Parallel MRI: Concepts and Methods - Advance Image Processing in Magnetic Resonance Imaging - page 58

Today SMASH remains an elegant theoretical development, although without real
widespread applicability.
The SENSE method proposed by Pruessmann et al. (8) was introduced in 1999
and is another parallel imaging technique that relies on the use of 2-D or 3-D
sensitivity profile information in order to reduce image acquisition times in MRI.
Similar to SMASH, the Cartesian version of SENSE requires the acquisition of
equally spaced k-space lines in order to reconstruct sensitivity-weighted, aliased
versions of the image. The aliasing is removed with the use of the sensitivity
profile information at each pixel. This is done by solving in the space domain the
linear system of equations defined by the subsampling pattern. This technique is
very similar to the approach described by Ra and Rim in 1993 (6), and only differs
in the sensitivity profile estimation, as well as in its use of numerical system
regularization strategies.
The general version of SENSE allows for data to be sampled along arbitrary
k-space trajectories, and has a number of advantages over the Cartesian version,
including the ability to minimize artifact, and maximize the SNR, effectively
achieving higher acceleration factors with acceptable quality. A very high com-
putational cost, however, accompanies the arbitrary k-space sampling in gener-
alized SENSE.
In 2000, the SPACE RIP (9) technique was introduced whereby flexibility is
allowed in the choice of the k-space lines acquired, given that the frequency-
encoded direction is kept unchanged. This allows one to maintain the advantages
of SNR and minimized artifacts of generalized SENSE, while having a consid-
erably smaller computational load. Similar to SENSE, SPACE RIP uses 2-D or
3-D sensitivity profile information in order to solve a linear system of equations.
In addition, it allows for total flexibility in the positioning of the coil array around
the object of interest.
Another technique, partially parallel imaging with localized sensitivities
(termed PILS), was described in 2000 (10), which requires the use of coils having
localized sensitivities and circumvents the need to estimate the sensitivity profiles.
PILS is simple in principle, but the condition that it imposes on the coil sensi-
tivities is impractical for a large number of applications.
Since 2001, the effort shifted toward optimizing the existing techniques
and a number of works have surfaced that describe better ways to estimate
the coil-sensitivity profiles, as well as to condition the reconstruction in the
various techniques. Generalized autocalibrating partially parallel acquisitions
(GRAPPA) (11), published in 2002, introduced a generalized autocalibrating
approach that can result in more robust SMASH-like reconstructions with
computational complexity far below more general SMASH approaches (12).
Currently, the trend has shifted toward finding a generalized formalism that
encompasses all the preceding techniques, as they all strive to solve the same
system of equations. In addition, more effort is spent on the design of coil
arrays that are optimized for use in parallel imaging. In the following section,
we describe the general problem of parallel MRI, and present the basics of
each of the techniques.