MRI Fundamentals

Week 4:
MR Image Formation

Sung-Hong Park
Bio and Brain Engineering, KAIST


 In this week, we will discuss how MR images are
formed.

• How to sequence gradient pulses to get an image

• Slice selection
• K-space
• Frequency encoding
• Phase encoding


 How to sequence gradient pulses to
get an image


 Free Induction Decay (FID)
Spin Excess Excitation 90°RF Relaxation

z RF z RF z z
Antenna
' ' ' (RF Coils)
'

FID
y y y y
' ' ' '
x x x x
' ' ' '

The RF pulse causes magnetic resonance. Signal

Amplitude
When the RF pulse stops, the net magnetization (𝑴)
starts to recover back to the initial state (relaxation). Envelope

MR signal called free induction decay (FID) is

Time
detectable during the relaxation. FID


 Sequencing of RF pulse, data sampling (ADC), and gradients
MR Image
RF

FID

Sampling
ADC
(analog-to-digital converter)
RF pulse (TX) and data sampling
𝐺𝑧 (ADC) (RX) are needed to get an MR
signal.
The FID signal cannot be directly
𝐺𝑥
? converted to an MR image (there is
𝐺𝑦 no time for spatial encoding).
How to sequence gradient pulses
relative to RF pulse (TX) and ADC (RX)
to form an MR image?


 Gradient Pulses and Precession Frequencies
Gradient pulses spatially modulate the
𝐺𝑧 ? magnetic field strength that is
proportional to the precession frequency.
𝐺𝑥 ? To get an tomographic MR image,
gradient pulses along X, Y, and Z should
𝐺𝑦 ? be combined with RF pulse (TX) and ADC
(RX) in a special way.
Slice selection K-space MR Image
(phase Y
encodin
g)

𝑍
X (Freq. encoding)


 Slice Selection


 Slice Selection
Slice selection determines the position of a slice to be excited by an RF pulse.
Slice selection gradient linearly changes the precession frequency of the magnetic
momentums along the Z-axis.
Remember, Larmor frequency (𝜔) depends on local magnetic field (𝜔 = 𝛾 𝐵).
𝜔 𝑧, slice thickness
Freq.

𝜔2 RF Time

𝜔0 𝑧
Space 𝐺𝑧
𝜔1 𝜔
Transmission RF excitation
frequency bandwidth 𝐺𝑧
Without the slice selection gradient, the RF pulse will excite all the body region (𝜔0 ).
In case of axial imaging, gradient along Z-axis (𝐺𝑧 ) is used as slice selection.
Since the ideal excitation profile in frequency domain is a rect function, RF pulse
shape in time domain is typically designed in a sinc-like pulse shape.


 Slice selection (cont.)
Strong gradient Weak gradient An RF pulse induces slice-selective
excitation in the specific position and
RF RF thickness determined by its transmission
frequency (𝜔1 ) and a bandwidth (𝜔).
𝐺𝑧1 𝐺𝑧2 • Transmission frequency: slice position
• Bandwidth: slice thickness (thk)
𝜔 The slice selection gradient strength also
Thin slice Thick slice affects the slice thickness (thk).
• 𝜔 = 𝛾 𝐵 = 𝛾Gzz
• thk = /𝛾Gz
𝐺𝑧1
𝐺𝑧2 The RF pulse shape and duration (and
𝑧 thus the RF bandwidth, ) are generally
predetermined as a matter of
𝜔 𝜔1 convenience.
RF excitation
bandwidth Strong gradient Weak gradient Then, the slice thickness and position
(from an operator) determines gradient
strength and transmission frequency.


 Slice selection (cont.)
Slice selection gradient enables us to selectively excite a slice of interest, but also
cause a problem of nonuniform precession frequencies within the excited slice.

𝜔 𝑧, slice thickness

RF
Slice refocusing gradient

𝐺𝑧
𝑧 z
'
𝜔0 𝜔 in phase
Transmission RF excitation y
frequency bandwidth 𝐺𝑧
x '
'
The phase induced by the nonuniform precession frequencies is proportional to
GZz/2, where  is the RF pulse duration.
The same phase with opposite polarity can be induced by applying a negative gradient
with half the area of the slice gradient, which is typically called slice refocusing
gradient, compensating for the phase induced by the slice-selection gradient.


 K-space


 K-space
FID Now we have selected slice position and
RF
Slice selection gradient
thickness to be excited.
𝐺𝑧 Slice refocusing gradient Only the magnetization in the slice is
tipped by RF pulse and has the
ADC
sampling transverse magnetization component.

Generally free induction decay (FID) is

not directly used for imaging. Why?
 FID at this point does not have spatial
information.
 Need time for spatial encoding by
using gradient coils.
(Exceptions : spiral imaging, radial
𝑧
imaging, etc)


 K-space (cont.)
MR signals after slice selection have spatially uniform precession
frequency (i.e., f0 : Larmor frequency) and thus can be described as
follows. ∞
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑓0 𝑡 𝑑𝑥𝑑𝑦
−∞
h(x,y) represents the final image to be generated. Then, it’s Fourier
transform 𝐻 𝑢, 𝑣 can be described as follows.
𝐹𝑇2𝐷{ℎ(𝑥, 𝑦)} = 𝐻 𝑢, 𝑣
∞
= ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑢𝑥 𝑒 −𝑗2𝜋𝑣𝑦 𝑑𝑥𝑑𝑦
−∞
If we can acquire 𝐻 𝑢, 𝑣 from MRI, then we can generate the image
ℎ(𝑥, 𝑦) by applying for inverse Fourier transform to the acquired data
𝐻 𝑢, 𝑣 .
We need to convert above 𝑆 𝑡 into 𝐻 𝑢, 𝑣 , i.e., convert 𝑒 −𝑗2𝜋𝑓0 𝑡 to
𝑒 −𝑗2𝜋𝑢𝑥 𝑒 −𝑗2𝜋𝑣𝑦 by gradient pulsing along the other two directions (X, Y).


 K-space (cont.)
Image, h(x,y) K-space, H(u,v)

2D FT

𝐹𝑇2𝐷{ℎ(𝑥, 𝑦)} = 𝐻 𝑢, 𝑣
∞
= ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑢𝑥 𝑒 −𝑗2𝜋𝑣𝑦 𝑑𝑥𝑑𝑦
−∞
The MRI data acquisition is to generate H(u,v), a frequency domain
determined as Fourier transform of the image domain, called K-
space.
u and v are typically replaced with kx and ky.


 K-space and Gradients

The role of gradient pulse is to modulate the precession freq. as a


function of spatial location
∞
RF
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑓0 𝑡 𝑑𝑥𝑑𝑦 Slice refocusing gradient
−∞
𝐺𝑧

𝑓 𝑥, 𝑦 = 𝛾(𝐵
ො 0 + 𝑥𝐺𝑥 + 𝑦𝐺𝑦) (where 𝛾𝐵
ො 0 = 𝑓0 ) 𝑡𝑥
𝐺𝑥 𝐺𝑥
Replace 𝑓0 with 𝑓 𝑥, 𝑦 . 𝑡y
∞ 𝐺𝑦 𝐺𝑦
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋ෝ𝛾(𝐵0 𝑡+𝑥𝐺𝑥 𝑡𝑥+𝑦𝐺𝑦𝑡𝑦) 𝑑𝑥𝑑𝑦
−∞ 𝑡
ADC ADC

∞
= 𝑒 −𝑗2𝜋ෝ𝛾𝐵0𝑡 ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋 ෝ 𝐺𝑥 𝑡𝑥 𝑥
𝛾
𝑒 −𝑗2𝜋 ෝ 𝐺𝑦 𝑡𝑦 𝑦
𝛾
𝑑𝑥𝑑𝑦
−∞


 K-space and Gradients (cont.)
Phase of the acquired data is modulated as a function of x and y.
∞
𝑆 𝑡 = 𝑒 −𝑗2𝜋ෝ𝛾𝐵0𝑡 ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋 ෝ 𝐺𝑥 𝑡𝑥 𝑥
𝛾 𝑒 −𝑗2𝜋 ෝ 𝐺𝑦 𝑡𝑦 𝑦
𝛾
𝑑𝑥𝑑𝑦
−∞

After demodulation of data, 𝑒 −𝑗2𝜋ෝ𝛾𝐵0𝑡 will be removed.

∞
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋 ෝ 𝐺𝑥 𝑡𝑥 𝑥
𝛾
𝑒 −𝑗2𝜋 ෝ 𝐺𝑦 𝑡𝑦 𝑦
𝛾
𝑑𝑥𝑑𝑦
−∞

Comparison with 𝐻 𝑢, 𝑣 .
∞
𝐻 𝑢, 𝑣 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑢𝑥 𝑒 −𝑗2𝜋𝑣𝑦 𝑑𝑥𝑑𝑦
−∞

The ADC data point corresponds to K-space location u and v as follows.

𝑘𝑥 = 𝑢 = 𝛾𝐺
ො 𝑥 𝑡𝑥 , 𝑘𝑦 = 𝑣 = 𝛾𝐺
ො 𝑦 𝑡𝑦


 K-space and Gradients (cont.)
The gradient strength and duration determines K-space location u and v.

𝑘𝑥 = 𝑢 = 𝛾𝐺
ො 𝑥 𝑡𝑥 𝑘𝑦 = 𝑣 = 𝛾𝐺
ො 𝑦 𝑡𝑦 

K-space, H(u,v) RF
Slice refocusing gradient
ky
𝐺𝑧

𝑡𝑥
kx 𝐺𝑥 𝐺𝑥
𝑡y
(u,v) 𝐺𝑦
𝐺𝑦
Sampling
one point
𝑡
ADC
How to coordinate Gx, Gy, and ADC to fill all the k-space points (u,v)?
 Frequency encoding and phase encoding.


 Frequency Encoding


 k-space coordinate
K-space Trajectory
Please keep in mind, with respect to spatial encoding, the coordinate in k-space at a
certain time is determined by the integral area of gradient as a function of time along
each axis. 𝑘𝑥 = 𝑢 = 𝛾𝐺
ො 𝑥 𝑡𝑥 𝑘 =𝛾 ̂Are
𝑘𝑦 = 𝑣 = 𝛾𝐺
ො 𝑦 𝑡𝑦 𝐺 a 𝑡
Pulse sequence provides a real time trajectory in k-space.
<Simple example>
k-space
Pulse sequence 𝑘𝑦
⑤
①②③ ④ ⑤ ⑥⑦⑧
⑥
𝐺𝑥 ⑦
time
⑧ ② 𝑘𝑥
𝐺𝑦 ①
time
④ ③


 Frequency Encoding
There are infinitely many ways of filling the K-space.
One simple and common way of filling the K-space is to fix gradient strength and
duration along one direction (e.g., Gy) and then apply a constant gradient along the
other direction (e.g. Gx) for a time period during which data sampling is performed.
Let’s assume Gy=0 for simplicity. 𝑘𝑥 = 𝑢 = K-space, F(u,v)
 𝛾𝐺
ො 𝑡
𝑥 𝑥 ky
𝑘𝑦 = 𝑣 =
𝛾𝐺
ො 𝑦 𝑡𝑦
RF 𝑢 = 𝛾𝐺
ො 𝑥 𝑡𝑥
Slice refocusing gradient

𝐺𝑧 kx
(u,v)
𝑡𝑥
𝐺𝑥 𝐺𝑥
This scheme fills only the right-hand side
ADC of the K-space.
Sampling
How to fill both sides of the K-space?


 Frequency Encoding and Gradient Echo
FID In k-space, the prephase gradient makes the
RF trajectory move left initially and then covers
Echo all the way to the right during data sampling.
𝐺𝑧
FID component decays rapidly because of
prephase gradient the field inhomogeneity from 𝐺𝑥 gradient.
𝐺𝑥 ② ③
① As the prephased signal is rewound by
sampling frequency encoding gradient, phase
ADC
coherence is recovered and then lost again.
0 TE
𝑘𝑦 This procedure is called frequency encoding.

Echo The echo formulated by prephasing

Sampling begins
③ 𝑘
② 𝑥 gradient and frequency-encoding gradient
① is called gradient echo.


 Frequency Encoding (another perspective)
To give spatial information, gradient coils change the frequency or phase of
precession spin depending on the position, and the signal is detected by receiver
coil. Then how do we obtain spatial information from the detected signal?
 Fourier transform (FT)

No spatial information

Without gradient FT
𝐵0 location
time
location

Spatial information
FT
With gradient
𝐵0 + 𝐺𝑥 𝑥 time
location
location


 Frequency Encoding (cont.)
FID

RF The frequency encoding gradient (𝐺𝑥

Echo
in this example) is also called readout
𝐺𝑧 gradient because the gradient is
turned on during data sampling.
prephase gradient
𝐺𝑥 ② ③ Signals from two objects at different
①
locations (e.g. along X) are detected as
ADC Sampling different frequency components in
0 TE time (K-space), which are converted
into their original locations in image
FT by Fourier transform.
𝐵0 + 𝐺𝑥 𝑥
Location
Location (X)
(X)


 Phase Encoding


 Phase Encoding
Now let’s change the gradient along the FID

last direction (Gy in this example) RF

Giving different spatial information along Echo
y-axis (i.e., Gy) prior to the sampling. 𝐺𝑧

The gradient strength (Gy) and timing (ty)

𝐺𝑥
determines K-space position, 𝑣 = ෝ𝛾𝐺𝑦 𝑡𝑦,
Adjusting the height of the horizontal line 𝑡𝑦
to be filled in K-space. 𝐺𝑦 𝐺𝑦

By repeating the RF excitation and ADC

sampling in multiple times with different ADC Sampling

𝑘𝑦
Gy values, we can fill all the K-space
region.
𝑣 =ෝ
𝛾𝐺𝑦 𝑡𝑦
This encoding scheme (along Y direction 𝑘𝑥

here) is called phase encoding.

We’ve completed pulse sequence diagram of gradient echo.


 Phase Encoding (another perspective)
FID
RF ① In phase after slice refocusing gradient

Echo ② Gradient makes the precession slower

𝐺𝑧
or faster, resulting in phase difference
along y-axis.
𝐺𝑥
③ Phase separation stops and all the spins
𝐺𝑦 precess at the same frequency.

①②③
Sampling
ADC
Rotating at 𝝎𝟎 = 𝜸𝑩𝟎 Slower Faster Rotating at 𝝎𝟎 but different phase
① ② ③

𝑩𝟎 𝑩𝟎 + 𝒚𝑮𝒚 𝑩𝟎

𝒚 𝒚 𝒚


 Mathematical Viewpoint of Frequency and Phase Encoding

Without frequency and phase encoding

∞
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑓0 𝑡 𝑑𝑥𝑑𝑦
−∞
Frequency encoding
𝑓 𝑥 = 𝛾(𝐵
ො 0 + 𝑥𝐺𝑥 )
∞
𝑆 𝑡 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋ෝ𝛾 𝐵0 𝑡+𝑥𝐺𝑥 𝑡𝑥
𝑑𝑥𝑑𝑦
−∞
∞
= 𝑒 −𝑗2𝜋ෝ𝛾𝐵0 𝑡 ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋ෝ𝛾𝑥𝐺𝑥 𝑡𝑥 𝑑𝑥𝑑𝑦
−∞
Phase encoding
𝜙𝑦 𝑦 = −2𝜋𝛾𝐺
ො 𝑦 𝑦𝑡𝑦 → additional phase
∞
𝑆 𝑡 = 𝑒 −𝑗2𝜋ෝ𝛾𝐵0 𝑡 ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋ෝ𝛾𝑥𝐺𝑥𝑡𝑥 𝑒 −𝑗2𝜋ෝ𝛾𝑦𝐺𝑦 𝑡𝑦 𝑑𝑥𝑑𝑦
−∞

After demodulation: ∞
𝐻 𝛾𝐺 ො 𝑦 𝑡𝑦 = ඵ ℎ 𝑥, 𝑦 𝑒 −𝑗2𝜋ෝ𝛾𝑥𝐺𝑥 𝑡𝑥 𝑒 −𝑗2𝜋ෝ𝛾𝑦𝐺𝑦 𝑡𝑦 𝑑𝑥𝑑𝑦
ො 𝑥 𝑡𝑥, 𝛾𝐺
∞
−∞ 𝑘𝑥 = 𝑢 = 𝛾𝐺
ො 𝑥 𝑡𝑥
𝐻 𝑢, 𝑣 = ඵ ℎ 𝑥, 𝑦  𝑒 −𝑗2𝜋𝑢𝑥
𝑒 −𝑗2𝜋𝑣𝑦
𝑑𝑥𝑑𝑦 𝑘𝑦 = 𝑣 = 𝛾𝐺
ො 𝑦 𝑡𝑦
−∞

∴ 𝐹𝑇2𝐷 ℎ 𝑥, 𝑦 = 𝐻 𝑢, 𝑣  The MR image is (inverse) 2DFT of K-space.


 Another Viewpoints of Frequency and Phase Encoding
The k-space has sinc-shaped data along both the frequency and phase encoding
directions.
Frequency encoding gradient is applied with a fixed strength but with varying time
points to be sampled, which modulates the K-space position.
In contrast, phase encoding gradient is applied with a fixed time duration but with
varying strength, which determines the K-space location.
These explain why they are called frequency encoding and phase encoding.
The desired MR image can be acquired by (inverse) 2D Fourier transform of K-space.
𝑘𝑦

𝑘𝑥 2D FT