Benchmarking and Exploring New Low-powerd Architectures for Ultrasound Processing
MetadataShow full item record
Small embedded systems are becoming more and more powerful through parallelization. In this thesis, we investigate how suitable such systems are for real-time medical signal processing. We start by analyzing and characterizing the performance of two recent energy-efficient many core embedded systems: Adapteva's Parallella board and NVIDIA's Jetson TK1 Development kit based on Tegra K1. Although, the Parallella board has a lower energy consumption, our application proved to be a challenge for it, due to memory constraints and the slow bus between the host and the Epiphany co-processor. We hence developed our imaging application for the TK1. Medical imaging applications are of particular interest since they are computationally demanding. Current diagnostic ultrasound techniques often suffer from poor image resolution. Most of these issues are due to what is referred to as multiple scattering or reverberation noise. This happens, for instance, when the signal tries to penetrate one or more soft layers (e.g. fatty skin tissue) to get to the image sought after (e.g. cancer cells). By using SURF (Second Order Ultrasound Field) imaging, a dual-frequency band technique where the conventional imaging pulse is manipulated by a second lower frequency pulse, a greater deal of information can be extracted from the propagating pulse. This information can in turn be used to create models that can counter the effects of destructive signals components. In addition, better noise suppression when using the same transmit and receive beam can be achieved. In this thesis, we investigate a novel method for achieving synthetic dynamic focusing on data obtained through static beamforming with the same transmit and receive beam. In particular, we show that it is possible to create a filter based on the simulated Westerwelt equation to achieve synthetic dynamic focusing by transversal filtering. We also compare our method to data obtained with dynamic aperture focusing. Our comparison shows that the synthetic dynamic focusing method to has a slight advantage. There is, however, a problem in the area surrounding the focus that needs to be resolved before the method is optimal. However, we do show that our method has great potential for use on low-powered GPUs such as NVIDIA's Tegra TK1. In particular, we show that we are able to meet real-time requirements: Our initial goal was 20 Frames Per Second, i.e 50 ms of processing per frame. However, our synthetic dynamic focusing algorithm on the Jetson TK1 is able to process a frame in 24 milliseconds! Our method is also tested on more powerful GPU PC hardware where we are able to process the same data set in 8.8 milliseconds. Our results thus prove that the increased power of embedded microprocessors and GPUs can be a valuable resource in ultrasound processing.