Deep-learning based iterative solver for the heterogeneous Helmholtz equation in 2D using a fully-learned optimizer. The lightweight network architecture is based on a modified UNet that includes a learned hidden state. The network is trained using a physics-based loss function and a set of idealized sound speed distributions with fully unsupervised training (no knowledge of the true solution is required).

Clone the repo in a local folder. Create an environment containing PyTorch, PyTorch Ligthning and various other libraries. If you don't want to install them by yourself, you can create such environment in anaconda using

conda env create -f environment.yml

However, you must obtain a copy of kWave and MATLAB by yourself. Depending on where k-Wave is installed, you may have to change some paths in some m files in the matlab folder.

To install the package use

pip install .

If you want to install in editable mode, use

pip install -e .

To retrain the network, first generate the training data by running generate_dataset.py. Delete or rename the last.ckpt file from the checkpoints folder, and then train the network using the following syntax:

python train.py --gpus 0,1

where the number of GPUs can be specified. You can visualize the training status in tensorboard, using

tensorboard --logdir logs/

And open http://localhost:6006/ on a browser.

The exact dataset used in the paper can be downloaded from this Google Drive link. The downloaded zip file needs to be extracted in the root folder of the repository. This dataset only have speed of sound maps in it.

For supervised learning, another dataset with pre-computed wavefields and denser sampling can be downloaded from this repository from Son Hai Nguyen.

from helmnet import IterativeSolver
from helmnet.support_functions import fig_generic
import numpy as np
import torch

solver = IterativeSolver.load_from_checkpoint(
    "trained_models/jcp_paper_trained_weights.ckpt", strict=False, test_data_path=None
)
solver.freeze()  # To evaluate the model without changing it
solver.to("cuda:0")

# Setup problem
source_location = [30, 128]
sos_map = np.ones((256, 256))
sos_map[100:170, 30:240] = np.tile(np.linspace(2,1,210),(70,1))

# Set model domain size (assumed square)
solver.set_domain_size(sos_map.shape[-1], source_location=source_location)

# Run example in kWave and pytorch, and produce figure
fig_generic(
    solver,
    sos_map,
    path="images/custom",
    source_location=source_location,
    omega=1,
    min_sos=1,
    cfl=0.1,
    roundtrips=10.0,
    mode="normal",
)

The model results are compared against k-Wave with the given CFL and rountrips parameters

To get the kWave and GMRES results on the test set you should

1. run parallel_sectral_gmres_solver.m
2. run parallel_kwave_solver.m

To get the model results on the test set run

python evaluate.py

All the figures can then be generated using

python produce_figures.py

For the real skull example, you'll need the file qure_ai-CQ500CT2-Unknown Study-CT0.625mm-CT000112.dcm from the qure.ai dataset in the examples folder.

@article{stanziola2021helmholtz,
    title = {A Helmholtz equation solver using unsupervised learning: Application to transcranial ultrasound},
    author = {Antonio Stanziola and Simon R. Arridge and Ben T. Cox and Bradley E. Treeby},
    journal = {Journal of Computational Physics},
    pages = {110430},
    year = {2021},
    issn = {0021-9991},
    doi = {https://doi.org/10.1016/j.jcp.2021.110430},
    url = {https://www.sciencedirect.com/science/article/pii/S0021999121003259}
}