Research map / 3D vision / Gaussian splatting
DDGS-CT: Direction-Disentangled Gaussian Splatting for Realistic Volume Rendering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DDGS, a direction-disentangled 3D Gaussian splatting method with radiodensity-aware initialization that renders more accurate and faster DRRs from CT volumes and improves 2D/3D registration.
Problem
Physics-based Monte Carlo DRR simulation is too slow for intraoperative use, while analytical differentiable DRR renderers ignore anisotropic X-ray effects such as Compton scattering.
Approach
Split Gaussian radiosity into an isotropic component and a direction-dependent component (using a low-degree spherical decomposition of ray angles with a shared basis) realized as separate Gaussian sets, plus a radiodensity-aware initialization that combines marching-cubes interface points with radiodensity-weighted voxel samples.
Key results
- On CTPelvic1K (10 scans), DDGS (L=1, k=8) reaches 39.08 avg PSNR / 0.986 SSIM vs 37.47 for 3DGS and 38.04 for X-Gaussian (k=32) with fewer points (42,625 vs 48,183).
- On Ljubljana registration, DDGS achieves mean TRE 0.80 mm vs 2.14 mm for DiffDRR, with mean DRR render time 7.46 ms vs 30.88 ms.
- Registration on CTPelvic1K gives DDGS rotation/translation errors that are lowest on most scans (e.g., 001 translation 1.285 mm vs 1.788 mm for 3DGS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use direction-disentangled isotropic and directional Gaussian contributions (ablation) | DDGS · CTPelvic1K scans 001–002 | PSNR: 36.17 -> 38.04; 37.32 -> 38.30 | Helped |
| add a directional component to direction-independent Gaussians (ablation) | DDGS · CTPelvic1K scans 001–002 | PSNR: 35.93 -> 38.04; 36.96 -> 38.30 | Helped |
| add an isotropic component to direction-dependent Gaussians (ablation) | DDGS · CTPelvic1K scans 001–002 | PSNR: 32.48 -> 38.04; 33.04 -> 38.30 | Helped |
| use learnable DDGS features rather than 3DGS-disentangled components (ablation) | DDGS · CTPelvic1K scans 001–002 | PSNR: 35.32 -> 38.04; 35.38 -> 38.30 | Helped |
| replace random initialization with RADS (ablation) | DDGS · CTPelvic1K scans 001–010, 30,000 iterations | PSNR: 38.89 -> 39.08 | Helped |
| replace even initialization with RADS (ablation) | DDGS · CTPelvic1K scans 001–010, 30,000 iterations | PSNR: 38.78 -> 39.08 | Helped |
| increase feature dimension from k=8 to k=16 (ablation) | DDGS · CTPelvic1K 001 | PSNR: 38.04 -> 38.41 | Helped |
| increase feature dimension from k=8 to k=32 (ablation) | DDGS · CTPelvic1K 001 | PSNR: 38.04 -> 38.27 | No clear effect |
| use DDGS instead of X-Gaussian | novel-view synthesis · CTPelvic1K-Dataset6 avg | PSNR: 38.04 -> 39.08 | Helped |
| use DDGS instead of 3DGS | novel-view synthesis · NAF-CT avg | PSNR: 44.21 -> 44.47 | Helped |
| replace DiffDRR rendering with DDGS | Ljubljana registration · Ljubljana | DRR render time (ms), mean: 30.88 -> 7.46 | Helped |
| replace DiffDRR with DDGS | Ljubljana registration · Ljubljana | TRE (mm), mean: 2.14 -> 0.80 | Helped |
Surprising findings
- Direction-dependent-only and direction-independent-only Gaussian variants perform substantially worse than the combined disentangled model, indicating both components are needed.
Limitations
- Requires an offline DRR renderer (e.g., DeepDRR) to produce realistic training targets.
- Direction-dependent function may fail to approximate complex multi-bounce light transport.
- Quality bounded by input CT scan quality and noise.
Experimental setup
Datasets: NAF-CT CTPelvic1K Ljubljana DeepFluoro
Benchmarks: NAF-CT CTPelvic1K Ljubljana DeepFluoro
Metrics: PSNR SSIM rotation error translation error target registration error DRR render time total optimization time
Baselines: 3DGS X-Gaussian DiffDRR GaSpCT
Builds on: 3D Gaussian splatting X-Gaussian DiffDRR iComMA DiffPose
Closest papers by meaning
- R$^2$-Gaussian: Rectifying Radiative Gaussian Splatting for Tomographic Reconstruction (NeurIPS 2024 · similarity 0.89)
- Volumetrically Consistent 3D Gaussian Rasterization (CVPR 2025 · similarity 0.89)
- DiffGS: Functional Gaussian Splatting Diffusion (NeurIPS 2024 · similarity 0.86)
- UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 2025 · similarity 0.85)
- 3DGUT: Enabling Distorted Cameras and Secondary Rays in Gaussian Splatting (CVPR 2025 · similarity 0.85)
- 6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering (ICLR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.