Asymmetric decoupled dual-domain learning for CT metal artifact reduction: Integrating Fast Fourier Convolution and...

Computed Tomography (CT) imaging is frequently compromised by metallic implants, which introduce severe streaking and shading artifacts that obscure anatomical details. While dual-domain networks have advanced Metal Artifact Reduction (MAR), they are limited by their reliance on architecturally homogeneous frameworks. This design ignores the cross-domain asymmetry in artifact manifestation: metal implants appear as broad geometric discontinuities in the sinogram domain but propagate as non-local streaks in the image domain. To bridge this topological discrepancy, we propose an Asymmetric Decoupled Dual-Domain Framework. First, we introduce a Spectral-Geometric Inpainting Module (SGIM) utilizing Fast Fourier Convolution to exploit global spectral consistency for reconstructing missing projections within large gaps. Second, an Anatomical Texture Refinement Module (ATRM) based on Vision Transformers is designed to capture long-range pixel dependencies and mitigate residual streaks. A Residual-Prior Guided Fusion strategy integrates geometrically corrected structural priors from the projection space into the image space. Extensive experiments on the Synthesized DeepLesion dataset demonstrate that our approach outperforms state-of-the-art baselines, achieving a Peak Signal-to-Noise Ratio (PSNR) of 46.49 dB and Structural Similarity (SSIM) of 0.9972. Additionally, the proposed framework exhibits superior generalization on clinical datasets, strong robustness against mask segmentation errors, and high computational efficiency suitable for clinical deployment.