a) b) c)
This study involves the comparison
of two image acquisition sequences, the spin-echo (SE) sequence and the
fast-spin-echo (FSE) sequence, used in magnetic resonance imaging (MRI).
The purpose of this study is to compare T1, T2, and
r
images calculated from the FSE and SE sequences. Fletcher demonstrated
a technique for tissue classification using images from the SE sequence
(1).
However, the data acquisition was a lengthy process. It is desirable
to use a fast imaging sequence in order to reduce patient discomfort.
Similar results for tissue classification may be obtained faster by using
a FSE sequence. The conventional SE sequence acquires one line of
k-space (spatial frequency domain) at a time while the FSE sequence can
acquire N lines of k-space at a time (2).
FSE scan times are therefore much faster, but the disadvantages include
a lower signal-to-noise ratio and reduced accuracy. The SE sequence requires
an imaging session of about ninety minutes. The FSE sequence provides
similar or slightly degraded results in twenty minutes or less.
The magnetic resonance imager generates
an image of the inside of your body by creating a magnetic field, sending
radio waves through your body, and then measuring the response. The
raw data collected is used to construct images with the help of computers.
Protons, electrons, and neutrons possess spin. The spin of a hydrogen nucleus
is comparable to a magnetic moment vector. A nucleus must have an
unpaired proton or an unpaired neutron, or both to be magnetic. The
hydrogen nucleus aligns itself in the direction of a strong magnetic field.
Then, radio waves stimulate the hydrogen nuclei, shown in Figure 1.
a) b) c)
Figure 1. The magnetization vector (a) aligns itself in the direction
of a strong magnetic field. The magnetization vector rotated 90 degrees
(b). The magnetization vector rotated 180 degrees (c). (3)
The gyromagnetic ratio, g, is 42.58 MHz / T for hydrogen (4). The factors that determine the rotation angle include g, t, the length of time the magnetic field is on, and the magnitude of the RF magnetic field, B1 (4). Equation 1 is used to determine the rotation angle (4).
q = 2p gt B1 Eq. 1
After stimulation by radio waves, the
hydrogen nucleus re-emits the absorbed energy in the form of radio waves.
A short-wave radio antenna and a receiver can detect the re-emitted waves.
The hydrogen spin-lattice relaxation time T1 and the spin-spin
relaxation time T2 define the rate at which the emitted signal
fades after stimulation (3).
T1 is the time the Z component of the magnetization changes
by a factor of e, shown in Figure 2. T1 represents the
re-emission of energy (3).
T2 represents the dephasing of oscillating hydrogen nuclei (3).
Dephasing occurs when a group of spins swing out of phase with each other
and their signals cancel. Their combined signal falls off at the
rate T2. T2 is always less than or equal to
T1 (4).
a)
b)
Figure 2. Spin-lattice relaxation time T1 (a) and the spin-spin relaxation time T2 (b). (4)
In order to separate the effects of
T1, and T2, a sequence of radio pulses is applied
to the tissue. These are called pulse sequences. A common pulse
sequence used in MRI is the spin-echo (SE) sequence. The sequence
produces a series of raw data images of different time of repetition (TR)
and time to echo (TE) parameters. First, a 90-degree pulse is applied,
followed by a 180-degree pulse. Figure 3 shows an echo arriving at
twice the time between the first and second pulses. It is the echo
of the first 90-degree decay signal. The 180-degree pulse acts as
a magnetic barrier that causes the echo.
Figure 3. The spin-echo sequence, showing a 90-degree pulse followed
by a decay signal, then a 180-degree pulse. An echo of the first
90-degree decay signal arrives at twice the time between the first and
second pulses. (3)
It is possible to apply more than one
180-degree pulse following a 90-degree pulse. TR is the time between
two complete sequences, or the time between two 90-degree pulses.
TE is the time between two echoes. Short TR and TE times result in
a strong T1-weighted image, while long TR and TE times produce
a strong T2-weighted image. The signal for a spin-echo
sequence is determined using equation 2 (4).
S = k (1 - e-TR/T1) e-TE/T2 Eq. 2
A conventional spin-echo sequence completes
one line of k-space per TR (2).
K-space is a two-dimensional frequency space. The middle lines of
k-space, which are low spatial-frequencies, have the greatest impact on
image contrast (2). The sequence
is repeated until all lines of k-space are filled. After the k-space
has been filled, a two-dimensional Fourier transform is applied to the
k-space data to produce an image. In Figure 4, as each echo is sampled,
each line of data is placed in a separate k-space (2).
Figure 4. A conventional spin-echo sequence completes only one line of k-space per TR. (2)
A fast spin-echo (FSE) sequence uses
a different phase encoding gradient for each echo generated. More
than one line of k-space is completed per TR (2).
The echo train length (ET) is the number of 180-degree pulses applied after
the initial pulse. Increasing the ET decreases the imaging time and affects
the contrast of the image (2).
Figure 5 shows a FSE sequence with an ET of four. The lines closer
to the middle of the k-space have a higher signal. This is shown
for echo number four, which has a phase encoding gradient near the middle
of the k-space.
Figure 5. FSE sequence with ET of 4. More than one line of k-space is completed per TR. (2)
A SE sequence requires an imaging session
of about ninety minutes. These long imaging times may cause patient
discomfort. A FSE sequence can take less than twenty minutes.
However, the images from a FSE sequence may result in decreased contrast
and blurring in T1 weighted images (2).
The FSE sequence could also introduce some artifacts to the image.
For SE and FSE sequences, a series
of images with a constant TR and varying TE and another series with a constant
TE and varying TR is generated. T2 images are generated
using the constant TR images. The T1 and r
images are generated using the constant TE images. The images are calculated
by fitting a curve to the data. Tissue classification is performed
by creating a three-dimensional histogram (3)
of the T1, T2, and r images.
Similar tissues are grouped in clusters in the three-dimensional histograms
(1),
which allows different tissue types to be classified. Multispectral
tissue classification (MTC) is the segmentation of tissue types in the
human body using medical images. MTC results from the processing
of hydrogen spin-lattice relaxation time T1, spin-spin relaxation
time T2, and spin density
r images
(1).
The experiment was carried out by preparing
a phantom of five test targets, each with different aim T1 and
T2 values. The phantom was made in the MRI lab at the
Chester F. Carlson Center for Imaging Science. The layout of the
test targets in the phantom is shown in Figure 6. The phantom is
a PVC cylinder five inches in diameter, with five 60 ml polyethylene bottles
as the test targets and two smaller polyethylene bottles filled with distilled
water inside. The T1 and T2 values for the
test targets were in the range of 0.25 s to 2 s for T1 and 60
ms to 200 ms for T2. These values are similar to T1
and T2 values for brain tissue. The test targets were
made by filling the five polyethylene bottles with the following solutions
(4):
0.0058 Ni (mole/L), 0.0007 Mn (mole/L), 0.0001 Mn (mole/L), 0.00022 Mn
(mole/L), and 0.000073 Mn (mole/L). The phantom was placed in a 1.5
T General Electric Signa imager at the Magnetic Resonance Imaging Center
at the University of Rochester. The precaution of removing metal
or magnetic materials was used before entering the imaging room.
Figure 6. Test target layout of the phantom. Numbers one through
five represent the five test targets. Numbers six through eight represent
distilled water. The purpose of vials six and seven is to keep the
rest of the targets from moving during the imaging session.
The phantom was imaged using a SE sequence
and three different FSE sequences with ET= 8, 16, and 32. The acquisition
times were 90, 30, 15, and 5 minutes, respectively. The parameters for
each sequence were: TR = 4000, 3000, 2000, 1500, 1000, 750, 500, 250 ms
at a constant TE = 15 ms, and then TE = 25, 50, 75, 100, 150, 200 at a
constant TR of 1000 ms. The acquisitions for the constant TR data were
all single echo images. The transmit gain (TG) and receiver gains
(R1 and R2) were fixed at the values determined for the TR = 4000 ms, TE
= 15 ms SE image. TG was set a 7.90 db, R1 = 6, and R2 = 14.
Increasing the ET decreases the time required for imaging. Due to
limitations of the imager, the minimum TR of the 16 ET FSE sequence was
267 ms instead of 250 ms, and the minimum TR of the 32 ET FSE sequence
was 517 ms. The parameters are presented in Table 1. Only thirteen
images were collected for the 32 ET FSE sequence. The total number
of images for all four sequences is fifty-five. The images were stored
as MR files and then converted to binary images.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| SE |
|
|
|
|
|
|
|
|
| N= 8 FSE |
|
|
|
|
|
|
|
|
| N=16 FSE |
|
|
|
|
|
|
|
|
| N=32 FSE |
|
|
|
|
|
|
|
Calculating T1, T2, and r images
An IDL program originally written in
Fortran by Gong, Li, and Hornak (5,
6)
was used to calculate T1, T2, and r
images for each imaging sequence used. A block diagram of the experiment
from data collection using spin-echo and FSE to tissue classification and
segmentation is shown in Figure 7.
The spin-echo images have the level
of quality we wish to achieve in the FSE images. Two-dimensional (2-D)
T2 vs. T1 histograms for each sequence were created
and compared (1). The histograms
are used to determine if clusters representing different tissue types,
or different test targets for this experiment, begin to overlap.
It is expected that the probability for overlapping clusters will increase
in the higher ET FSE 2-D histograms. If the clusters overlap, it
will be more difficult to classify different tissue types or the phantoms.
T1, T2, r
images and test target clusters in T2 vs. T1 histograms
are shown in Figure 8.
As N increased (and scan time decreased), the clusters in the 2-D histograms
that represented the test targets broadened and became less clearly defined
and their T1 values increased. Regions filled with distilled
water have T1 and T2 values that are higher than
the values for the test targets and their clusters are outside the region
of interest in the histograms. The apparent increase in noise of
the distilled water region is due to the limit of the calculation range
of the T1, T2, and r program.
Similar values calculated for r
confirms that the T1 T2 r
program is working correctly. This is because the test targets are
water-based and have spin-density r values close
to the background of distilled water (regions 6,7,and 8 in Figure 6).
The clusters in the 2-D T2 vs. T1 histogram for the
N = 32 FSE sequence in Figure 8
show that image segmentation is possible with decreased acquisition time.
Each test target can be separated by defining a boundary around each cluster.
The general trend for increasing N is for T1 values to increase,
and for clusters in the 2-D histograms to become broader. The increase
in T2 values for the test targets is minor as N increases.
However, the N = 8 FSE sequence does not fit the trend. For test
targets three and four (referring to Figure 6) of the N = 8 FSE sequence,
the T1 values actually decrease from the SE sequence.
Also, the T2 values for the N = 8 FSE sequence are higher than
corresponding T2 values in the other sequences. It is
not clear why this anomaly occurs for the N = 8 FSE sequence. Further
research is required to determine the cause. A problem encountered
with the IDL program for computing T1, T2, r
images involved the setting of the minimum and maximum T2 values
for the program to calculate. If the maximum set T2 value
is too low, the program sets the overflow pixels to zero. This was
most noticeable in the FSE sequence images, which had higher T2
values in the distilled water regions (regions 6, 7, and 8 in figure 6).
If the minimum T2 value is too low, the program would take much
longer to find the zero crossing point of the linear least-squares algorithm
or it would never find it. The IDL program calculates the T1,
T2, r images for each sequence in
about forty minutes on a SUN UNIX platform. Since processing time
per sequence is about forty minutes, there is still some room for optimization
of the code to run more efficiently.
The trend for increasing N of the FSE
sequences (increase in T1 values and similar T2 values
to the SE sequence) does not apply for the N = 8 FSE sequence. Therefore,
it would be difficult, if not impossible, to assign a correction factor
to the T1, T2, r images
from the FSE sequences to approximate T1, T2, r
images from the SE sequence. Future research should focus on determining
the cause of the anomaly in the trend, and on optimizing the IDL program
to compute T1, T2, r images
faster. In addition, the experiment should be repeated, this time
imaging the brain region of a human subject to determine if tissue classification
is possible with FSE sequences. On a positive note, the 2-D T2
vs. T1 histogram for the N = 32 FSE sequence demonstrates that
image segmentation of test target clusters is possible at an acquisition
time of only five minutes compared to the SE sequence acquisition time
of ninety minutes.