FID-Net: an interpretable fourier and image domain convolutional dictionary network for industrial CT metal artifact reduction

Abstract

In battery X-ray computed tomography (CT) imaging, artifacts emanating from internal metallic implants markedly distort electrode morphology, thereby impeding precise interpretation and quantification of internal structural features. Although sinogram‑domain correction techniques can mitigate these artifacts to some extent, they frequently introduce secondary artifacts due to localized correction errors and entail substantial computational overhead. Meanwhile, most deep learning metal artifact reduction (MAR) methods treat the task as a generic image‑restoration problem, overlooking the inherent physics of CT imaging and relying solely on off‑the‑shelf network components, which limits their interpretability. To overcome these challenges, we propose FID‑Net, an interpretable Fourier and image domain convolutional dictionary network for industrial CT metal artifact reduction. FID‑Net constructs spatial and spectral dictionaries in parallel and uses the fast Fourier transform as an efficient bridge to jointly capture global context and local details, iteratively refining dictionary coefficients to accurately model and remove metal artifacts while fully preserving electrode morphology. To validate the effectiveness of the method, we constructed a real battery CT dataset and conducted comparative experiments on extensive synthetic and real data, and the results show that the proposed FID‑Net demonstrates superior performance in MAR effectiveness and structural fidelity.