MRI Fundamentals

Week 3:
Signal Processing Theory

Sung-Hong Park
Bio and Brain Engineering, KAIST


 • To better understand how MR images are generated
and processed, you need to be familiar with basic
concepts of Fourier transform and Sampling.

• We will discuss concepts of Fourier transform and

Sampling in the viewpoint of MRI in this week.


 Fourier transform and MR Images


 Fourier transform

The Fourier transform decomposes a signal (typically a function of time)

N
S

into the frequency components that make it up, similarly to how a musical
chord is expressed as the amplitude (or loudness) of its constituent notes.
N

The Fourier transform of a signal is a complex-valued function of frequency,

S

whose absolute value represents the amount of that frequency component

present in the original function, and whose complex argument (phase angle)
is the phase offset of the basic sinusoid in that frequency.

https://en.wikipedia.org/wiki/Fourier_transform


 Fourier transform (1D)

𝐹 𝑢 = න 𝑓(𝑡)𝑒 −𝑗2𝜋𝑢𝑡 𝑑𝑡
−∞

𝑓 𝑡 = න 𝐹(𝑢)𝑒 𝑗2𝜋𝑢𝑡 𝑑𝑢
−∞

𝑓 𝑡 ↔ 𝐹(𝑢)
Time-domain Frequency-domain


 Fourier transform (1D)
ex)
𝑓 𝑡 = cos(𝑡)

A
𝑡𝑖𝑚𝑒(𝑡) Frequency (𝜔

𝑓 𝑡 = cos(3𝑡)
Fourier
Transform
B
𝑡𝑖𝑚𝑒(𝑡)
Frequency (𝜔

𝑓 𝑡 = 2cos(3𝑡)

C
𝑡𝑖𝑚𝑒(𝑡) Frequency (𝜔

Time-domain Frequency-domain

Decomposition of a signal into complex exponential signals  Fourier transform

N
S


 Fourier transform (2D)
∞ ∞

𝐹 𝑢, 𝑣 = න න 𝑓(𝑥, 𝑦)𝑒 −𝑗2𝜋(𝑢𝑥+𝑣𝑦) 𝑑𝑥𝑑𝑦

−∞ −∞

∞ ∞

𝑓 𝑥, 𝑦 = න න 𝐹(𝑢, 𝑣)𝑒 𝑗2𝜋(𝑢𝑥+𝑣𝑦) 𝑑𝑢𝑑𝑣

−∞ −∞

𝑓 𝑥, 𝑦 ↔ 𝐹(𝑢, 𝑣)
Image-domain Frequency-domain

• An image is generally formed by two independent variables, typically

N
S

denoted by x and y.
• 2D Fourier transform is the same as sequential application of 1D Fourier
N
S

transform for each of the two variables.


 Fourier transform (2D)
𝑦 𝑘𝑦 𝑘𝑦
𝑦
A B
2DFT 2DFT
𝑘𝑥 𝑘𝑥

1. 5

0. 5

𝑥
0

𝑥
-0. 5

-1

𝑘𝑦 𝑘𝑦
-1. 5

-2
-1. 5

-0. 5

-1. 5

-0. 5
0 50 100 150 200 250 300 350 400 450 500
0. 5

1. 5

0. 5

1. 5
𝑦 𝑦
-2

-1

-2

-1
0

2
0
0

C D

50
50

100
100

2DFT 2DFT
𝑘𝑥 𝑘𝑥

150
150

200
200

250
250

1. 5

0. 5

300
300

𝑥 𝑥
-0. 5

-1

-1. 5
-1. 5

-0. 5

𝑘𝑦
-2
0. 5

1. 5

𝑘𝑦
0 50 100 150 200 250 300 350 400 450 500
-2

-1

350
0

2
350

𝑦 𝑦
0

E F
400
50
400

450
100
450

2DFT 500
2DFT
150
500

𝑘𝑥 𝑘𝑥
200
250

450 400 350 300 250 200 150 100 50 0

-2

-1. 5
1
300

-1

-0. 5

𝑥
0. 5
0. 5

𝑥
1

1. 5

2
350

2D Fourier transform describes how much of each frequency component is needed to

-0. 5
400 420 440 460 480 500 520 540 560 580 600
N

400
S

produced the image.

450
500


 Fourier Transform and MRI Images (1)

Frequency Domain
(k-space)

2D Fourier Transform
↔

Images

 Slow signal variation in space (image)  low frequencies (center)

 Fast signal variation in space (image)  high frequencies (edge)


 Fourier Transform and MRI Images (2)

Frequency Domain
(k-space)

2D Fourier Transform
↔

Images

• The center of frequency domain (k-space) contains the low-spatial frequency

information (overall shades of white or gray), which defines image contrast, while the
peripheral parts of k-space contain the high-spatial-frequency components, which affect
spatial resolution (fine ripples).


 Properties of Fourier Transform


 1. Linearity

If 𝑓 𝑥, 𝑦 ֞ 𝐹(𝑢, 𝑣) 𝑔 𝑥, 𝑦 ֞ 𝐺(𝑢, 𝑣)

Linearity 𝑎1 𝑓 𝑥, 𝑦 + 𝑎2 𝑔 𝑥, 𝑦 ֞ 𝑎1 𝐹 𝑢, 𝑣 + 𝑎2 𝐺(𝑢, 𝑣)

where 𝑎1 and 𝑎2 are two constants


 2. Translation

If 𝑓 𝑥, 𝑦 ֞ 𝐹(𝑢, 𝑣)

Translation Property 𝑓 𝑥 − 𝑥0 , 𝑦 − 𝑦0 ֞ 𝐹 𝑢, 𝑣 𝑒 −𝑗2𝜋(𝑢𝑥0 +𝑣𝑦0 )


 3. Conjugation

If 𝑓 𝑥, 𝑦 ֞ 𝐹(𝑢, 𝑣)

Conjugation Property 𝑓 ∗ 𝑥, 𝑦 ֞ 𝐹 ∗ (−𝑢, −𝑣)

* If 𝑓 𝑥, 𝑦 is real, then

𝐹 𝑢, 𝑣 = 𝐹 ∗ (−𝑢, −𝑣) (Conjugate symmetry)

𝐹 𝑢, 𝑣  = 𝐹 (−𝑢, −𝑣) (Symmetric magnitude)

⦟𝐹 𝑢, 𝑣 = −⦟𝐹 (−𝑢, −𝑣) (Anti-symmetric phase)


 4. Scaling

If 𝑓 𝑥, 𝑦 ֞ 𝐹(𝑢, 𝑣)

1 𝑢 𝑣
Scaling Property 𝑓 𝑎𝑥, 𝑏𝑦 ֞ 𝐹( , )
𝑎𝑏 𝑎 𝑏

* If a=b=-1, then 𝑓 −𝑥, −𝑦 ֞ 𝐹(−𝑢, −𝑣)

 Reversing a signal in space also reverses its Fourier transform


 5. Separable Product

If 𝑓1 𝑥 ֞ 𝐹1 (𝑢) 𝑓2 𝑦 ֞ 𝐹2 (𝑣)

Separable Product 𝑓 𝑥, 𝑦 = 𝑓1 (𝑥) ∙ 𝑓2 (𝑦)

FT 𝑓 𝑥, 𝑦 ֞ 𝐹1 (𝑢) ∙ 𝐹2 (𝑣)

 Fourier transform of a separable signal is also separable.


 Basic Signal Functions


 • Complex signals can be simplified as linear combination of
certain basic functions.

• These basic functions are useful to understand the concept

of Fourier transform and sampling that we want to discuss in
this week.


 1. Unit Impulse Function
𝑝∆𝜏 (𝑡) 𝛿(𝑡)

∆𝜏 → 0
1/∆𝜏

𝑡 𝑡
−∆𝜏/2 ∆𝜏/2

Infinite at t=0
N
S
N

Zero when t≠0

S
N

Sum of total area is 1.

S
N

Often called Dirac delta function and denoted as 𝛿(𝑡).

S
N
S

Any pulse that has area 1 and is too short and too high to be displayed.
N
S

The definition of 𝛿 𝑡 doesn’t have to be based on a rectangular shape (e.g. Gaussian, sinc).
N
S

Mathematical function that does not exist in the real world.

N
S

Useful to understand the concept of sampling


 2. Unit step function

𝑢(𝑡)
0; 𝑡 < 0
1
1 𝑢 𝑡 = ; 𝑡=0
2
1; 𝑡 > 0
𝑡

𝑑𝑢(𝑡)
𝛿 𝑡 =
𝑑𝑡

If we integrate the delta function, we get the unit step function.

N
S


 3. Rect function

1
0 𝑖𝑓 𝑡 > 2
1 1
𝑟𝑒𝑐𝑡 𝑡 = ⨅ 𝑡 = 𝑖𝑓 𝑡 =
2 2
1
1 𝑖𝑓 𝑡 < 2

1 1
𝑟𝑒𝑐𝑡 𝑡 = 𝑢 𝑡 + −𝑢 𝑡−
2 2
1 1
=𝑢 𝑡+ ∙𝑢 −𝑡
2 2

𝑡 𝜏 𝜏
𝑟𝑒𝑐𝑡 =𝑢 𝑡+ − 𝑢 𝑡−
𝜏 2 2
𝜏 𝜏
=𝑢 𝑡+2 ∙𝑢 2−𝑡


 4. Sinc function

sin(𝜋𝑥)
𝑠𝑖𝑛𝑐 𝑥 =
𝜋𝑥

1 𝑡
𝛿 𝑡 = lim 𝑠𝑖𝑛𝑐( )
∆𝜏→0 ∆𝜏 ∆𝜏

Fourier transform of the Rect function is the Sinc function.

N
S
N

The delta function can be also defined using the Sinc function.
S
N

In some textbooks, sinc(x) is defined as sin(x)/x.

S


 Basic Fourier Transform Pairs (1D)

Signal (𝒕) Fourier Transform (𝒖)

𝛿(𝑡) 1
1 𝛿(𝑢)
𝛿(𝑡 − 𝑡0 ) −𝑗2𝜋𝑢𝑡0
𝑒
𝑒 𝑗2𝜋𝑢0 𝑡 𝛿(𝑢 − 𝑢0 )
sin[2𝜋𝑢0 𝑡] 1
[𝛿 𝑢 − 𝑢0 − δ 𝑢 + 𝑢0 ]
2𝑗
cos[2𝜋𝑢0 𝑡] 1
[𝛿 𝑢 − 𝑢0 + δ 𝑢 + 𝑢0 ]
2
𝑟𝑒𝑐𝑡(𝑡) 𝑠𝑖𝑛𝑐(𝑢)
𝑠𝑖𝑛𝑐(𝑡) 𝑟𝑒𝑐𝑡(𝑢)


 Basic Fourier Transform Pairs (1D)
  
 (t )  1   (t )e  j 2ut dt    (t )e  j 2u0 dt    (t )dt  1
N
S

  

 
 (t  t0 )  e  j 2ut 0
  (t  t0 )e  j 2ut dt    (t  t0 )e  j 2ut dt
0
N

 
S


 e  j 2ut0   (t  t0 ) dt  e  j 2ut0


sin( 2u0t )  (e j 2u0t  e  j 2u0t ) / 2 j   (u  u0 )   (u  u0 ) / 2 j

N
S

 1/ 2
rect (t )  sinc(u )  rect (t )e  j 2ut dt   e  j 2ut dt
N
S

 1 / 2

1 sin(u )
 (e  ju  e ju )   sinc (u )
 j 2u u


 Basic Fourier Transform Pairs (2D)

Signal (𝒙, 𝒚) Fourier Transform (𝒖, 𝒗)

𝛿(𝑥, 𝑦) 1
1 𝛿(𝑢, 𝑣)
𝛿(𝑥 − 𝑥0 , 𝑦 − 𝑦0 ) −𝑗2𝜋(𝑢𝑥0 +𝑣𝑦0 )
𝑒
𝑗2𝜋(𝑢0 𝑥+𝑣0 𝑦) 𝛿(𝑢 − 𝑢0 , 𝑣 − 𝑣0 )
𝑒
sin[2𝜋 𝑢0 𝑥 + 𝑣0 𝑦 ] 1
[𝛿 𝑢 − 𝑢0 , 𝑣 − 𝑣0
2𝑗
− δ 𝑢 + 𝑢0 , 𝑣 + 𝑣0 ]
cos[2𝜋 𝑢0 𝑥 + 𝑣0 𝑦 ] 1
[𝛿 𝑢 − 𝑢0 , 𝑣 − 𝑣0
2
+ δ 𝑢 + 𝑢0 , 𝑣 + 𝑣0 ]
𝑟𝑒𝑐𝑡(𝑥, 𝑦) 𝑠𝑖𝑛𝑐(𝑢, 𝑣)
𝑠𝑖𝑛𝑐(𝑥, 𝑦) 𝑟𝑒𝑐𝑡(𝑢, 𝑣)


 Sampling


 Continuous vs Discrete signals

x(𝑡) Continuous
Amplitude

A continuous signal (waveform) x(t)

N
Signal

S
0

has values for all points in some

-1
(possibly infinite) interval.
0 1 2 3 4 5 6 7 8 Time

Sampling
(Digitization)

x[𝑛]

N
Amplitude

S
Discrete 0 A discrete signal (data sequence)
Signal
x[n] has values for only discrete
-1

0 5 10 15 20 25 Sample
points.
numbers


 Point impulse function (Unit Impulse)
𝑝∆𝜏 (𝑡) 𝛿(𝑡)

∆𝜏 → 0
1/∆𝜏

𝑡 𝑡
−∆𝜏/2 ∆𝜏/2

Infinite at t=0
N
S
N

Zero when t≠0

S
N

Sum of total area is 1.

S
N

Often called Dirac delta function and denoted as 𝛿(𝑡).

S
N
S

Any pulse that has area 1 and is too short and too high to be displayed.
N
S

The definition of 𝛿 𝑡 doesn’t have to be based on a rectangular shape (e.g. Gaussian, sinc).
N
S

Mathematical function that does not exist in the real world.

N
S

Useful to understand the concept of sampling


 Shifted Impulse and Sampling

Shifted Impulse Sampling Property

𝑥 𝑡 𝛿(𝑡 − 𝜏)
𝛿(𝑡 − 𝜏) 𝑥(𝑡) = 𝑥 𝜏 𝛿(𝑡 − 𝜏)

𝑥 𝜏
𝑡 𝑡 (area)
𝑡
𝜏 𝜏 𝜏


 Sampling?

Sampling (digitization) : To electronically store and process continuous

N
S

signals using computers, we must transform them into collections of

numbers.

time Sample number

Continuous signal Discrete signal


 Sampling function (1D)

1D Sampling function
N
S

𝑆 𝑥; 𝑥 = ෍ 𝛿(𝑥 − 𝑘𝑥) … …
𝑘=−∞

−5𝑥−4𝑥−3𝑥 −2𝑥 −𝑥 0 𝑥 2𝑥 3𝑥 4𝑥 5𝑥

𝑓(𝑥)
𝑓 𝑥 ∙ 𝑆 (𝑥)

𝑥 𝑥


 Sampling function (2D)

Sampling in 2D: Picking off values on a grid or

N
𝛿𝑠 (𝑥, 𝑦;

S
∆𝑥, ∆𝑦) matrix of points

N
𝑦 Typical sampling points of MRI : 64 x 64 to 512 x 512

S
𝑥

∆𝑦

∆𝑥 N
S
2D Sampling function

∞ ∞

𝛿𝑠 𝑥, 𝑦; ∆𝑥, ∆𝑦 = ෍ ෍ 𝛿(𝑥 − 𝑚 ∙ ∆𝑥, 𝑦 − 𝑛 ∙ ∆𝑦)

𝑚=−∞ 𝑛=−∞


 Sampling (2D)

𝑦 𝑦
𝑓(𝑥, 𝑦) : 2D continuous signal

N
S
N
𝑓𝑆 𝑥, 𝑦; ∆𝑥, ∆𝑦 : 2D sampled signals

S
N
∆𝑥, ∆𝑦 : sampling periods

S
∆𝑦

N
∆𝑦
1/∆𝑥, 1/∆𝑦 : sampling frequencies

S
∆𝑥 𝑥 ∆𝑥 𝑥

Fewer samples : ∆𝑥 and ∆𝑦 ↑, resolution ↓, sampling frequency ↓, partial

N S

volume effects ↑, SNR ↑.


 Sampling (2D)

𝑓𝑠 𝑥, 𝑦 = 𝑓 𝑥, 𝑦 ∙ 𝛿𝑠 𝑥, 𝑦; ∆𝑥, ∆𝑦
∞ ∞

= ෍ ෍ 𝑓(𝑥, 𝑦) ∙ 𝛿(𝑥 − 𝑚 ∙ ∆𝑥, 𝑦 − 𝑛 ∙ ∆𝑦)

𝑚=−∞ 𝑛=−∞
∞ ∞

= ෍ ෍ 𝑓(𝑚 ∙ ∆𝑥, 𝑛 ∙ ∆𝑦) ∙ 𝛿(𝑥 − 𝑚 ∙ ∆𝑥, 𝑦 − 𝑛 ∙ ∆𝑦)

𝑚=−∞ 𝑛=−∞
∞ ∞

= ෍ ෍ 𝑓𝑑 [𝑚, 𝑛] ∙ 𝛿(𝑥 − 𝑚 ∙ ∆𝑥, 𝑦 − 𝑛 ∙ ∆𝑦)

𝑚=−∞ 𝑛=−∞

Given discrete signal 𝑓𝑑 [𝑚, 𝑛], we can calculate the continuous signal 𝑓𝑠 (𝑥, 𝑦).
N
S

Can we reconstruct the original signal 𝑓(𝑥, 𝑦) from its sampled version 𝑓𝑠 𝑥, 𝑦 ?
N
S


 Nyquist sampling theorem


 Sampling frequency

The sampling frequency or sampling rate, 𝑓𝑠 , is the average number of samples

N
S

obtained in one second (samples per second), thus 𝑓𝑠 = 1/𝑇.


 Fourier transform of sampling function (1D)

Fourier transform of 1D Sampling function

N
S

∞ ∞
1
𝑆 𝑥; 𝑥 = ෍ 𝛿(𝑥 − 𝑘𝑥)  𝑥
 ෍ 𝛿(𝑢 − 𝑘/𝑥)
𝑘=−∞ 𝑘=−∞

… …  … …

𝑥 𝑢
𝑥 1/𝑥

Fourier transform of an impulse train is another impulse train with

N
S

an impulse gap equal to inverse of the original impulse gap.


 Fourier transform of sampling function (2D)
Fourier transform of 2D Sampling function
N
S

∞ ∞
𝑆 𝑥, 𝑦; ∆𝑥, 𝑦 = ෍ ෍ 𝛿(𝑥 − 𝑚 ∙ ∆𝑥, 𝑦 − 𝑛 ∙ ∆𝑦)
𝑚=−∞ 𝑛=−∞ ∞ ∞
1
 𝑥𝑦
 ෍ ෍ 𝛿(𝑢 − 𝑚/∆𝑥, 𝑣 − 𝑛/∆𝑦)
𝑚=−∞ 𝑛=−∞

𝑦 𝑣


𝑥 𝑢
1/∆𝑦
∆𝑦

∆𝑥 1/∆𝑥

Fourier transform of an impulse train is another impulse train with

N
S

an impulse gap equal to inverse of the original impulse gap.


 Sampling Theorem (1D)

𝑓(𝑥) 𝑆 (𝑥) 𝑓𝑆 𝑥 = 𝑓 𝑥 ∙ 𝑆 (𝑥)

… …
𝑥 𝑥 𝑥
𝑥 𝑥



𝐹𝑆(𝑢)
𝐹(𝑢)

… …
𝑢
… … 𝑢
𝑢
1/𝑥 1/𝑥 1/𝑥
𝑚
The spectrum 𝐹𝑠 𝑢 is calculated by shifting the spectrum 𝐹 𝑢 to locations u =
N
S

∆𝑥
for all integer m, and adding all the shifted spectra (and dividing the result by ∆𝑥).


 Nyquist Sampling Theorem (1D)
Band-limited signal 𝑓(𝑥) 1 1
When 2W  =sampling freq. When 2W > =sampling freq.
𝑥 𝑥
𝑓(𝑥) 𝑓𝑆 𝑥 = 𝑓 𝑥 ∙ 𝑆 (𝑥) 𝑓𝑆 𝑥 = 𝑓 𝑥 ∙ 𝑆 (𝑥)

𝑥 𝑥 𝑥 𝑥
𝑥



𝐹𝑆(𝑢) Lowpass filter 𝐹𝑆(𝑢)
𝐹(𝑢)
Aliasing
… … … …
𝑢 𝑢 𝑢
−𝑊 𝑊
1/𝑥 1/𝑥 1/𝑥 1/𝑥

• The Nyquist theorem states that the sampling frequency (fS) must be at least twice as high as the
N
S

maximum bandwidth W of the original signal for the signal to be represented accurately.


 Sampling Theorem (2D)
𝑦 𝑓 𝑥, 𝑦 𝑆 𝑥, 𝑦 𝑦 𝑓𝑆 𝑥, 𝑦 =𝑓 𝑥, 𝑦 𝑆 𝑥, 𝑦

𝑥
∆𝑦
∆𝑦

∆𝑥 𝑥
𝑥
∆𝑥




𝐹 𝑢, 𝑣 𝐹𝑠 𝑢, 𝑣
𝑣 𝑣

…
𝑣
𝑢 1/∆𝑦 𝑢
𝑢 … …
1/∆𝑦

…
1/∆𝑥
1/∆𝑥 𝑚 𝑛
The spectrum 𝐹𝑠 𝑢, 𝑣 is calculated by shifting the spectrum 𝐹 𝑢, 𝑣 to locations ( , ),
N
S

∆𝑥 ∆𝑦
for all integer m and n, and adding all shifted spectra (and dividing the result by ∆𝑥∆𝑦).


 Nyquist Sampling Theorem (2D)
Band-limited signal 𝑓(𝑥, 𝑦) When 2W  1/𝑥 and 2Z  1/𝑦 When 2W > 1/𝑥 and 2Z > 1/𝑦
𝑦 𝑦 𝑦

∆𝑦
∆𝑦

𝑥
𝑥 ∆𝑥
𝑥


∆𝑥


𝐹 𝑢, 𝑣

𝐹𝑠 𝑢, 𝑣 𝐹𝑠 𝑢, 𝑣
𝑣 Lowpass filter 𝑣 Aliasing
…

…
𝑍
𝑢 𝑢 𝑢
−𝑊 𝑊 … … … …
−𝑍
…

…
• The Nyquist theorem states that the sampling frequency (fS) must be at least twice as high as the
N
S

maximum bandwidth W of the original signal for the signal to be represented accurately.


 Aliasing

Aliasing is an effect that causes different signals to become

N
S
indistinguishable when sampled.

N
It also refers to the distortion or artifact that results when

S
the signal reconstructed from samples is different from the
original continuous signal.

No Aliasing Aliasing

MRI phantom images