Splat-Based Metal Artifact Reduction in Cone-Beam CT via Compact Attenuation Modeling

Conventional cone-beam computed tomography (CBCT) reconstruction algorithms typically rely on the monochromatic assumption, which fails when high-attenuation objects like metal implants are present. This mismatch causes severe beam-hardening artifacts, such as dark streaks and shading. The authors aim to develop a computationally efficient, high-fidelity reconstruction framework that accounts for the polychromatic nature of X-rays to suppress these artifacts while preserving fine structural details.

Building upon Gaussian splatting, the authors incorporate a physically grounded polychromatic forward projection model. A key innovation is the use of a compact material parameterization: each Gaussian primitive is assigned a scalar parameter that controls its energy-dependent mass attenuation coefficient (MAC) via a quadratic Bezier curve. This low-dimensional representation captures the essential variations across biological tissues and metals, allowing for joint optimization of geometry and material properties without the need for manual metal segmentation or masks. The entire pipeline is differentiable, enabling the integration of image-domain priors like total variation (TV) regularization to further enhance reconstruction quality.

The proposed method significantly outperforms existing approaches, including standard FDK reconstruction, neural field-based methods, and previous metal artifact reduction (MAR) techniques. In both synthetic and real-world experiments, the framework effectively suppresses metal-induced streaking and shading while maintaining sharp structural boundaries and high-frequency textures. Because the model avoids the computational overhead of dense volumetric neural networks and does not require complex metal masking, it achieves faster convergence and more robust performance in challenging clinical and industrial imaging scenarios.