Abstract
Dynamic Magnetic Resonance Imaging (dMRI) aims to visualize time-dependent physiological processes within the human body. Important examples are functional cardiac imaging, dynamic contrast-enhanced (DCE) MRI or time-resolved angiography among others.
Since fast acquisition schemes already reached physical admissible limits in terms of nerve stimulation and SAR, in-vivo applications are always bounded by a trade-off between spatial and temporal resolution or spatial coverage depending on the time-scale of the motion under investigation. In order to improve the described limitations, undersampling strategies in combination with recent advances in iterative MRI reconstruction based on compressed sensing and parallel imaging are applied.
This project focuses on the fast implementation of a novel reconstruction technique termed "Infimal Convolution of Total Generalized Variation functionals (ICTGV)" as sophisticated regularization functional in a variational setting to recover diagnostically valuable image quality from undersampled dMRI data. The software presented in this work utilizes GPU functionality by CUDA to accelerate an existing MATLAB implementation in a fast parallel manner.
Two clinical relevant imaging scenarios are examined to compare the performance of the software to the MATLAB based CPU reference implementation. Furthermore, the utilization of a new vendor independent MRI raw data format is demonstrated, to facilitate integration into the clinical work-flow.
The results in terms of GPU execution times exhibit clinically acceptable reconstruction times and possible speedups up to a factor of 40. In conclusion usage limitations are discussed and further improvements are proposed.
Since fast acquisition schemes already reached physical admissible limits in terms of nerve stimulation and SAR, in-vivo applications are always bounded by a trade-off between spatial and temporal resolution or spatial coverage depending on the time-scale of the motion under investigation. In order to improve the described limitations, undersampling strategies in combination with recent advances in iterative MRI reconstruction based on compressed sensing and parallel imaging are applied.
This project focuses on the fast implementation of a novel reconstruction technique termed "Infimal Convolution of Total Generalized Variation functionals (ICTGV)" as sophisticated regularization functional in a variational setting to recover diagnostically valuable image quality from undersampled dMRI data. The software presented in this work utilizes GPU functionality by CUDA to accelerate an existing MATLAB implementation in a fast parallel manner.
Two clinical relevant imaging scenarios are examined to compare the performance of the software to the MATLAB based CPU reference implementation. Furthermore, the utilization of a new vendor independent MRI raw data format is demonstrated, to facilitate integration into the clinical work-flow.
The results in terms of GPU execution times exhibit clinically acceptable reconstruction times and possible speedups up to a factor of 40. In conclusion usage limitations are discussed and further improvements are proposed.
| Originalsprache | englisch |
|---|---|
| Qualifikation | Master of Science |
| Gradverleihende Hochschule |
|
| Betreuer/-in / Berater/-in |
|
| Publikationsstatus | Veröffentlicht - 2016 |
| Extern publiziert | Ja |