Deep learning-based ultrasonic image reconstruction for ultrafast echographic imaging
Ultrafast imaging based on plane wave (PW) is currently a very active area of research in ultrasound because of its capacity of reaching very high frame rate (several thousands/s in 2D). Obtaining ultrafast images using plane wave imaging however still remains a challenge in situations where extremely complex and fast phenomena happen, e.g. in the heart, due to the trade-off between image quality and frame rate: PW imaging indeed relies on compounding in order to preserve a good image quality, usually using multiple successive emissions, which in turn yields a decrease of the frame rate. One possible strategy to overcome this problem consists in reducing the number of acquired PW and developing a reconstruction method yielding high quality images despite the missing data.
Compressed sensing techniques are well suited to this type of reconstruction and have been recently applied to this problem at Creatis (Lorintiu et al., 2015; Lorintiu et al., 2016). However, compressed sensing-based reconstruction involves solving a minimization problem of very large dimensions, and the associated algorithms currently do not yield computing speeds compatible with ultrafast acquisitions.
In this context, the objective of this work is to develop an alternative method of reconstruction, based on a deep neural network (DNN), whose architecture will reduce the computation time by several orders of magnitude. The application of the DNN image reconstruction problems is a very recent topic (Jin et al., 2017; Metzler et al., 2017; Mousavi and Baraniuk, 2017) and its application to ultrasound images has only be very recently unexplored (Gasse et al. 2017).
As a consequence, the key points to be addressed in this work will be the following:
- Selection of the neural network type. In particular, the formulations based on convolutional networks, Generative adversarial networks, Compositional pattern producing networks will have to be examined.
- Once a type of network will have been selected, optimization of the network structure.
- Selection of input/output US data (i.e. raw radiofrequency, beamformed or envelope data) yielding the best performances
The development of DNNs adapted to the representation and reconstruction of ultrasound data will be based on PyTorch python library. The developed approach will be optimized and evaluated in terms of computing time/accuracy trade-off on experimental data acquired on ex vivo organs using the research ultrasound scanners available at Creatis.
Master’s in machine learning, showing a very good knowledge and experience of deep neural networks, as well as excellent skills in programming, training and testing such networks. Prior knowledge in ultrasound imaging is not a prerequisite, since the candidate will be trained in this field at Creatis. Some background and interest in the physics of ultrasound imaging would nevertheless be an asset.
- NB : the papers labelled with an asterisk (*) corresponds to works that have been performed at Creatis
*Gasse, M., Millioz, F., Roux, E., Garcia, D., Liebgott, H., Friboulet, D., 2017. High-Quality Plane Wave Compounding using Convolutional Neural Networks. IEEE Trans. Ultrason. Ferroelec. Freq. Cont. 64, 1637-1639.
Jin, K.H., McCann, M.T., Froustey, E., Unser, M., 2017. Deep Convolutional Neural Network for Inverse Problems in Imaging. IEEE Trans. Image Processing 26, 4509-4522.
*Lorintiu, O., Liebgott, H., Alessandrini, M., Bernard, O., Friboulet, D., 2015. Compressed sensing reconstruction of 3D ultrasound data using dictionary learning and line-wise subsampling. IEEE Trans. Med. Imaging 34, 2467‑2477.
*Lorintiu, O., Liebgott, H., Friboulet, D., 2016. Compressed sensing Doppler ultrasound reconstruction using block sparse Bayesian learning. IEEE Trans. Med. Imaging 35, 978-987.
Metzler, C., Mousavi, A., Baraniuk, R., 2017. Learned D-AMP: Principled Neural Network based Compressive Image Recovery, Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, USA, pp. 1770--1781.
Mousavi, A., Baraniuk, R.G., 2017. Learning to invert signal recovery via deep convolutional networks, IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), New Orleans, USA, pp. 2272-2276.