Abstract
I medisinsk ultralydavbildning, og i avbildningsystemer generelt, er det et stadig ønske om å oppnå høyere oppløsning og bedre bildekvalitet. I denne avhandlingen presenteres teknikker for å imøtekomme dette ved hjelp av mer avansert signalbehandling. Ved bruk av såkalte høyoppløselige metoder demonstreres vesentlige forbedringer i forhold til standard metoder brukt i ultralydavbildning.
En stor utfordring ved medisinsk ultralyd er at avbildningen foregår i sanntid. Høyoppløselige metoder har vært ansett som for komplekse eller for beregningskrevende for denne anvendelsen. I dette arbeidet foreslås det en høyoppløselig metode med svært lav kompleksitet, som er lett å implementere, og er spesielt egnet for ultralyd. I tillegg demonstreres det hvordan høyoppløselige metoder generelt kan anvendes på aktive avbildningssystemer på en robust måte.
Adaptive beamforming methods is a class of high resolution techniques that can improve the performance of imaging systems. Adaptive beamformers offer increased resolution over conventional methods as they take the recorded wavefield into account to find optimal beampatterns. These techniques have been exploited in fields like radar, sonar and seismology. In medical ultrasound, delay-and-sum beamforming is still the method of choice. This is a simple and robust method which suits the real-time requirements of such imaging systems. However, it offers limited resolution and sidelobe suppression compared to adaptive methods. The main goal of this project has been to study adaptive beamforming techniques for medical ultrasound imaging. Active systems, such as ultrasound, present specific challenges for these methods. Four papers concerning different aspects of adaptive beamforming for medical imaging have been included in this dissertation. We have also included a paper concerning adaptive beamforming in a single snapshot context, and a paper on blind source separation.
The first three papers concern the “minimum variance” beamformer. In the first paper we present how the method can be applied to medical ultrasound. We demonstrate increased resolution and suppression of sidelobes, both on simulated and experimental RF data. We also evaluate the robustness of the method, and show that increased robustness can be achieved by simple means. In the second paper, we investigate how the estimate of the spatial covariance matrix, used by the minimum variance beamformer, affects the statistics of speckle patterns. We show that an implementation based on a single snapshot of the wavefield gives very different speckle statistics compared to delay-and-sum. By averaging in depth, similar speckle as delay-and-sum is achieved, while the resolution of the method is retained. We exploit the high-resolution properties of the minimum variance beamformer in the third paper, and show that it can be used to decrease transducer size, increase frame rates or give higher penetration without sacrificing image quality compared to delay-and-sum. The minimum variance beamformer is significantly more complex than conventional methods, which is one of the reasons why it is not used in medical ultrasound systems. In the fourth paper we present a simplified adaptive beamformer, which better suits the real-time requirements of medical ultrasound. The method requires only a fraction of the number of computations compared to a full adaptive beamformer, but still gives significant improvements compared to delay-and-sum.
In the fifth paper we present a unifying framework to analyze the minimum variance beamformer in a single snapshot context. We show that implementations based on single snapshots may suffer from signal cancellation, a well-known phenomenon when sources are correlated. The framework allows us to construct an optimization which completely eliminates signal cancellation.
The last paper concerns the field of blind source separation. Blind source separation is a class of methods that can extract source information when we only have observations of mixtures of the sources available. We present a new approach to separation of convolutive mixtures by preprocessing the data using an array processing technique. We apply the method to the so-called “cocktailparty problem”.