|M.Sc Student||Peretz Nissim|
|Subject||Deep Learning Applied to Beamforming in Medical Ultrasound|
|Department||Department of Electrical Engineering||Supervisor||Professor Emeritus Arie Feuer|
|Full Thesis text|
Beamforming techniques have been applied in many fields such as radar, sonar and medical imaging. A beamformer is a spatial filter that has the ability to form a main beam in the direction of the desired signal, and place nulls in the directions of interferences. The basic idea is to impose delays to the transmitted/received RF signals, so that the desired signals are combined coherently at a focusing point, while the interferences are combined incoherently.
A recent trend is deep learning, which has been proven to be a promising tool, providing state-of-the-art results in numerous machine learning tasks. Deep learning has been successfully applied to inverse problems such as denoising, deconvolution and super-resolution. In particular, deep networks make use of large data sets to learn the unknown inverse mapping to the inverse problem.
In this thesis, we use deep neural networks for inverse problem in the context of ultrasound beamforming. More specifically, we reconstruct an image of a points reflecting object, from wideband RF signals, generated by translating a single transmitter/receiver element over a sensor array.
In the first part of this thesis, we explore the relationship between the inverse mapping and the sensor array transmission, for several classes of transmit/receive configuration. We investigate the concept of monostatic aperture, where the same element in the array is used both in transmit and receive modes, and the concept of multistatic aperture, where different elements act as transmitter/receiver pairs.
Motivated by the above, we propose to use deep neural networks to solve that inverse problem. The conventional beamforming process is a linear approximate solution to the inverse problem of the reconstruction of the reflectivity field. While such linearized model is attractive because of its simplicity, it has the defect of ignoring the basic nonlinear nature of the inverse scattering problem, e.g., multiple reflections.
In the second part of this thesis, we generalize the above idea and draw similarities between the pulsed-Doppler spectral estimation problem and spatial beamforming. In beamforing, in order to increase the lateral resolution, the effective aperture of the array must be increased by expanding the physical size of the array. In pulsed-Doppler ultrasound, this corresponds to increasing the number of transmitted pulses. We show that it is possible to decrease the number of transmissions, and by deep neural network spectral estimation still maintain acceptable spectral resolution and contrast.