BME 483 MRI Notes 3: page 1

Notes on MRI, Part III

1D Imaging – Frequency Encoding

Gradients act to setup a one-to-one correspondence between frequency and spatial position. This
is known as frequency encoding. For example:
∆ω(x) = γGxx
or
γ
∆f ( x ) = Gx x
2π
Graphically this is:
Low Frequency
B
Mag. Field Object
Strength
Low Frequency
High Frequency

High Frequency
1D Image
x Position x Position

As previously described, to create an image, we need to take the 1D (forward or inverse) FT of

the received signal:
Low Frequency

MR Signal Object
Fourier
Transform
High Frequency
time
1D Image
x Position

The pulse sequence to produce this might look like this:

RF

Gx

Data
Acq.
BME 483 MRI Notes 3: page 2

Because we have a constant gradient on while we are acquiring data, the frequency spatial
location relationship holds. The negative gradient at the beginning is due to the necessity to
“reverse time.” Previously, when we defined the relationship:

s (t ) = ∫ m( x) exp( −i 2πsx )dx

= F {m( x)} s =γG t / 2π

x

= M ( s = γG x t / 2π )

We ignored the fact that we where only acquiring the FT, M(s), for positive time in γG x t = 2πs

(time begins at the point of tipping the magnetization into the transverse plane). If we have a
negative gradient for a period of time (leading to negative accumulation of phase), then we have
effectively turned back the clock to a negative starting time for phase accumulation.

Alternate Methods for 1D Localization – Phase Encoding

While the constant gradient is an excellent method for 1D localization, it is not the only method.
It is important to note that it is the instantaneous phase of the magnetization at each sample that
is important and not the instantaneous frequency – that, where the magnetization is pointing the
complex plane, e.g. mxy,rot(x,y,t), is what is really important and not how fat it is moving. Recall,
we are sampling the received signal:
s (t ) = ∫∫ m xy,rot ( x,y,t )dxdy

Thus, we can look at other methods that will produce the same distribution of magnetization in
the rotating frame. Consider the following pulse sequence:

RF

Gx

Data
Acq.
In this pulse sequence the same amount of phase accumulation occurs between samples as in the
previous pulse sequence. When the gradient amplitude is zero, there is zero frequency offset
(∆B = ∆ω = 0) and thus not phase accumulation during these periods (there is, of course T2
BME 483 MRI Notes 3: page 3

decay, but we’re ignoring that for now). If you looked at a particular spin, it would have a stop-
action kind of appearance:

STOP

STOP

We could imaging putting a whole bunch of these frames together and it might have the
appearance of a continuous movement at a frequency ∆ω, but in this case, it is really occurring in
little jumps. If we take the FT of these sample, we will still get our 1D image.

Consider yet an alterative pulse sequence:

RF

Gx

Data
Acq.
Here, there is a differing amount of phase accumulation after each RF pulse. In this case, it
comes from having different G values for the same amount of time, but the impact is the same:

STOP

STOP

Sample 0 Sample 1 Sample 2

BME 483 MRI Notes 3: page 4

Again, if we take the FT of these sample, we will still get our 1D image. In both cases, the
samples are collected along a dimension that looks like time, but isn’t quite. Some in MRI call
this dimension pseudo-time. In these examples, the spatial location is not encoded in the
frequency, but rather is encoded into a specific sequence of phase accumulation between
samples. This is known as phase encoding.

2D Imaging
One way to do 2D imaging is to combine frequency encoding in one spatial dimension (e.g. x)
with phase encoding in another spatial dimension (e.g. y). Take the following pulse sequence:

RF

Gy

Gx

Data
Acq.
For a give phase encoding value, if we 1D FT the group of samples we’ll get a 1D image in x. If,
on the other hand, we take the nth sample from each group of samples (that is, each of these
differs only by the phase encoding amount), and take the 1D FT of this data, we’ll get a 1D
image of the object in y.

Thus, it we put the data in an array (each group along a row) and take the 2D FT of this data, we
will get a 2D image of the object! The above pulse sequence is known as the spin-warp pulse
sequence is it is the most used pulse sequence (imaging method) in MRI.

k-Space
We will now develop the general theory for 2D imaging. The received signal is:
s (t ) = ∫∫ m xy,rot ( x,y,t )dxdy
BME 483 MRI Notes 3: page 5

Again, please bear in mind that we are mixing coordinate systems here. For example, in
mxy,rot(x,y,t), the x,y in the argument refers to physical (x,y) locations in space, whereas the xy in
the subscript refers to a mini-coordinate frame to describe direction of the magnetization vector
at each point in space.

Let’s consider a spatially and temporally varying applied magnetic fields introduced by time-
varying gradient fields:
B( x, y, t ) = B0 + G x (t ) ⋅ x + G y (t ) ⋅ y

The instantaneous frequency at each point is space is then:

γB( x, y , t ) = γ (B0 + G x (t ) ⋅ x + G y (t ) ⋅ y )

which in the rotating frame is:

∆ω ( x, y , t ) = γ (G x (t ) ⋅ x + G y (t ) ⋅ y )

and the spatially variant phase distribution is:

φ ( x, y , t ) = ∫ γ (G x (τ ) ⋅ x + G y (τ ) ⋅ y )dτ
t

where time, t, begins with each RF pulse the bring magnetization from the longitudinal axis into
the transverse plane where it is observable.

The Signal Equation, revisited. For convenience, we will let C = 1 and we will define m(x,y) =
m0(x,y) = m0(r). We now get a revised version of the signal equation:

s (t ) = ∫∫ m xy,rot ( x,y,t )dxdy

= ∫∫ m( x,y ) exp( −iφ ( x, y , t ))dxdy

 t 
= ∫∫ m( x,y ) exp − i ∫ γ (G x (τ ) ⋅ x + G y (τ ) ⋅ y )dτ dxdy
 0 
 t t

= ∫∫ m( x,y ) exp − iγ  ∫ G x (τ )dτ ⋅ x + ∫ G y (τ )dτ ⋅ y  dxdy
 
 0 0 
BME 483 MRI Notes 3: page 6

Finally, we define two quantities:

γ
t

2π ∫0
k x (t ) = Gx (τ )dτ

γ
t

2π ∫0
k y (t ) = G y (τ )dτ

And substituting into the above signal equation:

s (t ) = ∫∫ m( x,y ) exp (− i 2π (xk x (t ) + yk y (t )))dxdy

= F2 D {m( x,y )} | u = k x (t ),v = k y ( t ) = M (k x (t ), k y (t ) )

That is, the signal is equal to the Fourier transform of the initial magnetization evaluated at
locations defined by kx and ky above. Reminder:
∞ ∞

∫ ∫ g ( x , y )e
− i 2π ( xu + vy )
G (u, v ) = F2 D {g ( x, y )} = dxdy
−∞ −∞

The signal equation says that samples of the received signal are equal to samples of the 2D
Fourier transform of the object. This make sense if we think about what exactly the expression
for the 2D FT means – the FT at any point (u,v) is the integral over the object modified by a
spatially variant (linear) rotation in the complex plane.

exp(-i2πy) exp(-i2πx)
BME 483 MRI Notes 3: page 7

exp(-i2π(x+y)) exp(-i0)

In MRI, the integration is performed by the integration of voltages in the RF coil. The phase
variation is performed by the gradients – by shifting the field (and thus frequency) in a spatially
linear fashion for a period of time, the magnetization will rotate to a new orientation (in the
complex plane). Thus MRI has exactly the same mechanisms as the FT operation.

K-space always begins at the origin (0,0). Why? After the excitation pulse, all spins across the
object are pointing in the same direction (e.g. exp(-i0) ) and the integral of this is the DC value of
the FT.

If we want to determine the object, we must fully sample its Fourier transform. A sequence of
samples can be viewed as samples along a pathway determined by “running integral” under the
gradient waveforms as defined by kx and ky above. The final object m(x,y) can be reconstructed
simply by taking the inverse 2D FT of the sampled Fourier data (k-space data):
m( x,y) = F2−D1 {M (k x (t ), k y (t ) )}

Does it make sense that our samples in time are actually samples of spatial frequency data?
Remember the 1D case – we sampled time, FT’ed to get a spectrum. Since there was a 1-1
correspondence between frequency and spatial position, the FT of the frequency data produces
time-domain data thus there is 1-1 correspondence between time and spatial-frequency.

Comments:
BME 483 MRI Notes 3: page 8

- Fourier space is called “k-space” in the MRI literature

- Gx(t) and Gy(t) control the k-space “trajectories” or paths on which sample locations fall.
- To create an image, we must sample M(u,v) densely enough to prevent aliasing and of a large
enough extent to have sufficient spatial resolution.

1D Imaging
We first examine the case of a 1D object m(x) = rect(x/W). The received signal will be the
Fourier transform M(u) = W sinc(Wu) evaluated at particular kx locations as dictated by the
integration of the Gx gradient waveform. Presented here is “pulse sequence” for a 1D imaging
experiment along with the k-space values and the received signal:

Recall the in the first MRI lecture we talked about taking the FT of the received signal to get a
1D view of the object. In this case, the received signal is a sinc function and thus the 1D FT of
the sinc function is a rect function, which is in fact the object. During the entire time that data is
acquired (see the Data Acq. line in the pulse sequence) the gradient is constant – during this
situation there is 1-1 correspondence between frequency and spatial position. This is know as
“frequency encoding” since spatial location is encoded as frequency.
BME 483 MRI Notes 3: page 9

Notice also that we use a negative gradient before the positive gradient. Without the negative
gradient, we can only acquire the positive spatial frequencies – or only ½ of the FT of the object.
This is the difference from the previous example – there we took the FT of s(t), but assumed that
we knew s(t) for all time – really we only know if for positive time.

2D Imaging using Projections

The first 2D imaging method implemented in MRI (by Paul Lauterbur while at SUNY Stony
Brook), used a series of 1D acquisition with the gradients in different directions. Please not that
by applying 1D gradients in x and y simultaneously, we get a single 1D gradient at an angle
θ=tan-1(Gy/Gx). Thus we can get 1D views of the object (or projections) from many different
angles.

We’ll discuss (in the section on computed tomography) the methods for reconstructing images
from 1D projections. For now, suffice it to say that if we acquire enough projections, we can
fully determine the underlying object. The pulse sequence used by Lauterbur is given here:
BME 483 MRI Notes 3: page 10

By sweeping through angles [0,2π) we acquire the full k-space (Fourier) data for the object.
There is also a variant on projection imaging in which the positive gradient is preceded by a
negative gradient – this will allow both positive and negative frequencies to be acquired along a
particular line in k-space. One advantage to this approach is that one only needs to sweep
through angles [0,π) in order to fully acquire the k-space data.

2D Spin-Warp Imaging
The most common acquisition used in MRI today is known as the “spin-warp” acquisition. The
pulse sequence is given here along with the corresponding k-space trajectory:
BME 483 MRI Notes 3: page 11

This is a repeated pulse sequence with a different y-gradient value for each RF excitation (TR
interval). Let’s look at the x-gradient – as in the case of 1D imaging, above, this gradient
encodes the x spatial position into frequency. This is often called the “freq
gradient or x, in this case, is known as the “frequency direction.”

The y-gradient is on briefly before each acquisition but is not on during data acquisition. Thus,
whatever encoding performed by the y-gradient is done. In this case, the y-gradient sets up a
spatially dependent phase distribution that remains fixed during the frequency encoding process.
In other words, the y-gradient encodes spatial position into the phase of the magnetization
(direction of the m vector), which is known as “phase encoding.” y, in this case, is known as the
“phase direction.” Below are depictions of the phase distribution set up for the 1, 0, +1 and +2
phase encoding steps. These correspond to –1, 0, +1 and +2 cycles of phase across the field of
view, respectively.
BME 483 MRI Notes 3: page 12

exp(-iγ(-1)∆GyTy) exp(-iγ(0)∆GyTy)

exp(-iγ(1)∆GyTy) exp(-iγ(2)∆GyTy)

In terms of parameters described in the above pulse sequence, we can define several parameters
of interest in the acquired space. The sample spacing and width of the k-space are:
γ
∆k x = G x ∆t
2π

γ
∆k y = ∆G y T y
2π

γ
Wkx = N x ∆k x = G xTread
2π

γ
Wky = N y ∆k y = 2G y ,max T y
2π
BME 483 MRI Notes 3: page 13

Sampling in k-space (spatial frequency domain). Previously we discussed sampling of the

object and its effect on the spectrum. Here we have the reverse – sampling in Fourier domain
and its effect on the reconstructed object. Again, we will perform our sampling by multiplying a
function times the 2D comb function. With sample spacing of ∆kx and ∆ky, in the kx and ky
directions, the sampled Fourier data is:

~  u v 
M (u, v ) = M (u , v)comb , 
 ∆k x ∆k y 
 

∞
= ∆k x ∆k y ∑ δ (u − n∆k
n , m = −∞
x , v − m∆k y )M (n∆k x , m∆k y )

The image (space) domain equivalent is:

~ ( x, y ) = m( x, y ) * *∆k ∆k comb(∆k u, ∆k v )
m x y x y

∞  n m 
= m(u , v) * * ∑ δ u −
n ,m = −∞ 
 ∆k x
,v −
∆k y




∞  
= ∑ mu − n , v − m 
 ∆k x ∆k y 
n , m = −∞  

Thus, sampling in the Fourier domain leads to replication in the image domain. Spacing of the
replicated image (object) is (1/∆kx,1/∆ky). The replicated images will not overlap the original
image if the highest spatial position in x is xmax ≤ 1
2 ∆kx and the highest spatial position in y is

y max ≤ 1
2 ∆ky . If this is not satisfied, then there will be spatial overlap in the images (or aliasing).

The field of view of an acquisition is typically defined as one over the k-space sample spacing:
FOVx = 1/∆kx and FOVy = 1/∆ky
and aliasing will not occur if xmax < ½ FOVx and ymax < ½ FOVy.
BME 483 MRI Notes 3: page 14

Point Spread Function.

Observe that the practical k-space is not of infinite extent, but rather is limited to Wkx and Wky.
The sampled k-space can be written as:

~  u v   u u 
M (u, v ) = M (u , v)rect , comb , 
W W   ∆k ∆k 
 kx ky   x y 
which results in an image of the following form:
~ ( x, y ) = m( x, y ) * *W W sinc (W x)sinc (W y ) * *∆k ∆k comb(∆k u, ∆k v )
m kx ky kx ky x y x y

The sinc functions are the point spread function and the comb function generates replicated
versions of the object (does the aliasing). Observe that the sinc functions have approximate
widths (in x and ) of ∆x = 1/Wkx and ∆y = 1/Wky. This defines, in essence, the spatial resolution
of an MRI acquisition – in order to get better (finer) spatial resolution, we need to acquire a
larger area in k-space.

Resolution of the FFT. Most forms of the FFT work this way – for an N point input function
the FFT will produce an N point output. Each output point corresponds to an integer number, n,
of cycles in exp(-i2πnx) across the object and go from n = [-N/2:N/2-1]. Observe the DFT of -
N/2 and N/2 are the same and thus this represents the entire, unaliased frequency domain of the
object. Thus F/2 (see below) = ½ of 1/∆t or F = 1/∆t. A similar argument can be made in
reverse to get T = 1/∆f.
BME 483 MRI Notes 3: page 15

Resolution and Object and Sample Spacing in MRI. For data acquired on a 2D rectilinear
grid in k-space and reconstructed with a 2D FFT the spatial resolution and Field of View
relationships are:
FOVx = 1/∆kx and FOVy = 1/∆ky
∆x = 1/Wkx and ∆y = 1/Wky