean CT data in clinical settings is challenging. This limitation can lead to decreased performance when applying these methods in clinical practice. Existing unsupervised MAR methods, whether based on learning or not, typically operate within a single domain, either in the image domain or the sinogram domain. In this paper, we propose an unsupervised MAR method based on the diffusion model, a generative model with a high capacity to represent data distributions. Specifically, we first train a diffusion model using CT images without metal artifacts. Subsequently, we iteratively utilize the priors embedded within the pre-trained diffusion model in both the sinogram and image domains to restore the degraded portions caused by metal artifacts. This dual-domain processing empowers our approach to outperform existing unsupervised MAR methods, including another MAR method based on the diffusion model, which we have qualitatively and quantitatively validated using synthetic datasets. Moreover, our method demonstrates superior visual results compared to both supervised and unsupervised methods on clinical datasets.

Figures (10)

• Figure 1: A typical diffusion model consists of a forward process and a reverse process. The forward process gradually adds noise to clean images, and the reverse process generates clean images from Gaussian noise.
• Figure 1: The presence of metallic implants (a) can lead to various physical effects on the sinogram (b), resulting in significant metal artifacts in the reconstructed images (c).
• Figure 1: (a) illustrates the overall workflow of our MAR method, DuDoDp, where each blue box shows one iteration to calculate $x_{t-1}$ from $x_t$ while the initial $x_T$ is sampled form a unit Gaussian distribution. (b) and (c) respectively illustrate the sinogram inpainting module and image fusion module in (a).
• Figure 1: The intermediate results (T=100) using the diffusion priors only in the sinogram domain is shown. The red parts indicate metallic implants. It can be observed that, in the early iterations, the discontinuity of inpainted sinogram lead to artifacts beyond the metal regions, and part of these artifacts persists even after the completion of iterations. The display window of CT images is [-175, 275] HU.
• Figure 1: Binary sinogram corresponding to different $\delta$ values and their reconstructed weight masks. With smaller $\delta$ values, the resulting weight masks are generally smaller (darker in the images).