Tri-Domain Filtering Transformer for CT Metal Artifact Reduction

Current computed tomography (CT) metal artifact reduction (MAR) methods struggle with real-world clinical applications due to their inability to adapt to diverse scanner parameterizations and imaging protocols. These methods often rely on rigid cross-domain operations, such as image-level concatenation or sequential domain conversion, leading to spectral-phase discontinuities and geometric misalignments that may limit artifact suppression in complex cases. To overcome this, we propose the tri-domain filtering transformer for CT MAR (TFT-MAR), which explores feature-granular tri-domain synergy (spatial, spectral, and sinogram-feature domains). First, spatial-spectral-sinogram mixing filtering (S3MF) replaces conventional image-level operations with feature-granular filters, where spatial modules suppress fine-grained artifacts, spectral operators enforce global intensity consistency, and Radon-aware units dynamically adapt to projection physics. Second, the cross-domain self-attention mechanism (CDSA) bridges these domains via physics-constrained token interaction, resolving parameter mismatches by synchronizing texture preservation and geometric constraints. Third, a dual-stream global-local gated feed-forward network (GLGFN) disentangles global spectral dependencies and local edge patterns, preventing high-frequency distortions during multidomain fusion. Rigorous experiments on the synthesized DeepLesion and clinical metal-implant benchmarks, along with radiologist study and extension to MRI super-resolution, demonstrate that TFT-MAR outperforms existing CT MAR techniques while exhibiting robust generalizability. It preserves microstructural details critical for diagnosis, as supported by radiologist evaluations that indicate improved clinical usability.