Computed tomography (CT) is widely used in medical diagnosis and image-guided treatment, but high-density metallic objects such as dental fillings, orthopedic implants, and vascular stents can cause beam hardening and photon starvation, producing severe streak artifacts that obscure surrounding anatomy. Existing methods often restore corrupted projections through interpolation or deep neural networks, but discontinuities between estimated and measured data may introduce secondary artifacts. Supervised approaches also rely on paired metal-corrupted and metal-free images, which are difficult to obtain clinically and may generalize poorly across different metal shapes, anatomical regions, and scanning protocols.
Researchers from the School of Information Science and Technology at ShanghaiTech University, China, recently published “Diffusion model-regularized implicit neural representation for computed tomography metal artifact reduction” in Quantitative Biology. The team proposed INR-DR, an unsupervised framework that combines CT physics constraints with diffusion-model priors to achieve robust metal artifact reduction without paired training data (Figure 1).
INR‑DR does not directly inpaint the corrupted projection data within metal traces. Instead, it models the CT image as a continuous coordinate function and enforces consistency with reliable measurements via a differentiable forward model. A pretrained unconditional diffusion model is used as a regularization prior (not as a direct generator), guiding the implicit representation toward realistic anatomy through one‑step denoising, while multiresolution hash encoding preserves fine structural details. This design balances global physics‑based fidelity with local anatomical plausibility.
Experiments on simulated and clinical dental CT data show that INR‑DR effectively suppresses streak artifacts and maintains tissue integrity for large, medium, and small metal implants, outperforming both traditional and supervised baselines. Ablation studies confirm the complementary roles of data consistency and the diffusion prior. The framework offers a general strategy for ill‑posed inverse problems such as sparse‑view and low‑dose CT, but clinical adoption will require faster per‑case optimization and further multicenter validation.
DOI: 10.1002/qub2.70031