Implementation of neural networks on ultrasound scanners

I. Scientic Context

The joint laboratory Image4US is interested in ultrasonic medical imaging (US) with a high number of channels, with the objective of making 3D imaging. It associates the company DB-SAS (SME, Nantes, France), specialized in electronics for ultrasound, and the Centre de Recherche en Acquisition et Traitement de l'Image pour la Santé (Creatis, Lyon, France) specialized in medical imaging.


Ultrasound is today the most widespread medical imaging modality in the world thanks to fast, transportable and non-invasive devices. While 2D imaging is commonly practiced in hospitals, 3D ultrasound is slow to become established in clinical practice. This is mainly due to the lack of image quality in terms of resolution, contrast and frame rate.


One of the reasons for this gap is the difficulty of obtaining high quality 3D images at high frame rates. In 2D, ultrafast imaging involves sending multiple plane waves or multiple divergent waves and then summing the individual images to obtain a high-quality image. However, the need to send many plane or divergent waves limits the frame rate.


II. Objectives of the internship

For several years, Creatis has been developing deep learning methods to further accelerate 2D ultrafast imaging, either in plane waves [1] or in divergent waves [2], allowing to gain a factor 10 on the frame rate. The two associated neural networks use the  radiofrequency signals from the ultrasound scanner. The second network has been recently modied to use demodulated radiofrequency signals, commonly called IQ (In phase/Quadrature) [3]. However, these methods have never been implemented directly on an ultrasound scanner, and have not been trained with the DB-SAS research ultrasound scanner. The main goal of this work is to implement these methods on the DB-SAS search ultrasound scanner.


The work will require several steps:

  • ˆ Understand the already existing neural networks [1, 2, 3]
  • ˆ Adapting the plane wave network to use IQ signals
  • ˆ Test the networks on the research ultrasound machine
  • ˆ Optimize these networks to increase the image rate as much as possible
  • Depending on the difficulties encountered, acquisitions with the DB-SAS ultrasound scanner will be planned to re-train the existing neural networks.

III. Required skills

  • ˆ Good Python programming skills (Pytorch, NumPy, SciPy)
  • ˆ Good knowledge in deep learning
  • ˆ Interest in biomedical imaging in general, ultrasound imaging in in particular

IV. Informations

  • Duration of the internship: 5 to 6 months
  • Location: Creatis laboratory, 21 Avenue Jean Capelle, Villeurbanne
  • ˆ Supervisors: Denis Friboulet (denis.friboulet@creatis.insa-lyon.fr), Fabien Millioz (fabien.millioz@creatis.insa-lyon.fr)
  • ˆ Send CV, cover letter and last transcript of marks

References

[1] M. Gasse, F. Millioz, E. Roux, D. Garcia, H. Liebgott, and D. Friboulet. High-quality plane wave compounding using convolutional neural networks. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , 64(10):16371639, 2017.
[2] J. Lu, F. Millioz, D. Garcia, S. Salles, W. Liu, and D. Friboulet. Reconstruction for diverging-wave imaging using deep convolutional neural networks. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , 67(12):24812492, 2020.
[3] J. Lu, F. Millioz, D. Garcia, S. Salles, D. Ye, and D. Friboulet. Complex convolutional neural networks for ultrafast ultrasound imaging reconstruction from inphase/quadrature signal. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control , 69(2):592603, 2022.