New AI Method Improves CT Images by Reducing Metal Artifacts

Computed tomography (CT) is essential for diagnosis and image‑guided treatment, yet high‑density metals such as dental fillings, orthopedic hardware, and stents create beam hardening and photon starvation that produce severe streak artifacts. These artifacts can obscure anatomy and compromise clinical decisions. Many correction methods rely on projection interpolation or supervised learning and can introduce secondary artifacts or require paired training data rarely available in practice. To help address this challenge, researchers have developed an unsupervised, diffusion‑regularized approach for metal artifact reduction.

Researchers from the School of Information Science and Technology at ShanghaiTech University (Shanghai, China) report a method published in Quantitative Biology. The team introduces INR‑DR, an unsupervised framework that combines CT physics constraints with diffusion‑model priors to achieve robust metal artifact reduction without paired training data. The approach is designed to generalize across different metal shapes, anatomical regions, and scanning protocols. Its goal is to improve reliability where existing techniques often fail.

INR‑DR avoids direct inpainting of corrupted projection data within metal traces. Instead, it represents the CT image as a continuous coordinate function and enforces consistency with reliable measurements through a differentiable forward model. A pretrained unconditional diffusion model serves as a regularization prior, guiding the implicit representation toward realistic anatomy via one‑step denoising rather than acting as a generator. Multiresolution hash encoding preserves fine structural details, balancing global physics‑based fidelity with local anatomical plausibility.

Experiments on simulated and clinical dental CT data show effective suppression of streak artifacts while maintaining tissue integrity across large, medium, and small metal implants. The framework outperformed both traditional methods and supervised deep‑learning baselines. Ablation studies confirmed the complementary contributions of data‑consistency enforcement and the diffusion prior. These findings support an unsupervised pathway to metal artifact reduction without paired exemplars.

The authors note that the same design offers a general strategy for ill‑posed inverse problems such as sparse‑view and low‑dose CT. Clinical adoption will require faster per‑case optimization and further multicenter validation. Designed to generalize across metals and protocols, the approach aims to deliver robust artifact reduction without paired training data.

Related Links
ShanghaiTech University