Use this URL to cite or link to this record in EThOS:
Title: Reconstruction of cardiac magnetic resonance images using short and long axis slices
Author: Basty, Nicolas
ISNI:       0000 0004 7966 3236
Awarding Body: University of Oxford
Current Institution: University of Oxford
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Cardiac Magnetic Resonance Imaging (MRI) provides high resolution images of the heart, facilitating its functional and anatomical analysis, and is often regarded as the gold standard for diagnosis. Slow imaging times as well as patient and heart motion constraints are making it impractical to image volumes at isotropic high resolution in cardiac MRI. As a result, multi-slice studies, composed of highly anisotropic slices acquired at different orientations, are usually acquired. Most clinical imaging protocols consist in a stack of short axis slices and a few (typically two or three) long axis slices orthogonal to the short axis stack. Post processing methods may be used to reconstruct an isotropic volume. Most of the methods presented in the literature, however, use imaging protocols which are not commonly followed in clinical practice. In this thesis, we developed three reconstruction methods for cardiac MRI reconstructions using long and short axis images as acquired in clinical practice. First, we showcase our work on a method which does not rely on machine learning or the use of a large database. Our method handles short and long axis data in a contrast independent manner, and is based on regularised least squares and the novel three dimensional directional total variation. We validated the algorithm using pre-clinical data, and showed improvements in reconstructions of in vivo data. We also explored hypothetical protocols for reconstructions, while keeping the number of slices constant. The best results came from balancing short and long axis slices. We subsequently shifted the research focus to deep learning, allowing a comparison of different methodologies. We used a large publicly available dataset of clinical acquisitions and separated methods into static and dynamic approaches. For the static approach, we present several architectures and show that our networks are able to recover high resolution long axis slices from synthetic data, as well as interpolated short axis stacks. We show that using a conditional Generative Adversarial Network gives the best qualitative and quantitative results, while other models show improved results that are overly smooth. For the dynamic data, we use Long Short-Term Memory. Including temporal context from adjacent frames present in cine data results in enhancement of results compared to the static method. For both methods, we see qualitative as well as quantitative improvements in peak signal-to-noise ratio and structural similarity over interpolations. Our methods could impact cardiac MRI acquisition and analysis in two ways. In the short term, they could be used to reconstruct missing or additional slices in current acquisitions. In the medium term, they could be used in combination with a fast, optimised acquisition protocol from which slices in any orientation could be obtained at post-processing.
Supervisor: Grau, Vicente Sponsor: RCUK Digital Economy Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available