Abstract Background We address computed tomography (CT) metal artifacts reduction (MAR) using a generative deep‐learning model in the imaging physics framework. Existing deep learning–based MAR methods, though promising, generally lack explicit physical modeling of artifact formation and rely heavily on data‐driven mappings. The absence of physics priors not only limits scalability, as they often require paired or task‐specific datasets, but also makes such methods prone to hallucination, anatomical distortion, and unstable artifact suppression. Purpose We propose a novel self‐supervised framework for CT MAR, integrating a lightweight multi‐layer perceptron (MLP)‐based beam‐hardening correction with a conditional latent diffusion model (LDM). By incorporating a physics‐informed correction step and an artifact‐reproducing simulation technique, the framework aims to enhance scalability across diverse scenarios, reduce hallucination effects, and improve structural fidelity in the reconstructed images. Methods The proposed MLP performs physics‐driven polynomial correction, serving as a simplified but efficient alternative to existing approaches. Also, the proposed MLP implicitly incorporates sinogram consistency into its optimization objective, allowing case‐specific adaptation and convergence toward the desired solution. Additionally, the learned MLP parameters are reused to simulate artifact‐contaminated images from artifact‐free scans, generating pseudo paired data for self‐supervised training without requiring real paired datasets. A conditional LDM is trained on these synthetic pairs to remove residual artifacts. Results By operating in a low‐dimensional latent space, the LDM significantly reduces inference time while maintaining high‐quality reconstructions. The proposed method is evaluated on both the SynDeepLesion dataset and real clinical data, demonstrating superior artifact removal and structural preservation compared to the existing state‐of‐the‐art MAR techniques. We particularly highlight the robustness, generalizability, and clinical applicability of the proposed framework. Conclusion We proposed a self‐supervised metal artifact reduction framework that combines MLP‐based beam‐hardening correction with a conditional latent diffusion model in the imaging physics framework. The MLP module provides physics motivated beam‐hardening corrected CT images, while the residual artifact simulation strategy enables fully self‐supervised training without the need for paired data. The proposed method demonstrated superior artifact suppression and structural preservation on both synthetic and clinical datasets, outperforming existing approaches.
Yun et al. (Sun,) studied this question.