Scalable Ultrafast Ultrasound Delay-and-Sum Beamforming design on an FPGA

Abstract

Ultrasound imaging has been established as a dominant imaging modality for evaluation and treatment of medical conditions, due to its non-invasive nature that is harmless for the patient. This imaging method is based on the ideas of the Doppler effect, producing and receiving high frequency acoustic waves (1-20 MHz), while supporting frame rates that were limited to 10-40 fps. However, since there is a plethora of dynamic responses observed inside the human body, the need to capture transient events has emerged, by increasing the frame rates of the application. This high frame rate imaging method is known as ultrafast ultrasound imaging. High frame rates could be supported by improving the beamforming rate of the ultrasound system, a component that often introduces bottlenecks during attempts of integrating ultrafast imaging. This thesis will focus on implementing an ultrafast beamforming kernel on an FPGA device, a potential candidate due to their configurable computational architecture that enables immense parallel, real-time processes. The main goal is to create a hardware kernel that receives raw ultrasound signals and beamforms them, compounds the contributions from different angles and saves them for further processing. While previous approaches have managed to design such kernels on FPGA devices, they either lack in adequate frame rate and beamforming rate, amount of probe channels or scalability. More specifically, we propose two ultrafast beamforming FPGA kernels; one that applies an interpolation where estimations equal to the mean value of the two adjacent ultrasound samples, and one that applies linear interpolation. The dataset used for beamforming was collected by the IEEE IUS 2016 Plane-wave Imaging Challenge in Medical UltraSound : PICMUS. While supporting an ultrafast beamforming process with 128 probe channels, the first version of the proposed kernel supports a pulse repetition frequency of 11.116 KHz and a beamforming rate of 2.173 GSamples/s. On the other hand, the second version of the proposed kernel supports a pulse repetition frequency of 10.014 KHz and a beamforming rate of 1.957 GSamples/s. The scalability for both of these designs was established, regarding the amount of utilized probe channels, the amount of plane waves, as well as an internal parameter of our beamformer, that will be explained later in this thesis