10/20/2022

‫ميحرلا نمحرلا هللا مسب هتاكربو هللا‬

‫ةمحرو مكيلع ملاسلا‬ Daftar Isi

• Introduction
Medical Imaging • Contrast
Basic Imaging Principles: Image
• Resolution
Quality
• Noise
• Signal-to-Noise Ratio
Amir Faisal, S.T., M.Eng., Ph.D. • Sampling
Teknik Biomedik • Other Effects
Jurusan Teknologi Produksi & • Accuracy
Industri
Institut Teknologi
Sumatera

Introduction
• The primary purpose of a medical imaging system is to create
images of the internal structures and functions of the human • Image quality depends on the particular imaging modality
body that can be used by medical professionals to diagnose used. With each modality, the range of image quality may be
abnormal conditions, determine the underlying mechanisms considerable, depending on the characteristics and setup of
that produce and control these conditions, guide therapeutic the particular medical imaging system, the skill of the operator
procedures, and monitor the effectiveness of treatment. handling the system, and several other factors, such as patient
• The ability of medical professionals to successfully characteristics and imaging time.
accomplish these tasks strongly depends on the quality of the
images acquired by the medical imaging system at hand, • Studying how these factors affect image quality is an
where by important and complicated task. This task is simplified by
‘‘quality’’ we mean the degree to which an image allows focusing on the following six important factors: (1) contrast, (2)
medical resolution, (3) noise, (4) artifacts, (5) distortion, and (6)
professionals to accomplish their goals. accuracy.

• The ability of medical professionals to discriminate among • In the overview to Part I, we presented images from
anatomical or functional features in a given image different modalities that have different contrast.
strongly depends on contrast.
• For example, the brain structures in the positron emission
• Contrast quantifies the difference between image tomography (PET) image in Figure I.4(c) are of higher
characteristics (e.g., intensities in shades of gray or color) of contrast than the same structures in Figure I.4(a) or (b),
an object (or feature within an object) and surrounding objects because the actual signal coming from the patient is of
or background. intrinsically higher contrast in PET.
• High contrast allows easier identification of individual objects
in an image, whereas low contrast makes this task difficult.

1
 10/20/2022

• Representative transverse • Sometimes, medical images are blurry and lack detail.
slice through the brain from
three different imaging • The ability of a medical imaging system to depict details
modalities: is known as resolution.
• (a) computed tomography, • High resolution systems create images of high
diagnostic quality.
• (b) magnetic resonance
imaging, and • Low resolution systems create images that lack fine detail.
• (c) positron emission • For example, while the PET image in Figure I.4(c) has higher
tomography. contrast than the CT image in Figure I.4(a) or the MRI image
in Figure I.4(b), it has poorer spatial resolution.

• A medical image may be corrupted by random fluctuations • Most medical imaging systems can create image features
in image intensity that do not contribute to image quality. that do not represent a valid object within, or characteristic of,
• This is known as noise. The source, amount, and type of the patient.
noise depends on the particular imaging modality used. • These features are known as artifacts, and can
• Object visibility is often reduced by the presence of frequently obscure important features or be falsely
noise, because the noise masks image features. interpreted as abnormal findings.
• Nuclear medicine images tend to have the highest noise, • Medical images should not only make desired features
as reflected in the PET image in Figure I.4(c). visible but should also give an accurate impression of their
shape, size, position, and other geometric characteristics.

• Unfortunately, medical imaging systems frequently • Fundamentally, we are interested in the accuracy of medical
introduce distortion of these important factors. images in the context of a clinical application, where
• Distortion in medical images should be corrected in order ‘‘accuracy’’ means both conformity to truth and clinical utility.
to improve the diagnostic quality of these images. • The user of a medical imaging system is very interested in
• Ultimately, the quality of medical images should be judged adjusting the system to produce images of the highest
on their utility in the context of a specific clinical application. possible quality, while maintaining a safe environment for the
• For example, medical images that increase the chance of patient.
tumor detectability in nuclear medicine should be preferable to • In order to achieve this, methods must be developed
images with poor tumor detectability. for evaluating image quality.

2
 10/20/2022

Contrast
• Since image quality depends primarily on the six previously • Contrast refers to differences between the image intensity of
discussed factors, we need to mathematically quantify these an object and surrounding objects or background.
factors and systematically study their influence on image • This difference, or image contrast, is itself the result of
quality. the inherent object contrast within the patient.
• The purpose of this chapter is to provide a fundamental • In general, the goal of a medical imaging system is to
exposition on contrast, resolution, noise, artifacts, accurate- ly portray or preserve the true object contrast in the
distortion, and accuracy. image.
• More detailed discussions on these factors, and how they • Particularly for detection of abnormalities, a medical
affect image quality for a specific imaging modality, will be imaging system that produces high contrast images is
preferable to a system that produces low contrast images,
provided later. since anatomical and functional features are easier to
identify in high contrast images.

Modulation
• Use of a periodic signal and its modulation is an effective way • In general, �� refers to the contrast of the periodic signal
to quantify contrast. � �, � relative to its average value.
• The modulation �� of a periodic signal � �, � , with • We assume here that � �, � has nonnegative values, in
maximum and minimum values �max and �min, is defined which case 0 ≤ �� ≤ 1. Of importance, �� = 1 only when
by �min = 0.
�m ax −�m in • Thus, in practice, the usual presence of a nonzero
• �� =
max +�mi ‘‘background’’ intensity in a medical image reduces image
�
n
• Modulation quantifies the relative amount by which the contrast. If �� = 0 (in which case, �min = �max ), we say that
amplitude (or difference) �max − �min Τ2 of � �, � � �, � has no contrast.
stands out from the average value (or background) �max +
• If � �, � and � �, � are two periodic signals with
�min Τ2. the same average value, we say that � �, � has more
contrast than
� �, � if �� > �� .

Modulation Transfer Function

• The way a medical imaging system affects contrast can be
investigated by imaging a sinusoidal object � �, � of the • Figure 3.1 depicts four instances of � �, � for the cases when
form �� = 0, 0.2, 0.5, 1.
• � �, � = � + � sin 2���0 � • Notice that as modulation increases, it becomes much easier
• where � and � are two nonnegative constants such that A ≥ to distinguish differences in shades of gray in the image of �
�. �, � ; in other words, contrast increases.
• This is a sinusoidal object that varies only in the x direction • We are now interested in determining how an linear shift-
with spatial frequency �0 . invariant (LSI) imaging system with point spread function
• Notice that �max = � + � and �min = � − �, so the (PSF) ℎ �, � affects the modulation of � �, � ; that is, we
modulation of are interested in mathematically relating the modulation mg of
� �, � is given by the output � �, � to the modulation �� of the input � �,
�
• �� = � � .

3
 10/20/2022

• In order to simplify our discussion, we assume that ℎ �, � • From (3.5), �max = �� 0,0 + � � �0 , 0 and �min = �� 0,0 −
is circularly symmetric [see (2.104)]. Since � � �0 , 0 , in which case the modulation of � �, � is given by
� �2�� −�2�� � � �0 � �0 , 0
�0 �
• � �, � = � + � sin 2���0 � = � + ��0� − • �� =, 0 = ��
2� �� 0,0 � 0,0
� • The modulation �� of � �, � depends on the spatial frequency
• the output � �, � of the system is given by �0 .
• � �, � = �� 0,0 + � � �0 , 0 sin 2���0 � • The way an LSI medical imaging system affects
modulation, and therefore contrast, is illustrated in Figure
• Notice that the recorded image � �, � of the sinusoidal 3.2.
object
� �, � is also sinusoidal with (the same) frequency �0 .

• The output modulation �� is a scaled version of the input

modulation �� , the scaling factor being the magnitude
spectrum
� �0 , 0 of the medical imaging system under consideration.
• If � 0,0 = 1 and � �0 , 0 < 1, then �� < �� , and since
both signals � �, � and � �, � have the same average
value, the output � �, � will have less contrast than the
input � �, � .
• The ratio of the output modulation to the input modulation as
a function of spatial frequency is called the modulation
transfer function (MTF), and [by rearranging (3.6)] is given by
• MTF � �� � �,0
= � 0,0
Basic principles for determining the modulation of the output of an LSI medical
=
� imaging system from the modulation of the input, when the input object is
�
sinusoidal.

• This shows that the MTF of a medical imaging system is, in • The MTF quantifies degradation of contrast as a function
essence, the ‘‘frequency response’’ of the system, and it can of spatial frequency.
be directly obtained from the Fourier transform of the PSF of
the system (remember that � �, 0 = ℱ2� ℎ �, � . • For most medical imaging systems,
• Since � �, 0 = � −�, 0 , the MTF is usually considered • 0 ≤ MTF � ≤ MTF 0 = 1, for every �
only at nonnegative frequencies. • with the MTF becoming significantly less than unity (or
• Since the MTF, which characterizes contrast, can be even zero) at high spatial frequencies.
mathematically related to the PSF, which characterizes
blurring (or resolution), we can assume that blurring reduces
contrast.

4
 10/20/2022

Example 3.1
• A typical MTF is depicted in
Figure 3.3. Note that 0 • At spatial frequency 0.6 mm−1, the MTF takes a value of
≤ MTF � ≤ 1, with the 0.5.
maximum value attained at
� • This means that the contrast of a sinusoidal object at spatial
= 0.
frequency 0.6 mm−1 is reduced by half when imaged
• Note also that the MTF monoto- through this system.
nically decreases to zero with
increasing frequency, becoming • Moreover, since the MTF is zero at all spatial frequencies
zero at spatial frequencies larger than 0.8 mm−1, any sinusoidal input with frequency
larger than 0.8 mm−1. larger than
• What can we learn about the 0.8 mm−1 will be imaged as a constant output of zero
contrast behavior of an imaging contrast.
system with this MTF? A typical MTF of a medical imaging system.

• As indicated above, we can think of loss of contrast as

the result of the blurring action of a medical imaging
system.
• This is illustrated in Figure 3.4, which depicts the outputs of
three radiographic imaging systems with increasingly
poorer MTFs. Impact of the MTF of
• Here, a poorer MTF is one that drops to zero at lower a medical imaging
spatial frequencies. A poorer MTF results in less contrast. system on contrast.
Clearly, a poorer
• It should be noted here that the PSF of a medical imaging MTF results in lower
system need not be isotropic—that is, equivalent in all (2-D contrast.
or
3-D) directions.

Local Contrast

• In a non-isotropic system, the profile through the PSF changes • The identification of some specific object or feature within
with orientation; thus, the system has an orientation- an image is only possible if its value differs from that of
dependent response. surrounding areas.
• In the non-isotropic (2-D) case, the MTF is • The definition of modulation or contrast for sinusoidal
rotationally dependent, and (3.7) can be generalized signals can be adapted for use in this situation as well.
to
• MTF � �� � �, � �, • It is common in many imaging modalities (e.g., nuclear
� in which case �� = �
= medicine) to refer to an object of interest (e.g., a tumor in
= � 0,0 �� � 0,0
� � the � 0,0
• For a typical non-isotropic medical imaging system,
� �,�
• 0 ≤ MTF �, � = ≤ MTF 0,0 = 1, for
every �, �

5
 10/20/2022
liver) as the target, as illustrated in Figure 3.5.
• Suppose that the target has a nominal image intensity
of �� .

6
 10/20/2022

• Suppose that the target has a

nominal image intensity of • The difference between the target and its background
�� . is captured by the local contrast, defined as
�𝑡 − ��
• The target (i.e., the tumor) is
surrounded by other tissues •�= ��
(i.e., the liver tissue), called
the background, which may • The definition of local contrast in (3.12) differs from the
obscure our ability to see or definition of modulation in (3.1) for sinusoidal signals in that
detect the target. the intensities �� and �� may be selected locally—for
• Suppose that the background example, within the liver—and they need not be the maximum
has a nominal image intensity and minimum intensities within the image as a whole.
of �� . Local contrast
scenario.

Example 3.2
• For example, �� could be taken to be the average image • Consider an image showing an organ with intensity ��
intensity within the tumor, whereas �� could be taken as and a tumor with intensity �� > �� .
the average image intensity within the liver.
• What is the local contrast of the tumor? If we add a constant
• The tumor intensity could be larger than �� or less than intensity �� > � to the image, what is the local contrast?
�� . Is the local contrast improved?
� −�
• If �� is less than �� , then � is negative and, in • By definition, the local contrast of the tumor is � = � 𝑡 𝑜
certain circumstances, it may be convenient to report its 𝑜
absolute value. • If we add a constant intensity �� to the image, the
intensities of the background and the target become �� =
�� + �� , and �� =
�� + �� .

Resolution
• The local contrast of the processed image is • Resolution is another basic measure of image quality.
�𝑡 + �� − �𝑜 + �� �𝑡 −�𝑜
•�� ′ = = =� • For our purposes, resolution can be thought of as the ability of a
𝑜
��� +�� �� �� medical imaging system to accurately depict two distinct events
� +� � +�
� �
• So, the local contrast is worse if we add a constant intensity in space, time, or frequency as separate.
�� • In this case, we talk about spatial, temporal, or
to the image. spectral resolution, respectively.
• Resolution can also be thought of as the degree of smearing,
or blurring, a medical imaging system introduces to a single
event in space, time, or frequency.

7
 10/20/2022

Line Spread Function

• These two ways of looking at resolution are related, because • As suggested above, we can think of resolution as the degree
the less smearing a system introduces, the closer in space, of smearing, or blurring, a medical imaging system introduces
time, or frequency two events can be and still be to a single event (e.g., a point) in space.
distinguished as separate.
• This is the traditional PSF, and the response of an imaging
• Therefore, a high resolution medical imaging system is system to a point source (i.e., a point impulse) is often used
characterized by low smearing, whereas a low to characterize resolution.
resolution system is characterized by high smearing.
• As an alternative, we can consider the response of a
• We mostly focus on spatial resolution, although a brief medical imaging system to a line impulse
discussion of temporal and spectral resolution is also
provided.

• Consider an LSI medical imaging system with isotropic • Then, the output � �, � of the system will be given by
PSF
ℎ �, � that is normalized to • � �,=�∞ ∞
ℎ 𝜂��, 𝜂 �‫ � ׬‬− ��, � − 𝜂∞ ∞
1. ‫�𝜉 ׬‬
‫� ׬‬ = ‫׬‬ൣ ℎ ��, 𝜂 �� ሺ� −
−∞ −∞ −∞ −∞
• Suppose that a line source, passing through the origin of the ∞
𝜉 ሻ ��� ൧ �𝜂ℎ �, 𝜂 �𝜂
spatial domain, is imaged through the system. = ‫׬‬−∞

• We will mathematically represent this line source by the • where, in the third equality, we have used the 1-D analog of (2.6).
line impulse � �, � = ��� �, � [see (2.11)]. • The resulting image � �, � is only a function of �, say � � .
• Since the system is isotropic, it is sufficient to consider the • This is known as the line spread function (LSF) of the system
response to a vertical line through the origin; from (2.11), under consideration, and it can be used to quantify resolution.
we see that for this case, � �, � = 𝛿 � . • This function is directly related
∞
to the PSF ℎ �, � since, from
(3.13), we have � −∞
� = ‫ ׬‬ℎ �, 𝜂 �𝜂

• Notice that, since the PSF ℎ �, � is assumed to be • The values of the transfer function along the horizontal line that
isotropic, passes through the origin of the frequency domain are adequate for
determining the 1-D Fourier transform of the LSF, and the LSF
� � is symmetric [i.e., � � = � −� ], and since itself.
∞
the PSF is normalized
−∞ to 1, 1 = �� � � ‫׬‬. • Since the PSF ℎ �, � is assumed to be isotropic, the
transfer function is isotropic as well.
• Moreover, the 1-D Fourier transform � � of the LSF � • Thus, the LSF is adequate for determining the PSF of the system.
� is related to the transfer function � �, � of the
system, since
∞
• � � =1ℱ � � −∞ = ‫ ׬‬−�2� �� • Indeed, from the LSF � � , we can calculate the 1-D
� Fourier transform � � , and � �, 0 = � � .
���
� � � =
∞ ∞
‫∞ ׬‬−‫׬‬−∞ ℎ �, 𝜂 ���𝜂
−�2����
�, 0 =�
� • However, since the transfer function � �, � is isotropic, the values of
� �, � along any line that passes through the origin of the
frequency domain will be the same as the values of � �, � .

7
 10/20/2022

Full Width at Half Maximum

• Given the LSF (or the PSF) of a medical imaging system, its
resolution can be quantified using a measure called the full width • The points move closer to each other as we go from (a) to (d). In a
at half maximum (FWHM). linear system, the observed profile is the sum of the individual
profiles from the two points.
• This is the (full) width of the LSF (or the PSF) at one-half
its maximum value. • In practice, a residual ‘‘dip’’ between the points must occur in the
(summed) profile for the two points to be visualized or resolved
• The FWHM is usually expressed in millimeters. as separate.
• Provided there is no geometric scaling, the FWHM equals the • The profile depicted in Figure 3.6(c) shows the separation
minimum distance that two lines (or points) must be separated distance at which the two points are just distinguishable.
in space in order to appear as separate in the recorded image. • Notice that, in this case, the two points are separated by the
• This is depicted in Figure 3.6, where we see the profiles of FWHM. Therefore, a decrease in the FWHM indicates an
two points through a 1-D imaging system with PSF ℎ � . improvement in resolution.

Resolution and Modulation Transfer Function

• Another way to quantify the resolution of a medical imaging
system is as the smallest separation (in mm) between two
adjacent maxima (or minima) in a sinusoidal input that can
be resolved in the image.
• Consider the case when the input of a medical imaging system
is sinusoidal of the form � �, � = � sin 2𝜋�� , with
amplitude � and frequency �.
• From (3.5) and (3.7), the output of the system is given by
• � �, � = MTF � � 0,0 � sin 2𝜋��
An example of the effect of system resolution on the ability to differentiate two points.
The FWHM equals the minimum distance that the two points must be separated in order to
be distinguishable.

Example 3.3
• Notice that the separation between two adjacent maxima (or minima) • The MTF depicted in Figure 3.3
of the sinusoidal input � �, � is 1Τ�. becomes zero at spatial
• The recorded image � �, � is also sinusoidal, with 1Τ� frequencies larger than 0.8 mm−1.
being the separation between two adjacent maxima (or minima) • What is the resolution of this
as well. system?
• Answer: The resolution of a system
• However, the amplitude of the output image equals the amplitude with such an MTF is 1/(0.8 mm−1) =
of the input multiplied by the MTF at spatial frequency �. 1.25 mm.
• In practice, MTF � ≠ 0, for every � ≤ �� , and MTF � = 0, for • Fine structures of an object to be
every imaged by such a system with
spatial frequencies larger than 0.8
� > �� , for some spatial cut-off frequency �� , in which case, � mm−1 cannot be seen at the output
�, �
= 0, for every � > �� .
• In this case, the resolution of the system will be 1Τ�� . of the system.

8
 10/20/2022

• From our discussion, we find that the MTF can be • If the MTF curves are of different shapes, the situation is
used to compare two competing medical imaging systems
effectively complicated.
more
in terms of their contrast and resolution. • Figure 3.7 depicts two MTF curves that correspond to two
• If the MTFs of the two systems under consideration are of a com- peting medical imaging systems: SYSTEM 1 and
similar shape but have a different cut-off frequency �� , we SYSTEM 2.
can conclude that the system with higher MTF values will be • SYSTEM 1 has better low frequency contrast, and is thus
better in terms of contrast and resolution. better for imaging coarse details, while SYSTEM 2 has a better
• For example, the radiographic imaging system depicted in high frequency contrast, and is thus better for imaging fine
the first panel in Figure 3.4 is better in terms of contrast and details.
resolution than the system depicted in the third panel. • Since contrast, as quantified by the MTF, is a function of
spatial frequency, we can make this sort of frequency-by-
frequency comparison.

• Spatial resolution, as described here, is not frequency-

dependent, so it becomes harder to directly compare MTFs • Therefore, the MTF equals • MTF curves of two competing
in the context of ‘‘better resolution.’’ the (normalized) magnitude of medical imaging systems.
• As described above, the FWHM of the PSF or LSF is the the 1-D Fourier transform of
most direct metric of resolution. the LSF.
• Further understanding of a system’s resolution from its MTF • The next example shows that
usually comes from MTF values at higher spatial this relationship can be used
frequencies and the cut-off frequency �� . to determine the FWHM of a
• The MTF can be directly obtained� �
from the LSF. Indeed, from medical imaging system
(3.7) and (3.16), MTF � =� 0 , for every �. directly from its MTF.

Example 3.4
Subsystem Cascade

• The PSF, LSF, or MTF can be described by a mathematical 0 0

0

function by either fitting observed data or by making

simplifying assumptions about its shape. Assume that the MTF • This results in FWHM = 2 ln 2Τ��.
2
of a medical imaging system is given by MTF � = � −���
• What is the FWHM of this system?
• By using the 1-D inverse Fourier transform and 2the fact that
� � = MTF � , we have � � = � −��� . The FWHM
will then be given by FWHM = 2� , where � is such that
2 1 2
� −��� = .

9
 10/20/2022

• Medical
• Fromimaging systemswe
our discussion, arefind
often
thatmodeled
the MTFascan
a cascade
be of • If the MTF curves are of different shapes, the situation is
LSI subsystems, as introduced in Section 2.3.4.
effectively more
• Accordingly, the recorded image � � can
be modeled as the convolution of the input
object � � with the PSF of the first
subsystem, followed by the convolution with
the PSF of the second subsystem, etc. (see
Section 2.3.4).
• For example, in the case of � subsystems with PSFs
ℎ1 �, � , ℎ2 �, � , … , ℎ� �, � ,
• � �, � = ℎ� �, � ∗ ⋯ ℎ2 �, � ∗ ℎ1 �, � ∗ � �

9
 10/20/2022

• The quality of the overall system, in terms of contrast and • Notice that the overall FWHM R is dominated by the largest
resolution, can be predicted by considering the quality of (i.e., the poorest resolution) term.
each subsystem.
• Thus, small improvements in any given subsystem’s
• If resolution is quantified using the FWHM, then the FWHM resolution do not often yield noticeable improvements in
of the overall system can be determined approximately from overall system resolution.
the FWHMs �1 , �2 , … , �� of the individual subsystems, by
• The following example shows that, when a medical imaging
•�= � 2
1 + �2 + ⋯ ��
2 2
system is composed of subsystems with Gaussian PSFs,
(3.22) gives the exact value of the FWHM of the overall
system.

Example 3.5

• Consider a 1-D medical imaging system with PSF ℎ • ℎ2 � 1

exp −�1
2
= 1
� =
2
composed of two subsystems with Gaussian PSFs of the form 2�� 2� 2 2����2
��2 �2

1 2 1 • After algebraic manipulation of (3.24) and (3.25),

−� 2
• ℎ1 � = exp 2 and ℎ2 �−�= exp 2
2��𝜎
1 1 2�� 2 2��𝜎2 2�� • �1 = 2��1 2 ln 2 and �2 = 2��2 2 ln 2
• What is the FWHM of this system? • Following Example 2.4, it can be shown that the PSF ℎ �
• Answer: The FWHMs �1 and �2 associated with the of the overall system is given by
two
subsystems ℎ1 and ℎ2 , respectively, are given by �1 = 2�1 1 2
1
and • ℎ � = ℎ1 � ∗ ℎ2 � −�
2
�2 = 2�2 , where �1 and �2 are such that = exp ��22+�
�
2 2 1 2
1 2𝜋 ��1 +��2
− 1
� 21
• ℎ1 � = exp 2 = and
2��𝜎
1 1 2�� 12 2��𝜎

• The FWHM � of this system is given by � = 2�0 , where • If contrast and resolution are quantified using the MTF, then the
�0 is such that MTF of the overall system will be given by
2
−0
�
•ℎ� 0 = exp 2 2=
• MTF �, � = MTF1 �, � MTF1 �, � ⋯ MTF� �, �
2 �� +��
0.5 1 2 • in terms of the MTFs MTF� �, � , � = 1, 2, … , � of the individual
• From which we obtain � = 2 ��
2 2 2 ln 2 subsystems.
1
+ ��2 • This is a direct consequence of (3.9) and the fact that the
frequency response � �, � of the overall system is given
by
• � �, � = �1 �, � �2 �, � … �� �, � ,
• which is the product of the frequency responses �� �, � , �
= 1, 2, … , �, of the individual subsystems.

10
 10/20/2022

• Figure 3.8 depicts the MTF curves of three subsystems of • MTF curves of three
a medical imaging system together with the MTF curve of subsystems of a medical
the overall system. imaging system and the MTF
• If one subsystem has a small value of the MTF at some curve of the overall system.
spatial frequency, then the MTF of the overall system will be
small at that frequency as well.
• In other words, the MTF of the overall system will always
be less than the MTF of each subsystem.

• This follows from (3.30), which suggests that, for every � = • From the previous discussion, the resolution of a medical
1, 2, … , �, imaging system can be specified by either its PSF or its
• MTF �, � ≤ MTF� �, � , for every �, �, LSF.
• provided that MTF� �, � ≤ 1, for every � = 1, 2, … , �. • In some medical imaging systems, we are able to create
an object consisting of a very small ‘‘point’’ or ‘‘line’’ of
• Therefore, the overall quality of a medical imaging system, in some material.
terms of contrast and resolution, will be inferior to the quality • For example, we may create a very small and highly
of each subsystem. radioactive point or line source for measuring the resolution of
a nuclear medicine camera.
• When a medical imaging system is mathematically modeled
using an LSI system, its output to these objects will be the
PSF and the LSF, respectively.

• This provides a practical way for calculating these two • In a non-isotropic system, the profile through the PSF changes
important response functions. with orientation; thus, the system has an orientation-
• From our discussion, we can see that spatial resolution and dependent resolution.
image contrast are tightly linked, since the Fourier transforms • A good example of this situation is in ultrasound imaging
of the PSF and the LSF, which are measures of resolution, systems where the range resolution (along the transducer
yield the MTF, which is a measure of contrast. axis) is usually substantially better than the lateral resolution
• Indeed, spatial resolution can be thought of as the ability of an (orthogonal to the transducer axis).
imaging system to preserve object contrast in the image, • It is also possible for a medical imaging system to be linear
since blurring, due to poor resolution, is what actually reduces but not shift-invariant.
contrast.

11
 10/20/2022

Resolution
Tool
• Then, the resolution is spatially dependent. • Images with better resolution are preferable to images
• For example, this situation exists in ultrasound systems with poorer resolution, since they contain more image
because the acoustic energy spreads out with details.
increasing distance from the transducer. • Resolution can be quantified in terms of the ability of a
• This makes the FWHM increase with depth, which system to image details of a given test pattern.
corresponds to a degradation of lateral resolution. • For example, one common way to measure resolution for a
• Magnification in pinhole collimators, or focused collimators in particular system is to image the so-called resolution tool or
nuclear medicine, also leads to spatially dependent resolution, bar phantom, such as the one depicted in Figure 3.9.
in which case the image plane resolution is related, in a • This tool is composed of groups of parallel lines of a
precise geometric way, to actual resolution within the object. certain width, separated by gaps having the same width as
the line width (yielding an overall duty cycle of 50 percent).

• A resolution tool or • Each group is characterized by the density of such

bar phantom. lines, measured in line pairs per millimeter (lp/mm).
• The tool is imaged through the system under consideration,
and system resolution is reported as the frequency (in lp/mm)
of the finest line group that can be resolved at the output.
• For example, the resolution might be 6–8 lp/mm for a
projection radiography system and 2 lp/mm for a CT scanner.

Temporal and Spectral Resolution

• The preceding discussion on spatial resolution and the • Conceptually, we could create a frequency distribution histo-
concepts embodied in the PSF, LSF, and FWHM apply gram of the number of observed events as a function of time or
equally well to temporal and spectral resolution. energy from a single-time or single-energy process (i.e., a
• Temporal resolution is the ability to distinguish two events point impulse in time or energy) to yield the equivalent of a
in time as being separate. PSF.
• Spectral resolution is the ability to distinguish two • As with the PSF, the ideal system response would be a
different frequencies (or, equivalently, energies). delta function and the actual FWHM (in time or energy)
would quantify the resolution.
• In other words, the concept of PSF applies equally well
to events in time or frequency as it does to events in
space.

12
 10/20/2022

Noise
• An unwanted characteristic of medical imaging systems is • The discrete nature of their arrival leads to random
noise. fluctuations, called quantum mottle, which give an x-ray image
• Noise is a generic term that refers to any type of random a textured or grainy appearance.
fluctuation in an image, and it can have a dramatic impact • In MRI, radio frequency pulses generated by nuclear spin
on image quality; image quality decreases as noise systems are sensed by antennas connected to amplifiers.
increases. • Since these signals have very low power, they compete
• The source and amount of noise depend on the imaging me- with signals being generated in the antenna from natural
thermal vibrations.
thod used and the particular medical imaging system at
hand. • Thermal vibrations are unpredictable—that is, random—and
therefore comprise one source of noise in magnetic
• For example, in projection radiography, x-rays arrive at the resonance images. The effect of increasing noise is shown in
detector in discrete packets of energy, called quanta or Figure 3.10.
photons.

• The source of noise in a medical imaging system depends

on the physics and instrumentation of the particular modality.
• The main objective of this section is to provide an
introductory exposition of the tools used to mathematically
characterize noise.
• A general way to characterize noise is to consider it as
the numerical outcome of a random event or experiment.
• In nuclear medicine, certain radioactive sources emit
The effect of noise on image quality: image quality decreases gamma ray photons that are received and recorded by a
rapidly with increasing noise contamination. detector.

Random
Variables
• Although governed by fixed physical properties (such as • The numerical quantity associated with a random event
photon energy and decay rate), the specific nature of or experiment is called a random variable.
radioactive decay is random: photons are emitted at random
times in random directions. • Different repetitions of the experiment may produce
different observed values—that is, the experiment has a
• We think of the noise as the deviation from a nominal value random outcome.
that would be predicted from purely deterministic arguments.
• A random variable is mathematically described by ��𝑁 𝜂
• This deviation, arising from the random nature of radioactive , its probability distribution function (PDF), given by
emissions, accounts for the noise that is present in all
nuclear medicine images. • ��𝑁 𝜂 = Pr 𝑁 ≤ 𝜂
• where Pr[·] denotes probability.

13
 10/20/2022

Continuous Random Variables

• The PDF gives the probability that random variable N will take • If ��𝑁 𝜂 is a continuous function of ��, then 𝑁 is a
on a value less than or equal to η. continuous random variable.
• Notice that 0 ≤ ��𝑁 𝜂 ≤ 1, ��𝑁 −∞ = 0, ��𝑁 ∞ =
• This random variable is uniquely specified by its
1, and probability density function (pdf),
��𝑁 ��1 ≤ ��𝑁 ��2 , for ��1 ≤ ��2 . � 𝑃𝑁 𝜂
• ��𝑁 𝜂 =
�𝜂
• Any pdf satisfies the following three properties:
• ��𝑁 𝜂 ≥ ∞ 0, ‫׬‬−∞ ��𝑁 𝜂 ��� = ∞ 1, ��𝑁 𝜂 = ‫׬‬−∞ ��𝑁 � ��

• In practice, the pdf of a random variable may not be known. • The mean can be thought of as the average value of the
• Instead, a random variable is often characterized by random variable, whereas the standard deviation can
its expected value be thought of as the ‘‘average’’ variation of the values of
∞ the random variable about its mean.
• �𝑁 = � 𝑁 = 𝜂−∞
‫׬‬ 𝑁𝑝 𝜂 �𝜂
• The larger the standard deviation, the ‘‘more random’’ the
• also called its mean, and its variance random variable.
2
• ��
𝑁 = Var 𝑁 = � 𝑁𝑁− � 2 • As the standard deviation approaches zero, the observed
• where E[·] and Var[·] denote expectation and variance, values of the random variable more tightly cluster around the
respectively. The square root ��𝑁 of the variance is called mean and, in the limit, the random variable becomes a
the standard deviation of ��. constant that equals ��� .

Uniform Random ariable Gaussian Random Variable

V
• A random variable N issaid to • and the expected value and
be uniform over the rval variance are given by • If the pdf of a random variable 𝑁 is given by
1 2Τ2��2
inteb] if its pdf is of the form
[a, 0 for 𝜂 < � • 𝑝𝑁 𝜂 = � − ��−�
2� 2
��−� �𝜎
• ��𝑁 𝜂 • 𝑝𝑁 𝜂 = ൞ for � ≤ 𝜂 ≤ �
1 �−
�
• then 𝑁 is a Gaussian random variable.
, for � ≤ 𝜂 � 1 for 𝜂 < � 1
= ቐ �−� •𝑃 𝜂 = + erf 𝜂 −�
< 0, otherwise 2 ,
• and the expected value and 𝑁 𝜎
variance are given by • where erf � denotes the error function given by the integral
2 1 2 �
• �𝑁 =� + �and ��
2
𝑁 =
� −�
• erf � =−� Τ‫׬‬2 �
2
��
2𝜋 0
12

14
 10/20/2022

Example
3.6
• The expected value and variance are given by • Distribution functions can be found by integrating
• �𝑁 = � and𝑁�� 2 = �� 2 their corresponding density function using (3.37).
• Can (3.44) be proved by direct integration?
• The integral in (3.45) cannot be evaluated in closed form, but
it is usually tabulated in mathematics handbooks and can be • The distribution function of a Gaussian random variable
numerically evaluated using most mathematics, statistics, and with mean � and variance ��2 is
engineering software packages. 𝜂 𝜂 1 − ��−� 2 �𝜏
•𝑝 ��𝑁 𝜂 = ‫׬‬ 𝜂 ��� �
2 Τ2𝜎
−∞ 𝑁 =‫׬‬
��−𝜇
−∞ 2����2
1 2Τ 1 2 �− 0 Τ2 1
= 𝜎 � −�
‫�� ׬‬ 2 𝜂 −�
+ erf 𝜎 ��� = � ‫׬‬ =
2����2 −∞
𝜂−� 2𝜋 −∞ 𝜎 2
+ erf 𝜎

Discrete Random Variables

• The mean and variance do not uniquely specify a random
• When the random variable 𝑁 takes only values ��1 , ��2 , …
variable. This means that, given �𝑁 and �� 2 , there , 𝜂� , it is said to be a discrete random variable.
𝑁
might be
more than one pdf that produces the same mean and variance. • This random variable is uniquely specified by the probability
mass function (PMF) Pr 𝑁 ≤ ��� , for � = 1, 2, … , �,
• However, in the case of a Gaussian random variable, the pdf is
uniquely specified by its mean and variance. Usually, noise in where Pr ሾ𝑁
≤ ��� ሿ is the probability that random variable 𝑁 will take on the
medical imaging systems is the result of a summation of a particular value ��� . The PMF satisfies the following three
large number of independent noise sources. properties: 0 ≤ Pr 𝑁 ≤ ��� ≤ 1, for � = 1, 2, … , �, σ�
�=1
�
• According to the central limit theorem of probability, a Pr 𝑁 ≤ 𝜂
�
random variable that is the sum of a large number of = 1, and ��𝑁 𝜂 = Pr 𝑁 ≤ ���all=�� σ �� ≤𝜂 Pr 𝑁 ≤ 𝜂
�
independent causes tends to be Gaussian. Therefore, it is
often natural to model noise in medical imaging system by • It is permissible for � → ∞, meaning that there will be an infinite
means of a Gaussian random variable. (but still countable) number of possible outcomes.

Poisson Random Variable

• In the case of a discrete random variable, the mean value • Let 𝑁 be a discrete random variable that takes values 0, 1, . . . ,
and variance are given by and has the PMF Pr 𝑁 = � =�!
𝑘
−�
� � , for � = 0, 1, … ,
• �𝑁 = � 𝑁 = �=1 σ� � 𝜂 Pr 𝑁 ≤ 𝜂 , and
� • where � > 0 is a real-valued parameter. 𝑁 is said to be a
• ��2
2
𝑁 = Var 𝑁 = � 𝑁
2 = σ�
�=1 ��� Pr 𝑁 ≤
Poisson random variable, and2
it turns out that its mean equals
− �𝑁 − �𝑁 ��� , its variance and = ��𝑁 = �.
�𝑁
• respectively. • Poisson random variables play an important role in medical
• Notice that the integrals in (3.38) and (3.39) have been imaging systems, i.e., in radiographic and nuclear medicine
replaced by sums in (3.50) and (3.51). imaging. They are used to statistically characterize the
distribution of photons counted per unit area by an x-ray
image intensifier, or to characterize the photon counts
produced by a radiotracer in nuclear medicine.

15
 10/20/2022

Example Example
3.7 3.8
• In x-ray imaging, the Poisson random variable is used to • For the Poisson process of Example 3.7, the time that the
model the number of photons that arrive at a detector in time first photon arrives is a random variable, say �. What is the
�, which pdf
is a random variable referred to as a Poisson process and given 𝑝� 𝜏 of random variable �?
the notation 𝑁 � . • Assume that the first photon arrives in the interval � < � < � +
�� 𝑘 ∆�. For very small ∆�, we have Prob � < � < � + ∆� .
• The PMF of 𝑁 � is given by Pr 𝑁 � = � �!= −� � �
For this to happen, it must be true that no photons arrived in
• where � is called the average arrival rate of the x-ray photons. the interval
• What is the probability that there is no photon detected in time 0, � and exactly one photon arrived in the interval �, � + ∆� .
�?
0 • Prob No photon detected in interval � = � −��
• Pr 𝑁 � = 0 =��
0!
� −�� = � −�� �1∆� −�∆
• Prob One photon detected in interval �, � + ∆� 1!= � � =
�∆�� −�∆�

• Using the Taylor series expansion for the exponential function • As ∆� → 0, all approximations get tighter, and so we
2 �3 �
• � � = 1 + �2!+ 3! + +⋯ recognize by dividing both sides of the above equation by
• and the fact that ∆� is very small (and will go to zero in the ∆� that
limit) • 𝑝� � = �� −�� , � ∈ 0, ∞
permits the approximation
• The random variable � corresponding to this pdf is called
• Prob One photon detected in interval �, � + ∆� ≈ �∆� the exponential random variable.
• where second-order terms in ∆� have been dropped.
• Putting all this together yields 𝑝� �
∆� ≈ Prob No photon detected in
interval � ∙
Prob One photon detected in interval �, � + ∆� ≈ � −�∆�
�∆�

Independent Random Variables

• It is usual in imaging experiments to consider more than
one random variable at a time. • Consider the collection of random variables ��1 , ��2 ,
• The theory that is required in order to characterize a collection … , ��� , having the pdf’s ��1 𝜂 , ��2 𝜂 , … , ���
of random variables follows from the theory of a single 𝜂 , respectively.
random variable, with parallel definitions of distribution, • The sum of these random variables S is another
density, and mass functions. In this course, we require only a random variable having another pdf, ��� 𝜂 .
simplified discussion related to sums of independent random
variables. • It is always the case that the mean of S is precisely the sum
• A collection of random variables are independent if of the means of ��1 , ��2 , … , ��� .
knowledge of some of the random variables (i.e., making a • That is, �� = �1 + �2 + ⋯ + ��,
partial observation) tells you nothing, statistically speaking,
about the remaining random variables.

16
 10/20/2022

Example Example
• where
3.7 �1 , … , �� are the means of 3.8
the pdf’s given above. Independence
is not required for this to be true.

16
 10/20/2022

• When the random variables are independent, we can go a • The facts that the variances add and that the pdf of the sum
step further. can be determined easily are powerful results following from
• In this case, the variance of � is the sum of the individual the independence of random variables.
2 2 2 2
variances, ��� = ��1 + ��2 + ⋯ + ��� • We will use these facts to derive approximate noise and signal-
• We emphasize that the variances, not the standard to-noise ratio (SNR) expressions for CT imaging.
deviations, are added.
• Also, in the case of independence, it is possible to
determine the pdf of � by ��� 𝜂 = ��1 𝜂 ∗ ��2 𝜂 ∗
⋯ ∗ ��� 𝜂
• where ∗ is convolution.

Example 3.9
• Consider the sum � of two independent Gaussian random • Using the fact that the convolution of two Gaussian waveforms
variables ��1 and ��2, each having a mean of zero and yields a Gaussian waveform, we conclude that the pdf of �
variance of �� 2 . has a Gaussian form.
• What are the mean, variance, and pdf of the resulting • Furthermore, since the Gaussian pdf is characterized
random variable? completely by its mean and variance, we conclude
that
• The mean of � is zero since �� = �1 + �2 = 0 + 0 = 1 −��2
= 1 −��2
0 and • ��� exp 2 exp
𝜂 =
2� 4�� 4��2
2 2 2 2 2 2
�𝑆 ��2

•��
the2variance 2 2� 𝑆2
= 𝜎 + 𝜎of � is 2�� since ��� = ��1 + = 2𝜎 �𝜎

Signal-to-Noise Ratio
• The output of a medical imaging system is a random variable • A useful way to quantify this is by means of the signal-to-
� (or a collection of random variables), composed of two noise ratio (SNR).
components, � and ��.
• The SNR describes the relative ‘‘strength’’ of signal �
• Component �, which is usually referred to as signal, is the with respect to that of noise ��.
(deterministic or non-random) ‘‘true’’ value of �, whereas 𝑁 is
a random fluctuation or error component due to noise. • Higher SNR values indicate that � is a more accurate
representation of �, whereas lower SNR values indicate that
• The identification of an abnormal condition within the human g is less accurate.
body most often depends on how ‘‘close’’ an observed value
� of �, characteristic to that condition, is to its true value �. • Therefore, higher image quality requires that the output of
a medical imaging system be characterized by high SNR.

17
 10/20/2022

• The ‘‘signal’’ is that it is the modulation or contrast in the

image, whereas ‘‘noise’’ is the unwanted, random fluctuations.
• Blurring reduces contrast and thus SNR; noise also reduces
SNR.
• Figure 3.11 depicts the effects of both blur and noise on
SNR; as one moves from the upper left to the lower right,
SNR decreases.
The effect of increasing
blur and noise on SNR.

Amplitude SNR Example 3.10

• Most frequently, the SNR is expressed as the ratio of • In projection radiography, the number of photons � counted
signal amplitude to noise amplitude: per unit area by an x-ray image intensifier follows a Poisson
� Amplitude distribution as in (3.52).
• SNR = � Amplitude • In this case, we may consider signal � to be the average
𝑁
• We refer to SNR� as the amplitude SNR. photon count per unit area (i.e., the mean of �) and noise 𝑁 to
be the random variation of this count around the mean, whose
• Exact definition of the amplitude SNR depends on specifying amplitude is quantified by the standard deviation of �.
what we mean by ‘‘signal amplitude’’ and ‘‘noise amplitude.’’
• What is the amplitude SNR of such a system?
• Thus, the amplitude SNR is case dependent, and its definition
must be specifically adapted to the particular situation at
hand.

Power SNR
• From (3.52) to (3.54), it follows that the amplitude SNR is
given by (�� = �� 2 = �) • Another way to express the SNR is as the ratio of signal
� power to noise power:
• SNR � =�𝐺 = �
= • SNR =
powe r �
� 𝜎
𝐺
�
p power 𝑁
• This quantity is known as the intrinsic SNR of x-rays.
• We refer to SNRp as the power SNR.
• The greater the average number μ of photons, the larger
the amplitude SNR, and the smaller the relative amplitude • Notice that exact definition of the power SNR depends on
of random fluctuations in G. specifying what we mean by ‘‘signal power’’ and ‘‘noise power.’’
• Therefore, a higher x-ray exposure generally improves the • Thus, as with the amplitude SNR, the power SNR is case-
quality of radiographic images. however, the greater dependent, and its definition must be specifically adapted to
exposure to ionizing radiation means more risk of radiogenic the particular situation at hand.
cancer.

18
 10/20/2022

Example 3.11
• If � �, � is the input to a noisy medical imaging system with
PSF ℎ �, � , then the output at �, � may be thought of as • Usually, it is assumed that there is no correlation between
a random variable � �, � composed of signal ℎ �, � ∗ � noise values and that for every �, �
�, � and noise 𝑁 �, � , with mean �𝑁 �, �2 and variance • �𝑁 �, � = 0 and ��𝑁 �, � = ���� .
��𝑁 �, � .
• This type of noise is known as white noise.
• What is the power SNR of such a system?
• White noise is a crude approximation of reality, but it is
• In this framework, the power SNR at the output of this system conve- nient to use since it leads to mathematically simple
is given by
models.
∞ ∞
• SNR p = ‫׬‬ ‫׬‬
∞− −∞
ℎ �,� ∗� �
2
• In most cases, it is more accurate to consider correlated
,� 𝜎���� noise and assume some mathematical expression for such
2
𝑁

2
• where the noise power is specified by the variance correlation.
���� .

• In this case, if we assume that the noise mean and variance do • From (3.63) and Parseval’s theorem, given by (2.96), we have
not depend on �, � (a common assumption that 2, ���
2 �,� � ‫׬‬ ‫׬‬
∞
‫׬‬
∞
‫׬‬ S NR p � , � NP S � , � ����
characterizes ����
−∞ −∞
so-called wide-sense stationary noise), then it can be • SNRp = −∞ −∞
∞ ∞ = ∞ ∞
shown NPS �,� ‫׬‬ ‫׬‬
that ‫׬‬ ‫���� ׬‬
∞−∞
− −∞ −∞NPS �,� ����
‫ ∞ ׬‬−‫׬‬− ∞ ℎ � , � ∗ � ∞ ∞

�,�
2
����
• where
∞ ∞
, where � �,� � MTF �,�
• SNR p = ‫∞∞׬‬−‫׬‬
����∞−∞
NPS �,�
•=SNRp �, �
�,�
2 2 2 � �, � 2
�2 0,0
=
1 0� 0� NPS NPS �,�
•lim
NPS �, � = � ൤ቚ ‫׬‬ ‫ ׬‬ሾ𝑁 �,�
�0 ,�0 →∞
�, � − • is called the frequency-dependent power SNR.
−�0
4�0 �0 −�0
2
��� ሿ exp −�2𝜋 ��_�� ����ቚ ൨ is known as • The frequency-dependent power SNR quantifies, at a given
the noise power frequency, the relative ‘‘strength’’ of signal to that of noise at
spectrum (NPS). the output of the LSI system under consideration.

Differential
SNR
• From (3.66) and (3.67), SNR p �, � provides a • Consider an object (or target) of interest placed on
relationship between contrast, resolution, noise, and a background.
image quality.
• Let �� and �� be the average image intensities within the
• For a given output noise level (i.e., a fixed NPS) and a given target and background, respectively.
input � �, � , better contrast and resolution properties
(i.e., a larger MTF) result in better image quality (i.e., a • A useful choice for SNR is obtained by taking the ‘‘signal’’ to be
higher output power SNR). the difference in average image intensity values between the
target and the background integrated over the area A of the
target, and by taking the ‘‘noise’’ to be the random fluctuation of
image intensity from its mean over an area A of the
background.

19
 10/20/2022

Example 3.12
• This leads to the differential signal-to-noise ratio • Consider the case of projection radiography.
(SNRdiff), given by
� �𝑡 −�� • We may take �� to be the average photon count per unit area
in the background region around a target, in which case, �� =
�� ,
• SNR diff = where �� is the mean of the underlying Poisson distribution
��
� �
• where ��� � is the standard deviation of image intensity governing the number of background photons counted per
values from their mean over an area A of the background. unit area.
• From (3.12), we have • Notice that, in this case, ��� � = �� �
• SNR diff = ��
���� • What is the average number of background photons
� � counted per unit area, if we want to achieve a desirable
• which relates the differential SNR to contrast. differential SNR?

Decibels
�� ��
• From (3.69), SNRdiff = ��
= � �� • The SNR is sometimes given in decibels (dB).
�
�
• When the SNR is the ratio of amplitudes, such as with
• From (3.70), the differential SNR is approximately proportional to
contrast as well as to the square root of the object area multiplied the amplitude SNR or the differential SNR, then
by radiation exposure (characterized by the average photon count • SNR in dB = 20 × log10 SNR ratio of amplitudes . (3.72)
per unit area � ). To achieve a desirable differential SNR, it is
required S� 2 • When the SNR is the ratio of powers, such as with the power
NR di f f
= ��
that 2
�
SNR, then
�
• This relationship was first suggested by Albert Rose and is known
as the Rose model. To maintain good image quality (i.e., to obtain • SNR in dB = 10 × log10 SNR ratio of amplitudes .
images with high SNR), high radiation dose is required when
viewing small, low-contrast objects.

Sampling
• To electronically sense, store, and process continuous
signals using computers, we must transform them into • More precisely, given a 2-D continuous signal � �, � ,
collections of numbers. rectangular sampling generates a 2-D discrete signal �� �,
• This transformation, called discretization or sampling, means � , such that
that we only retain representative signal values and discard • �� �, � = � �∆�, �∆� , for �, � = 0,1, …
the rest.
• In (3.74), ∆� and ∆� are the sampling periods in the x
• We focus our attention on the so-called rectangular and y directions, respectively. This is illustrated in Figure
sampling scheme.
3.12.
• According to this scheme, a 2-D continuous signal is
replaced by a discrete signal whose values are the values of • Notice that �� �, � forms an array of numbers that contains
the continuous signal at the vertices of a 2-D rectangular grid. the values of the 2-D continuous signal � �, � at the
discrete points
�∆�, �∆� .

20
 10/20/2022

• The inverses 1Τ∆� and 1Τ∆� of ∆� and ∆� are referred to • So we ask: Given a 2-D continuous signal � �, � , what are
as the sampling frequencies in the � and � direction, the maximum possible values for ∆� and ∆� such that � �,
respectively. � can
• As shown below, these sampling frequencies are related to be reconstructed from the 2-D discrete signal �� �, � ,
the comb and sampling functions. obtained by (3.74)?
• Ideally, we would like to keep as few samples as possible. • As illustrated in Figure 3.12, we may be tempted to
• This strategy minimizes the number of detectors (or the perform coarse sampling as opposed to fine sampling.
scanning time needed by a single detector) and reduces • However, we will end up (in the extreme case) with only one
signal storage and subsequent processing requirements. sample, which is clearly not enough to represent a
continuous signal in general (unless the signal is constant).

Illustration of coarse
• Sampling a continuous signal with too few samples results in
(left) and fine (right)
rectangular sampling
a type of signal corruption called aliasing, in which higher
schemes.
frequencies ‘‘take the alias of’’ lower frequencies.
Although coarse • A signal sampled with too few samples cannot be
sampling results in reconstructed from its discrete representation; the best
fewer samples, it possible continuous reconstruction from these samples will be
may not allow corrupted by aliasing.
reconstruction of the
original continuous
• The visual appearance of this artifact depends on the
signal from these signal’s spectrum, but generally it appears as new high-
samples. frequency patterns where none should exist.

• The spatial frequency of these patterns is always lower than

it should be but may be relatively high compared with other
spatial information in the image.
• An example of this phenomenon is depicted in Figure 3.13.
The improperly sampled image, as shown in Figure 3.13(b), is
corrupted by a high-frequency texture that is not present in the
original image.
• When proper anti-aliasing is used, sampling will produce
an image with reduced resolution but no aliasing artifacts,
as shown in Figure 3.13(c). Obviously, it is vital that
medical imaging systems sample without aliasing.
(a) Original chest x-ray image and sampled images, (b) without, and (c) with
anti-aliasing. Inserts show a zoomed region.

21
 10/20/2022

Signal Model for Sampling

• A slowly varying signal could be reconstructed from fewer
samples than a rapidly varying signal, which may require • Consider multiplying a continuous signal � �, � by the
finer sampling at regions of rapid signal variation. sampling function. We have
• Since signal variation in space is directly associated • �� �, � = � �, � ��� �, �; ∞
σ∞
�=−∞� �, � �� ሺ�
with ∆�, ∆� = σ �=−∞
∞
σ∞ � �∆�, �∆� ��ሺ� − �∆�, �
frequency content, there must be a direct relationship between − �∆�, � − �∆�ሻ = �=−∞
σ�=−∞ �, � 𝛿 � − �∆�, � − �∆�
∞
σ∞
the �
the appropriate values of sampling periods ∆� and ∆� and − �∆�ሻ = �=−∞ �

σ�=−∞
frequency content of the signal under consideration. • where we have used (3.74) for the final
• Here, we prove this intuition mathematically, and derive a step.
recipe for choosing appropriate values for ∆� and ∆� by using • It follows that, given the discrete signal �� �, � = �
the
comb function comb �, � , given by (2.13), and its close �∆�, �∆� , we can calculate the continuous signal �� �,
relative, � regardless of the sampling periods ∆� and ∆�.
the sampling function ��� �, �; ∆�, ∆� , given by (2.14).

• Thus, if we can reconstruct � �, � from �� �, � , then � • Therefore,

�, � can also be reconstructed from �� �, � .
• �� �, � = � �, � ∗ comb �∆�, �∆� = � �, � ∗
• The most important point here is that, in order to understand the σ�=−∞
∞ σ�=−∞
∞ 𝛿 �∆� − �, �∆� − �1 � �, � ∗
effects of sampling, we need only look at the continuous signal
=
∆�∆�

σ∞ σ∞ 𝛿 � − �Τ∆� , � −

�Τ∆�
� =
�,���,
. � and its relationship to �
� �=−∞
=−∞
1
σ∞ ∞
�=−∞ σ�=−∞� �, � ∗ 𝛿 � − �Τ∆� , � − �Τ∆� =
• Since �� �, � is the product of two functions, its Fourier ∆�∆
trans- �
1 ∞ ∞
form is the convolution of Fourier transforms of two functions. σ∞ ∞
∆�∆� �=−∞ σ�=−∞ൣ‫׬‬
−∞ ‫׬‬
−∞ � �, � ∗ ��ሺ� − �Τ∆� − 𝜉 , � −
1
• This is a consequence of the product property (2.92) of the �Τ∆� − ��ሻ �𝜉���൧ = σ∞ σ∞ � �, � ∗ �� ሺ� − �Τ∆�
,�− �=−∞ �=−∞
Fourier transform. ∆�∆
�Τ∆ �

�ሻ

• where we have used the fact that the Fourier transform of • This is illustrated in Figure 3.14.
the sampling function is given by
• If the shifted spectra in �� �, � do not overlap, then the
• ℱ2� ��� �, �; ∆�, ∆� = comb �∆�, �∆� original spectrum � �, � of � �, � , and thus � �, �
• From (3.76), it is clear that the spectrum �� �, � of �� �, itself, can be recovered by filtering �� �, � to ‘‘pick’’ one of
� is calculated by shifting the spectrum � �, � of � �, � the (equivalent) spectra.
to locations • It is customary to use a low-pass filter [see (2.102)] to capture
�Τ∆� , �Τ∆� , for all � and �, adding all shifted the spectrum centered at the origin, like the one outlined with
spectra and dividing the result by ∆�∆�. a gray box in Figure 3.14(b).
22
 10/20/2022

• (a) The spectrum � �, � of a band-limited continuous signal

� �, � with cut-off frequencies � and �.
• (b) The spectrum �� �, � of signal �� �, � obtained by sampling
�� �, � with sampling periods � < 1Τ2 � and � < 1Τ2 �. In this
case, � �, � can be perfectly reconstructed from �� �,
� . The spectrum � �, � takes value one within the light
gray area and
zero outside.
• (c) The spectrum �� �, � of �� �, � obtained by sampling
� �, � with sampling periods � < 1Τ2 � and � < 1Τ2 �. In this
case, �� �, � experiences aliasing and � �, �
cannot be reconstructed perfectly in general.

Nyquist Sampling Theorem

• In order that the spectra in �� �, � do not overlap, it • If ∆� > 1Τ2 �or ∆� > 1Τ2 � , there is overlap of ‘‘high’’
is first necessary that the spectrum of � �, � be zero frequencies of � �, � in �� �, � , producing aliasing.
outside a rectangle in frequency space.
• Aliasing is illustrated in Figure 3.14(c). In this case, the
• Such signals are called band-limited. spectrum � �, � cannot be recovered from the
• If the highest frequencies present in � �, � in the � spectrum
and � �� �, � and, therefore, � �, � cannot be reconstructed
directions are � and �, respectively, then if from its samples �� �, � .
1 1
• ∆� ≤ and ∆� ≤ • In summary, A 2-D continuous band-limited signal � �, � , with
2� 2� cut-off frequencies � and �, can be uniquely determined
• the spectrum � �, � can be reconstructed from the from its samples �� �, � = � �∆�, �∆� , if and only if
spectrum
the sampling 1
�� �, � , in which case � �, � can be reconstructed periods ∆� and ∆� satisfy ∆�1 and ∆� ≤2�
from �� �, � ≤ 2�
and thus from the samples �� �, � .

• This is known as the sampling theorem (or the Nyquist • From (3.79), the minimum number of samples required
sampling theorem, after its discoverer). for aliasing-free sampling is directly proportional to the
• To avoid aliasing, the maximum allowed values for ∆� and cut-off frequencies � and �.
∆� • This verifies the fact that slowly varying signals, which
are given by are
1 1 characterized by small values of � and �, require fewer
• ∆� max = and ∆� max = • Aliasing-free sampling requires band-limited continuous signals.
2� 2�
• and are known as the Nyquist sampling periods.

23
 10/20/2022
samples than rapidly varying signals, which
are characterized by large values of � and
�.
• The undersampled signals tend to show high-
frequency artifacts because the overlap of high-
frequency spectra in the aliased Fourier
transform artificially boosts high-frequency
content.

24
 10/20/2022

Anti-Aliasing Filters
• In medical imaging systems, there is an inherent trade-off • In this case, the image is degraded by blurring rather
between the number of samples acquired by a detector and than aliasing artifacts, which is usually preferable.
image quality. Acquiring a large number of samples
typically produces the highest image resolution but is • Such a low-pass filter is called an anti-aliasing filter, and
almost always either expensive or time-consuming. must be applied before sampling.
• Reducing the number of samples may lead to aliasing, • Anti-aliasing filtering is often an inherent part of a
which will introduce unwanted artifacts into the image. medical imaging instrument because the true continuous
signal is blurred by the impulse response of the system.
• An alternative is to first filter the continuous signal using a
low- pass filter, and then sample with fewer samples. • It is useful to consider two sources of blurring, that caused
by the inherent physics and geometry of the system and that
caused by the detectors themselves.

• The physics and geometry cause only a small amount of blur-

ring and a high frequency signal is presented to the • Since the detectors are not point samplers, it is incorrect to
detectors. model the discretization of a medical imaging system as the
• Although this situation is highly dependent on the modality, multiplication of the incident signal � �, � with the
it would be fairly accurate to say that very little anti-aliasing sampling function ��� �, �; ∆�, ∆� .
is accomplished by the physics and geometry of the system. • Instead, the effect of detector integration can be modeled
• Anti-aliasing is more often accomplished by the detectors, by first convolving � �, � with the PSF ℎ �, � of the
which (in most cases) integrate the incident signal over the detector.
area of each detector.
• The overall detector system PSF is the result of the cascade of
• This additional integration, which is still part of the overall
PSF of the system, provides additional low-pass filtering prior PSFs that characterize the individual components of
to sampling. resolution, including the detector’s geometric resolution.

• Here we focus on a specific component of resolution: • Since integration is a low-pass process, this resembles the
any sampling process that is not point sampling. anti- aliasing filtering approach described above: The
• The sampling is area sampling rather than point sampling. continuous signal � �, � is first filtered by a low-pass filter
• This occurs when the detector is intrinsically digital (i.e., it is with PSF ℎ �, � , from which �� �, � is obtained by
a pixelated detector) or its output is an array of pixels. sampling.
• The PSF of the area sampling component can be • Note that the part of the overall PSF due to the other individual
represented by a rect function [(2.16)], and the sampling components of resolution can be included in this analysis
process can be modeled by �� �, � = �� �∆�, �∆� simply by considering ℎ �, � to be the overall PSF of the
system rather than just the PSF of the area sampling
• �� �∆�, �∆� = ℎ �, � ∗ � �, � ��� �, �; component.
∆�, ∆�

25
 10/20/2022

Example 3.13
• Consider a medical imaging system with sampling period • If an anti-aliasing filter is used, the highest frequency of the
∆in both the � and � directions. image being sampled is the cut-off frequency of the filter (here,
• What is the highest frequency allowed in the images so that we assume the transfer function of the filter has value 0
the sampling is free of aliasing? If an anti-aliasing filter, whose outside the cut-off frequency).
PSF is modeled as a rect function, is used and we ignore all • The anti-aliasing filter is modeled as a rect function.
the side lobes of its transfer function, what are the widths of the • Its transfer function is given by the sinc function (we ignore the
rect function? magnitude of the rect function, since it does not change the
• From the Nyquist sampling theorem, we know that if the cut- off frequency), � �, � = sinc ∆� �, ∆� �
image contains frequency higher than 1Τ2 ∆, then there will be • where ∆� and ∆� are the widths of the filter in the � and �
aliasing. So the highest allowed frequency is 1Τ2 ∆. directions, respectively.

• If we ignore the side lobes of � �, � , the cut-off

frequencies of the filter are the first zeros of the sinc function, • Figure 3.13 shows the impact of sampling with and without
which are 1Τ∆� and 1Τ∆� . anti- aliasing.
• For the given sampling periods ∆, we must • When the sampling period is too small, there is spectral
overlap and the resulting image has artifacts that are most
have
1 1 1
typically
•
1 ≤ and ≤ seen as high-frequency patterns, as are visible in Figure
∆� ∆� 3.13(b).
2∆ 2∆

• Which is equivalent to ∆� ≥ 2∆ and ∆� ≥ • Anti-aliasing low-pass filters the image before sampling,
2∆ making the image blurrier but devoid of artifacts, as is apparent
in
Figure 3.13(c).

Other Artifacts
Effects
• A problem that frequently affects image quality is the creation
• Artifacts of image features known as artifacts that do not represent valid
• Distortion anatomical or functional objects.
• Artifacts can obscure important targets, and they can be
falsely interpreted as valid image features.
• They can impair correct detection and characterization of
features of interest by adding ‘‘clutter’’ to images.
• Artifacts are caused by a variety of reasons and can appear
at any step of the imaging process.

25
 10/20/2022

• In projection radiography, artifacts can be generated by the x- • An example of this artifact is depicted in Figure 3.15(b).
ray source, by restricting the x-ray beam in order to avoid ex-
posing parts of a patient that need not be imaged, and by non- • Another artifact is the so-called beam hardening artifact.
uniformities in the x-ray image intensifier over the imaging • This artifact shows up as broad dark bands or streaks in the
area. image, and it is due to significant beam attenuation caused
by certain materials.
• In CT, artifacts may arise from patient motion, which produces • An example of this artifact is depicted in Figure 3.15(d).
streak artifacts throughout the image, known as motion
artifacts; Figure 3.15(a) shows an example. • Finally, a common artifact in CT is the so-called ring
artifact, illustrated in Figure 3.15(c).
• Another typical artifact in CT is known as a star artifact. This • This artifact is caused by detectors that go out of calibration
artifact is generated by the presence of metallic materials in and do not properly record incoming data.
the patient, which results in incomplete projections.

• There are many reasons for medical images to be corrupted • Examples of artifacts in
by artifacts. CT: (a) motion artifact, (b)
• Evaluation and possibly removal of artifacts should be part star artifact, (c) ring artifact,
of any high quality medical imaging system. and (d) beam hardening
• Good design, proper calibration, and maintenance of and partial voluming
medical imaging systems may control and even eliminate artifact.
artifacts.
• The specific artifacts that appear in each particular
imaging modality will be discussed in subsequent
chapters.

Distortion
• (a) Size distortion in a
• Medical imaging systems often introduce distortion, radiographic imaging system
another factor affecting image quality. due to magnification: although
the sizes of the two dark objects
• Distortion is geometrical in nature and refers to the inability of are different, their projections
a medical imaging system to give an accurate impression of are the same.
the shape, size, and/or position of objects of interest. • (b) Shape distortion in a
• In projection radiography, for example, size distortion can be radiographic imaging system
present due to magnification caused by the distance of the x- due to x-ray beam divergence:
although the two dark objects
ray source from the object being imaged. are the same, the shapes of
• This is illustrated in Figure 3.16(a). Notice that although the their projections are different.
sizes of the two dark objects are different their projections
are the same.

26
 10/20/2022

• On the other hand, shape distortion can be generated as • Unfortunately, distortion can be very difficult to determine
a result of unequal magnification of the object being and correct.
imaged. • In order to evaluate distortion, knowledge of the actual
• One common cause of shape distortion is the fact that shape and size of the object being imaged is required.
anatomical structures lie at different levels within the • Moreover, a good understanding of the imaging geometry
human body. is necessary.
• In projection radiography, shape distortion is also caused by • Developing methods for correcting distortion is very
the divergence of the x-ray beam. This is illustrated in Figure important for increasing image quality and improving
3.16(b). Notice that although the two dark objects are the diagnosis.
same the shapes of their projections are different. • Distortions introduced by each particular imaging modality
will be discussed in later chapters.

Accuracy
• We have focused on quantitative metrics of image quality. • Fundamentally, we are interested in the accuracy of
• The image quality ultimately must be judged in the context of medical images in the context of these clinical applications.
a specific clinical application. • Here, ‘‘accuracy’’ means both conformity to truth (i.e., freedom
• Medical images are used for from error) and clinical utility.
• Diagnosis (‘‘Is the disease present?’’), • In practice, we are usually interested in quantitative
• Prognosis (‘‘How will the disease progress, and what is the expected accuracy and diagnostic accuracy.
outcome?’’),
• Treatment planning (‘‘Which treatment will work best?’’), and
• Treatment monitoring (‘‘Is the treatment reversing the disease, and
to what extent?’’).

Quantitative Accuracy
• Sometimes, we are interested in the numerical value of a
given anatomic or functional feature within an image. • It is helpful to separate the component of error due to bias
• For example, we may wish to measure tumor dimensions from from that due to imprecision.
a radiograph, or estimate glucose metabolic rate from a • If we can establish that the measurement is precise (i.e.,
nuclear medicine image. reproducible), then we can correct for systematic errors
• In such situations, we need to know the error in through the use of a calibration standard that converts the
our measurement. measured value to the true value.
• This error, or difference from the true value, arises from two • In practice, error usually arises from both components, and
sources: bias, which represents a systematic, reproducible our measurements are never error-free.
difference from the truth, and imprecision, which represents
a random, measurement-to-measurement variation.

27
 10/20/2022

• We can conceptualize a medical image as representing

some parameter (or set of parameters) of interest in a • In a clinical setting, we are interested in two parameters:
patient. • Sensitivity, also known as the true-positive fraction; this is the fraction
of patients with disease who the test (e.g., the medical image) calls
• For the sake of simplicity, the diagnostic process involves a test abnormal.
that extracts this parameter (or set of parameters), which is
then used to classify a given patient as either normal or • Specificity, also known as the true-negative fraction; this is the fraction
diseased. of patients without disease who the test (e.g., the medical image) calls
normal.
• We further conceptualize two Gaussian distributions of the
parameter (or set of parameters), one for normal patients • In practice, sensitivity and specificity are established in a
and another for diseased patients, with (usually) some group of patients through the use of a 2 × 2 contingency table,
degree of overlap. as shown in Figure 3.17.
• This implies that some normal patients may be classified
as diseased and vice versa.

• Here, a and b are, • In this case, the sensitivity and specificity are given by
respectively, the number � �
of • sensitivity = and specificity =
diseased and normal patients �+
�
�+�

who the test calls abnormal, • Typically, the ultimate diagnosis that confirms the presence
whereas c and d are, or absence of disease is through a more-invasive test than
respectively, the number of medical imaging, a ‘‘gold standard’’ that is used to verify the
diseased and normal patients accuracy of (non-invasive) medical imaging procedures.
who the test calls normal. • The diagnostic accuracy (DA) is the fraction of patients that
are diagnosed correctly, and is given by
�+�
• DA = �+�+�+�

• In order to maximize diagnostic accuracy, both sensitivity • Probability distributions of

and specificity must be maximized. the test results for normal
• In practice, because of overlap in the distribution of parameter subjects and patients with
values between normal and diseased patients, a threshold certain disease.
must be established to call a study ‘‘abnormal,’’ in Figure 3.18.
• A lower threshold implies that more studies will be called
abnormal, thus increasing sensitivity but decreasing
specificity.
• A higher threshold implies that fewer studies will be called
abnormal, thus increasing specificity but decreasing
sensitivity.

28
 10/20/2022

• One way of graphically depicting this relation is via a receiver • The threshold also depends on the prevalence or proportion
operating characteristic (ROC) curve, which plots sensitivity of all patients who have the disease, because in practice we
as a function of 1 − specificity. are interested in two other parameters:
• In practice, the threshold must be chosen as a balance • Positive predictive value (PPV), which is the fraction of patients
called abnormal who actually have the disease
between sensitivity and specificity. • Negative predictive value (NPV), which is the fraction of persons
• The choice of threshold for a specific test depends on the called normal who do not have the disease
� �
relative cost-of-errors in calling a normal patient abnormal or • Notice that PPV = and�+�
NPV =
�+�
a diseased patient normal. • and that both depend on prevalence (PR), which is given by
�+�
• PR = �+�+�+�

Example 3.14
• Diagnostic accuracy alone cannot tell how good a • The contingency table for the • From the table, we have a = b
diagnostic method is. test is shown below. = 0, c = 10, and d = 90.
• This example shows that a bad diagnostic method can • So the sensitivity,
still achieve a high DA. specificity, and diagnostic
• Consider a group of 100 patients, among which 10 are accuracy of the test are
�
diseased and 90 are normal. We simply label all patients as • sensitivity = �+ = 0
normal. �
�
• Construct the contingency table for this test and determine • specificity =
�+
=1
�
the sensitivity, specificity, and diagnostic accuracy of the test. � +�
• DA = �+�+�
= 0,9
+�

• It can be seen that although all patients with the disease • Image quality characterizes the performance of a
are diagnosed wrong, we still have a relatively high medical imaging system and directly affects clinical
diagnostic accuracy of 0.9. utility.
• This is simply because the patients without the disease • It is assessed through a combination of specific
comprise the majority of patients in the group being performance parameters as measured in images.
studied.
1. Image quality refers to the degree to which an image allows
• This would be considered a flawed experimental design. a radiologist to accomplish the clinical goals of the imaging
study.
2. The six most important factors influencing image quality are
contrast, resolution, noise, artifacts, distortion, and
accuracy.

29
 10/20/2022

3. Contrast refers to the difference in image intensity of an 7. Aliasing is caused by improper sampling of a
object or target and surrounding objects or background. continuous signal, yielding artifacts in the resulting
4. Resolution is the ability of an imaging system to distinguish digital image.
and depict two signals that differ in space, time, or energy 8. Artifacts are false signals in an image that do not
as distinct. represent any valid structural or functional signal in the
5. Noise is any random fluctuation in an image; noise patient.
generally interferes with the ability to detect a signal in an 9. Distortion is any geometric inaccuracy in size or shape.
image.
10.Quantitative accuracy refers to the accuracy, compared
6. Continuous signals are transformed by sampling or with the truth, of numerical values obtained from an image;
discretization into discrete signals in order to be diagnostic accuracy refers to the accuracy of interpretations
digitally represented. and conclusions about the presence or absence of disease
drawn from image patterns.

‫نيملعلا برلا هلل دمحلا هتاكربو هللا‬

‫ةمحرو مكيلع ملسلاو‬

Thank you for your attention!

Engineering is essentially a problem-

solving.

30