Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
153 tokens/sec
GPT-4o
7 tokens/sec
Gemini 2.5 Pro Pro
45 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
38 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis (2403.04116v3)

Published 7 Mar 2024 in eess.IV and cs.CV

Abstract: X-ray is widely applied for transmission imaging due to its stronger penetration than natural light. When rendering novel view X-ray projections, existing methods mainly based on NeRF suffer from long training time and slow inference speed. In this paper, we propose a 3D Gaussian splatting-based framework, namely X-Gaussian, for X-ray novel view synthesis. Firstly, we redesign a radiative Gaussian point cloud model inspired by the isotropic nature of X-ray imaging. Our model excludes the influence of view direction when learning to predict the radiation intensity of 3D points. Based on this model, we develop a Differentiable Radiative Rasterization (DRR) with CUDA implementation. Secondly, we customize an Angle-pose Cuboid Uniform Initialization (ACUI) strategy that directly uses the parameters of the X-ray scanner to compute the camera information and then uniformly samples point positions within a cuboid enclosing the scanned object. Experiments show that our X-Gaussian outperforms state-of-the-art methods by 6.5 dB while enjoying less than 15% training time and over 73x inference speed. The application on sparse-view CT reconstruction also reveals the practical values of our method. Code is publicly available at https://github.com/caiyuanhao1998/X-Gaussian . A video demo of the training process visualization is at https://www.youtube.com/watch?v=gDVf_Ngeghg .

Citations (14)

Summary

  • The paper’s main contribution is X-Gaussian, which integrates a radiative Gaussian point cloud model with DRR and ACUI to drastically speed up X-ray view synthesis.
  • It employs a Radiative Gaussian Point Cloud Model that simplifies intensity modeling by leveraging isotropic X-ray radiation, avoiding complex spherical harmonics.
  • Empirical results show significant improvements in PSNR and SSIM, outperforming NeRF methods and enabling efficient sparse-view CT reconstruction.

Efficient X-ray Novel View Synthesis with Radiative Gaussian Splatting

Introduction to X-Gaussian

In the domain of X-ray novel view synthesis, the recent work presents an innovative technique called X-Gaussian based on a 3D Gaussian splatting framework. The development of this method is driven by the inadequacies of existing approaches, notably those derived from Neural Radiance Fields (NeRF), which are hampered by lengthy training durations and sluggish inference speeds. The X-Gaussian method, through the introduction of a Radiative Gaussian Point Cloud Model and an Angle-pose Cuboid Uniform Initialization (ACUI) strategy, not only accelerates the training and inference processes significantly but also attains a remarkable improvement in synthesis quality.

Radiative Gaussian Point Cloud Model

A pivotal aspect of X-Gaussian is its Radiative Gaussian Point Cloud Model. This model is inspired by the isotropic nature of X-ray imaging and is tailored to correctly predict the radiation intensity of 3D points. Unlike approaches that depend on spherical harmonics for modelling light reflections in the RGB domain, the radiative model simplifies the representation by focusing on the inherent radiation intensity properties of points, independent of the viewing direction. This innovation allows for a more accurate and efficient rendering of X-ray views.

Differentiable Radiative Rasterization

Coupled with the Radiative Gaussian Point Cloud Model is the introduction of a Differentiable Radiative Rasterization (DRR) process, implemented with a CUDA kernel. This rasterization technique, specifically designed for X-ray view synthesis, leverages the isotropy of X-ray radiation to optimize both the quality and speed of rendering. Notably, the DRR technique facilitates a substantial boost in inference speed, achieving over 73 times the speed of state-of-the-art NeRF-based methods while significantly reducing the training time.

Angle-pose Cuboid Uniform Initialization

The initialization process of the 3D Gaussian point clouds is another critical aspect where X-Gaussian introduces efficiency improvements. The proposed ACUI strategy utilizes the parameters of the X-ray scanner to directly compute camera information and uniformly initializes the positions of Gaussian points within a cuboid that envelops the scanned object. This methodological refinement not only eliminates the need for time-consuming structure-from-motion algorithms but also considerably shortens the training time without sacrificing the quality of the synthesized views.

Empirical Validation

The efficiency and effectiveness of X-Gaussian have been empirically validated across several datasets involving different human organs. In comparison to existing state-of-the-art methods, X-Gaussian consistently outperforms in terms of synthesis quality, measured in significant improvements in PSNR and SSIM metrics. Furthermore, the application of X-Gaussian in sparse-view CT reconstruction demonstrates its practical utility, where it enables superior reconstruction results with a significantly reduced number of views.

Conclusion and Future Directions

X-Gaussian represents a significant advancement in the field of X-ray novel view synthesis, setting a new benchmark for both performance and efficiency. The introduction of a radiative Gaussian point cloud model together with the DRR process and ACUI strategy underscores the potential of specialized approaches in handling specific imaging modalities like X-ray. Looking forward, extending the benefits of Gaussian splatting to other medical imaging modalities and integrating machine learning advancements could further broaden the impact of this work in clinical applications and beyond.

Github Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com