@MASTERSTHESIS\{IMM2004-03154,
    author       = "H. \'{O}lafsd\'{o}ttir",
    title        = "Registration and Analysis of myocardial perfusion {MRI}",
    year         = "2004",
    month        = "feb",
    keywords     = "Motion-compensation, registration, myocardial perfusion {MRI,} active appearance models, minimum description length, cubic splines, contour extraction tool, perfusion assessment, perfusion map",
    pages        = "170",
    school       = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "",
    type         = "",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/3154-full.html",
    abstract     = "This thesis presents the registration and analysis of myocardial
perfusion magnetic resonance images.  Registration is carried out
by a method based on Active Appearance Models (AAM). Extensions to
the basic {AAM} include clustering of texture vectors in temporal
dimension in addition to slice-coupled modelling.

A training set for the {AAM} is generated semi-automatically.  This
involves implementation of a tool to allow manual marking of
points on object outlines and interpolation of the points by cubic
splines to generate a full shape contour. Given the shape
contours, point correspondences are obtained by Minimum
Description Length (MDL) shape modelling. Requirements for the {MDL} 
approach imply that the cardiac shape cannot be optimised
directly.  Consequently, it is splitted into five contours at
anatomical and pseudo landmarks and each contour set is optimised
with influence from the full shape.

After registration of the set of images, perfusion assessment is
carried out by providing pixel-wise signal intensity curves from
the myocardium.  Subsequently, perfusion maps are generated for
three parameters derived from the curves: Maximum upslope, peak
and time-to-peak.

Qualitative and quantitative validation is carried out on the
presented methods.  Data from 10 patients with acute myocardial
infarction is provided for this purpose. {MDL} optimisations
resulted in 10 {--}14\% improvement in terms of the description
length and 10 - 28\% improvement in terms of the total model
variance.  A leave-one-out cross validation was carried out for
the registration method.  This resulted in mean point to curve
distance of 1.25 ± 0.36 pixels. Perfusion maps for maximum
upslope and time-to-peak positioned ischemic segments in most of
the patients."
}