Cardiovascular disease is the number one cause of death in the
western world. Cardiac imaging is an established approach to diagnosing
cardiovascular disease and plays an important role in its interventional
treatment. Three-dimensional imaging of the heart and the cardiovascular
system is now possible with X-ray computed tomography, magnetic
resonance, positron emission tomography, single photon emission
tomography, and ultrasound to name just the main imaging modalities.
While cardiac imaging capabilities are developing rapidly, the
images are mostly analyzed visually and therefore qualitatively.
Clinical ability to quantitatively analyze the acquired image
data is still not sufficiently available in routine clinical care.
Large amounts of acquired data are not fully utilized because
of the tedious and time-consuming character of manual analyses.
This is even more so when three-dimensional image data need to
be processed and analyzed. Image segmentation is a pre-requisite
to quantitative analysis and thus developing methods for highly
automated three-dimensional cardiac image segmentation is of primary
importance.
Enhancements in cardiac gating and imaging gradient speeds continue
to improve MR image quality and allow for visualization of finer
detail of cardiac and extracardiac structures. This has led to
an expanded role for cardiac MR imaging in the diagnosis and management
of children and young adults with congenital heart disease. Highly
automated image segmentation methods, such as the Active Appearance
Model (AAM), will allow additional quantitative data to be incorporated
into management decisions. In this proposal, the AAM method of
image segmentation will be applied to two main areas that continue
to pose diagnostic challenges for clinicians - right ventricular
function in postoperative tetralogy of Fallot patients and assessment
of the aorta in patients with suspected connective tissue disease.
These two groups should demonstrate the strength of the AAM algorithm
in automatically providing clinically useful information from
cardiovascular MR images.
While the majority of MR image analysis of patients with congenital
heart disease is qualitative, an increasing amount of quantitative
data is being extracted from the images. Tracing the endocardial
borders has provided accurate measurement of right and left ventricular
ejection fraction and stroke volumes in both normal subjects and
patients with ventricular dysfunction. Additional tracing of the
epicardial borders has also generated reproducible ventricular
mass measurements. Recently, phase velocity flow mapping across
the mitral valve has been used to assess left ventricular diastolic
function, although, to our knowledge, indices of diastolic and
systolic time dependent ventricular volume changes have not been
measured.
Similar to the largely qualitative approach used for intracardiac
MR image analysis, MR images of the aorta are often assessed qualitatively,
such as patients undergoing study to rule out or localize an aortic
aneurysm. Quantitative measurements derived from MR images of
the aorta that replicate the echocardiographic dimensions suggested
by Roman et al. have been routinely obtained at our institution
but data supporting this approach in the literature are limited.
Reproducible measurements of the distal thoracic aorta and standards
by which to judge these dimensions are currently not available.
While these data are available using standard MR imaging techniques,
the laborious methods needed in multiple patients to obtain these
values have precluded their availability in the literature.
To date, automated algorithms to generate quantitative data from
MR images have seen limited clinical use either because they are
cumbersome to implement and use or they lack robust reproducibility.
Most radiologists have a sufficient caseload so that hand tracing
the large number of images generated from a 4-D acquisition cannot
be expected. Additionally, longitudinal follow-up of patients
over time suffers from poor reproducibility of the manual measurements.
Image segmentation performed in a highly automated fashion based
on an individual, patient-specific model should address the problems
and improve patient care.
