Postdoctoral position: Deep learning-based reconstruction for 3D ultrasonic imaging

1.      Deep learning-based reconstruction for 3D ultrasonic imaging

1.2.    Context:

3D data acquisition for ultrasonic imaging uses probes made of a matrix of sensors. For reasons of physical space, connectivity and control, only a small fraction of these sensors can be activated at the same time. Furthermore this type of acquisition leads to a very important data stream, which limits the imaging speeds less than 20 frames/s. One strategy to overcome these problems is to reduce the number of acquired ultrasound lines and develop method for reconstruction the the so-obtained subsampled volume.

1.3.    Objective:

The compressed sensing techniques are well suited to this type of reconstruction and have been applied to this problem at Creatis [1-2]. However, compressed sensing-based reconstruction involves solving a minimization problem of very large dimensions. The numerical solution of this problem has led to the development of many iterative algorithms based on convex relaxation techniques or greedy algorithms. Unfortunately, none of these algorithms currently achieves computing speeds compatible with real-time acquisition of the ultrasound images.

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 [3-6] and its application to ultrasound images is still unexplored.

As a consequence, a number of key points will have to be addressed in this work:

  • The selection of the neural network type. In particular, the formulations based on convolutional networks, simple, variational or recurring autoencoders [5-8] will have to be examined.
  • The selection of a sub-sampling strategy, which has to be adapted to the ultrasound acquisition and must also allow to optimize the reconstruction DNN [9-10]

1.4.    Methodology:

The development of DNNs adapted to the representation and reconstruction of ultrasound data will be based on Keras and Theano python libraries. The developed approach will be optimized and evaluated in terms of computing time/accuracy trade-off, first on 3D ultrasound data from numerical simulations and then on experimental data acquired on ex vivo organs using the research ultrasound scanner available at Creatis.

1.5.    References:

[1]    O. Lorintiu, H. Liebgott, M. Alessandrini, O. Bernard, and D. Friboulet, "Compressed sensing reconstruction of 3D ultrasound data using dictionary learning and line-wise subsampling," IEEE Transactions on Medical Imaging, 2015.
[2]    J. Richy, D. Friboulet, A. Bernard, O. Bernard, and H. Liebgott, "Blood Velocity Estimation Using Compressive Sensing," IEEE Transactions on Medical Imaging, vol. 32, pp. 1979-1988, 2013.
[3]    K. Gregor and Y. LeCun, "Learning Fast Approximations of Sparse Coding," in International Conference on Machine Learning (ICML-10), Haifa, Israel, 2010, pp. 399-406.
[4]    P. Sprechmann, A. M. Bronstein, and G. Sapiro, "Learning Efficient Sparse and Low Rank Models," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 37, pp. 1821-1833, 2015.
[5]    A. Mousavi, A. B. Patel, and R. G. Baraniuk, "A Deep Learning Approach to Structured Signal Recovery," ArXiv, pp. 1-8, 2015.
[6]    K. Kulkarni, S. Lohit, P. Turaga, R. Kerviche, and A. Ashok, "ReconNet: Non-Iterative Reconstruction of Images from Compressively Sensed Random Measurements," ArXiv, pp. 1-10, 2016.
[7]    D. P. Kingma and M. Welling, "Auto-Encoding Variational Bayes," ArXiv, pp. 1-14, 2014.
[8]    J. T. Rolfe and Y. Lecun, "Discriminative recurrent sparse auto-encoders," in International Conference on Learning Representations (ICLR), Scottsdale, Arizona, USA, 2013, pp. 1-15.
[9]    L. Baldassarre, Y. H. Li, J. Scarlett, B. Gözcü, I. Bogunovic, and V. Cevher, "Learning-Based Compressive Subsampling," IEEE Journal of Selected Topics in Signal Processing, vol. 10, pp. 809-822, 2016.
[10]  J. Bigot, C. Boyer, and P. Weiss, "An Analysis of Block Sampling Strategies in Compressed Sensing," IEEE Transactions on Information Theory, vol. 62, pp. 2125-2139, 2016.
 

2.      Application:

2.1.    Profile

PhD in machine learning, showing a very good experience regarding approaches based on deep neural networks.

2.2.    Documents

Curriculum + cover letter to be sent to denis.friboulet@creatis.insa-lyon.fr  and fabien.millioz@univ-lyon1.fr before November 15, 2016

2.3.  Funding:

Funding for this post-doc is provided by the Labex PRIMES. The expected net monthly salary is € 2,100.