Abstract
Visualizing vascular flow is essential for understanding cardiovascular pathologies, improving diagnosis, and monitoring patients. Ultrasound contrast imaging uses microbubbles to enhance the scattering of the blood volume, enabling real-time visualization of blood flow. Recent developments in vector flow imaging have expanded the imaging capabilities of ultrasound by temporally resolving fast arterial flow. However, a key challenge that remains is the limitation in spatial resolution.
This thesis investigates the application of neural networks to ultrasound radiofrequency (RF) data to achieve super-resolution contrast-enhanced ultrasound imaging. Additionally, the demand for large datasets to train deep learning-based computational ultrasound imaging methods necessitates the development of a simulation tool that can reproduce the physics of ultrasound wave interactions with tissues and microbubbles.
Chapter 1 highlights recent advances in ultrasound imaging, including microbubble technology, ultrafast imaging, and ultrasound localization microscopy (ULM). ULM has recently gained much attention as an ultrasound super-resolution method. However, ULM relies on low concentrations of microbubbles in the blood vessels, resulting in long acquisition times.
Chapter 2 presents an alternative super-resolution approach, based on direct deconvolution of single-channel ultrasound RF signals with a one-dimensional dilated convolutional neural network (CNN). Because the neural network effectively deconvolves the RF data, the conventional link between pulse length and resolution is broken. Therefore, Chapter 3 explores the effect of transmit pulse characteristics on super-resolution performance.
However, these initial studies rely on synthetic data generated by a simplified simulator, hindering the translation to in-vitro and in-vivo studies. Chapter 4 addresses this issue by developing a physically realistic simulator, PROTEUS, incorporating phenomena such as hemodynamics, medium inhomogeneity, nonlinear propagation, and nonlinear microbubble behaviour. Chapter 5 applies this simulator to various contrast-enhanced ultrasound scenarios.
Chapter 6 addresses the need for accurate system characterization. A complete transducer characterization is performed, culminating in a virtual transducer that can be employed to generate realistic training data with ground-truth labels for future studies.
This thesis investigates the application of neural networks to ultrasound radiofrequency (RF) data to achieve super-resolution contrast-enhanced ultrasound imaging. Additionally, the demand for large datasets to train deep learning-based computational ultrasound imaging methods necessitates the development of a simulation tool that can reproduce the physics of ultrasound wave interactions with tissues and microbubbles.
Chapter 1 highlights recent advances in ultrasound imaging, including microbubble technology, ultrafast imaging, and ultrasound localization microscopy (ULM). ULM has recently gained much attention as an ultrasound super-resolution method. However, ULM relies on low concentrations of microbubbles in the blood vessels, resulting in long acquisition times.
Chapter 2 presents an alternative super-resolution approach, based on direct deconvolution of single-channel ultrasound RF signals with a one-dimensional dilated convolutional neural network (CNN). Because the neural network effectively deconvolves the RF data, the conventional link between pulse length and resolution is broken. Therefore, Chapter 3 explores the effect of transmit pulse characteristics on super-resolution performance.
However, these initial studies rely on synthetic data generated by a simplified simulator, hindering the translation to in-vitro and in-vivo studies. Chapter 4 addresses this issue by developing a physically realistic simulator, PROTEUS, incorporating phenomena such as hemodynamics, medium inhomogeneity, nonlinear propagation, and nonlinear microbubble behaviour. Chapter 5 applies this simulator to various contrast-enhanced ultrasound scenarios.
Chapter 6 addresses the need for accurate system characterization. A complete transducer characterization is performed, culminating in a virtual transducer that can be employed to generate realistic training data with ground-truth labels for future studies.
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 10 Dec 2024 |
Place of Publication | Enschede |
Publisher | |
Print ISBNs | 978-90-365-6404-5 |
Electronic ISBNs | 978-90-365-6405-2 |
DOIs | |
Publication status | Published - Dec 2024 |