4DRGS: 4D Radiative Gaussian Splatting for Efficient 3D Vessel Reconstruction from Sparse-View Dynamic DSA Images

TL;DR

This paper presents 4DRGS, a novel framework for fast, high-quality 3D vessel reconstruction from sparse-view dynamic DSA images. It represents vessels as a set of 4D radiative Gaussian kernels with time-invariant geometry and time-varying attenuation predicted by a Dynamic Neural Attenuation Field, enabling efficient X-ray rasterization-based image synthesis and attenuation voxelization. Two key innovations—accumulated attenuation pruning and bounded scaling activation—improve reconstruction accuracy and suppress artifacts. Empirical results on real patient data show 4DRGS achieves state-of-the-art performance with training times an order of magnitude shorter than NeRF-based methods, highlighting its potential for clinical use.

Abstract

Reconstructing 3D vessel structures from sparse-view dynamic digital subtraction angiography (DSA) images enables accurate medical assessment while reducing radiation exposure. Existing methods often produce suboptimal results or require excessive computation time. In this work, we propose 4D radiative Gaussian splatting (4DRGS) to achieve high-quality reconstruction efficiently. In detail, we represent the vessels with 4D radiative Gaussian kernels. Each kernel has time-invariant geometry parameters, including position, rotation, and scale, to model static vessel structures. The time-dependent central attenuation of each kernel is predicted from a compact neural network to capture the temporal varying response of contrast agent flow. We splat these Gaussian kernels to synthesize DSA images via X-ray rasterization and optimize the model with real captured ones. The final 3D vessel volume is voxelized from the well-trained kernels. Moreover, we introduce accumulated attenuation pruning and bounded scaling activation to improve reconstruction quality. Extensive experiments on real-world patient data demonstrate that 4DRGS achieves impressive results in 5 minutes training, which is 32x faster than the state-of-the-art method. This underscores the potential of 4DRGS for real-world clinics.

Figures (5)

• Figure 1: Overview of DSA imaging and vessel reconstruction. (a) DSA images are generated by subtracting fill-run X-ray images from their mask-run counterparts. (b) We model vessels as a set of 4D radiative Gaussians. (c) The final 3D vessel volume is reconstructed via attenuation voxelization.
• Figure 2: The overall pipeline of 4DRGS. We model vessels as a set of 4D radiative Gaussian kernels (\ref{['sec:4d_Radiative_Gaussian_Modeling']}) and optimize them with image losses (\ref{['sec:model_optimization']}). 3D vessel volume is reconstructed via attenuation voxelization (\ref{['sec:vascular_reconstruction']}).
• Figure 3: 3D vessel reconstruction of different methods with CD($\mathrm{mm}$)/HD($\mathrm{mm}$) values shown at the top right of each image.
• Figure 4: 2D DSA image synthesis of different methods at test frames. PSNR($\mathrm{dB}$)/SSIM values averaged over the test set are shown at the top right of each image.
• Figure 5: Qualitative results of ablation study. (a) 3D vessel reconstruction. Top row: 3D visualization with CD($\mathrm{mm}$)/HD($\mathrm{mm}$) values shown at the top right of each image. Bottom row: sagittal slice of reconstructed volume. (b) 2D DSA image synthesis at test frame. PSNR($\mathrm{dB}$)/SSIM values averaged over the test set are shown at the top right of each image.