GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View Tomographic Reconstruction

yikuang yuluo, Yue Ma, Shen Kuan, Tongtong Jin, Wang Liao, Yangpu Ma, Fuquan Wang

Keywords: 3D Gaussian Splatting ; CT Reconstruction ; Graph-Based Radiative Gaussian Splatting ;

Abstract: Computed tomography (CT) reconstruction under sparse-view settings remains highly challenging due to severe artifacts. Recently, 3D Gaussian Splatting (3DGS) has shown promise for this task, but existing methods often rely on view-averaged gradient magnitudes, which easily cause needle-like artifacts in sparse views. To overcome this limitation, we propose GR-Gaussian, a graph-based 3DGS framework. It explicitly leverages a CT-specific prior, where regions of the same tissue or material have similar attenuation coefficients, forming a natural structural relationship among neighboring points. This structure motivates a graph-based representation, which guides gradient refinement to suppress needle-like artifacts. To exploit this structure, GR-Gaussian introduces (1) a Denoised Point Cloud Initialization strategy that mitigates initialization errors, and (2) a Pixel-Graph-Aware Gradient strategy that leverages graph-based density differences to refine gradient computation, improving splitting accuracy and density representation. Experiments on X-3D and real-world datasets validate the effectiveness of GR-Gaussian, achieving PSNR improvements of 0.67 dB and 0.92 dB, and SSIM gains of 0.011 and 0.021. These results highlight the importance of embedding domain-specific structural priors for accurate CT reconstruction under challenging sparse-view conditions.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 8639