Splat-based Metal Artifact Reduction in Cone-Beam CT via Polychromatic Modeling

Cone-beam CT has long struggled with metal-induced streaking and beam hardening, and the usual monochromatic reconstruction assumptions break down badly around high-attenuation materials. A new splat-based approach tackles that problem by folding a polychromatic X-ray projection model, material-dependent attenuation, and system response modeling into a Gaussian Splatting pipeline.

The practical angle is important: the method is self-calibrating, so it does not rely on hand-authored metal masks or heavy prior assumptions. During training, it jointly optimizes the reconstruction parameters and the X-ray spectrum itself, which is a more physics-aware route than simply asking a network to inpaint the damage after the fact.

The team also built a high-fidelity synthetic CBCT data generation pipeline, validated against a Monte-Carlo X-ray simulation toolbox, and released new datasets with severe metal artifacts. That matters because artifact-heavy medical imaging datasets are hard to come by, and reconstruction research tends to bottleneck on data quality as much as model quality.

For graphics and rendering folks, the interesting bit is the crossover: this is the first splat-based method aimed at beam-hardening reduction in CBCT, and it shows how splatting-style representations can be pushed beyond view synthesis into physically grounded inverse problems. The method is reported to outperform current state-of-the-art approaches on both synthetic and real-world data, which could make it a useful reference point for anyone working on differentiable rendering, inverse imaging, or...