Diffusion prior-guided implicit neural representation for metal artifact reduction in sparse-view CT reconstruction

Objective. Sparse-view computed tomography (SVCT) reduces radiation exposure, but it inevitably leads to severe streak artifacts, which become even more pronounced in the presence of metallic implants. Most existing methods address SVCT and metal artifact reduction (MAR) separately, making the joint SVCT and MAR (SVMAR) problem suboptimal. While a few methods have been proposed to tackle SVMAR jointly, most are supervised and require large paired datasets that are difficult to obtain in clinical practice. To overcome these limitations, we propose a self-supervised framework that leverages a denoising diffusion probabilistic model and an implicit neural representation (INR) to address the SVMAR problem.Approach. First, an INR is optimized to produce an initial reconstruction using sparse-view measurements by sampling only rays outside the metal trace. This initial estimate is then used to accelerate the reverse diffusion process. During reverse diffusion, we alternately perform: (i) a MAR step, where diffusion priors guide the inpainting of metal-trace regions in the measurements using Poisson blending; and (ii) an SVCT step, where data fidelity is enforced by refining the INR with both diffusion priors and the MAR-corrected sinograms obtained from the MAR step.Main result. Experiments on both simulation and clinical datasets show that the proposed method reconstructs high-quality images without requiring large paired datasets. Quantitative evaluations demonstrated that the proposed method outperformed existing methods including IndudoNet+ on the OOD dataset across 160-, 80-, and 40-view settings, with peak signal-to-noise ratio (PSNR) improving from 36.28 to 45.39 dB, 32.69-44.22 dB, and 31.37-41.90 dB, and structural similarity index measure (SSIM) increasing from 0.955 to 0.988, 0.920-0.983, and 0.906-0.973, respectively.Significance. By leveraging diffusion priors within a self-supervised INR framework, the method provides a practical and generalizable solution for real-world SVMAR scenarios where ground-truth images are unavailable.