Acoustic Simulation with Deep Learning for Low-intensity Transcranial Focused Ultrasound Digital Twins

Overview

This repository contains the official implementation of "Acoustic Simulation with Deep Learning for Low-intensity Transcranial Focused Ultrasound Digital Twins", accepted to the Workshop on Digital Twin for Healthcare 2025 (at MICCAI 2025).

Features

• CNN-based Autoencoder and U-Net
• Swin Transformer-based U-Net
• Python codes for training, evaluation, loading dataset
• Pre-trained model weights

Note: Dataset is not provided due to privacy concerns.

0. Installation

Clone this repository: git clone https://github.com/CMME-Lab/LIFUSimul-DL.git
Install all prerequisites with pip install -r requirements.txt

1. Preparing dataset

We do not provide the dataset due to privacy concerns.
For your experiments, please prepare the dataset in the following format.

Instructions

• All data must be defined in the HDF5 file format. The keys for each data point within the HDF5 file must be sortable in order by the natsorted function.
• The data must be organized sequentially by subject, and it is assumed that each subject has the same number of data points.
  • This is to ensure that the same ratio of train/valid data is extracted for each subject.
  • To modify this behavior, please adjust the split_dataset function in dataset.py.
• Place the following files in the parent directory:
  • ff_train.hdf5, ff_test.hdf5 (Acoustic free-field)
  • ct_train.hdf5, ct_test.hdf5 or mr_train.hdf5, mr_test.hdf5 (Skull images)
  • td_train.hdf5, td_test.hdf5 (Transducer placement)
  • target_train.hdf5, target_test.hdf5 (Intracranial acoustic field)
    Afterwards, modify the default value of the data_path argument in config.py to ensure the model always references the correct dataset location.

Note: For reproducibility, it is assumed that the transducer placement data has already undergone Fourier feature embedding. Please refer to the fourier_feature_embed function in utils.py to prepare your data by completing the embedding according to its format.

Note: Compute maximum and minimum value of your acoustic free-field, and replace the value of ff_max_value and ff_min_value in MinMaxScaling (utils.py) for proper scaling.

2. Training

Run the training process using train.py.

Example usage :
python train.py --run_name my_experiments --modality ct --model swin --num_epoch 100 --decay_epoch 100 --init_model --cuda

3. Evaluation

Run the evaluation using test.py.

Example usage :
python test.py --run_name my_experiments --modality ct --model swin --cuda --plot

Authors

• Minjee Seo, School of Mathematics and Computing (Computational Science and Engineering), Yonsei University, Seoul, Republic of Korea
• Minwoo Shin, Department of Software, Yonsei University, Wonju, Republic of Korea
• Gunwoo Noh, School of Mechanical Engineering, Korea University, Seoul, Republic of Korea
• Seung-Schik Yoo, Department of Radiology, Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA
• Kyungho Yoon, School of Mathematics and Computing (Computational Science and Engineering), Yonsei University, Seoul, Republic of Korea

Acknowledgement

The work was supported by the National Research Foundation of Korea (NRF) funded by the Korean government (MSIT) under Grants RS-2024-00335185 and RS-2023-00220762.

License

MIT License

Contact

For any queries, please reach out to Minjee Seo.