@TechReport{Dartmouth:TR96-290,
  author = {Y. Joy Ko and Michael B. Taylor},
  title = {{MRI On the Fly:  Accelerating MRI Imaging Using LDA Classification with LDB Feature Extraction}},
  institution = {Dartmouth College, Computer Science},
  address = {Hanover, NH},
  number = {PCS-TR96-290},
  year = {1996},
  month = {June},
  URL = {http://www.cs.dartmouth.edu/reports/TR96-290.ps.Z},
  comment = {
    A Senior Undergraduate Honors Thesis in Computer Science.<p>
    Postscript is 10 MB uncompressed!<br>
    You can also retrieve this file in two pieces:
    <a href="http://www.cs.dartmouth.edu/reports/TR96-290a.ps.Z">A</A>,
    <a href="http://www.cs.dartmouth.edu/reports/TR96-290b.ps.Z">B</A>
  },
  abstract = {
           To improve MRI acquisition time, we explored the uses of linear
    discriminant analysis (LDA), and local discriminant bases (LDB) for
    the task of classifying MRI images using a minimal set of signal
    acquisitions.
           Our algorithm has both off-line and on-line components.  The
    off-line component uses the k-basis algorithm to partition a set of
    training images (all from a particular region of a patient) into
    classes.  For each class, we find a basis by applying the best basis
    algorithm on the images in that class.  We keep these bases to be used
    by the on-line process.  We then apply LDB to the training set with
    the class assignments, determining the best discriminant basis for the
    set.  We rank the basis coordinates according to discriminating power,
    and retain the top M coordinates for the on-line algorithm.  We keep
    the top M coordinates, which index the basis functions with the most
    discriminating capability, for on-line purposes.  Finally, we train
    LDA on these transformed coordinates, producing a classifier for the
    images.
           With the off-line requirements complete, we can take advantage
    of the simplicity and speed of the on-line mechanism to acquire an
    image in a similar region of the patient.  We need acquire only the M
    important coordinates of the image in the discriminant basis to create
    a ``scout image.''  This image, which can be acquired quickly since M
    is much much smaller than the number of measurements needed to fill in
    the values of the 256 by 256 pixels, is then sent through the map
    furnished by LDA which in turn assigns a class to the image.
    Returning to the list of bases that we kept from the k-bases
    algorithm, we find the optimal basis for the particular class at hand.
    We then acquire the image using that optimal basis, omitting the
    coefficients with the least truncation error.  The complete image can
    then be quickly reconstructed using the inverse wavelet packet
    transform.
           The power of our algorithm is that the on-line task is fast and
    simple, while the computational complexity lies mostly in the off-line
    task that needs to be done only once for images in a certain region.
    In addition, our algorithm only makes use of the flexibility of MRI
    hardware, so no modifications in hardware design are needed.
  }
}