M.Sc Thesis

M.Sc StudentSenouf Ortal
SubjectImproving Ultrasound Imaging with Deep Neural Networks
DepartmentDepartment of Electrical and Computers Engineering
Supervisors PROF. Alexander Bronstein
DR. Michael Zibulevsky
Full Thesis textFull thesis text - English Version


Ultrasound is a common medical imaging modality, mainly because it is cost-effective, flexible, and does not require the use of ionizing radiation. In recent years, many efforts within ultrasound research have been targeted at developing faster, more compact and portable ultrasound devices without compromising the image quality. This goal has not been yet achieved, mostly due to the inherent trade-offs between different ultrasound imaging parameters such as frame-rate, size of transducer, resolution, signal-to-noise-ratio and contrast. Existing model-based methods trying to improve one parameter of the imaging, or even resolve one trade-off, usually harm the quality with respect to other parameters.  

The recent revival of artificial neural networks, or more specifically, deep neural networks, and their phenomenal performance in many tasks, including classification, detection and image restoration, had lead us to believe that applying them to the ultrasound signal processing pipeline would provide a better or even optimal trade-off of the different imaging parameters.

In light of that, at this work, we present a step towards replacing the traditional ultrasonic pipeline with a learning based one.

First, we show that by training an end-to-end convolutional neural network (CNN), we can approximate conventional US post-processing algorithms and provide a major speed-up in run-times. Moreover, we show that a CNN can reconstruct a CT-quality image from ultrasound images within the same run-time range.  Trying to tackle the inaccessibility of real CT-US paired data we make an attempt to approximate ultrasound simulators in an unsupervised way.

Second, we optimize artifact correction for different high frame-rate ultrasound configurations, by training an end-to-end CNN on pairs of raw fast acquisition, distorted US signals and their corresponding slow acquisition, higher quality signals. We show that our learning based method achieves a significant artifact reduction without affecting other image qualities.