Abstract: This paper proposes a deep learning-based metal artifact reduction method (DRDNet), which effectively suppresses metal artifacts in CT images by combining dictionary learning and deep unfolding techniques. The proposed method assumes that images affected by metal artifacts can be decomposed into artifact-free components and metal artifact template. A multi-stage neural network is employed to iteratively optimize the decomposition process. Specifically, FNet extracts metal artifact features, while TNet generates artifact-free images. By incorporating a composite loss function (combining L1 Loss and Edge Loss), the method preserves structural details while removing artifacts. Experiments conducted on the SynDeepLesion dataset demonstrate that DRDNet outperforms existing methods in terms of both PSNR and SSIM metrics. Moreover, it exhibits superior artifact reduction performance across metal masks of varying sizes and shapes.