AS-Mamba: Asymmetric Self-Guided Mamba Decoupled Iterative Network for Metal Artifact Reduction

TL;DR

AS-Mamba tackles metal artifact reduction in CT by decoupling high-frequency directional streaks from low-frequency beam-hardening shading using an asymmetric Mamba-based branch and a frequency-domain Dual Enhancement Network. A Self-Guided Contrastive Regularization strategy is embedded in an iterative refinement loop to bridge distribution gaps across clinical data. Empirical results on synthetic DeepLesion and clinical dental CBCT datasets show AS-Mamba achieves state-of-the-art PSNR/SSIM and qualitative fidelity, with lower STD and higher CNR in clinical images. The work demonstrates how integrating physical priors with deep networks improves MAR performance and robustness, with potential for unsupervised extensions and 3D reconstruction.

Abstract

Metal artifact significantly degrades Computed Tomography (CT) image quality, impeding accurate clinical diagnosis. However, existing deep learning approaches, such as CNN and Transformer, often fail to explicitly capture the directional geometric features of artifacts, leading to compromised structural restoration. To address these limitations, we propose the Asymmetric Self-Guided Mamba (AS-Mamba) for metal artifact reduction. Specifically, the linear propagation of metal-induced streak artifacts aligns well with the sequential modeling capability of State Space Models (SSMs). Consequently, the Mamba architecture is leveraged to explicitly capture and suppress these directional artifacts. Simultaneously, a frequency domain correction mechanism is incorporated to rectify the global amplitude spectrum, thereby mitigating intensity inhomogeneity caused by beam hardening. Furthermore, to bridge the distribution gap across diverse clinical scenarios, we introduce a self-guided contrastive regularization strategy. Extensive experiments on public andclinical dental CBCT datasets demonstrate that AS-Mamba achieves superior performance in suppressing directional streaks and preserving structural details, validating the effectiveness of integrating physical geometric priors into deep network design.

Figures (10)

• Figure 1: Comparison of inference speed and image quality on the DeepLesion dataset. The x-axis represents the average inference time per image, and the y-axis denotes the PSNR(dB) value. Our method achieves superior reconstruction quality with highly competitive inference speed compared to other state-of-the arts.
• Figure 2: Analysis of metal artifact mechanisms. (a) Physical decomposition of the corrupted image $f_{CT}$ into the artifact term $f_{MA}$ and ideal term $f_{E_0}$. (b) Mechanism dissection: The intensity profile (left) quantifies the low-frequency hardening distortion, where blue and red lines represent the Reference and MA, respectively. The schematic (right) illustrates the anisotropy of high-frequency streaks, where singularities from boundary points $y, z \in \partial D$ propagate along linear trajectories that structurally align with the Mamba scan path.
• Figure 3: Overall framework of AS-Mamba, consisting of a frequency-decoupled reconstruction module and an iterative refinement module: (1) The frequency-decoupled module targets specific artifact components. The High Frequency Reduction Network (HFRN) utilizes Mamba to suppress directional streaks in the high-frequency domain $X_{HF}$ , while the Dual Enhancement Network (DEN) corrects beam hardening in the low-frequency domain $X_{LL}$. (2) The iterative module ensures global structural consistency. The Iterative MANet progressively refines the initial reconstruction $U^{(0)}$ to recover anatomical details, enhanced by Self-guided Contrastive Regularization.
• Figure 4: Detailed architecture of the proposed core sub-modules. (a) The Iterative MANet focuses on progressive image refinement. (b) The DEN operates in the Fourier domain to rectify the amplitude spectrum. (c) The HFRN designed to capture long-range dependencies for suppressing directional streak artifacts.
• Figure 5: Overview of the data acquisition system. (a) The Smart3D-X dental CBCT device equipped with a cone-beam X-ray generator and a flat panel detector. (b) Schematic diagram of the axial scanning geometry. The system features a source-to-object distance of 200 mm and a source-to-image distance of 640 mm.