Data Processing Methods in MRS - Advance Image Processing in Magnetic Resonance Imaging - page 409

Depending on the characteristics of the undesired peaks, different signal
processing methods can be applied to suppress them. These techniques range
from time-domain filtering (in which unwanted peaks are separated in frequency
from the peak of interest), discharging or weighting the FID data points in
different ways (when nuisance peaks have larger line widths than the peak of
interest) to a more or less complicated modeling of the spectrum baseline.
Time-domain filtering has been proposed to suppress the water peak, which
is usually located at zero frequency. High-pass digital filters can be used in this
regard [24,25]. Marion [25] proposed a linear-phase low-pass filter to extract
purely

water signals and subtract the filter output from the data. The proposed
filter coefficients h were sine-bell-shaped

m

( mM

M

−− ( 12 )/ )

h = cos π (12.7)

m

or Gaussian-shaped

he
m

= −−−

( 4 mM ( 1 )) /

22

M

(12.8)

with M being the filter order. Linear-phase filters are proposed to avoid phase
distortion in the unfiltered components, whereas high-order M (17 65)

values were suggested [25] to reduce the influence on the peak of interest. Because
the first and last (M1)/2 values cannot be calculated, they must be extrapolated
using linear predictions [25] or modeling [26]. Extrapolating the data may intro-
duce distortions, and a careful compromise between ideal filter response and the
number of extrapolated points has to be taken into account for filter design.
Another approach to water suppression is based on modeling of time-domain
water signals [27,28]. Because of partial water suppression performed using
special sequences, the water peak is far from the theoretical lineshape. Therefore,
several damped exponentials are usually used to fit it. In modeling the water peak,
user intervention is usually necessary to define the frequency region of the water
peak and the number of fitting exponentials.

< M <

12.4 FREQUENCY-DOMAIN METHODS

Accurate quantification of in vivo spectra parameters can be obtained by fitting
the observed MRS spectrum with known lineshapes. In the ideal case, the MRS
spectrum consists of a superimposition of pure complex Lorentzian lines [29,30].
In the real case, due to acquisition imperfections, the ideal lineshape is distorted,
and the model should usually include a mixture of Lorentzian ( L ) and Gaussian
( G ) curves [21,31]. The following is a general model of an MRS spectrum:

K P

∑ β ∑

Xf () = ( L ( , ) v f + β G ( , ) v f + cf

p

+ w () (12.9)

Lk k G k k p

k = 1 p = 1