Multidimensional Image Analysis of Cardiac Function in MRI
by Michael Lynch
Department of Electronic Engineering,
Dublin City University,
Supervisor: Prof. Paul F. Whelan
Examining Board Chair: Prof. Barry McMullin
Examining Board Members: Prof. Rasmus Larsen (DTU, Technical University
of Denmark), Prof. Alistair Sutherland (DCU).
Cardiac morphology is a key indicator of cardiac health. Important metrics that are currently in clinical use are left-ventricle cardiac ejection fraction, cardiac muscle (myocardium) mass, myocardium thickness and myocardium thickening over the cardiac cycle. Advances in imaging technologies have led to an increase in temporal and spatial resolution. Such an increase in data presents a laborious task for medical practitioners to analyse.
In this thesis, measurement of the cardiac left-ventricle function is achieved by developing novel methods for the automatic segmentation of the left-ventricle blood-pool and the left ventricle myocardium boundaries. A preliminary challenge faced in this task is the removal of noise from Magnetic Resonance Imaging (MRI) data, which is addressed by using advanced data filtering procedures. Two mechanisms for left-ventricle segmentation are employed.
Firstly segmentation of the left ventricle blood-pool for the measurement of
ejection fraction is undertaken in the signal intensity domain. Utilising the
high discrimination between blood and tissue, a novel methodology based on a
statistical partitioning method offers success in localising and segmenting the blood pool of the left ventricle. From this initialisation, the estimation of the outer wall (epi-cardium) of the left ventricle can be achieved using gradient information and prior knowledge.
Secondly, a more involved method for extracting the myocardium of the left-ventricle is developed, that can better perform segmentation in higher dimensions. Spatial information is incorporated in the segmentation by employing a gradient-based boundary evolution. A level-set scheme is implemented and a novel formulation for the extraction of the cardiac muscle is introduced. Two surfaces, representing the inner and the outer boundaries of the left-ventricle, are simultaneously evolved using a coupling function and supervised with a probabilistic model of expertly assisted manual segmentations.
Finally, to fully utilise all data presented from a single 4D cardiac (3D + t) MRI scan a novel level-set segmentation process is developed that delineates and tracks the boundaries of left ventricle. By encoding prior knowledge about cardiac temporal evolution in a parametric framework, an expectation-maximisation algorithm tracks the myocardium deformation and iteratively updates the level-set segmentation evolution in a non-rigid sense.
Both methods for the extraction of cardiac functions have been tested on patient data and provide positive qualitative and quantitative experimental results when compared against expertly assisted segmentations.
© Michael Lynch, 2006