README.md

September 10, 2025 · View on GitHub

TOGS: Gaussian Splatting with Temporal Opacity Offset for Real-Time 4D DSA Rendering

arXiv Paper

Shuai Zhang 1, Huangxuan Zhao 3, Zhenghong Zhou 1, Guanjun Wu 2,Chuansheng Zheng 3,
Xinggang Wang 1,Wenyu Liu 1,📧

1 School of Electronic Information and Communications, Huazhong University of Science and Technology
2 School of Computer Science &Technology, Huazhong University of Science and Technology
3 Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology

(📧 corresponding author)

Abstract

img

We propose TOGS, a Gaussian splatting method with opacity offset over time, which can effectively improve the rendering quality and speed of 4D DSA. We introduce an opacity offset table for each Gaussian to model the temporal variations in the radiance of the contrast agent. Additionally, we introduced a Smooth loss term in the loss function to mitigate overfitting issues that may arise in the model when dealing with sparse view scenarios. This model achieves stateof-the-art reconstruction quality under the same number of training views. Additionally, it enables real-time rendering while maintaining low storage overhead.

Installation

git clone https://github.com/hustvl/TOGS.git
conda create -n togs
conda activate togs
pip install -r requirements.txt
python ./submodules/diff-gaussian-rasterization/setup.py install
python ./submodules/simple-knn-main/setup.py install

Training

Dataset preparation

Organize medical data into a form similar to D-NeRF.

  1. Use the ./processing_data/get_image.py script to extract image and angle information from DICOM data.
  2. Use the ./processing_data/get_json.py script to get transforms_train.json and transforms_test.json.
├── data
│   | medicaldata 
│     ├── imgname1.png
│     ├── imgname2.png
│     ├── imgname3.png
│     ├── imgname4.png
│     ├── ...
│     ├── transforms_train.json
│     ├── transforms_test.json

Training

python train.py -s ./data/medicaldata --eval --yaml_file /data5/zhangshuai/TOGS/arguments/30.yaml

usage: python train.py -s [data path] --eval --yaml_file [configfile path] (30: Number of training views)

Rendering

python render.py -m ./output/30_2024-08-23_13:53:56

usage: python render.py -m [model save path]

Evaluation

python metrics.py -m ./output/60_2024-01-01_21:36:19

usage: python metrics.py -m [model save path]

Visualization

Number of training views: 30, testing views: 103, PSNR: 31.80.

video

Visualizing 3D point clouds:

Acknowledgement

This project is built upon 3DGS.

Citation

@ARTICLE{11020652,
  author={Zhang, Shuai and Zhao, Huangxuan and Zhou, Zhenghong and Wu, Guanjun and Zheng, Chuansheng and Wang, Xinggang and Liu, Wenyu},
  journal={IEEE Journal of Biomedical and Health Informatics}, 
  title={TOGS: Gaussian Splatting With Temporal Opacity Offset for Real-Time 4D DSA Rendering}, 
  year={2025},
  volume={29},
  number={9},
  pages={6793-6805},
  keywords={Rendering (computer graphics);Three-dimensional displays;Biomedical imaging;Training;Neural radiance field;Image reconstruction;Real-time systems;Contrast agents;Deformation;Shape;Gaussian Splatting;4D DSA novel view synthesis;NeRF;medical imaging;real-time rendering},
  doi={10.1109/JBHI.2025.3575613}}