Salience-Enhanced Network for Metal Artifact Reduction in CT Imaging

Computed Tomography (CT) technology is widely used in clinical diagnosis; however, metal implants often introduce severe artifacts in CT images. To address this issue, metal artifact reduction (MAR) techniques have emerged. In the image domain, existing MAR methods have made some progress, but few have focused on utilizing artifact information to guide models effectively. To bridge this gap, we propose a salience-enhanced metal artifact reduction network. The framework is comprised of two stages: In the first stage, a Siamese network extracts common artifact features from different datasets while distinguishing non-artifact features, and Grad-CAM is utilized to generate heatmaps that highlight salient artifact features. In the second stage, a baseline network based on a transformer module, guided by these salience maps, concentrates on artifact regions for precise artifact correction. We evaluated our model's performance using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). The experimental results achieved state-of-the-art performance, demonstrating the efficacy of the proposed method.