| dc.description.abstracteng | Magnetic resonance imaging (MRI) is a non-invasive technique widely used for imaging of humans and animals. It came into existence based on several fundamental inventions made in the 1970s. Since then, MRI has evolved into one of the leading tools in biomedical research and clinical diagnostics. Compared to other medical imaging modalities such as X-ray computed tomography (CT) or positron emission tomography (PET), MRI uses neither ionizing radiation nor radioactive materials and offers a rich set of image contrasts based on signal relaxation times T1 and T2, spin diffusion or more complex quantities such as tissue perfusion. Moreover, the acquisition of images with multiple contrasts in the same anatomical region enables the capability to obtain quantitative maps of the underlying relaxation parameters, which renders the MRI scanner not only a camera but also a scientific measuring instrument. This thesis specifically addresses the quantitative mapping of T1 relaxation times.
Quantitative T1 mapping normally consists of a suitable magnetization preparation (e.g., inversion), followed by the acquisition of a predefined number of images of the relaxation process. After data acquisition and image reconstruction, the relaxation model is then fit to the images to obtain the parameter maps. However, the acquisition of multiple images with a spatial and temporal resolution that is sufficient for clinical use with only one preparation may be difficult or even impossible in presence of short relaxation times. One way to overcome this problem is a segmentation of the data acquisition process. By acquiring complementary data subsets at different segmentations, the temporal resolution can be highly improved. For segmented acquisitions with preparation by inversion, a sufficient delay time is required between the end of one segment and a following inversion for the next acquisition. This delay period largely prolongs the total data acquisition time and restricts the in vivo applications of T1 mapping techniques. Therefore, novel ways that allow for the reconstruction of accurate high-resolution T1 maps with less data sampled than required by the Nyquist criterion, known as undersampling, would be highly desirable.
In the last two decades, several techniques have been proposed to address the general problem of the long MRI acquisition times which are caused by the point-by- point data acquisition scheme in Fourier space. One approach is parallel imaging which uses multiple receive coils to acquire data in parallel. By exploiting complementary spatial information from these spatially distinct receiver coils, parallel imaging typically allows for a moderate reduction of spatial encoding steps (i.e., acceleration) by a factor of 2-4. Another effort to improve imaging speed is the development of non-Cartesian MRI. Especially the adoption of radial sampling has gained a lot of interest as it is inherently robust to motion and at least partly tolerant to undersampling. More specifically, the combination of highly undersampled radial fast low angle shot (FLASH) acquisitions, parallel imaging and image reconstruction by nonlinear inversion (NLINV) which jointly estimates the image and all coil sensitivity maps has achieved real-time MRI at millisecond resolution. Moreover, when calculating quantitative maps of relaxation parameters, a lot of redundancy may be found in the data of image series with different contrasts. Model-based reconstruction techniques promise to exploit such redundancies
by directly reconstructing parameter maps from raw data. Although this approach comes at a cost of increased complexity and computational demand, it has been successfully demonstrated in preliminary applications to T2 mapping.
The main focus of this thesis is to develop new methods for fast and accurate T1 mapping at high spatial resolution by taking advantage of both the above mentioned real-time MRI developments and the concept of a model-based reconstruction. The former method reconstructs images from highly undersampled radially encoded data sets, which then may be followed by a pixel-wise fitting to obtain parameter maps. The latter, on the other hand, bypasses the intermediate steps of image reconstruction and pixel-wise fitting by estimating parameter maps directly from the undersampled raw data with use of a known signal model.
As far as organization of this thesis is concerned, Chapter 2 discusses the basic principles of MRI, while Chapter 3 briefly explains undersampled radial FLASH acquisitions and NLINV reconstruction algorithms as the main components of real-time MRI. Chapter 4 introduces the three most commonly used experimental approaches to quantitatively measure the T1 relaxation process. The main achievements of this thesis are presented in Chapters 5 to 7.
Chapter 5 deals with the development of a single-shot high-resolution T1 mapping technique based on real-time MRI. Apart from the adaptation of the highly undersampled radial FLASH data acquisition technique, the image reconstruction involved a modified NLINV-based algorithm. After optimization of acquisition and reconstruction parameters, the single-shot T1 mapping method could be further developed to a simultaneous multi-slice technique which yields T1 maps of multiple slices within a single experiment, i.e., within a few seconds as needed for a single inversion-recovery experiment.
Chapter 6 presents a clinically relevant extension of the above method to single-shot diastolic myocardial T1 mapping where systolic images are automatically masked out prior to pixel-wise fitting. Robustness and reproducibility of this and the previous method have been validated with use of a numerical phantom, and experimental phantom and human subjects in vivo.
Chapter 7 introduces and evaluates the proposed model-based reconstruction technique which estimates the parametric maps of a suitable signal model and all coil sensitivities directly from the undersampled raw data. A golden-angle data acquisition scheme together with random RF spoiling was applied to efficiently sample the data. The joint estimation of unknowns is formulated as a nonlinear inverse problem where a priori information (i.e., sparsity constraints) of the parameter maps is incorporated into the reconstruction process to improve T1 precision. Validations again included numerical simulations, an experimental phantom and in vivo studies of healthy human subjects. A comparison of the model-based reconstruction to the method introduced in Chapter 5 is also presented. | de |