SGAMARN: A GAN Framework for Metal Artifact Reduction in CT Imaging.

Metallic objects cause severe artifacts degrading CT image quality, complicating interpretation, and influencing clinical diagnosis in radiation therapy. This paper introduces a novel Generative Adversarial Network (GAN) framework developed for metal artifact reduction (MAR), incorporating a transformer generator and a conditional discriminator, jointly optimized to improve artifact removal while maintaining anatomical accuracy. We adopt a modified U-Net structure for the generator, featuring hierarchical encoder stages, a bottleneck, and an asymmetric decoder. The encoder utilizes transformer blocks and self-attention mechanisms to extract multi-scale features, while pooling cascading facilitates dense connectivity and gradient flow. The decoder, optimized for computational efficiency, reconstructs artifact reduced images. The discriminator, on the other hand, leverages convolutional blocks with instance normalization to detect residual artifacts and guide the generator towards structural fidelity and artifact suppression.Ablation experiments were performed to determine the effectiveness of the module. Evaluation of the proposed network has been conducted for training using the Deeplesion synthasized Dataset. The Structural Similarity Index Method (SSIM) and the Peak Signal to Noise Ratio (PSNR) were measured for quantitative results which demonstrate that SGAMARN outperforms traditional and deep learning MAR techniques in artifact suppression, structural fidelity, and robustness achieving 7.6 percent improvement in PSNR.