Structure Aware Metal Artifact Reduction for Computed Tomography

Abstract

Metal implants can cause severe artifacts in CT images, including streaking and photon starvation effects that degrade diagnostic visibility. Building on hybrid convolution–attention feature modeling exemplified by CTLformer, this paper introduces a residual learning approach for metal artifact reduction that combines image-domain denoising with sinogram-guided attention cues. The method uses a guidance branch derived from corrected sinograms to steer attention toward artifact-affected regions while preserving bone and soft-tissue edges. Evaluations on a clinical dataset of 4,800 scans (about 120,000 slices) show improvements of 0.9–1.5 dB in PSNR and 0.012–0.020 in SSIM compared with U-Net, CNN-based MAR, and transformer-only baselines. Artifact index metrics also decrease by 10%–18% in implant-adjacent regions.