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

Zhentao Liu*, Ruyi Zha*, Huangxuan Zhao, Hongdong Li, and Zhiming Cui

Arxiv | Project Page

Code for IPMI 2025 Oral paper. We present 4DRGS, the first Gaussian splatting-based framework for efficient 3D vessel reconstruction from sparse-view dynamic DSA images. Our method achieves impressive results with sparse input (30 views) in minutes, highlighting its potential to support real-world medical assessment while reducing radiation exposure.

Updated Feature

• [2025-11-09] We now support LEAP toolbox for FDK reconstruction. TIGRE toolbox may encounter a CUDA error as reported in issue #3. You can select the desired toolbox in arguments/__init__.py via ModelParams.fdk_toolbox.
• [2025-08-07] tiny-cuda-nn now comes with a just-in-time (JIT) compilation mode. We have updated this feature in scene/field.py by setting model.jit_fusion = tcnn.supports_jit_fusion(), which provides some speed improvements. Note that tinycudann>=2.0 is required. Results in our paper is reported with tinycudann==1.7.

Setup

First clone this repo. And then set up an environment and install packages. C++ Compiler is required. We used Visual Studio 2019 for Windows and GCC 8.3.0 for Linux.

git clone https://github.com/ShanghaiTech-IMPACT/4DRGS.git
cd 4DRGS
conda create -n 4DRGS python=3.8
conda activate 4DRGS
pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
git clone https://github.com/CERN/TIGRE.git
cd TIGRE
pip install .
cd ..
git clone --recursive https://github.com/nvlabs/tiny-cuda-nn
cd tiny-cuda-nn/bindings/torch
python setup.py install
cd ../../..
pip install submodules/diff-Xray-gaussian-rasterization-voxelization
pip install submodules/simple-knn

Please refer to LEAP toolbox for your best installation. The following is what I do.

git clone https://github.com/LLNL/LEAP.git
cd LEAP

## for windows user ##
.\etc\win_build.bat
copy /y .\win_build\bin\Release\libleapct.dll.
## for linux user ##
sh ./etc/build.sh
cp ./build/lib/libleapct.so .

python manual_install.py

Data-Preparation

We provide case2 in our paper, and you can find it in this data link, including fill run, mask run, reference reconstructed volume from DSA scanner, reference mesh, and geometry description json file. You may use it for quick validation.

Training

After downloading the data, you could run the following command to train your model.

python train.py -m=output/case2_30v_30k -s=./dataset/case2 --Nviews=30

In this way, you would train a model with 30 input views on case2 for 30k iteration, finished in tens of minutes. You can also train a fast version in several minutes as follows.

python train.py -m=output/case2_30v_10k -s=./dataset/case2 --Nviews=30 --iteration=10000 --ADC_until_iter=5000

Testing

Use the following commands to test your trained model. It would conduct multi-view rendering, fix-view rendering, and 3D vessel reconstruction.

python test.py -m=output/case2_30v_30k -s=./dataset/case2 --Nviews=30 --render_2d --render_fixview --VQR
python test.py -m=output/case2_30v_10k -s=./dataset/case2 --Nviews=30 --iteration=10000 --render_2d --render_fixview --VQR

Related Links

• Traditional FDK reconstruction is implemented based on TIGRE toolbox and LEAP toolbox
• The first 3DGS-based framework for CT reconstruction: R²-Gaussian
• The first 3DGS-based framework for DSA image synthesis: TOGS
• NeRF-based framework for DSA reconstruction: VPAL, TiAVox
• It is recommended to observe medical data in nii format with ITK-SNAP or 3D Slicer.

Our method is developed based on the amazing open-source code: 3DGS and R²-Gaussian.

Thanks for all these great works.

Contact

There may be some errors during code cleaning. If you have any questions on our code or our paper, please feel free to contact with the author: [email protected], or raise an issue in this repo. We shall continue to update this repo. TBC.

Citation

If you think our work and repo are useful, you may cite our paper.

@article{4DRGS,
  title={4DRGS: 4D Radiative Gaussian Splatting for Efficient 3D Vessel Reconstruction from Sparse-View Dynamic DSA Images},
  author={Liu, Zhentao and Zha, Ruyi and Zhao, Huangxuan and Li, Hongdong and Cui, Zhiming},
  journal={arXiv preprint arXiv:2412.12919},
  year={2024}
}