A Deep Learning Approach to Ultrasound Image Recovery

Dimitris Perdios¹, Adrien Besson¹, Marcel Arditi¹, and Jean-Philippe Thiran^{1, 2}

¹Signal Processing Laboratory (LTS5), Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland

²Department of Radiology, University Hospital Center (CHUV), Switzerland

Paper accepted at the IEEE International Ultrasonics Symposium (IUS 2017).

Based on the success of deep neural networks for image recovery, we propose a new paradigm for the compression and decompression of ultrasound (US) signals which relies on stacked denoising autoencoders. The first layer of the network is used to compress the signals and the remaining layers perform the reconstruction. We train the network on simulated US signals and evaluate its quality on images of the publicly available PICMUS dataset. We demonstrate that such a simple architecture outperforms state-of-the art methods, based on the compressed sensing framework, both in terms of image quality and computational complexity.

Paper updates:

Please note that the accepted preprint version (v1) has been updated. Each version can be found here.

• v1: accepted preprint
• v2: fix typos in the abbreviations used in Table II w.r.t. the text
• v3: fix fig. 2(f) display dB range and corresponding interpretation

Installation

1. Install Python 3.6 and optionally create a dedicated environment.

2. Clone the repository.

   git clone https://github.com/dperdios/us-rawdata-sda
   cd us-rawdata-sda

3. Install the Python dependencies from requirements.txt.

   • Note 1: by default, it will install the GPU version of TensorFlow. If you do not have a compatible GPU, please edit requirements.txt as follow:
     • Comment the line starting with tensorflow-gpu.
     • Uncomment the line starting with tensorflow.
   • Note 2: TensorFlow 1.3 has not been tested yet, hence requirements.txt will install 1.2.1.

     pip install --upgrade pip
     pip install -r requirements.txt

4. Download the PICMUS dataset, which will be used as the test set.

   python3 setup.py

   This will download, under datasets/picmus17, the following acquisitions, provided by the PICMUS authors:

   • dataset_rf_in_vitro_type1_transmission_1_nbPW_1
   • dataset_rf_in_vitro_type2_transmission_1_nbPW_1
   • dataset_rf_in_vitro_type3_transmission_1_nbPW_1
   • dataset_rf_numerical_transmission_1_nbPW_1

   It will also download, under datasets/picmus16 two in vivo acquisitions provided by the PICMUS authors (from the first version of the challenge), namely:

   • carotid_cross_expe_dataset_rf
   • carotid_long_expe_dataset_rf

Usage

Trained networks

To reproduce the results on the PICMUS dataset described above, using the trained networks provided under networks/ius2017, use the following command:

python3 ius2017_results.py

This will compute the beamformed image for all the trained networks on every PICMUS acquisitions. The corresponding images will be saved as PDF under results/ius2017.

Note 1: This can take some time since PICMUS acquisition for each measurement ratio will be beamformed using a non-optimized delay-and-sum beamformer.

Note 2: The codes used to generate the CS results cannot be provided for the moment. Sorry for the incovenience.

Once the results have been computed, it is possible to lauch the following command for a better visualization:

python3 ius2017_add_figure.py

This will save a PDF image, under results/ius2017/*_metrics.pdf, for each PICMUS acquisition (see above) displaying the results in terms of PSNR for each method (SDA-CL, SDA-CNL and CS) and measurement ratio.

Training the networks (not yet available)

We are currently investigating an appropriate way to release and distribute the training set that has been numerically generated using the open-source k-Wave toolbox. The code is however already provided. Once the training set is available, it will be possible to re-train the networks using the following command:

python3 ius2017_train.py

License

The code is released under the terms of the MIT license.

If you are using this code, please cite our paper.

Acknowledgements

This work was supported in part by the UltrasoundToGo RTD project (no. 20NA21_145911), funded by Nano-Tera.ch.