Nature Communications

ArXiv paper

Large-scale 3D photoacoustic (PA) imaging has become increasingly important for both clinical and pre-clinical applications. Limited by cost and system complexity, only systems with sparsely-distributed sensors can be widely implemented, which desires advanced reconstruction algorithms to reduce artifacts. However, high computing memory and time consumption of traditional iterative reconstruction (IR) algorithms is practically unacceptable for large-scale 3D PA imaging. Here, we propose a point cloud-based IR algorithm that reduces memory consumption by several orders, wherein the 3D PA scene is modeled as a series of Gaussian-distributed spherical sources stored in form of point cloud. During the IR process, not only are properties of each Gaussian source, including its peak intensity (initial pressure value), standard deviation (size) and mean (position) continuously optimized, but also each Gaussian source itself adaptively undergoes destroying, splitting, and duplication along the gradient direction. This method, named the sliding Gaussian ball adaptive growth (SlingBAG) algorithm, enables high-quality large-scale 3D PA reconstruction with fast iteration and extremely low memory usage. We validated SlingBAG algorithm in both simulation study and in vivo animal experiments.

As a by-product of our research on SlingBAG, we also developed a novel approach: utilizing the forward process of SlingBAG as a fast method to generate simulated photoacoustic data. This simulation method completes in mere seconds what k-Wave requires tens of hours to compute, while reducing memory consumption by orders of magnitude. The simulation results achieve a similarity of over 99.9% with those generated using k-Wave.
The detailed usage guidance for simulation can be found here:
Ultrafast Simulation Method for 3D Photoacoustic Imaging
If you find this simulation tool useful and want to use it for simulation or just as a forward model in reconstruction algorithms, you can simply cite our article "Sliding Gaussian ball adaptive growth: point cloud-based iterative algorithm for large-scale 3D photoacoustic imaging", thanks!

The overview framework of Sliding Gaussian Ball Adaptive Growth algorithm. (a) The SlingBAG pipeline. (b) The forward simulation of differentiable rapid radiator. (c) Adaptive growth optimization based on iterative point cloud.

Comparison of reconstruction results of rat liver between UBP and SlingBAG.

Comparison of reconstruction results of hand vessels between UBP and SlingBAG using sparse 576 sensors.

The example in the provided codes is for the reconstruction of simulated hand vessel with 196 elements planar array (detailed in the article), if you want to reconstrut your own data, please replace the sensor location and sensor data files in the train_196_elements_coarse_recon.ipynb and train_196_elements_fine_recon.ipynb. Besides, the boundary set of the Gaussian balls should be modified carefully to match the reconstruction area. Sorry for all the inconvenience, we promise that the SlingBAG will soon be much more user-friendly, and we hope this guidance may help you. Good luck my friends!
If you have any questions while using SlingBAG for 3D reconstruction, please be free to contact us. Best wishes!

git clone https://github.com/JaegerCQ/SlingBAG.git
cd SlingBAG

conda create -n SlingBAG python=3.10 -y
conda activate SlingBAG
conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia

pip install -r requirements.txt

Warning: for Windows, it needs setuptools <= 72.1.0; for Linux, it needs gcc >= 9.1.

Run train_196_elements_coarse_recon.ipynb.

Run train_196_elements_fine_recon.ipynb.

Run point_cloud_to_voxel_grid_shader.ipynb.

utils/differentiable_rapid_raditor_kernel_v4_fine.cu is the latest version of the CUDA kernel for the fine reconstruction stage. Compared to utils/differentiable_rapid_raditor_kernel_v3_fine.cu, it incorporates shared memory access, optimizes computational efficiency, and accelerates the reconstruction process. In train_196_elements_fine_recon.ipynb, simply replace utils/differentiable_rapid_raditor_kernel_v3_fine.cu with utils/differentiable_rapid_raditor_kernel_v4_fine.cu to use the updated version.

@article{li2025slingbag,    
  title={SlingBAG: point cloud-based iterative algorithm for large-scale 3D photoacoustic imaging},    
  author={Li, Shuang and Wang, Yibing and Gao, Jian and Kim, Chulhong and Choi, Seongwook and Zhang, Yu and Chen, Qian and Yao, Yao and Li, Changhui},    
  journal={Nature Communications},    
  year={2025},    
  publisher={Nature Publishing Group UK London}    
}

We are deeply grateful to Professor Chulhong Kim and Dr. Seongwook Choi for providing the invaluable in vivo experimental data. We are also grateful to the authors of 3D Gaussian Splatting for their great work as well as open source codes that inspire us a lot.