Cambridge Books Online

http://ebooks.cambridge.org/

Fundamentals of Medical Imaging

Paul Suetens

Book DOI: http://dx.doi.org/10.1017/CBO9780511596803

Online ISBN: 9780511596803

Hardback ISBN: 9780521519151

Chapter

Appendix A - Linear system theory pp. 219-231

Chapter DOI: http://dx.doi.org/10.1017/CBO9780511596803.010

Cambridge University Press

 Appendix

A Linear system theory

Introduction We denote even and odd signals by se (x) and so (x),

This appendix summarizes a number of fundamen- respectively. Obviously, the product of two even sig-
tal definitions from linear system and Fourier theory. nals is even, the product of two odd signals is even,
Only what is relevant for this book is discussed. More and the product of an even and an odd signal is odd.
information can be found in the specific textbooks on From the definition it is also clear that
this topic [37, 38, 39].
+∞
+∞
se (x) dx = 2 se (x) dx (A.4)
−∞ 0
Signals
and
Definitions and examples
+∞
A signal represents the measurable change of some so (x) dx = 0. (A.5)
quantity with respect to one or more independent −∞
variables such as time or spatial position. Mathemat-
ically, a signal can be represented as a function. In Any signal can be written as the sum of an even and
medical imaging, the signals are multidimensional. an odd part:
Modern acquisition systems acquire three- (3D) and    
s(x) s(−x) s(x) s(−x)
even four-dimensional (4D) data. The signal can then s(x) = + + −
be written as 2 2 2 2
= se (x) + so (x). (A.6)
s = f (
r , t ) = f (x, y, z, t )
∀ x, y, z, t ∈ R and s ∈ C. (A.1) A signal is periodic if

s(x + X ) = s(x) ∀ x ∈ R. (A.7)

The value of the function is usually real, but it can be
complex.
The smallest finite X that satisfies this equation is
Signals have some particular properties. The most
called the period. If no such X exists, the function is
important for this book are defined as follows.
aperiodic.
A signal is even if
A complex function can be written in Cartesian
representation. For 2D signals, we have
s(−x) = s(x) ∀ x ∈ R. (A.2)
s(x, y) = u(x, y) + iv(x, y), (A.8)
A signal is odd if
where u(x, y) and v(x, y) are the real part and imag-
s(−x) = −s(x) ∀ x ∈ R. (A.3) inary part, respectively. A complex function can also
be written in polar representation as
[37] R. N. Bracewell. The Fourier Transform and Its Applications. New
York: McGraw-Hill, second edition, 1986. s(x, y) = |s(x, y)| eiφ(x,y) , (A.9)
[38] E. Oran Brigham. The Fast Fourier Transform and its Appli-
cations. Englewood Cliffs, NJ: Prentice-Hall International, first
edition, 1988. where
[39] A. Oppenheim, A. Willsky, and H. Nawab. Signals and Sys- 
tems. Upper Saddle River, NJ: Prentice-Hall International, second |s(x, y)| = u 2 (x, y) + v 2 (x, y) (A.10)
edition, 1997.

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

Exponentials Sinusoid Figure A.1 Some of the most important

8 1 signals in linear system theory. (a) The
7 0.8 exponentials ex (solid) and e−x (dashed).
0.6 (b) sin(x) (dashed) and cos(x) (solid). (c)
6
0.4 Rectangular pulse. (d) Triangular pulse. (e)
5 0.2 Normalized Gaussian with µ = 0. (f)
4 0
sinc(x).

3 –0.2
–0.4
2
–0.6
1 –0.8
0 –1
–2 –1 0 1 2 –2p –p 0 p 2p
(a) (b)
Rectangular Triangular
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
–2L –L 0 L 2L –2L –L 0 L 2L
(c) (d)
Normalized Gaussian (σ = 1) Sinc
0.4 1
0.35 0.8
0.3
0.6
0.25
0.4
0.2
0.2
0.15
0.1
0

0.05 0.2
0 0.4
–3 –2 –1 0 1 2 3 –5p–4p –3p –2p –5p 0 1 2 3 4 5
(e) (f)

is the modulus or the amplitude and When the constant a > 0, the exponential func-
tion increases continuously with increasing x (solid
 
v(x, y) line); when a < 0, it decreases toward zero with
φ(x, y) = arctan (A.11) increasing x (dashed line).
u(x, y)
• Complex exponential or sinusoid (Figure A.1(b))
is the argument or the phase of s.
A number of signals are used extensively in system A ei(2πkx+φ)
theory and are important enough to have a unique = A (cos(2π kx + φ) + i sin(2πkx + φ)).
name. Here are some of them (see also Figure A.1). (A.13)
• Exponential (Figure A.1(a))
A sinusoid is characterized by three parameters: its
modulus or amplitude A , spatial frequency k, and
exp(ax) = eax . (A.12) phase φ. The term i is the imaginary unit; that is
220

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

i2 = −1. The real and imaginary parts of a sinu- nor periodic. To be compatible with the theory of
soid are, respectively, a cosine (solid line) and sine single-valued functions, it is common to use the mean
function (dashed line). of the value immediately left and right of the disconti-
• Unit step function (also called Heaviside’s function) nuity. For example, the values at the discontinuities of
the rectangular pulse equal 1/2. For more information,

 for x < x0 we refer to [37, 38].

 0

1
u(x − x0 ) = for x = x0
The Dirac impulse
(A.14)

 2

1 for x > x0 . The Dirac impulse, also called impulse function or δ-
function, is a very important function in linear system
The constant x0 denotes the location of the step. theory. It is defined as
The function is discontinuous at x0 .
• Rectangular function (Figure A.1(c)) δ(x − x0 ) = 0 for x  = x0 ,

+∞
 (A.19)
 for |x| < L δ(x − x0 ) dx = 1,
1
x  
1
−∞
 = for |x| = L (A.15)
2L 
 2 with x0 a constant. The value of a Dirac impulse is

0 for |x| > L. zero for all x except in x = x0 , where it is undefined.
However, the area under the impulse is finite and is by
The constant 2L is the width of the rectangle. definition equal to 1. A Dirac impulse can be consid-
Because the nonzero extent of the function is finite, ered as the limit of a rectangular pulse of magnitude
the function is also called a rectangular pulse. 1/ε and spatial extent ε > 0 such that the area of the
• Triangular function (Figure A.1(d)): pulse is 1:

  
1 x
 x  1 − |x| for |x| < L δ(x) = lim  . (A.20)
 = L (A.16) ε→0 ε ε
2L 
0 for |x| ≥ L.
When ε becomes smaller, the spatial extent decreases,
Note that the base of the triangular pulse is the amplitude increases, but the area remains the same.
equal to 2L. Clearly, the Dirac impulse is not a function in the strict
mathematical sense. Its rigorous definition is given by
• Normalized Gaussian (Figure A.1(e)) the theory of generalized functions or distributions,
which is beyond the scope of this text [40].
1
e−(x−µ)
2 /2σ 2
Gn (x) = √ . (A.17) Using Eq. (A.20), it is clear that
2πσ

+∞
+∞  
1 x
The Gaussian is normalized (i.e., its integral for all δ(x)s(x) dx = lim  s(x) dx,
x is 1). The constants µ, σ , and σ 2 are the mean, the −∞ ε→0 −∞ ε ε
standard deviation, and the variance, respectively. (A.21)
• Sinc function (Figure A.1(f))
and consequently the following properties hold:

sinc(x) =
sin(x)
. (A.18)
• sifting let s(x) be continuous at x = x0 , then
x

+∞
According to L’Hôpital’s rule, sinc(0) = 1. s(x) δ(x − x0 ) dx = s(x0 ); (A.22)
−∞
Note that the rectangular, the triangular, the normal-
ized Gaussian, and the sinc function are all even and [40] R. F. Hoskins. Generalised Functions. New York: McGraw-Hill
aperiodic. The step function is neither even nor odd Book Company, 1979. 221

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

• scaling dynamic relationships that involve (sets of) difference

equations.

+∞
With respect to their model, systems can be linear
A δ(x) dx = A, (A.23)
−∞ or nonlinear. A system is linear if the superposition
principle holds, that is,
this is a special case of sifting.
L{c1 s1 + c2 s2 } = c1 L{s1 } + c2 L{s2 }
The definition of the impulse function can be extended
to more dimensions by replacing x by
r . The properties ∀ c1 , c2 ∈ R, (A.27)
are analogous; for example, the sifting property in 2D
becomes with s1 and s2 as arbitrary signals. For example, the
amplifier introduced above is linear because



+∞

s(
r ) δ(
r −
r0 ) d
r = s(
r0 ). (A.24) L{c1 s1 + c2 s2 } = A(c1 s1 + c2 s2 )

−∞
= c1 A s1 + c2 A s2
The impulse function is crucial for a thorough under- = c1 L{s1 } + c2 L{s2 }. (A.28)
standing of sampling, as discussed on p. 228.
A system is nonlinear if the superposition principle
Systems does not hold. For example, a system whose output is
the square of the input is nonlinear because
Definitions and examples
A system transforms an input signal (also called exci- L{c1 s1 + c2 s2 } = (c1 s1 + c2 s2 )2
tation) into an output signal (also called response).
Mathematically this can be written as  = (c1 s1 )2 + (c2 s2 )2 . (A.29)

so = L{si }, (A.25) In this text, only linear systems are dealt with.
A system is time invariant if its properties do not
where si and so are the input and output signals, respec- change with time. Hence, if so (t ) is the response to
tively.∗ The term L is an operator and denotes the the excitation si (t ), so (t − T ) will be the response to
action of the system. A system can be complex and it si (t − T ). Analogously, a system is shift invariant if
can consist of many diverse parts. In system theory, its properties do not change with spatial position: if
however, it is often considered as a black box, and the so (x) is the response to the excitation si (x), so (x −
detailed behavior of the different components is irrel- X ) will be the response to si (x − X ). We will denote
evant. As a simple example, consider an amplifier. It linear time-invariant systems as LTI systems and linear
consists of many electrical and electronic parts, but shift-invariant systems as LSI systems.
its essential action is to amplify any input signal by a The response to a Dirac impulse is called the
certain amount, say A. Hence, impulse response. From Eq. (A.22) it follows that

+∞
so (t ) = L{si (t )} = A si (t ). (A.26)
si (x) = si (ξ ) δ(x − ξ ) dξ . (A.30)
−∞
The process of finding a mathematical relationship
between the input and the output signal is called mod- Let h(x) be the impulse response of a LSI system. Based
eling. The simplest is an algebraic relationship, as in on the superposition principle (A.27), so (x) can then
the example of the amplifier. More difficult are con- be written as
tinuous dynamic relationships that involve (sets of)
+∞
differential or integral equations, or both, and discrete
so (x) = L{si } = si (ξ )L{δ(x − ξ )} dξ
−∞
∗ We also use s to represent an odd signal. However, this should
+∞
o
222 cause no confusion because the exact interpretation is clear from = si (ξ ) h(x − ξ ) dξ . (A.31)
the context. −∞

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

A similar equation holds for a LTI system: • multiplication, the product of the mirrored and
shifted function s1 (x − ξ ) with s2 (ξ ) is the colored

+∞ part in Figure A.2(c),
so (t ) = si (τ ) h(t − τ ) dτ . (A.32)
−∞ • integration, the area of the colored part is the
convolution value in point x.
The integral in Eqs. (A.31) and (A.32) is a so-called
convolution and is often represented by an asterisk: The convolution function is found by repeating the
previous steps for each value of x.
so = si ∗ h. (A.33) Convolution can also be defined for multidimen-
sional signals. For 2D (two-dimensional) signals,
we have
The function h is also known as the point spread func-
tion or PSF (see Figure 1.4). Because of its importance
s1 (x, y) ∗ s2 (x, y)
in this book, convolution will first be discussed in some
more detail.


+∞

= s1 (x − ξ , y − ζ ) s2 (ξ , ζ ) dξ dζ , (A.36)
Convolution −∞

Given two signals s1 (x) and s2 (x), their convolution is

or equivalently
defined as follows:

+∞ s2 (x, y) ∗ s1 (x, y)
s1 (x) ∗ s2 (x) = s1 (x − ξ ) s2 (ξ ) dξ , (A.34)
−∞


+∞

= s2 (x − ξ , y − ζ ) s1 (ξ , ζ ) dξ dζ . (A.37)
or equivalently −∞

+∞
The graphical analysis shown above can be extended
s2 (x) ∗ s1 (x) = s1 (ξ ) s2 (x − ξ ) dξ . (A.35)
−∞ to 2D. The convolution values are then represented by
volumes rather than by areas.
The result of both expressions is identical, as is clear The convolution integrals (A.34)–(A.37) have
when substituting ξ by x − ξ . many properties. The most important in the context
A graphical interpretation of convolution is given of this book include the following.
in Figure A.2. The following steps can be discerned: • Commutativity
• mirroring , changing ξ to −ξ ,
• translation over a distance equal to x, s1 ∗ s2 = s2 ∗ s1 . (A.38)

Figure A.2 Graphical interpretation of

1 1 the convolution of a rectangular pulse
s1 (x) with a triangle s2 (x). Changing the
s2(x) s1(x) s2(–x) s1(x) independent variable to ξ does not
change the functions (a). After s2 (ξ ) is
mirrored (b), it is translated over a
distance x, both functions are multiplied
–1 0 1 x –1 0 1 x and the result is integrated (c). The area of
(a) (b) the overlapping part is the result for the
chosen x. The convolution s1 (x) ∗ s2 (x) is
shown in (d).
s1(x) *s2(x)
1

s2(x–x) s1(x) 1/2

–1 0 x 1 x –1 0 1 2 x
223
(c) (d)

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

• Associativity transfer function, that is, So (k) = Si (k)H (k), and

calculating the inverse Fourier transform of So (k).
(s1 ∗ s2 ) ∗ s3 = s1 ∗ (s2 ∗ s3 ) = s1 ∗ s2 ∗ s3 . In linear system theory, the transfer function H (k)
(A.39) is often used instead of the PSF h(x) because of
its nice mathematical and interesting physical prop-
• Distributivity erties. The relationship between the PSF h(x) and
the transfer function H (k) is given by the Fourier
s1 ∗ (s2 + s3 ) = s1 ∗ s2 + s1 ∗ s3 . (A.40) transform (A.42). Because of its importance in med-
ical imaging, the Fourier transform is discussed in
more detail in the next section. Note however that
Response of a LSI system the Fourier transform is not the only possible trans-
Let us first consider the response of a LSI system to a form. There are many others (Hilbert, Laplace, etc.),
sinusoid. Using Eq. (A.31) with si (x) = A e2πikx yields although the Fourier transform is by far the most

important in the theory of medical imaging.
+∞
so (x) = A e2πik(x−ξ ) h(ξ ) dξ
−∞

+∞
The Fourier transform
=Ae 2πikx
e −2πikξ
h(ξ ) dξ Definitions
−∞
Let k and r be the conjugate variables in the Fourier
= A e2πikx H (k), (A.41) domain and the original domain, respectively. The
forward Fourier transform (FT) of a signal s(r) is
with H (k) the so-called Fourier transform of the PSF defined as
h(x):
+∞

S(k) = F{s(r)} = s(r) e−2πirk dr. (A.45)
+∞ −∞
H (k) = e−2πikξ h(ξ ) dξ . (A.42)
−∞
The operator symbol F (calligraphic F ) is used as the
notation for the transform. Uppercase letters are used
The function H (k) is also called the transfer function.
for the result of the forward transform. Analogously,
It can be shown that any input signal si (x) can
the inverse Fourier transform (IFT) is defined as
be written as an integral of weighted sinusoids with
different spatial frequencies:
+∞
s(r) = F −1 {S(k)} = S(k) e+2πirk dk. (A.46)

+∞ −∞
si (x) = Si (k) e2πikx dk, (A.43)
−∞ It can be shown that for continuous functions s,

where Si (k) is the Fourier transform of si (x). The s(r) = F −1 {F{s(r)}}. (A.47)
signal si (x) is the so-called inverse Fourier transform
(because of the + sign in the exponent instead of the From the definitions (A.45) and (A.46), it follows that
− sign in Eq. (A.42)) of Si (k). for an even function se (r),
Using Eq. (A.41) and the superposition principle,
the output signal so is then F{se (r)} = F −1 {se (r)}. (A.48)

+∞
If r is time with dimension seconds, k is the temporal
so (x) = Si (k) H (k) e2πikx dk. (A.44)
−∞ frequency with dimension hertz. Related to the tem-
poral frequency is the angular frequency ω = 2π k with
Summarizing, the output function so of a LSI system dimension radians per second. In this case, the base
can be calculated in two ways: either by convolving function of the forward FT describes a rotation in the
the input function si with the PSF, that is, so = si ∗ h clockwise direction with angular velocity ω. If r is spa-
224 (Eq. (A.33)), or in the k-space or frequency domain tial position with dimension mm, k is spatial frequency
by multiplying the Fourier transform of si with the with dimension mm−1 .

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

In this definition, the original signal and the result The forward FT of a rectangular pulse is a sinc function
of the transform are one dimensional. In medical whose maximum amplitude is equal to the area of the
imaging, however, the signals are often multidimen- pulse. The first zero crossing occurs at
sional and vectors must be used in the definitions.
1
Forward: k= . (A.53)

+∞ 2L



S(k) = F{s(
r )} = s(
r ) e−2πik·
r d
r . Thus, the broader the width of the rectangular pulse in
−∞
the original domain, the closer the first zero-crossing
Inverse: lies near the origin of the Fourier domain or the more

+∞
“peaked” the sinc function is (see Figure A.1(c) and(f)).
−1
=
r


e+2πik·

s(
r ) = F {S(k)} S(k) dk. (A.49) Example 2
−∞
The forward FT of the product of a step function

r and k
are the conjugate variables,
r being spatial (Eq. (A.14)) and an exponential (Eq. (A.12)) (we
position and k
spatial frequency. Although only one assume a > 0) is
integral sign is shown, it is understood that there are
+∞
as many as there are independent variables. The orig- F{u(x) e−ax } = u(x) e−ax e−2πikx dx
inal signal and its transform are known as a Fourier −∞
transform pair denoted as
+∞
= e−(a+2πik)x dx
s(r) ←→ S(k). (A.50) 0
1
=
In general, the result of the forward FT of a signal a + 2π ik
is a complex function. The amplitude spectrum is the a 2π k
modulus of its FT, while the phase spectrum is the = 2 −i 2 .
a + 4π k2 2 a + 4π 2 k 2
phase of its FT. Both spectra show how amplitude and (A.54)
phase vary with spatial or temporal frequencies. Often,
the phase spectrum is considered irrelevant, and only The result is complex; according to Eqs. (A.8) and
the amplitude spectrum is considered. Note, however, (A.9), we have the following.
that a signal is completely characterized if and only if
both the amplitude and phase spectrum are specified. a
Real part: .
a 2 + 4π 2 k 2
Examples Imaginary part: −
2π k
.
a 2 + 4π 2 k 2
Example 1 (A.55)
1
The FT of a rectangular pulse (Eq. (A.15)), scaled with Modulus: √ .
amplitude A is a 2 + 4π 2 k 2
 
 
+∞   2π k
x x Phase: −arctan .
F A = A e−2πikx dx a
2L −∞ 2L

+L This transform pair is a mathematical model of the
= A e−2πikx dx filter shown in Figure 4.21.
−L
A Example 3
=− (e−2πikL − e+2πikL ).
2π ik The forward FT of the Dirac impulse. Direct applica-
(A.51) tion of the sifting property (A.22) gives

Using Eqs. (A.13) and (A.18), we finally obtain
+∞

F{δ(x − x0 )} = δ(x − x0 ) e−2πikx dx
  −∞
x
A ←→ 2AL sinc(2πkL). (A.52) 225
2L = e−2πikx0 . (A.56)

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

The FT of a Dirac impulse at x0 is complex: in the Table A.1 Important Fourier transform pairs
in linear system theory
amplitude spectrum, all spatial frequencies are present
with amplitude 1. The phase varies linearly with k with Image space Fourier space
slope −2πx0 . 1 δ(k)
A difficulty arises when calculating the IFT: δ(x) 1

+∞ cos(2πk0 x) 1 (δ(k + k ) + δ(k − k ))
2 0 0
−2πikx0 +2πikx
s(x) = e e dk sin(2πk0 x) i (δ(k + k ) − δ(k − k ))
2 0 0
−∞ x )

+∞ ( 2L 2L sinc(2π Lk)
x ) L sinc 2 (π Lk)
= cos(2π k(x − x0 )) dk ( 2L
−∞ 2 2 2
Gn (x) e−2π k σ

+∞
+i sin(2πk(x − x0 )) dk. (A.57)
−∞

Because its integrand is odd, the second integral is

Properties
• Linearity If s1 ←→ S1 and s2 ←→ S2 , then
zero. The first integral has no meaning, unless it is
interpreted according to the distribution theory. In
this case, it can be shown that c1 s1 + c2 s2 ←→ c1 S1 + c2 S2 ∀c1 , c2 ∈ C.
(A.61)

+∞
+∞
cos(2πk(x − x0 )) dk = e+2πik(x−x0 ) dk This can easily be extended to more than two
−∞ −∞
signals.
= δ(x − x0 ). (A.58)
• Scaling If s(x) ←→ S(k), then
Hence,  
1 k
s(ax) ←→ S a ∈ R0 . (A.62)
δ(x − x0 ) ←→ e−2πikx0 . (A.59) |a| a

Example 4 • Translation If s(x) ←→ S(k), then

The forward FT of a cosine function is
s(x − x0 ) ←→ e−2πix0 k S(k) x0 ∈ R. (A.63)
F{cos(2πk0 x)}

+∞ Thus, translating a signal over a distance x0 only
= cos(2πk0 x) e−2πikx dx modifies its phase spectrum.
−∞
• Transfer function and impulse response (or PSF)

+∞  e+2πik0 x + e−2πik0 x

are a FT pair. Indeed, Eq. (A.42) shows that
= e−2πikx dx
−∞ 2

+∞ h(x) ←→ H (k). (A.64)
1
= e−2πi(k−k0 )x dx
2 −∞

In imaging, the FT of the PSF is known as the
1 +∞ −2πi(k+k0 )x optical transfer function (OTF). The modulus of the
+ e dx
2 −∞ OTF is the modulation transfer function (MTF). As
1 1 mentioned in Chapter 1, the PSF and OTF char-
= δ(k − k0 ) + δ(k + k0 ). (A.60) acterize the resolution of the system. If the PSF is
2 2
expressed in mm, the OTF is expressed in mm−1 .
The spectrum of a cosine function consists of two Often, line pairs per millimeter (lp/mm) is used
impulses at spatial frequencies k0 and −k0 . In gen- instead of mm−1 . The origin of this unit can easily
eral it can be shown that a periodic function has a be understood if an image with sinusoidal inten-
discrete spectrum (i.e., not all spatial frequencies are sity lines at a frequency of 1 period per millimeter
226 present), whereas an aperiodic function has a contin- or 1 lp/mm, that is, one dark and one bright line
uous spectrum. Table A.1 shows a list of FT pairs used per millimeter, is observed (Figure A.3). This line
in this book. pattern can be written as sin(2π x), x expressed in

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

• The FT of a real signal is Hermitian

= S(
S(−k)
if s(x) ∈ R,
¯ k) (A.66)

where S¯ denotes the complex conjugate of S (i.e.,

the real part is even and the imaginary part is odd).
From Eqs. (A.6), (A.13), and (A.45), we obtain

+∞
S(k) = s(x) e−2πikx dx
−∞

+∞
= [se (x) + so (x)]
−∞
· [cos(2π kx) − i sin(2πkx)] dx

+∞
= se (x) cos(2π kx) dx
Figure A.3 Image with sinusoidal intensity lines at a frequency of −∞
1 lp/mm.
+∞
−i so (x) sin(2π kx) dx. (A.67)
−∞
mm. The Fourier transform of this function con-
sists of two impulses at spatial frequency 1 mm−1 The first integral is the real even part of S(k),
and −1 mm−1 . This then explains why the fre- and the second is the imaginary odd part of S(k).
quency units mm−1 and lp/mm can be used as Hence, to compute the FT of a real signal, it suffices
synonyms. to know one half-plane. The other half-plane can
The resolution of an imaging system is some- then be computed using Eq. (A.66).
times characterized by the distinguishable number Equation (A.67) further shows that if a function
of line pairs per millimeter. It is clear now that is even (odd), its FT is even (odd). Consequently,
this is a limited and subjective measure, and that if a function is real and even, its FT is real and
it is preferable to show the complete OTF curve even, whereas if a function is real and odd, its FT
when talking about the resolution. Nevertheless, is imaginary and odd.
it is common practice in the technical documents • Parseval’s theorem
of medical imaging equipment and in the medical
+∞
+∞
literature simply to list an indication of the resolu- |s(x)|2 dx = |S(k)|2 dk. (A.68)
tion in lp/mm at a specified small amplitude (in %) −∞ −∞
of the OTF.
• Separability In many cases, a 2D FT can be calcu-
• Convolution On p. 224 it was concluded that an
output function so of a LSI system can be calculated lated as two subsequent 1D FTs. The transform is
in two ways: (1) so = si ∗ h in the image domain then called separable. For example,
or (2) F −1 {So (k) = Si (k)H (k)} in the Fourier
domain. In general, if s1 ←→ S1 and s2 ←→ S2 , F{sinc(x) sinc(y)}
then


+∞
sin(x) sin(y) −2πi(kx x+ky y)
= e dx dy
s1 ∗ s2 ←→ S1 · S2 x y
(A.65) −∞
s1 · s2 ←→ S1 ∗ S2 .
+∞ sin(x) −2πikx x
= e dx
This is a very important property. The convolution −∞ x

+∞
of two signals can be calculated via the Fourier sin(y) −2πiky y
transform by calculating the forward–inverse FT · e dy
−∞ y 227
of both signals, multiplying the FT results, and
calculating the inverse–forward FT of the product. = F{sinc(x)} F{sinc(y)}. (A.69)

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

• Another important property of a 2D FT is the Sampling

projection theorem or central-slice theorem. It Equation (A.1) represents an analog continuous sig-
is discussed in Chapter 3 on X-ray computed nal, which is defined for all spatial positions and can
tomography. have any (real or complex) value:

Polar form of the Fourier transform s(x) ∀ x ∈ R. (A.76)

Using polar coordinates
In practice, the signal is often sampled, that is, only
x = r cos θ discrete values at regular intervals are measured:
(A.70)
y = r sin θ,
ss (x) = s(nx) n ∈ Z. (A.77)
Eq. (A.49)
The constant x is the sampling distance. Information



+∞ may be lost by sampling. However, under certain condi-
tions, a continuous signal can be completely recovered
S(kx , ky ) = s(x, y) e−2πi(kx x+ky y) dx dy (A.71)
from its samples. These conditions are specified by the
−∞ sampling theorem, which is also known as the Nyquist
criterion. If the Fourier transform of a given signal is
can be rewritten as
band limited and if the sampling frequency is larger
than twice the maximum spatial frequency present in
S(kx , ky )
the signal, then the samples uniquely define the given

2π
+∞ signal. Hence,
= s(r, θ ) e−2πi(kx r cos θ +ky r sin θ) r dr dθ
0 0 

π
+∞ S(k) = 0 ∀ |k| > kmax and
= s(r, θ ) e−2πi(kx r cos θ+ky r sin θ) |r| dr dθ . if
0 −∞  1 > 2kmax
(A.72) x
then ss (x) = s(nx) uniquely defines s(x).
The factor r in the integrand is the Jacobian of the (A.78)
transformation:
  To prove this theorem, sampling is defined as a
 ∂x ∂x  
  multiplication with an impulse train (see Figure A.4):
  ∂r ∂θ  cos θ −r sin θ 
J =  =
 sin θ
 ∂y ∂y . r cos θ 
  ss (x) = s(x) · (x), (A.79)
∂r ∂θ
= r (cos2 θ + sin2 θ ) = r. (A.73) where (x) is the comb function or impulse train:

The polar form of the inverse FT is obtained analo- +∞


gously. Let (x) = δ(x − nx). (A.80)
n=−∞
kx = k cos φ
(A.74) The sampling distance x is the distance between any
ky = k sin φ,
two consecutive Dirac impulses. Note that this for-
mula is a formal notation because the product is only
then
valid as an integrand.
Based on Eq. (A.80) and using the convolution the-
s(x, y)



orem, the Fourier transform Ss (k) can be written as
π +∞
follows:
= S(k, φ) e+2πi(xk cos φ+yk sin φ)
|k| dk dφ.
0 −∞
228 (A.75) Ss (k) = S(k) ∗ F{(x)}. (A.81)

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

s(x) |s(k)| Figure A.4 A signal with an infinite spatial

extent (a) and its band-limited Fourier
transform (b). The sampled signal (e) is
obtained by multiplying (a) with the
impulse train (c). The spectrum (f) of the
sampled signal is found by convolving the
original spectrum (b) with the Fourier
transform of the impulse train (d). This
results in a periodic repetition of the
original spectrum.

x kmax k
(a) (b)

III (x) { III(x)}

∆x

x
K= 1 k
∆x
(c) (d)

ss(x) |ss(k)|

x kmax K K k
2
(e) (f)

It can be shown that Hence,

+∞

F{(x)} = K δ(k − lK ), (A.82) Ss (k) = K (S(k) + S(k − K ) + S(k + K )
l=−∞ + S(k − 2K ) + S(k + 2K ) + · · · ). (A.84)
which is again an impulse train with consecutive
impulses separated by the sampling frequency Because

1 K
K = . (A.83) S(k) = 0 ∀ |k| ≥ 229
x 2

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

s(x) |s(k)| Figure A.5 A signal with a finite spatial

extent (a) is not band limited (b). The sampled
signal (e) is obtained by multiplying (a) with
the impulse train (c). The spectrum (f) of the
sampled signal is found by convolving the
original spectrum (b) with the Fourier
transform of the impulse train (d). This results
in a periodic repetition of the original
spectrum. Because of the overlap, aliasing
cannot be avoided.

x k
(a) (b)

III (x) { III(x)}

∆x

x
K= 1 k
∆x
(c) (d)

ss(x) |ss(k)|

1
2∆x

x k
(e) (f)

it follows that example, note that a patient always has a limited spa-
  tial extent, which implies that the FT of an image of the
x body is never band limited and, consequently, aliasing
K S(k) = Ss (k)  , (A.85)
K is unavoidable. Several practical examples of aliasing
are given in this textbook.
and consequently s(x) can be recovered from Ss (k). Numerical methods calculate the Fourier trans-
If the signal s(x) is not band limited or if it is form for a limited number of discrete points in the
band limited but 1/x ≤ 2kmax , the shifted replicas of frequency band (−kN , +kN ). This means that not only
S(k) in Eq. (A.84) will overlap (see Figure A.5). In that the signal but also its Fourier transform is sampled.
case, the spectrum of s(x) cannot be recovered by mul- Sampling the Fourier data implies that it yields shifted
tiplication with a rectangular pulse. This phenomenon replicas in the signal s, which may overlap. To avoid
230 is known as aliasing and is unavoidable if the origi- such overlap or aliasing of the signal, the sampling
nal signal s(x) is not band limited. As an important distance k must also be chosen small enough. It can

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016
 Appendix A: Linear system theory

easily be shown that this condition can be satisfied if In both cases, m, p = 0, 1, . . . , M − 1 and n, q =
the number of samples in the Fourier domain is at least 0, 1, . . . , N − 1. Here, M and N need not be equal
equal to the number of samples in the signal domain. because both directions can be sampled differently.
In practice they are chosen equal. However, for a particular direction, the number of
Based on the preceding considerations, the discrete samples in the spatial and the Fourier domain is
Fourier transform (DFT) for 2D signals can be written the same.
as (more details can be found in [38]): Direct computation of the DFT is a time-
consuming process. However, when the number of
−1 M −1   samples is a power of two, a computationally very fast

N  −2πi
mp nq
M +N
S(mkx , nky ) = s(px, qy) e . algorithm can be employed: the fast Fourier transform
q=0 p=0 or FFT. The FFT algorithm has become very important

N −1 M
 −1 
mp nq
 in signal and image processing, and hardware versions
2πi M +N
s(px, qy) = S(mkx , nky ) e . are frequently used in today’s medical equipment. The
n=0 m=0 properties and applications of the FFT are the subject
(A.86) of [38].

231

Cambridge Books Online © Cambridge University Press, 2009

Downloaded from Cambridge Books Online by IP 85.122.26.89 on Wed Mar 16 13:15:09 GMT 2016.
http://dx.doi.org/10.1017/CBO9780511596803.010
Cambridge Books Online © Cambridge University Press, 2016