TG-Field: Geometry-Aware Radiative Gaussian Fields for Tomographic Reconstruction

TL;DR

TG-Field tackles ultra-sparse-view CBCT and dynamic CT reconstruction by introducing a geometry-aware Gaussian deformation framework. It regularizes radiative Gaussians with a multi-resolution hash encoder and extends to time-varying scenes with time-conditioned representations, a spatiotemporal attention block, and a motion-flow network. A semantic-consistency loss from pretrained visual foundation models reinforces cross-view coherence. Across synthetic and real datasets, TG-Field achieves state-of-the-art reconstruction quality under ultra-sparse views, reducing artifacts and preserving fine anatomical detail, with clear clinical potential.

Abstract

3D Gaussian Splatting (3DGS) has revolutionized 3D scene representation with superior efficiency and quality. While recent adaptations for computed tomography (CT) show promise, they struggle with severe artifacts under highly sparse-view projections and dynamic motions. To address these challenges, we propose Tomographic Geometry Field (TG-Field), a geometry-aware Gaussian deformation framework tailored for both static and dynamic CT reconstruction. A multi-resolution hash encoder is employed to capture local spatial priors, regularizing primitive parameters under ultra-sparse settings. We further extend the framework to dynamic reconstruction by introducing time-conditioned representations and a spatiotemporal attention block to adaptively aggregate features, thereby resolving spatiotemporal ambiguities and enforcing temporal coherence. In addition, a motion-flow network models fine-grained respiratory motion to track local anatomical deformations. Extensive experiments on synthetic and real-world datasets demonstrate that TG-Field consistently outperforms existing methods, achieving state-of-the-art reconstruction accuracy under highly sparse-view conditions.

Figures (7)

• Figure 1: Visual comparison of point clouds obtained by different initialization methods: (a) cubic uniform sampling cai2025radiative, (b) FDK-based initialization r2_gaussian, (c) our iteration-based initialization, and (d) reference. Our method clearly captures more accurate and detailed geometric structures.
• Figure 2: Our TG-Field framework first initializes a point cloud via a two-stage iterative refinement. The refined coordinates are encoded using a hash encoder to capture spatial geometric features, which are then decoded into optimized Gaussian splats by multiple MLP decoders. Additionally, we employ semantic consistency regularization with pretrained visual foundation models to maintain multi-view semantic consistency. Finally, these Gaussian splats are rendered into arbitrary-view X-ray projections and voxelized into a CT volume. For dynamic scenarios, we incorporate a spatiotemporal attention module and a Motion-Flow module to effectively capture temporal variations and fine-grained local movements.
• Figure 3: Ablation of the hash encoder (HE) and semantic regularization (SR) on static CT reconstruction.
• Figure 4: Qualitative comparisons of static CT reconstruction results across different view numbers (5, 10, and 20 views). Compared to SAX-NeRF and $\text{R}^2$-Gaussian, our method consistently provides sharper structural details and fewer artifacts, especially under extremely sparse-view conditions.
• Figure 5: Visual ablation of the motion-flow module (MF) for 4D CT reconstruction.