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

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.

Digital subtraction angiography (DSA) is a widely recognized gold standard for diagnosing vascular diseases. DSA imaging involves two rotational cone-beam X-ray scans: the mask run, performed before the contrast agent injection, and the fill run, performed after the injection. Subtracting X-ray images in the fill run from those in the mask run yields 2D DSA images, which highlight blood flow marked by the contrast agent while removing background tissues. However, significant vessel overlap in DSA images hinders accurate anatomical assessment. Therefore, reconstructing 3D vessel structures from DSA images is essential for clear visualization to support medical diagnosis. Existing DSA systems typically capture hundreds of images (133 in our study) for precise reconstruction based on the Feldkamp-Davis-Kress (FDK) algorithm, exposing patients and radiographers to significant radiation. In this work, we aim to achieve high-quality reconstruction efficiently with dozens of images to reduce radiation exposure.

In this paper, we introduce 4D radiative Gaussian splatting (4DRGS), the first Gaussian splatting-based framework for efficient 3D vessel reconstruction from sparse-view dynamic DSA images. A key observation is that vessels maintain static structures during DSA scanning process, while their attenuation varies over time due to the contrast agent flow. Therefore, we represent vessels as a set of 4D radiative Gaussian kernels, where each kernel acts as a local Gaussian-shaped time-varying attenuation distribution. The static vessel structures are modeled with time-invariant geometry parameters, including position, rotation, and scale. To mimic the temporal attenuation changes, we introduce a compact neural network dubbed dynamic neural attenuation field (DNAF). It takes kernel position and timestamp as input, and predicts the central attenuation. We splat these Gaussian kernels to synthesize DSA images via X-ray rasterization and optimize them by minimizing the disparities with real captured images. After training, 3D vessel volume is reconstructed via attenuation voxelization. Two innovations are further introduced to enhance reconstruction quality: (1) accumulated attenuation pruning to remove non-vessel kernels and (2) bounded scaling activation to reduce needle artifacts. Experiments on real-world data demonstrate our method's effectiveness for both 3D vessel reconstruction and 2D DSA image synthesis. Notably, 4DRGS achieves impressive results in 5 minutes and converges in 13 minutes, offering a speedup of 32x and 12x compared to the state-of-the-art (SOTA) method VPAL.