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.  This thesis proposes learning-based methods to tackle such trade-offs encountered in the ultrasound signal processing pipeline.


First, we propose new artifact correction methods for different high frame-rate ultrasound acquisition configurations. We frame artifact correction as a supervised learning problem and train a convolutional neural network on pairs of raw-signals acquired from low and high frame-rates. We demonstrate that such learning-based methods achieve significant artifact reduction without effecting the image quality.

Second, we propose an end-to-end differentiable Rx pipeline for ultrasound signal processing, by rendering the dynamic focusing as a differentiable resampling-interpolation step. This allowed us to design transmit patterns that are optimal for a chosen ultrasound acquisition mode. We demonstrate that the frame-rate vs. reconstruction accuracy trade-offs can be pushed further by jointly designing Tx beampatterns with a learning-based Rx beamformer. We also demonstrated superiority of the learned beampatterns on standalone beamformers, such as delay-and-sum; and the superiority of learned Rx beamformer on several high frame-rate acquisition modes, generalizing across anatomies. This makes both of these blocks useful for practical deployment.