Real-Time, Dynamic, and Highly Generalizable Ultrasound Image Simulation-Guided Procedure Training System for Musculoskeletal Minimally Invasive Treatment.
Here, we propose a Real-time, Dynamic, and highly Generalizable UltraSound Image Simulation (RDG-USIS) algorithm, specifically designed to enhance training in minimally invasive procedures.
The RDG-USIS:
Our developed ultrasound image simulation-guided minimally invasive procedure training system integrates the proposed RDG-USIS algorithm. It generates high-quality ultrasound images from CT scans (see module indicated by the red circle). It supports real-time, dynamic alignment with other multimodal imaging data, significantly enhancing 3D spatial understanding and surgical accuracy during ultrasound-guided training.
The generation of convolutional images requires the following input: a nii.gz format mask file that has been segmented by totalsegmentator, and the modification_mask_label function in the cov_img/nii_deal.py file needs to be called for preprocessing.
python cov_img/get_sim_us.pyInstall dependencies:
pip install -r requirements.txtThe project is only compatible with multi-GPU DDP mode for training.
CUDA_VISIBLE_DEVICES=0,1,2,3 torchrun --master_port=12345 --nnodes=1 --nproc_per_node=4 train.py --dataroot ./datasets/test --name test --model cycle_gan --use_distributed --lambda_ssim 5After the article is accepted, we will open-source the high-quality US-CT dataset that we have designed and collected, which will have a positive impact on the community.
If you find this repo useful for your research, please consider citing our papers:
@inproceedings{wang2025real,
title={Real-Time, Dynamic, and Highly Generalizable Ultrasound Image Simulation-Guided Procedure Training System for Musculoskeletal Minimally Invasive Treatment},
author={Wang, Xiandi and Jiang, Zekun and Tang, Mengqi and Han, Ying and Pu, Dan and Li, Kang},
booktitle={International Workshop on Human-AI Collaboration},
pages={35--43},
year={2025},
organization={Springer}
}