Dove – Physics of Medical Imaging 10/8/2004

Page 1
Notes on Computerized Tomography
51:060 Bioimaging Fundamentals
Edwin L. Dove

Table of Contents
Notes on Computerized Tomography................................................................................. 1
1. Introduction............................................................................................................. 3
2. Calculation of tomographic images ........................................................................ 6
2.1 Basic concepts....................................................................................................... 6
2.2 Projection Details and the Radon Transform........................................................ 9
2.3 Sampling ............................................................................................................. 13
2.4 Image reconstruction........................................................................................... 13
2.4.1 Brute Force................................................................................................... 13
2.4.2 Iterative Reconstruction............................................................................... 14
2.4.3 Backprojection Method................................................................................ 16
2.4.4 Fourier Reconstruction................................................................................. 17
2.4.5 Filtered Backprojection (FBP)..................................................................... 18
2.4.5.1 Back Projection..................................................................................... 18
2.4.5.2 Filtered backprojection method ............................................................ 21
2.4.5.3 Discrete FBP......................................................................................... 22
2.4.5.4 Implementation ..................................................................................... 24
2.4.6 Fan beam FBP.............................................................................................. 25
2.5 CT number .......................................................................................................... 28
2.6 X-ray detectors.................................................................................................... 29
Dove – Physics of Medical Imaging 10/8/2004
Page 2
2.7 Imaging in three dimensions............................................................................... 29
2.7.1 Multiple 2D acquisitions.............................................................................. 29
2.7.2 Spiral CT...................................................................................................... 30
2.7.3 Cone-beam CT............................................................................................. 30
2.8 Image quality ...................................................................................................... 31
2.8.1 Image contrast.............................................................................................. 31
2.8.2 Spatial resolution ......................................................................................... 31
2.8.3 Noise ............................................................................................................ 32
2.8.4 Artifacts........................................................................................................ 33
2.8.4.1 Under-sampling..................................................................................... 33
2.8.4.2 Beam hardening .................................................................................... 33
2.8.4.3 Partial volume effects ........................................................................... 33
3. Appendix I – Fourier analysis................................................................................... 35
3.1 Introduction:........................................................................................................ 35
3.2 Response of LTIL system to eigenfunctions ...................................................... 35
3.3 Fourier Analysis of Periodic Signals .................................................................. 36
3.4 Calculating the Fourier Coefficients................................................................... 39
3.5 Fourier Analysis of Long and Possibly Non-periodic Waveforms..................... 40
3.6 Practical Procedure for Performing Fourier Analysis......................................... 41
3.7 Frequency Analysis of Random Signals............................................................. 43
References..................................................................................................................... 47

Dove – Physics of Medical Imaging 10/8/2004
Page 3
1. Introduction
In this chapter we consider systems that provide important tomographic or three-
dimensional (3D) capability. The tomogram is effectively an image of a slice taken
through a 3D volume. Ideally, it is free of the effects of intervening structures, thus
providing distinct improvement on the ability to visualize structures of interest. A CT
image is the result of the superposition of all planes normal to the direction of
propagation. Essentially, the system has infinite depth of focus.
Even though a CT image provides information not available in a traditional X-ray, the
information is not provided at zero cost. Consider the following table in which the
radiation dose of one CT slice is compared to the doses delivered in other common
medical procedures.
Table 1.1 Radiation dose received by the patient in several X-ray procedures. (from Sung et al.)

1 R= 0.01 Sv
(ICRP allows 50 mSv=5R per worker/5 year; 1 mSv=100mR/person/5year)
Are these doses dangerous? The answer to this question is not straightforward and
simple, but rather involves, in a complex way, many factors or parameters. The
biological effects of radiation have been studied for many years, and based on
epidemiological studies, the answer to this question seems to depend on five parameters:
1. Threshold: There is no single threshold that serves as an upper limit of acceptable
exposure. In general, since radiation effects are accumulative over a lifetime, the
threshold should be as low as possible based on an analysis of the benefits of
exposure versus the iatrogenic effects of radiation.
2. Exposure time: The severity of the biological effects depends on the duration of
exposure. A given dose will produce fewer effects if divided than if it were given
in a single exposure.
3. Exposure area: The larger the body areas exposed, the greater the chances of
damage. Those areas of the body not directly near the part being studied are
highly shielded.
4. Individual sensitivity: There is a wide variation in the radiosensitivity of various
species. For this reason, the lethal dose is expressed in statistical terms. For
humans, the lethal dose is 450 rads for whole body irradiation, that is, 50% of the
population will die if exposed to 450 rads over a 30-day period.
X-ray procedure Dose (per exposure)
Chest 20 mR
Brain 250 mR
Abdomen 550 mR
Dental 650 mR
Breast 54 mR
Xeromammography 200 mR
CT/slice 1000 mR
Dove – Physics of Medical Imaging 10/8/2004
Page 4
5. Cell sensitivity: Within the same individual, the most sensitive cells to radiation
damage are those cells that are rapidly dividing. Furthermore, specialized cells
are more sensitive than non-specialized cells. Cells that are very sensitive to
radiation effects include egg cells, sperm cells, white blood cells, immature red
blood cells, and epithelial cells. Cells less sensitive include muscle and nerve
cells.
If a dose of radiation of unusually large quantity (over 100 rads) is delivered to the body
in a very short time, biological effects that occur within hours or days after irradiation are
referred to as short-term effects. The signs and symptoms of these effects include nausea,
vomiting, malaise, and ultimately fever, shock and possible death. These effects are
collectively known as the acute radiation syndrome.
In diagnostic radiation, long-term and subtler effects caused by low dose radiation may
become manifest years after the exposure. Of these effects, the carcinogenic and genetic
effects have received the most attention. There is mounting evidence now to indicate that
diagnostic X-rays can induce these effects. The National Council on Radiation
Protection and Measurement established the maximal allowable dose per person per year
is 5 R, or about 1/5 of the typical dose delivered in one CT image. In other words, a CT
study of more than 5 slices typically exceeds the maximal allowable radiation dose in one
year. The need for a large CT study must be medically compelling, especially in the
young, and in those in childbearing years.
So what are the reasons for using CT rather than a traditional X-ray? There are primarily
three reasons. When we look at a chest X-ray (Figure 1.1), certain anatomic features are
immediately apparent. For example, the ribs show up as light structures because they
attenuate the X-ray beam more strongly than the surrounding soft tissue, so the film (or
detector) receives less exposure in the shadow of bones. The air-filled lungs show up as
darker regions.
Dove – Physics of Medical Imaging 10/8/2004
Page 5

Figure 1.1 Typical chest X-radiograph.
A short calculation illustrates the type of structure that one could expect to see with this
sort of conventional radiograph. The linear attenuation coefficients for the energy
spectrum of a typical diagnostic X-ray beam for air, bone, muscle and blood are

air
-1
bone
-1
muscle
-1
blood
0
0.48 cm
0.18 cm
0.178 cm
µ
µ
µ
µ
=
=
=
=

Assuming that the tissue transmission follows Beer’s law that states:

0
( ) exp( ) I x I x µ = − (1.1)
then the following contrasts can be calculated for a slab of soft tissue with a 1-cm cavity
in it:
Table 1.2 Contrast in a transmission radiogram
Dove – Physics of Medical Imaging 10/8/2004
Page 6
Material in cavity
0
( ) I x
I

Difference (%) with
respect to muscle
Air 1.0 +20
Blood 0.837 +0.2
Muscle 0.835 0
Bone 0.619 -26
Modern X-ray film allows contrasts of the order of 2% to be easily seen, so a 1-cm
diameter air-filled trachea can be visualized. However, blood in vessels and other soft
tissue details (e.g., heart anatomy) cannot be seen on conventional radiographs. Thus, the
first reason for using CT is to improve contrast to enable visualization of different types
of soft tissue with X-ray energies.
The second reason for using CT is to enable 3D visualization (i.e., enable depth
visualization). In a conventional radiogram, the 3D structure is collapsed onto a 2D film
of detector.
The third reason is to allow visualization of a slice of the anatomy along any imaging
plane. Conventional X-ray images produce a 2D image of the anatomy perpendicular to
the X-ray beam. CT imaging allows visualization of a slice of anatomy at any oblique
angle.
2. Calculation of tomographic images
2.1 Basic concepts
Consider the case of an X-ray passing through a slice of tissue with four attenuation
coefficients as shown in Figure 2.1. The output intensify I is given by
( )
0 1 2 3 4
exp I I x µ µ µ µ ( = − + + +
¸ ¸
(2.1)
where x is the distance traveled by the X-ray beam. Taking the logarithm of Equation
(2.1) and rearranging, we obtain the projection function

4
1
( )
i
i
p x µ
=
=

(2.2)
The projection function represents the summation of the attenuation coefficients along a
given X-ray path.
Dove – Physics of Medical Imaging 10/8/2004
Page 7

Figure 2.1. A simplified X-ray beam passing through a segment of tissue with different linear
attenuation coefficients (from Cho et al.).
In X-ray CT the contrast between one tissue type and another is associated with the
different attenuation coefficients of the materials involved. The projection data taken at
different views are the basis for tomographic image reconstruction. An example is shown
in Figure 2.2. If the tissue block is rasterized (i.e., sampled into rows and columns), then
for a particular row i the projection function across all m columns becomes:

1 2 3
( )
i i i i im
p x µ µ µ µ = + + + + L (2.3)


Figure 2.2. A 2D matrix of linear attenuation coefficients representing a
segment of tissue (from Cho et al.).
The scan represented in Figure 2.2 is a scan taken at zero degrees with respect to the
horizontal. In standard X-ray-based CT practice, a radiation source scans linearly along
an object at a given view angle θ , and after completion of the scan, the source and
detector pair rotate a small angle θ ∆ and repeat the linear scan. This process is
continued until the completion of 180˚ rotation. Thus for a parallel-beam data collection
Dove – Physics of Medical Imaging 10/8/2004
Page 8
scheme, a total of 180˚/N scans with an angular displacement of θ ∆ is collected for a
scan angle of 180˚. This is illustrated in Figure 2.3.

Figure 2.3 Parallel beam and angular scanning. Line integral data sets from different angles
measuring the amount of radiation passing through the body along each ray are the basic data sets to
be used for image reconstruction. Each ray represents the line integrals along the path.
For the final image reconstruction in an image of NxN, each beam is backprojected in the
image matrix with some from of spatial or spatial-frequency filtering. The goal of CT
imaging is to derive the unknown attenuation coefficients based only on the set of
projection functions. The basic hardware for performing this procedure is illustrated in
Figure 2.4
Dove – Physics of Medical Imaging 10/8/2004
Page 9

Figure 2.4 Basic scanning system for computerized tomography (from Macovski).

2.2 Projection Details and the Radon Transform
Consider the parallel-beam geometry in Figure 2.5. The function ( , ) f x y represents the
linear attenuation coefficients.

Figure 2.5 Parallel beam geometry and coordinate system (from unpublished material).
Dove – Physics of Medical Imaging 10/8/2004
Page 10
The X-ray beam makes an angle θ with the x-axis. The un-attenuated intensity of the X-
ray beams is
0
I . A new coordinate system ( , ) r s is defined by rotating ( , ) x y
counterclockwise over the angle θ . The coordinate change obeys the following
transformation formulae:


cos sin
sin cos
x r
y s
θ θ
θ θ
− ( ( (
=
( ( (
¸ ¸ ¸ ¸ ¸ ¸
(2.4)
and

cos sin
sin cos
r x
s y
θ θ
θ θ
( ( (
=
( ( (

¸ ¸ ¸ ¸ ¸ ¸
(2.5)
For a fixed angle θ the measured intensity profile as a function of r is shown in Figure
2.6 and is given by

( )
0
0
( ) exp ( , )
exp ( cos sin , sin cos
s
s
I r I f x y s
I f r s r s s
θ
θ θ θ θ
(
= − ∆
(
¸ ¸
(
= − − + ∆
(
¸ ¸


(2.6)

Figure 2.6 Intensity profile (from unpublished material).
For each measured intensity, we can compute the corresponding attenuation line sum.
Each intensity profile is then transformed into an attenuation profile as
( )
( )
( ) ln ( cos sin , sin cos )
s
I r
p r f r s r s s
I
θ
θ
θ
θ θ θ θ = − = − + ∆

(2.7)
Dove – Physics of Medical Imaging 10/8/2004
Page 11
The function ( ) p r
θ
is called the projection of the function ( , ) f x y along the angle θ . A
typical attenuation profile (or projection) is shown in Figure 2.7. Usually the projection
profile is measured for θ ranging from 0 to π .

Figure 2.7. Typical projection profile (from unpublished material).
Stacking all the projections ( ) p r
θ
together results in a 2D data set ( , ) p r θ , which is
called a sinogram. The transformation of a function ( , ) f x y into the sinogram ( , ) p r θ is
called the Radon transform, named after Johann Radon, who some contend is the founder
of tomographic reconstruction. The Radon transform for a given θ is given by
{ } ( , ) ( , ) ( cos sin , sin cos ) p r f x y f r s r s s θ θ θ θ θ = ℜ = − + ∆

(2.8)
To recapitulate, the Radon Transform computes projections of an image along specified
directions. A projection of a two-dimensional function ( , ) f x y is a line integral in a
certain direction. For example, the line integral of ( , ) f x y in the vertical direction is the
projection of ( , ) f x y onto the x-axis; the line integral in the horizontal direction is the
projection of ( , ) f x y onto the y-axis. As an illustrative example, the function ( , ) f x y in
Figure 2.8 has a projection onto the x′ axis (s in Equation (2.8)) found by integrating
along the y′ direction (r in Equation (2.8)). The projections are at an angle of q with
respect to the original orientation (q is θ in Equation (2.8)).

Dove – Physics of Medical Imaging 10/8/2004
Page 12

Figure 2.8 Geometry of the Radon Transform (from MATLAB help files).
Another way of viewing this is to consider the parallel projections illustrated in Figure
2.9 in which parallel beams of X-rays passing through an object produce the projection
function ( )
q
R x′ , which plays the role of ( ) p r
θ
in our development.

Figure 2.9 Parallel beam projections through an object (from MATLAB help).
Dove – Physics of Medical Imaging 10/8/2004
Page 13

2.3 Sampling
In practice there is a limited number M of projections or views and a limited number of
detector samples. Under these limits, the sinogram ( , ) p r θ given by Equation (2.8) then
becomes ( , ) p n r m θ ∆ ∆ with M rows and N columns. Given a particular beam aperture, it
is possible to calculate the minimum number of detector samples required to prevent
distortion due to aliasing. This minimum is known as the Nyquist criterion. This criterion
states that if s ∆ is the beam width, then the maximum sampling distance is
2
s ∆
, or at
least two samples per beam width are required.
The number of views is not so straightforward. A general rule is followed that states that
the number of views should be
2
π
times the number of detector samples. For example,
if the field of view is 50 cm and the beam with is 1 cm, then 100 detector samples and
150 views are required.
2.4 Image reconstruction
Image reconstruction is the process of estimating an image ( , ) f x y from a set of
projections ( , ) p r θ . Several algorithms exist to accomplish this task. These are:
• Brute force inversion technique
• Iterative reconstruction technique
• Backprojection technique
• Fourier-based technique
• Filtered backprojection technique
Of these, the last is the technique that is used in the majority of the modern medical
image processing machines. We will, however, consider some of the other methods in
some detail in order to develop an understanding of the process.
2.4.1 Brute Force
With this method, the projection set defines a system of simultaneous linear equations
that in theory can be solved using algorithms from linear algebra. Consider the following
2x2 image with 6 projections:

12
8
w x
y z
(
(
¸ ¸

13 11 9 7
Dove – Physics of Medical Imaging 10/8/2004
Page 14
These can be organized into a system of equations given as:

1 1 0 0 12
0 0 1 1 8
1 0 1 0 11
0 1 0 1 9
1 0 0 1 7
0 1 1 0 13
w
x
y
z
( (
( (
(
( (
(
( (
(
=
( (
(
( (
(
( (
¸ ¸
( (
( (
¸ ¸ ¸ ¸
(2.9)
This is an over-determined system of equations; there are 5 equations and 4 unknowns.
As a consequence, some of the rows and columns are linearly dependent. The pseudo-
inverse can be used to solve for the unknowns. This approach is not practical for real
systems, which can have hundreds of simultaneous equations for a single slice.
2.4.2 Iterative Reconstruction
The iterative reconstruction (also known as algebraic reconstruction technique-ART) was
the original reconstruction method used to solve the tomographic problem. The method
consists of three steps:
1. Make an initial guess at the solution
2. Compute projections based on the guess
3. Refine the guess based on the weighted difference between the desired (real)
projections and the actual (calculated) projections:

1
( )
i i
p p g desired actual
+
= + − (2.10)
For example, if we have our favorite small image and its projections
12
8
w x
y z
(
(
¸ ¸

13 11 9 7
then:
1. Initial guess and projections
0 0 0
0 0 0
(
(
¸ ¸

0 0 0 0
Dove – Physics of Medical Imaging 10/8/2004
Page 15
2. Refine horizontal ( 0 θ =
o
) estimates (assume
1
2
g = )


0 (12 0) / 2 6
0 (12 0) / 2 6
0 (8 0) / 2 4
0 (8 0) / 2 4
w
x
y
z
+ − ( ( (
( ( (
+ −
( ( (
= =
( ( ( + −
( ( (
+ −
¸ ¸ ¸ ¸ ¸ ¸

6 6 12
4 4 8
(
(
¸ ¸

10 10 10 10
3. Refine vertical ( 90 θ =
o
) estimates
6 (11 10) / 2 6.5
6 (9 10) / 2 5.5
4 (11 10) / 2 4.5
4 (9 10) / 2 3.5
w
x
y
z
+ − ( ( (
( ( (
+ −
( ( (
= =
( ( ( + −
( ( (
+ −
¸ ¸ ¸ ¸ ¸ ¸

6.5 5.5 12
4.5 3.5 8
(
(
¸ ¸

10 11 9 10
4. Refine diagonal ( 45 θ = ±
o
) estimates
6.5 (7 10) / 2 5
5.5 (13 10) / 2 7
4.5 (13 10) / 2 6
3.5 (7 10) / 2 2
w
x
y
z
+ − ( ( (
( ( (
+ −
( ( (
= =
( ( ( + −
( ( (
+ −
¸ ¸ ¸ ¸ ¸ ¸


5 7 12
6 2 8
(
(
¸ ¸

13 11 9 7
Dove – Physics of Medical Imaging 10/8/2004
Page 16
We are now done; the image data and projections match.
The ART was the original reconstruction method used in medical imaging. The method
works, but is slow and susceptible to noise. With more noise, the convergence is very
slow, and is not guaranteed.
2.4.3 Backprojection Method
It is obvious that we desperately need a rigorous mathematical answer to the burning
question that we all have: given the sinogram (family of projections) ( , ) p r θ , what is the
original image ( , ) f x y ? As ( , ) p r θ is the Radon transform of ( , ) f x y , this means that
what we actually need is an expression of the inverse Radon transform
{ }
1
( , ) ( , ) f x y p r θ

= ℜ
(2.11)
The projection theorem, also called the central slice theorem or the central section
theorem, provides an answer to this burning question. Let ( , )
x y
F k k be the 2D Fourier
transform of ( , ) f x y

( ) { }
( , ) ( , ) exp 2
x y x y
F k k f x y j k x k y dxdy π
∞ ∞
−∞ −∞
= − +
∫ ∫
(2.12)
The variables
x
k and
y
k are orthogonal frequencies with units cycles (or radians) per
distance. Let ( ) P k
θ
be the 1D Fourier transform of the projection ( ) p r
θ

( ) { }
( ) ( ) exp 2 P k p r j kr dr
θ θ
π

−∞
= −

(2.13)
The variable k is frequency (cycles or radians per distance). Allowing θ to vary from 0
to π and stacking all ( ) P k
θ
together, we obtain a 2D dataset ( , ) P k θ . The projection
theorem (or central slice theorem-CST) then states that

( , ) ( , )
x y
P k F k k θ =
(2.14)
with
2 2
cos
sin
x
y
x y
k k
k k
k k k
θ
θ
¦
=
¦
¦
=
´
¦
= +
¦
¹
(2.15)
In words, the set of 1D Fourier transform in r of the Radon transform of a function
is the 2D Fourier transform of that function. This means that if all the Fourier
Dove – Physics of Medical Imaging 10/8/2004
Page 17
transforms of the projections ( ) P k
θ
of a function ( , ) f x y are known, then it is
possible to calculate the value of the function ( , ) f x y for each point in the plane by
calculating the inverse Fourier transform of ( , ) P k θ . Proof of the CLT is
straightforward (but a bit tedious) and in all CT textbooks.

2.4.4 Fourier Reconstruction
The computational problem with the Central Section Theorem is that a 2D inverse
transform is required. In addition, CT requires various coordinate system shifts and
interpolations that complicate the calculations even further. To make this point more
clear, consider the following outline of the algorithm thus far. The reconstruction scheme
based on the CST consists of the following steps:
1. Calculate the 1D Fourier transform of all projections { }
1
( ) ( ) F p r P k
θ θ
= .
Practically this is the FFT of the projection function in the θ direction, for allθ .
2. Place all ( ) P k
θ
on a polar grid to obtain ( , ) P k θ , which after resampling in
Cartesian space, is equal to ( , )
x y
F k k

Figure 2.10. Polar (a) and Cartesian (b) sampling grid.
3. Calculate the 2D inverse Fourier transform of ( , )
x y
F k k to yield ( , ) f x y .
Practically this is the inverse FFT along each row (or column) for each column (or
row). The result is the desired image.
In practice, re-sampling from polar to rectangular coordinates involves considerable
interpolation, which makes the resultant estimate of ( , ) f x y noisy. In other words, the
spatial interpolation produces enough noise in the image to make this direct
implementation of the CST unacceptable for clinical diagnostic purposes. There are two
Dove – Physics of Medical Imaging 10/8/2004
Page 18
ways around this problem. The first is to use smaller values for θ ∆ and s ∆ , which will
decrease the noise introduced by the spatial interpolation scheme. Unfortunately, this
will also increase the hardware requirements of the machines, it will cause many more
calculations to be performed, both of which can lead to round-off and truncation errors.
The second method is to use the filtered backprojection method.
2.4.5 Filtered Backprojection (FBP)
2.4.5.1 Back Projection
For reference, Figure 2.5 is reproduced below:

Figure 2.5 (repeated for reference)
In back projection the measurements obtained at each projection are projected back along
the same line (or equivalently sameθ ). Thus the measured values are “smeared” across
the unknown density line as if a line of wet ink is drawn across the line. This is
illustrated in Figure 2.11 for the case of an object consisting of a single point.
Dove – Physics of Medical Imaging 10/8/2004
Page 19

Figure 2.11 Projections of a point at the origin back projected.
In this illustration, each projection is identical. Mathematically the back projection of a
single measured projection along an unknown density is given by:
( , ) ( ) ( cos sin ) b x y p r x y r dr
θ θ
δ θ θ = + −

(2.16)
where ( , ) b x y
θ
is the back-projected density due to the projection ( ) p r
θ
. Essentially we
are stating that we know that the point of density is somewhere along the line so that a
“crude’ reconstruction will result if we assign the measured value along the entire line.
The ( ) δ L function is the Dirac delta function that has the following properties:

1 if 0
( )
0 otherwise
r
r δ
= ¦
=
´
¹
(2.17)
Adding up these densities at all angles, we obtain the laminogram:

0
0
( , ) ( , )
( ) ( cos sin )
b
f x y b x y d
p r x y r drd
π
θ
π
θ
θ
δ θ θ θ

−∞
=
= + −

∫ ∫
(2.18)
The laminogram represents a distorted picture of the anatomy through which the X-ray
beam passes. We will now correct this distortion.
Each projection of a delta function at the origin is also a delta function. If these delta
functions are back-projected, the resulting response (known as the impulse response) is
simply:
Dove – Physics of Medical Imaging 10/8/2004
Page 20

0
( ) ( ) ( cos( ) )
1

b
h r r r r drd
r
π π
δ δ θ φ θ
−∞
= − −
=
∫ ∫
(2.19)
The result is derived after knowing something about generalized functions; however, the
lack of knowledge of generalized functions does not invalidate the result, which is correct
by all accounts. Operations on generalized functions can be “esoteric.” For example,
generalized functions predict that

( )
[ ( )]
( )
n
n
x x
f x
f x
δ
δ

=



where
n
x are the roots of ( ) f x .
Knowing that the response to an impulse is
1
r
and assuming linear system theory
applies, then the following relationship holds:

1
( , ) ( , ) **
b
f x y f x y
r
= (2.20)
The double asterisk (**) signifies convolution. Practically this means that the actual
image ( , ) f x y is blurred by the
1
r
term. Thus, the image that one calculates with the
backprojection method (the laminogram) is a blurred version of the actual image. The
early reconstructions were based on Equation (2.20), with marginal results at best.
The Fourier transform of Equation (2.20) is

( , )
( , )
b
F k
F k
k
θ
θ = (2.21)
since the 2D Fourier transform of
1
r
is
1
k
.
The
1
r
blurring must be removed. There have been two methods for accomplishing this
task. The first is to calculate the 2D Fourier transform of Equation (2.20), weight the
result with the 2D Fourier transform of the reciprocal of
1
r
, and then perform the
inverse 2D Fourier transform of the product. The computational requirements and
potential for noise introduction (due to re-sampling, etc.) are enormous. The second
method is to use the filtered backprojection algorithm.
Dove – Physics of Medical Imaging 10/8/2004
Page 21
2.4.5.2 Filtered backprojection method
It is desirable to be able to use the elegant simplicity of back projection and to undo the
1
r
blur without requiring 2D transforms. This is accomplished with our powerful
friend, the Central Section Theorem.
We begin by restating the back-projection relationship for the laminogram ( , )
b
f x y as

0
( , ) ( , ) ( cos sin )
b
f x y p r x y r drd
π π
θ δ θ θ θ
−∞
= + −
∫ ∫
(2.22)
Restructuring this into a Fourier transform mode and applying the CST to substitute the
inverse transform of ( , ) P k θ
( )
0
( , ) ( , ) exp 2 ( cos sin )
b
f x y F k i kr dk x y r drd
π
θ π δ θ θ θ
∞ ∞
−∞ −∞
(
= + −
(
¸ ¸
∫ ∫ ∫
(2.23)
Performing the integration over r, we obtain
| |
0
( , ) ( , ) exp 2 ( cos sin )
b
f x y F k j k x y dkd
π
θ π θ θ θ

−∞
= +
∫ ∫
(2.24)
This equation is actually quite important. To appreciate this statement, we can rewrite
the 2D Fourier transform in polar form, yielding
( )
2
0 0
( , ) ( , ) exp 2 cos sin f x y F k j k x y kdkd
π
θ π θ θ θ

= + (
¸ ¸ ∫ ∫
(2.25)
This can be modified to conform to the form of Equation (2.24) yielding
( )
0
( , ) ( , ) exp 2 cos sin f x y F k j k x y k dkd
π
θ π θ θ θ

−∞
= + (
¸ ¸ ∫ ∫
(2.26)
It is necessary to use k instead of k because the integration includes negative values.
Comparing Equation (2.26) with Equation (2.24), we see that they differ only by the k
weighting term, as expected. Substituting { } ( ) p r
θ
ℑ for ( , ) F k θ in Equation (2.24), and
multiplying and dividing by k , we obtain
Dove – Physics of Medical Imaging 10/8/2004
Page 22

{ }
( )
0
( )
( , ) exp 2 cos sin
b
g r
f x y j k x y k dk
k
π
θ
π θ θ

−∞

= + (
¸ ¸ ∫ ∫
(2.27)
This equation states that the transform of each projection has been weighted by k along
each radial line in Figure 2.5 (or Figure 2.11). This accounts for the blurred
reconstruction of ( , )
b
f x y . The “fix” for this blurring is to pre-weight the transform of
the projection prior to back projecting, that is

{ }
( )
0
( )
( , ) exp 2 cos sin
p r k
f x y j k x y k dk
k
π
θ
π θ θ

−∞

= + (
¸ ¸ ∫ ∫
(2.28)
This reconstruction is known as the filtered back-projection system, and is the current
standard algorithm for calculating a tomographic image. This system involves only a
one-dimensional Fourier transform that is fast (because the FFT algorithm is actually
used).
The scheme is:
1. Each projection is individually transformed,
2. The transformed projection is weighted with k ,
3. The result is inverse transformed, and
4. The result is back projected.
2.4.5.3 Discrete FBP
Of course digital computers require discrete representations of Equation (2.28). Going
from continuous space to discrete space is usually straightforward:

become

,
dθ becomes θ ∆ , and dk becomes k ∆ . Conversion would be straightforward if it were
not for the pesky k term. This term is the ramp filter, which is not stable (there are
values of k for which things are not well behaved). The instability makes this filter
unusable in practice. To see this, consider the inverse Fourier transform of the filter:

( )
1
2 2
{ } exp 2
1

2
F k k j kr dk
r
π
π


−∞
=
= −

(2.29)
Note that the inverse Fourier transform of the filter has a singularity at r=0. Is this
important? Unfortunately it is; the CT image “blows up” for each θ , which is bad if the
goal is to derive an image that faithfully reproduces the anatomy.
Dove – Physics of Medical Imaging 10/8/2004
Page 23
What to do? We know from our discussion on sampling that the measured data are
limited to frequencies below the Nyquist frequency
N
k as shown in Figure 2.12(a).

Figure 2.12: Decomposition of Ram-Lak filter (a) into the difference between a block (b) and a
triangle (c).
The resulting filter (graphed in Figure 2.12(a)) is known as the Ram-Lak filter, named
after its inventors Ramachandran and Lakshiminarayanan. The Ram-Lak filter can be
written as the difference of a block and a triangle (Figure 2.12(b) and (c)). The inverse
Fourier transform then gives the following:

2 2
sin(2 ) 1 cos(2 )
( )
2
N N N
k k r k r
q r
r r
π π
π π

= − (2.30)
The resulting filter function has no singularity at r=0. To make the filter discrete,
substitute
n
r n r = ∆ and
1
(2 )
n
k
r
=

. Thus Equation (2.30) becomes:

( )
2
2
1
if =0
4( )
( ) 0 if is even
1
if is odd
( )
n
n
r
q r n
n
n r π
¦
¦

¦
¦
=
´
¦
¦−
∆ ¦
¹
(2.31)
Usually frequencies below
N
k are corrupted by aliasing and noise. Applying a
smoothing filter (usually the Shepp-Logan filter, or other low-pass filters such as Hann,
Hamming, or other) suppresses the highest spatial frequency and reduces noise. A
generalized Hamming window is given by

( ) 1 cos for
( )
0 for
N
k
k N
N
k k
H k
k k
π
α α ¦ + − ≤
¦
=
´

¦
¹
(2.32)
Dove – Physics of Medical Imaging 10/8/2004
Page 24
For 0.5 α = , we have the Hann (or Hanning) window, and for 0.54 α = , we have the
Hamming window. Figure 2.13 shows the results of applying the Hann smoothing filter
to the Ram-Lak filter.

Figure 2.13 Product of a smoothing filter and the Ram-Lak filter.
After all of this development, the complete formula of discrete FBP can be written by
combining a discrete convolution using the kernel in Equation (2.31) and a discrete back
projection given in Equation (2.22) as
( ) ( )
1 1
0 0
( , ) , cos sin
M N
i j m n i m j m n
m n
f x y p r q x y r r θ θ θ θ
− −
= =
(
= ⋅ + − ∆ ∆
(
¸ ¸
∑ ∑
(2.33)
2.4.5.4 Implementation
Actual implementation of the FBP algorithm might follow this outline (assuming parallel
projections):
1. Design a smoothing filter (Shepp-Logan, Hann, Hamming, or other) in the Fourier
(frequency) domain;
2. Calculate 1D FFT of the projections for each θ ;
3. Filter the projections by multiplying the filter with the Fourier transformed
projections (remember: convolution in the space domain is multiplication in the
frequency domain);
4. Determine image size; this involves either a desired size, or a calculation of the
diagonal similar to diag=2*ceil(N/sqrt(2))+1, where N is the specified number
rows, or if unspecified, N = 2*floor(size(P,1)/(2*sqrt(2))) where P is the
projection matrix;
5. Back-project; this may involve interpolation if the size of the image is different
from the size of the projection matrix. The interpolation scheme could be nearest
neighbor, linear, spline, or other;
6. Scale the image elements (pixels) if necessary;
7. Add ad hoc image processing to prepare the image for ultimate clinical use. This
ad hoc processing is usually vender specific, and often defines the individual
characteristics of the particular CT machine.
Dove – Physics of Medical Imaging 10/8/2004
Page 25
2.4.6 Fan beam FBP
The CT equipment has undergone various “generations” of development as illustrated in
Figure 2.14.

Figure 2.14 Generations of X-ray CT machines (from Cho)
The first generation was Hounsfield’s configuration and consisted of a pencil beam
source and detector, and then linear translation across the patient. The second generation
consisted of a small fan beam source, linear detector, and translation and rotation about
the patient. The third generation consisted of a wider fan beam (about 30°) where the fan
beam could cover the entire object under study. This generation of machines acquired
data with rotation only; no translation was necessary. The fourth generation of machine
consisted of a rotating fan beam, and a stationary detector array.
There is a “fifth generation” machine that consists of an electronically steered electron
beam swept across a target ring of tungsten. This electron-beam CT (EBCT) shows high
spatial and temporal resolution – there are no moving parts. This is illustrated in Figure
2.15.
Dove – Physics of Medical Imaging 10/8/2004
Page 26

Figure 2.15 EBCT geometry
From Imatron’s literature:
Single Slice mode scanning utilizes 100 millisecond sweeps along one of the four
target rings. This sequence provides two distinct acquisition features: Continuous
Volume Scanning (CVS) and Step Volume Scanning (SVS). In a CVS sequence
data is acquired by moving the patient continuously through the x-ray beam path.
In a SVS sequence timed scans are acquired while the patient moves
incrementally through the scanner. In SVS diagnosticians can utilize real time
ECG triggering to provide a volume acquisition of the heart at a prescribed phase
of the cardiac cycle. The ability of the ECG to trigger the Imatron EBT scanner at
high heart rates, without skipping a beat, provides extremely high temporal
resolution and high patient compliance without exposure to unnecessary dose.
In Multi-Slice sequence scanning, the electron beam sweeps along one to four
target rings in rapid succession covering up to 8 centimeters of the body without
moving the patient through the scanner. This unique functionality, along with
real time ECG cardiac triggering provides cine and flow capabilities. This very
fast scanning speed and temporal resolution is ideal for evaluating ventricular
function, joint motion, airway physiology and blood flow analysis. The
exceptional scan speed of the Imatron EBT scanner provides the diagnostician
with critical diagnostic information when imaging cardiac, pediatric, trauma and
geriatric patients where speed of acquisition is essential.
The images from the EBCT machine are impressive. For example, the Left Anterior
Descending (LAD) coronary artery looks like:
Dove – Physics of Medical Imaging 10/8/2004
Page 27

Figure 2.16 LAD artery with lesion; from EBCT.
The reconstruction algorithms that we have developed are all based on the assumption
that the data have been acquired with parallel-beam geometry. Data from modern
machines are actually not ordered in parallel subsets, but rather in fans as shown in
Figure 2.17.

Figure 2.17 Fan-beam geometry (a) and required measurement range (b).
The fan angle is the formed by the fan. How can the mathematical models developed in
the previous section be used in this new geometry? Three are two approaches;
Dove – Physics of Medical Imaging 10/8/2004
Page 28
1. Record the fan beam data, and interpolate into parallel data. This process is
known as re-binning. These data then require the same algorithm as
developed above.
2. Adopt the FBP algorithm. This approach uses a modified convolution back-
projection system of Gullberg and Denton (see refs.). Here the convolution
kernel is slightly different, and the back projection involves a quadratic
weighting factor rather than the uniform weighting of the parallel rays. This
approach is widely used in existing scanners.
2.5 CT number
The reconstruction methods studied so far indicate how a 2D function can be
reconstructed from its line integrals. These line integrals are the sum of the function in
different directions. In the case of X-ray attenuation, however, we measure the exponent
of the desired line integral as

( )
0
z dz
I I e
µ −

= (3.1)
Note that there is a nonlinear relationship between the measured projections and the
desired line integral. This nonlinearity is removed if we use the log of the measured
transmission where

0
ln ( )
I
z dz
I
µ =

(3.2)
Unfortunately, this is not going to work. Equation (3.1) was derived with the assumption
that the source is mono-energetic, and the beam is infinitesimally small. Most machines
compensate for this distortion on an ad hoc basis by using a nonlinear function of the log
of the measurements.
All machines report derived linear attenuation coefficients using a CT number. The CT
number of a pixel with attenuation coefficient
x
µ is:
CT Number
x water
water
K
µ µ
µ

= (3.3)
If the scale factor K=1000, then the units of the CT Number is Hounsfield Units (HU).
The following table shows data for various materials.
Dove – Physics of Medical Imaging 10/8/2004
Page 29
Table 2.1 CT Number and µ for various materials at 60 keV.
Material CT Number (HU) µ (cm
-1
)
Bone 808 0.38
Water 0 0.21
Muscle -48 0.20
Fat -142 0.18
Air -1000 0.0
2.6 X-ray detectors
There are three kinds of X-ray detectors used in modern CT equipment. These are
scintillation detectors, gas ionization chambers, and solid-state (semiconductor) detectors.
Scintillation detectors, used in the earliest CT machines, consist of a scintillation crystal
(NaI, CaF
2
) in combination with a photo-multiplier tube. The advantages of this
technique are high quantum efficiency, and fast time response. The disadvantage is the
low packing density.
Gas ionization chambers consist of a pressurized gas chamber (Xe or Kr) with two
electrodes. Some of the incident X-ray photons cause ionization of the gas (by the photo-
electric absorption). The resulting free electrons and ions drift towards the anode and
cathode, giving rise to measurable electric signals. Gas ionization chambers have a low
quantum efficiency and slow response time. The advantage is high packing density.
The newest scanners typically use semiconductor detectors. A solid-state conductor
consists of a scintillation crystal (CsI) and a photo-diode. The photo-diode is based on a
pn-junction. When a scintillation photon passes through the depletion layer (boundary
layer containing no mobile carriers), electron-hole pairs are produced, giving rise to a
measurable electric current. These detectors have a very fast response time, high
quantum efficiency, a good stability, and good reliability. On the other hand, they are
susceptible to scattered radiation and thus require external collimation.
2.7 Imaging in three dimensions
2.7.1 Multiple 2D acquisitions
The most straightforward way to image an entire volume is to scan number of
consecutive slices; this is called a sequential scan. This is illustrated in Figure 2.18(a) for
a 360˚ acquisition.
Dove – Physics of Medical Imaging 10/8/2004
Page 30

Figure 2.18 Acquisition schemes: (a) multiple 2D acquisition, (b) 360° nearest neighbor,
(c) 360° linear interpolation.
β represents the angular position of the X-ray tube, and z its axial position relative to the
patient. Data are acquired for discrete axial positions
i
z and for angular tube positions
β ranging from 0 to 2π .
In order to acquire a complete 3D data set containing all information needed to calculate
a reconstruction, and in order to not miss any lesions or other anatomical details, the axial
sampling density must obey the Nyquist criterion. Let’s assume a rectangular slice
thickness (or profile) of z ∆ . This means that the real data is convolved with an axial
block of width z ∆ . We come to the conclusion that the maximum distance between
slices is
2
z ∆
, or in words, we need to acquire at least two slices per slice thickness.
2.7.2 Spiral CT
A technique that is widely used is spiral (or helical) CT. The X-ray source rotates
continuously around the patient, but simultaneously the patient is slowly translated
through the gantry. The tube describes a helical orbit relative to the patient. Figure
2.18(b) illustrates this type of data collection scheme. In this figure, data are
simultaneously collected in β and in z. The table feed (TF in the figure) is the axial
distance over which the table translates during the time needed by the source to complete
a 360˚ orbit. The pitch ratio is the ratio between the table feed and the slice thickness.
In practice, a pitch ratio between 1.0 and 2.0 is used; for example, a slice thickness of 2
mm is combined with a table feed of 3 mm (pitch=1.5). A higher pitch, and thus faster
scanning, has several advantages: lower dose, less tube load, and less motion artifacts.
2.7.3 Cone-beam CT
Some experimental machines have a 2D detector array in order to measure the entire
volume on one single orbit of the X-ray source. Unfortunately, this technique has not
lived up to the original expectations. This technique could use some smart biomedical
engineer to continue its development.
Dove – Physics of Medical Imaging 10/8/2004
Page 31
2.8 Image quality
2.8.1 Image contrast
Image contrast depends on several factors, the major factors being:
1. Tube current (mAs) per slice (the more current supplied to the tube, the better the
contrast, but also the more radiation delivered to the patient).
2. Spectrum of the X-ray tube. The effective linear attenuation coefficient of the
tissue depends on the incident spectrum and therefore the contrast depends on the
incident spectrum. The narrower the spectrum (more like mono-energetic
photons) the better the contrast.
3. The dynamic range of the hardware limits the contrast.
4. Nonlinearities such as scatter, etc. limit the image contrast.
2.8.2 Spatial resolution
The spatial resolution of a CT image depends on several factors, the most important
being:
1. Size of the focal spot, which is the area on which the electrons hit the anode.
2. Size of the detector elements; the size of the detector elements and the size of the
focal spot determine the beam width.
3. Pre- and post-patient (axial) collimators, which determine the slice thickness.
4. Reconstruction kernel (or convolution filter), which can be tuned to enhance high
frequencies (for determining edges), or low frequencies (for attenuate noise and
oth er artifacts)
5. Better resolution can be obtained by:
a. Increasing the number of detectors,
b. Using a smaller pixel (or voxel) size, at the expense of more noise,
c. Using a reconstruction kernel with less smoothing, at the expense of more
noise.
A typical modern CT machine has sub-mm in-plane resolution, and up to 1 mm axial
resolution.
Dove – Physics of Medical Imaging 10/8/2004
Page 32

Figure 2.19 Example of image artifacts: (a) test phantom, (b) second phantom, (c) noise, (d) detector
under-sampling, (e) view under-sampling, (f) beam hardening, (g) scatter, (h) nonlinear partial
volume effect, and (i) object motion. (unpublished results)
2.8.3 Noise
The main source of noise is quantum noise. The number of photons passing through the
patient and striking the detectors is susceptible to quantum fluctuations, and can be
described by the Poisson distribution. The variance of the number of photons is
proportional to the actual number of photons; therefore an increase in the number of
photons used to improve contrast will result in more noise. Increasing the energy of the
X-ray beam can reduce noise.
Noise in the projections in CT exhibits a characteristic structure. A high noise level
results in alternating dark and light thin streaks radiating from locations with high
attenuation (bright areas) as shown in Figure 2.19(c).
Dove – Physics of Medical Imaging 10/8/2004
Page 33
2.8.4 Artifacts
2.8.4.1 Under-sampling
Taking too few samples results in two distinct phenomena (shown in Figure 2.19):
1. Aliasing due to the number of detectors being too small.
2. If the number of views is too small, then streaks appear.
2.8.4.2 Beam hardening
Low energy photons are preferentially absorbed. This causes the X-ray beam too
“harden” as it passes through the body. As a consequence, hardening artifacts appear,
such as “cupping” and streaks (Figure 2.19(f)).
2.8.4.3 Partial volume effects
Due to a finite beam aperture, every measurement represents an intensity that is averaged
over the beam aperture. This average intensity is log-converted to give an effective beam
attenuation. This value is always an underestimate of the actual attenuation. The larger
the attenuation gradients perpendicular to the beam, the larger the underestimate. This
results in streaks tangent to edges. This phenomenon is illustrated in Figure 2.19 (h).
Another example is shown in Figure 2.20 in which the white arrows point to the streak
artifacts due to the partial volume effect. The large white “blob” is highly absorbent
contrast material in the patient’s stomach.

Dove – Physics of Medical Imaging 10/8/2004
Page 34

Figure 20. Image with artifactual streaks due to the partial volume effect.









Dove – Physics of Medical Imaging 10/8/2004
Page 35
3. Appendix I – Fourier analysis
3.1 Introduction:
The development of Fourier analysis has a long history involving many individuals. The
Babylonians used "trigonometric series" to predict astronomical events. In 1748 L. Euler
investigated the motion of a vibrating violin string and found that the vibrations along the
length of the string were periodic, and the magnitude of the vibrations was sinusoidal.
Euler preformed the same calculations that we will perform in the next section. In 1753
Bernoulli criticized the use of trigonometric series to represent vibrations on strings. His
criticism was based on his own belief that it was impossible to represent signals with
corners (i.e., discontinuous slopes) using trigonometric series.
Jean Baptiste Joseph Fourier entered this controversy around 1800. Fourier was
interested in diffusing the heat generated in cannons used by Napoleon Bonaparte. In
1807 Fourier submitted a manuscript for publication describing the Fourier Series and
Fourier Transform. The manuscript went to four reviewers, including Laplace and
Lagrange. Only Lagrange objected to the theory that signals can be represented as
trigonometric series. The manuscript was never published. The theory was finally given
the "light of day" in Fourier's book The Analytical Theory of Heat published in 1822.
The ideas offered by Fourier have had an enormous impact in engineering, science, and
mathematics. Circuits are design based on Fourier analysis. Blood pressure transducers
are designed based on Fourier analysis. The list of engineering systems designed based
on Fourier analysis is never-ending; the list grows daily.
3.2 Response of LTIL system to eigenfunctions
It is easy to show that if the input to a linear, time-invariant, lumped (LTIL) system is a
sinusoid, then the output will also be a sinusoid of the same frequency, but scaled in
magnitude and shifted in phase. If the input to a LTIL system is a complex exponential,
the output is also a complex exponential with only a change in (possibly complex)
amplitude. A signal for which the system output is a (possibly complex) constant times
the input is referred to as an eigenfunction of the system, and the amplitude factor is
referred to as the system's eigenvalue.
We know that a LTIL system is always characterized by the convolution of its impulse
response h(t) and the input u(t). Let's let u(t)=exp(st). Thus we have:

∫ ∫

∞ −

∞ −
− −
= = τ τ τ τ
τ τ
d e h e d e h t y
s st t s
) ( ) ( ) (
) (
(4.1)
By the definition of the Laplace transform, this becomes:

τ s
e s H t y ) ( ) ( = (4.2)
Dove – Physics of Medical Imaging 10/8/2004
Page 36

The function H(s) is a complex constant the value of which depends on a complex
number s=σ+jω. The value of H(s) and the system impulse response are a Laplace
Transform pair. Thus, the constant H(s) for a specific value of s is the eigenvalue
associated with the eigenfunction exp(st).
We can do an example to illustrate this concept. Let

t s t s t s
e c e c e c t u
3 2 1
3 2 1
) ( + + = (4.3)
Then by superposition we know that

t s t s t s
e s H c e s H c e s H c t y
3 2 1
) ( ) ( ) ( ) (
3 3 2 2 1 1
+ + = (4.4)
Extending this concept leads to a convenient expression for the response of a LTIL
system. If the input is written as


=
n
t s
n
n
e c t u ) ( (4.5)
then the output will be written as


=
n
t s
n n
n
e s H c t y ) ( ) ( (4.6)

Thus if the input can be represented as a series of complex exponentials, then the output
can also be represented as a series of complex exponentials. Furthermore, each
coefficient in this representation of the output is obtained as the product of the
corresponding coefficient of the input c
n
and the system's eigenvalue H(s
n
).

3.3 Fourier Analysis of Periodic Signals
The value of the variable s in the above treatment is an arbitrary complex number. In
Fourier analysis the value of s is restricted to purely imaginary numbers, that is, s= jω.
Thus in Fourier analysis only complex exponentials exp(jωt) are considered. We know
however that complex exponentials are related to real functions. Euler proved that
) sin( ) cos(
0 0
0
t j t e
t j
ω ω
ω
+ = (4.7)
Thus, complex exponential are periodic with period T
0
given by
Dove – Physics of Medical Imaging 10/8/2004
Page 37

0
1
0
2
0 f
T = =
ω
π
(4.8)
Let’s look at an example of what this means.
Consider a periodic signal x(t) with fundamental frequency ω
0
of 2π. Let’s express the
signal as


− =
=
3
3
2
) (
k
t jk
k
e c t x
π
(4.9)
where c
0
=1, c
-1
=c
1
=
1
/
4
, c
2
=c
-2
=
1
/
2
, and c
3
=c
-3
=
1
/
3
. Substituting these values and collecting
terms, we get
) (
3
1
) (
2
1
) (
4
1
1 ) (
6 6 4 4 2 2 t j t j t j t j t j t j
e e e e e e t x
π π π π π π − − −
+ + + + + + = (4.10)
Using Euler's relation in Eq. 7, Eq. 10 can be written as
t t t t x π π π 6 cos
3
2
4 cos 2 cos
2
1
1 ) ( + + + = (4.11)

Thus we see that the summation of a series of complex exponentials is equivalent to
the summation of real periodic signals.

To generalize this result, Fourier showed that x(t) can be written as



−∞ =

=
n
t jnw
n
e c t x
0
) ( (4.12)
The constant c is, in general, complex. This can be rewritten as



=


+ + =
1
0
] [ ) (
0 0
n
t jn
n
t jn
n
e c e c c t x
ω ω
(4.13)
This can be further simplified by writing the coefficient c
n
in rectangular form as
c
n
= A
n
+ jB
n
where A
n
and B
n
are real. In this case, Eq. 13 can be rewritten as:
Dove – Physics of Medical Imaging 10/8/2004
Page 38

(
¸
(

¸

− + =
∑ ∑

=

= 1 1
0 0 0
sin cos 2 ) (
n n
n n
t n B t n A A t x ω ω (4.14)
This is usually rewritten in the more familiar form

∑ ∑

=

=
+ + =
1 1
0 0 0
sin cos ) (
n n
n n
t n b t n a a t x ω ω
(4.15)
Equations 13 and 15 are equivalent if the following holds

2 2
n n
n
b
j
a
c − =
(4.16)
This is the Fourier Series in the trigonometric form. The fundamental frequency is ω
0

rads/sec. The frequency nω
o
rads/sec is said to be the n-th harmonic frequency. The
Fourier coefficients a
n
and b
n
are sometimes called the spectral coefficients of x(t). These
coefficients are a measure of the portion of the signal x(t) that is at each harmonic of the
fundamental component.
Dirichlet proved that under certain conditions, x(t) equals its Fourier Series
representation, except at isolated values of t for which x(t) is discontinuous. At these
values, the infinite series converges to the average value on either side of the
discontinuity.
The Dirichlet conditions for the existence of the Fourier Series are:
1. Over any period, x(t) must be absolutely integrable, that is, the integral of x(t)
over one period must be finite.
2. In any finite interval of time, x(t) must have a finite number of maxima and
minima.
3. In any finite interval of time, there are only a finite number of discontinuities.

Parseval showed that for a continuous-time periodic signal x(t), then




−∞ =
=
T
k
k
a dt t x
T
2 2
| | | ) ( |
1
(4.17)
The term on the left-hand side of this equation is the average power (energy per unit of
time) in one period of the signal x(t). Thus, Parseval's theorem proved that the total
average power in a periodic signal equals the sum of the average powers in all of its
harmonic components. Stated in another way, the average power of a periodic signal is
Dove – Physics of Medical Imaging 10/8/2004
Page 39
conserved across the Fourier Series. Thus, one can work in the time domain or the
frequency domain, and power will not be lost.
3.4 Calculating the Fourier Coefficients
In order to take advantage of the power of Fourier analysis, the unknown coefficients
a
n
and b
n
must be found. There are at least two methods that can be used to find the
coefficients, but in practice, one uses the Fast Fourier Transform (more later). The
original method of Fourier, however, was based on the orthogonality of trigonometric
functions. Two functions are said to be orthogonal if their inner product is zero. For
example,


=
T
tdt m t n 0 sin cos ω ω (4.18)


¦
¹
¦
´
¦ ≠
=
T
T
n m
tdt m t n
otherwise
2
if 0
cos cos ω ω (4.19)


¦
¹
¦
´
¦ ≠
=
T
T
n m
tdt m t n
otherwise
2
if 0
sin sin ω ω (4.20)
The Fourier Series is

∑ ∑

=


+ + =
1 1
0 0 0
sin cos ) (
n n
n n
t n b t n a a t x ω ω
(4.21)
If we multiply both sides of this equation by x(t) and integrate over T
0
(one period) and
take advantage of the fact that integrating a sine or cosine over one period is zero, then
we get that


=
0
0 0
0
) (
1
T
dt t x
T
a
(4.22)
Thus the first term in the Fourier Series is simply the average value of the waveform.
If we multiply both sides of the Fourier Series by cos(mω
0
t) and sin(mω
0
t), and take
advantage of the orthogonality of sinusoids, we get:


=
0
0
0
0
) cos( ) (
2
T
n
dt t n t x
T
a ω
(4.23)
Dove – Physics of Medical Imaging 10/8/2004
Page 40


=
0
0
0
0
) sin( ) (
2
T
n
dt t n t x
T
b ω
(4.24)
Thus we have an easy way to calculate the Fourier Series coefficients for any periodic
waveform x(t). Unfortunately there are two problems with this formulation. First most
biomedical signals are not periodic. Second, performing this integration for each
harmonic can be time consuming. This problem grows quadratically as T
0
increases.
3.5 Fourier Analysis of Long and Possibly Non-periodic Waveforms
Given a time function f(t), one can predict the response of a linear system (filter) by first
converting the time function to the frequency domain, and then altering the signal in the
new domain. If the signal is periodic with period T
0
seconds, or if a computer is used to
sample the function with a sampling window of T
0
seconds in width, then f(t) can be
written as a Fourier Series.
Euler's relations for the cosine and sine can be used to develop a practical approach for
calculating the Fourier Series coefficients c
n
. Remember that Euler said that
j
jx jx
x
jx jx
x
2
) exp( ) exp(
sin and
2
) exp( ) exp(
cos
− −
=
− +
=
Using the definition of c
n
in Eq. 16, the Fourier Series coefficients can be written
as


− =
0
0
0
0
) exp( ) (
1
T
n
d jn f
T
c τ τ ω τ (4.25)
The c
n
coefficients are found from solving this equation, or from the Fast Fourier
Transform (FFT) algorithm.
The FFT algorithm was proposed by Cooley and Tukey in the early 1960's (see Brigham,
E. O. “The Fast Fourier Transform,” Prentice-Hall, 1974). This algorithm calculates the
complex coefficients c
n
at a blazingly fast speed. The algorithm is easily implemented in
software using only a few multiplications and many shifts; this is the reason that the
algorithm is fast. The algorithm is also easily implemented in hardware. One can buy an
FFT on a chip for just a few dollars.

Dove – Physics of Medical Imaging 10/8/2004
Page 41
3.6 Practical Procedure for Performing Fourier Analysis
The procedure for handling large amounts of data, and for performing Fourier analysis of
non-periodic signals, is as follows:
1. Sample the waveform for T
0
seconds at a sampling rate of f
s
samples per second.
Thus the total number of samples is N
p
= T
0
f
s
. The value of T
0
is the sampling
window. The fundamental frequency is that frequency that produces one complete
cycle in this sampling window. Thus
0
1
0
2
0 f
T = =
ω
π
.
2. Assemble the N
p
samples data
0,
data
1
, … into an array. The array will usually be
complex with the imaginary parts all zero. The real parts will be the sampled data.
Thus the data array will be
D =
data j
data j
data j
data j
N
N
p
p
0
2
2
1
0
0
0
0
+
+
+
+

¸

(
¸
(
(
(
(
(
(


M
(4.26)
3. Calculate the values of c
n
using the FFT algorithm. By convention, one must
scale the results (output) of the FFT by
1
/N
p
, where N
p
is the number of points
sampled in T
0
seconds. Thus the c
n
coefficients from the FFT must be multiplied
by
1
/N
p
, or equivalently, the points in the vector of data passed to the FFT
algorithm must be scaled by N
p
. In the conventional FFT algorithm, the number
N
p
must be a power of two (2), such as 32, 64, 128, 256, 1024, etc. The results
from the FFT will be arranged as n=0 to N
p
/2, and in most implementations, these
are followed in reverse order as n=-(N
p
/2) to -1. The coefficients from the FFT
algorithm (assuming that scaling has already been applied) are thus arranged as a
vector C of complex numbers given by
C =
c
c
c
c
c
N
N
p
p
0
1
2
2
1
M
M

¸

(
¸
(
(
(
(
(
(
(
(
(
(
(4.27)
Thus, only the first N
p
/2 points in C are useful. Once the C coefficients (or
equivalently the a
n
and b
n
coefficients) are obtained, one can reconstruct the time
function from the Fourier Series equation (Eq. 20).

Dove – Physics of Medical Imaging 10/8/2004
Page 42
In sum, the FFT yields a scaled version of the c
n
coefficients. From these, one can filter
the time signal f(t) by altering the c
n
coefficients in the frequency domain. One can then
use the Fourier Series representation to reconstruct the time waveform after filtering.
Often the values of c
n
are displayed as a graph with the magnitude of c
n
on the ordinate,
and frequency on the abscissa. This graph is sometimes called the line spectrum.
There are many FFT algorithms available. All of the available FFT routines require that
the number of points (samples) passed to the routine be a power of 2. If this is not the
case, one can either pad the data sequence with zeros, or truncate the sequence to a power
of two, or use the integration formula for calculating the a and b coefficients directly.
The IMSL library has an easy-to-use algorithm that is callable from FORTRAN, C, or
C
++
. The MATLAB package has a discrete FT (DFT) algorithm that calculates the
coefficients for sampled data. The corresponding frequencies of the DFT are
summarized in the following table. The first two columns are digital frequencies, and the
last two columns are analog frequencies. The two are related by ω=WT
k
where ω is the
analog frequency and W is the digital frequency. If x(t)=sin(ωt), then x[k]=sin(WT
k
).
Note that in the Table the variable T
k
is the time step (in seconds) between data samples,
and the variable N is the total number of samples. The frequency Hz-sec is obtained by
multiplying the frequency in Hz by the time T
k
, the inter-sample time, in seconds.

Frequencies of DFT Components in Different Units
Hz-sec Rad Hz rad/sec
Symbol N ω f W
Frequency of F
1
1
/
N
2π/
N
1
/
NT
2π/
NT
Frequency of F
m
m
/
N
2πm/
N
m
/
NT
2πm/
NT
Frequency of F
N/2
0.5 π 1
/
2T
π/
T
Sampling Frequency 1.0 2π 1
/
T
2π/
T

The frequency of the final component, F
N/2
, of the DFT of any real series is exactly one-
half the sampling frequency in accordance with Shannon's Sampling Theorem. The
period of the F
m
component is just the reciprocal of the frequency in all cases. Thus, the
period of F
m
is
N
/
m
samples or
NT
/
m
seconds.
Dove – Physics of Medical Imaging 10/8/2004
Page 43
3.7 Frequency Analysis of Random Signals
The above analysis is not the same as finding the power spectrum. Spectral estimation is
the determination of the spectral content of a random process based on a finite set of
observations from the process. Formally, the power spectral density (PSD), which is
denoted by P
xx
(f), of a wide-sense stationary random process x
n
is defined as

(Kay, S.M.
“Modern Spectral Estimation,” Prentice-Hall, Englewood Cliffs, 1988)
{ } ) = t R FT f P
xx xx
( ) ( (4.28)
FT{} represents the Fourier transform. The function R
xx
{t} is the autocorrelation
function given by
{ } ) + ( ) ( = )

t f f E t R
xx
τ τ ( (4.29)
where E{} represents the expected value (expectation) operator. The superscript *
represents the complex conjugate operator. The PSD represents the distribution of power
with frequency of the random process.
For finite data lengths, the autocorrelation function cannot be found exactly - only
estimates are available. For this case, the definition given for the autocorrelation function
is changed to the following:
| |
2
)} ( {
1
) ( t x FT
M
f P
xx
= (4.30)
This new definition says to take the magnitude squared of the Fourier transform of the
data (calculated with the FFT) and then divide by the record length. This approach is the
one used by MATLAB, which assumes that M, the number of samples, is large. There
are many other methods for estimating the PSD. In contrast to the methods for estimating
the line spectrum, no one method is universally accepted for estimating the PSD.
Often the values of P
xx
(f) are displayed on a graph with the magnitude (in dB) of the
PSD on the ordinate and frequency (in Hz) on the abscissa. This is called the
periodogram and is similar to a probability distribution function that describes how the
power of a random signal is distributed over frequency. This is not the same as the line
spectrum that describes how the time signal is decomposed into its sinusoidal
components.
The area of the periodogram for positive frequencies gives the total power in the random
signal. The units of power are squared units of the sampled data x
n
. The units of power
density are therefore power per hertz. It turns out that R
xx
(0) is the variance of the
Dove – Physics of Medical Imaging 10/8/2004
Page 44
random signal, which, if the signal is Gaussian, is also the total power.

(Stearns, S.D.,
David, R.A. “Signal Processing Algorithms,” Prentice-Hall, Englewood Cliffs, 1988).
In sum, Fourier analysis is based on the idea that one can decompose a signal (either a
function of time, or space) into the summation of sine and cosine waves. Fourier theory
supports the statement that any continuous signal can be exactly written as the sum of a
constant, an infinite number of weighted sine waves, and an infinite number of cosine
waves. For example, if the signal ( ) f x is periodic, then Fourier analysis states:
( ) ( )
0 0 0
1 1
( ) cos sin
n n
n n
f x a a n x n n x ω ω
∞ ∞
= =
= + +
∑ ∑
(4.31)
If ( ) f x is not periodic, then one can write that the Fourier Transform of ( ) f x is:
( ) ( ) ( ) exp F f x j x ω ω

−∞
= −

(4.32)
The notion of Fourier transform is graphically illustrated below in Figure A in which the
complex signal is decomposed into its constituent sinusoids of different frequencies by
applying either Equation (4.31) or Equation (4.32).

Figure A – Decomposition of signal into summation of sine and cosine functions.
If the signal is discrete, then the Fourier Transform becomes the Discrete Fourier
Transform (DFT) that is given below (copied from MATLAB help files)

Dove – Physics of Medical Imaging 10/8/2004
Page 45
If the signal is real, as ours will most likely be, then the above equation becomes (again,
from MATLAB):

where

There is also a two-dimensional Fourier transform that is given by the following discrete
equation (again copied from MATLAB):


As an example, consider the 2D figure below.

Figure B – Test figure from MATLAB
Dove – Physics of Medical Imaging 10/8/2004
Page 46
The result of applying the 2D DFT to the test figure is shown below.

Figure C – 2D DFT of object in Figure B.
The actual calculation of the DFT is performed with an algorithm called the Fast Fourier
Transform (FFT) invented by Tukey and Cooley. The MATLAB function fft() and ifft()
calculate the forward FFT and inverse FFT, respectively. Try the following MATLAB
code:
load sunspot.dat
year = sunspot(:,1);
wolfer = sunspot(:,2);
plot(year,wolfer)
title('Sunspot Data')

Y = fft(wolfer);

N = length(Y);
Y(1) = [];
power = abs(Y(1:N/2)).^2;
nyquist = 1/2;
freq = (1:N/2)/(N/2)*nyquist;
figure, plot(freq,power), grid on
xlabel('cycles/year')
title('Periodogram')

period = 1./freq;
figure, plot(period,power), axis([0 40 0 2e7]), grid on
ylabel('Power')
xlabel('Period(Years/Cycle)')


[mp index] = max(power);
period(index)
Dove – Physics of Medical Imaging 10/8/2004
Page 47
References
1. Berger, S., W. Goldsmith, and E. Lewis (eds.) Introduction to Bioengineering,
Oxford, New York, 1996. This reference is complete, but it is written at a
beginning graduate level. The section on medical imaging is brief, but useful.
2. Brown, B., R. Smallwood, D. barber, P. Lawford, and D. Hose. Medical Physics
and Biomedical Engineering, Institute of Physics Publishing, Bristol, 1999. This
was to be our textbook for the semester. There are about 30 copies in the IMU
bookstore, but we could not get more than 30 from the publisher. Too bad, so
sad.
3. Macovski, A. Medical Imaging Systems, Prentice-Hall, Englewood Cliffs, 1883.
This small text contains the mathematical treatment of imaging that we all need to
know. Unfortunately, the book is out of print.
4. Shung, K. M. Smith, B. Tsui. Principles of Medical Imaging, Academic Press,
San Diego, 1992. This is a wonderful brief text from which many of the images
are drawn. Unfortunately, this text is out of print.
5. Webb, S. (ed) The Physics of Medical Imaging, Institute of Physics Publishing,
Bristol, 1992. This is a complete text covering all aspects of medical imaging.
Unfortunately, this text is out pf print.

Dove – Physics of Medical Imaging

10/8/2004

2.7 Imaging in three dimensions............................................................................... 29 2.7.1 Multiple 2D acquisitions.............................................................................. 29 2.7.2 Spiral CT...................................................................................................... 30 2.7.3 Cone-beam CT ............................................................................................. 30 2.8 Image quality ...................................................................................................... 31 2.8.1 Image contrast.............................................................................................. 31 2.8.2 Spatial resolution ......................................................................................... 31 2.8.3 Noise ............................................................................................................ 32 2.8.4 Artifacts........................................................................................................ 33 2.8.4.1 Under-sampling..................................................................................... 33 2.8.4.2 Beam hardening .................................................................................... 33 2.8.4.3 Partial volume effects ........................................................................... 33 3. Appendix I – Fourier analysis................................................................................... 35 3.1 Introduction:........................................................................................................ 35 3.2 Response of LTIL system to eigenfunctions ...................................................... 35 3.3 Fourier Analysis of Periodic Signals .................................................................. 36 3.4 Calculating the Fourier Coefficients................................................................... 39 3.5 Fourier Analysis of Long and Possibly Non-periodic Waveforms..................... 40 3.6 Practical Procedure for Performing Fourier Analysis......................................... 41 3.7 Frequency Analysis of Random Signals ............................................................. 43 References..................................................................................................................... 47

Page 2

Dove – Physics of Medical Imaging

10/8/2004

1. Introduction
In this chapter we consider systems that provide important tomographic or threedimensional (3D) capability. The tomogram is effectively an image of a slice taken through a 3D volume. Ideally, it is free of the effects of intervening structures, thus providing distinct improvement on the ability to visualize structures of interest. A CT image is the result of the superposition of all planes normal to the direction of propagation. Essentially, the system has infinite depth of focus. Even though a CT image provides information not available in a traditional X-ray, the information is not provided at zero cost. Consider the following table in which the radiation dose of one CT slice is compared to the doses delivered in other common medical procedures.
Table 1.1 Radiation dose received by the patient in several X-ray procedures. (from Sung et al.)

X-ray procedure Chest Brain Abdomen Dental Breast Xeromammography CT/slice

Dose (per exposure) 20 mR 250 mR 550 mR 650 mR 54 mR 200 mR 1000 mR

1 R= 0.01 Sv (ICRP allows 50 mSv=5R per worker/5 year; 1 mSv=100mR/person/5year)

Are these doses dangerous? The answer to this question is not straightforward and simple, but rather involves, in a complex way, many factors or parameters. The biological effects of radiation have been studied for many years, and based on epidemiological studies, the answer to this question seems to depend on five parameters: 1. Threshold: There is no single threshold that serves as an upper limit of acceptable exposure. In general, since radiation effects are accumulative over a lifetime, the threshold should be as low as possible based on an analysis of the benefits of exposure versus the iatrogenic effects of radiation. 2. Exposure time: The severity of the biological effects depends on the duration of exposure. A given dose will produce fewer effects if divided than if it were given in a single exposure. 3. Exposure area: The larger the body areas exposed, the greater the chances of damage. Those areas of the body not directly near the part being studied are highly shielded. 4. Individual sensitivity: There is a wide variation in the radiosensitivity of various species. For this reason, the lethal dose is expressed in statistical terms. For humans, the lethal dose is 450 rads for whole body irradiation, that is, 50% of the population will die if exposed to 450 rads over a 30-day period.
Page 3

Dove – Physics of Medical Imaging

10/8/2004

5. Cell sensitivity: Within the same individual, the most sensitive cells to radiation damage are those cells that are rapidly dividing. Furthermore, specialized cells are more sensitive than non-specialized cells. Cells that are very sensitive to radiation effects include egg cells, sperm cells, white blood cells, immature red blood cells, and epithelial cells. Cells less sensitive include muscle and nerve cells. If a dose of radiation of unusually large quantity (over 100 rads) is delivered to the body in a very short time, biological effects that occur within hours or days after irradiation are referred to as short-term effects. The signs and symptoms of these effects include nausea, vomiting, malaise, and ultimately fever, shock and possible death. These effects are collectively known as the acute radiation syndrome. In diagnostic radiation, long-term and subtler effects caused by low dose radiation may become manifest years after the exposure. Of these effects, the carcinogenic and genetic effects have received the most attention. There is mounting evidence now to indicate that diagnostic X-rays can induce these effects. The National Council on Radiation Protection and Measurement established the maximal allowable dose per person per year is 5 R, or about 1/5 of the typical dose delivered in one CT image. In other words, a CT study of more than 5 slices typically exceeds the maximal allowable radiation dose in one year. The need for a large CT study must be medically compelling, especially in the young, and in those in childbearing years. So what are the reasons for using CT rather than a traditional X-ray? There are primarily three reasons. When we look at a chest X-ray (Figure 1.1), certain anatomic features are immediately apparent. For example, the ribs show up as light structures because they attenuate the X-ray beam more strongly than the surrounding soft tissue, so the film (or detector) receives less exposure in the shadow of bones. The air-filled lungs show up as darker regions.

Page 4

48 cm-1 µ muscle = 0.1 Typical chest X-radiograph. bone. The linear attenuation coefficients for the energy spectrum of a typical diagnostic X-ray beam for air. A short calculation illustrates the type of structure that one could expect to see with this sort of conventional radiograph.178 cm -1 Assuming that the tissue transmission follows Beer’s law that states: I ( x) = I 0 exp(− µ x) (1.1) then the following contrasts can be calculated for a slab of soft tissue with a 1-cm cavity in it: Table 1.Dove – Physics of Medical Imaging 10/8/2004 Figure 1.18 cm-1 µ blood = 0.2 Contrast in a transmission radiogram Page 5 . muscle and blood are µ air = 0 µ bone = 0.

. Calculation of tomographic images 2.g. In a conventional radiogram.e.619 Difference (%) with respect to muscle +20 +0.837 0.Dove – Physics of Medical Imaging 10/8/2004 Material in cavity I ( x) Air Blood Muscle Bone I0 1. The second reason for using CT is to enable 3D visualization (i. Taking the logarithm of Equation (2.1) and rearranging. Page 6 . The third reason is to allow visualization of a slice of the anatomy along any imaging plane. enable depth visualization).0 0. so a 1-cm diameter air-filled trachea can be visualized. the 3D structure is collapsed onto a 2D film of detector.2) The projection function represents the summation of the attenuation coefficients along a given X-ray path.835 0. CT imaging allows visualization of a slice of anatomy at any oblique angle.1 Basic concepts Consider the case of an X-ray passing through a slice of tissue with four attenuation coefficients as shown in Figure 2. blood in vessels and other soft tissue details (e. The output intensify I is given by I = I 0 exp  − ( µ1 + µ 2 + µ3 + µ 4 ) x    (2. However. heart anatomy) cannot be seen on conventional radiographs. 2.2 0 -26 Modern X-ray film allows contrasts of the order of 2% to be easily seen. Conventional X-ray images produce a 2D image of the anatomy perpendicular to the X-ray beam.. we obtain the projection function p ( x) = ∑ µi i =1 4 (2.1) where x is the distance traveled by the X-ray beam. Thus. the first reason for using CT is to improve contrast to enable visualization of different types of soft tissue with X-ray energies.1.

In standard X-ray-based CT practice.). The projection data taken at different views are the basis for tomographic image reconstruction. sampled into rows and columns).1. A 2D matrix of linear attenuation coefficients representing a segment of tissue (from Cho et al. This process is continued until the completion of 180˚ rotation.2 is a scan taken at zero degrees with respect to the horizontal.3) Figure 2. a radiation source scans linearly along an object at a given view angle θ . A simplified X-ray beam passing through a segment of tissue with different linear attenuation coefficients (from Cho et al. and after completion of the scan. If the tissue block is rasterized (i. In X-ray CT the contrast between one tissue type and another is associated with the different attenuation coefficients of the materials involved. then for a particular row i the projection function across all m columns becomes: pi ( x) = µi1 + µi 2 + µi 3 + L + µim (2. Thus for a parallel-beam data collection Page 7 .Dove – Physics of Medical Imaging 10/8/2004 Figure 2.e.2.2. The scan represented in Figure 2. An example is shown in Figure 2.. the source and detector pair rotate a small angle ∆θ and repeat the linear scan.).

a total of 180˚/N scans with an angular displacement of ∆θ is collected for a scan angle of 180˚. Each ray represents the line integrals along the path. The goal of CT imaging is to derive the unknown attenuation coefficients based only on the set of projection functions.4 Page 8 . Figure 2. The basic hardware for performing this procedure is illustrated in Figure 2.3 Parallel beam and angular scanning. For the final image reconstruction in an image of NxN.Dove – Physics of Medical Imaging 10/8/2004 scheme.3. Line integral data sets from different angles measuring the amount of radiation passing through the body along each ray are the basic data sets to be used for image reconstruction. This is illustrated in Figure 2. each beam is backprojected in the image matrix with some from of spatial or spatial-frequency filtering.

Page 9 .2 Projection Details and the Radon Transform Consider the parallel-beam geometry in Figure 2.4 Basic scanning system for computerized tomography (from Macovski). Figure 2. The function f ( x. 2.5 Parallel beam geometry and coordinate system (from unpublished material). y ) represents the linear attenuation coefficients.5.Dove – Physics of Medical Imaging 10/8/2004 Figure 2.

4) sin θ   x  cosθ   y    (2. s) is defined by rotating ( x.6 Intensity profile (from unpublished material).5) For a fixed angle θ the measured intensity profile as a function of r is shown in Figure 2. transformation formulae: The coordinate change obeys the following  x  cos θ  y  =  sin θ    and  r   cosθ  s  =  − sin θ    − sin θ   r  cos θ   s    (2.Dove – Physics of Medical Imaging 10/8/2004 The X-ray beam makes an angle θ with the x-axis.7) Page 10 . r sin θ + s cosθ ) ∆s   s  (2. we can compute the corresponding attenuation line sum. The un-attenuated intensity of the Xray beams is I 0 . y ) counterclockwise over the angle θ . A new coordinate system (r . Each intensity profile is then transformed into an attenuation profile as pθ (r ) = − ln Iθ (r ) = ∑ ( f (r cos θ − s sin θ . y )∆s   s    = I 0 exp−  ∑ ( f (r cos θ − s sin θ .6 and is given by   Iθ (r ) = I 0 exp  −∑ f ( x. r sin θ + s cos θ ) ) ∆s Iθ s (2. For each measured intensity.6) Figure 2.

For example.8)) found by integrating along the y′ direction (r in Equation (2.8)). the line integral of f ( x. r sin θ + s cosθ )∆s (2. The transformation of a function f ( x.θ ) = ℜ { f ( x. Page 11 . A projection of a two-dimensional function f ( x.Dove – Physics of Medical Imaging 10/8/2004 The function pθ (r ) is called the projection of the function f ( x. Figure 2.θ ) is called the Radon transform. named after Johann Radon.7. the function f ( x. A typical attenuation profile (or projection) is shown in Figure 2.8)). the line integral in the horizontal direction is the projection of f ( x. the Radon Transform computes projections of an image along specified directions. who some contend is the founder of tomographic reconstruction. Stacking all the projections pθ (r ) together results in a 2D data set p (r . Usually the projection profile is measured for θ ranging from 0 to π . The Radon transform for a given θ is given by p (r . The projections are at an angle of q with respect to the original orientation (q is θ in Equation (2.8 has a projection onto the x′ axis (s in Equation (2. y )} = ∑ f (r cos θ − s sin θ . y ) in the vertical direction is the projection of f ( x.7. y ) into the sinogram p (r . y ) along the angle θ .θ ) . y ) in Figure 2. y ) onto the y-axis. y ) is a line integral in a certain direction.8) To recapitulate. As an illustrative example. y ) onto the x-axis. which is called a sinogram. Typical projection profile (from unpublished material).

Page 12 .Dove – Physics of Medical Imaging 10/8/2004 Figure 2. which plays the role of pθ (r ) in our development.8 Geometry of the Radon Transform (from MATLAB help files).9 Parallel beam projections through an object (from MATLAB help). Figure 2.9 in which parallel beams of X-rays passing through an object produce the projection function Rq ( x′) . Another way of viewing this is to consider the parallel projections illustrated in Figure 2.

the sinogram p (r . then 100 detector samples and 150 views are required. Given a particular beam aperture.θ ) given by Equation (2. however. the projection set defines a system of simultaneous linear equations that in theory can be solved using algorithms from linear algebra.4. Consider the following 2x2 image with 6 projections:  w x  12  y z 8   13 11 9 Page 13 7 . m∆θ ) with M rows and N columns. it is possible to calculate the minimum number of detector samples required to prevent distortion due to aliasing. For example.1 Brute Force With this method. This minimum is known as the Nyquist criterion. Several algorithms exist to accomplish this task.3 Sampling In practice there is a limited number M of projections or views and a limited number of detector samples.8) then becomes p (n∆r . A general rule is followed that states that the number of views should be π times the number of detector samples. Under these limits. The number of views is not so straightforward. 2. This criterion states that if ∆s is the beam width. These are: • • • • • Brute force inversion technique Iterative reconstruction technique Backprojection technique Fourier-based technique Filtered backprojection technique Of these. consider some of the other methods in some detail in order to develop an understanding of the process. y ) from a set of projections p (r .θ ) . then the maximum sampling distance is ∆s .4 Image reconstruction Image reconstruction is the process of estimating an image f ( x.Dove – Physics of Medical Imaging 10/8/2004 2. We will. or at 2 least two samples per beam width are required. 2. 2 if the field of view is 50 cm and the beam with is 1 cm. the last is the technique that is used in the majority of the modern medical image processing machines.

Dove – Physics of Medical Imaging 10/8/2004 These can be organized into a system of equations given as: 1 0  1  0 1  0  1 0 0 1 0 1 0 1 1 0 0 1 0 12  1   w  8     0   x  11   =   1  y   9    1  z   7     0 13    (2. The pseudoinverse can be used to solve for the unknowns. As a consequence.10) For example. Make an initial guess at the solution 2. some of the rows and columns are linearly dependent. which can have hundreds of simultaneous equations for a single slice.9) This is an over-determined system of equations. The method consists of three steps: 1. there are 5 equations and 4 unknowns. Initial guess and projections 0 0 0 0 0 0   0 0 0 0 11 9 7 Page 14 . if we have our favorite small image and its projections  w x  12  y z 8   13 then: 1.4.2 Iterative Reconstruction The iterative reconstruction (also known as algebraic reconstruction technique-ART) was the original reconstruction method used to solve the tomographic problem. This approach is not practical for real systems. Compute projections based on the guess 3. 2. Refine the guess based on the weighted difference between the desired (real) projections and the actual (calculated) projections: p i +1 = p i + g (desired − actual ) (2.

5 8   10 11 9 10 4.5        z   4 + (9 − 10) / 2  3.5 + (13 − 10) / 2   7   = =   y   4.5 + (7 − 10) / 2   2 5 7  12 6 2 8   13 11 9 Page 15 7 .5 5.5  x   6 + (9 − 10) / 2  5. Refine horizontal ( θ = 0o ) estimates (assume g = 1 ) 2  w  0 + (12 − 0) / 2   6   x   0 + (12 − 0) / 2   6   = =   y   0 + (8 − 0) / 2   4        z   0 + (8 − 0) / 2   4  6 6  12  4 4 8   10 10 10 10 3.5 + (7 − 10) / 2   5   x  5.5  = =   y   4 + (11 − 10) / 2   4.5 3.Dove – Physics of Medical Imaging 10/8/2004 2. Refine diagonal ( θ = ±45o ) estimates  w  6.5 12  4.5  6. Refine vertical ( θ = 90o ) estimates  w  6 + (11 − 10) / 2   6.5 + (13 − 10) / 2   6         z   3.

we obtain a 2D dataset P (k .15) In words. With more noise. the convergence is very slow.Dove – Physics of Medical Imaging 10/8/2004 We are now done. this means that what we actually need is an expression of the inverse Radon transform f ( x. provides an answer to this burning question. what is the original image f ( x. Let F (k x .12) The variables k x and k y are orthogonal frequencies with units cycles (or radians) per distance. The method works. k y ) be the 2D Fourier transform of f ( x.θ ) .11) The projection theorem. The ART was the original reconstruction method used in medical imaging. k y ) (2. y ) . y ) exp −2π j ( k x x + k y y ) dxdy { } (2. Let P (k ) be the 1D Fourier transform of the projection pθ (r ) θ ∞ P (k ) = θ −∞ ∫ pθ (r ) exp {−2π j ( kr )}dr (2. y ) ? As p (r . and is not guaranteed. k y ) = −∞ −∞ ∫∫ f ( x. y ) ∞ ∞ F (k x .θ ) is the Radon transform of f ( x.3 Backprojection Method It is obvious that we desperately need a rigorous mathematical answer to the burning question that we all have: given the sinogram (family of projections) p (r . the set of 1D Fourier transform in r of the Radon transform of a function is the 2D Fourier transform of that function. This means that if all the Fourier Page 16 .θ ) = F ( k x .θ ) . 2. The projection θ theorem (or central slice theorem-CST) then states that P ( k .θ )} (2. but is slow and susceptible to noise. Allowing θ to vary from 0 to π and stacking all P (k ) together. the image data and projections match.4.14) with   k x = k cos θ   k y = k sin θ  2 2 k = k x + k y  (2. also called the central slice theorem or the central section theorem.13) The variable k is frequency (cycles or radians per distance). y ) = ℜ −1 { p (r .

θ ) . is equal to F (k x . Calculate the 2D inverse Fourier transform of F (k x . The reconstruction scheme based on the CST consists of the following steps: 1. y ) noisy. y ) .Dove – Physics of Medical Imaging 10/8/2004 transforms of the projections P (k ) of a function f ( x. Polar (a) and Cartesian (b) sampling grid. for all θ . In practice. consider the following outline of the algorithm thus far. which after resampling in θ Cartesian space. Practically this is the inverse FFT along each row (or column) for each column (or row). CT requires various coordinate system shifts and interpolations that complicate the calculations even further. 2. k y ) to yield f ( x. k y ) Figure 2. The result is the desired image. 2.10. To make this point more clear. In addition. Place all P (k ) on a polar grid to obtain P (k . There are two Page 17 . 3. In other words.θ ) . then it is θ possible to calculate the value of the function f ( x. the spatial interpolation produces enough noise in the image to make this direct implementation of the CST unacceptable for clinical diagnostic purposes. Proof of the CLT is straightforward (but a bit tedious) and in all CT textbooks. Calculate the 1D Fourier transform of all projections F1 { pθ (r )} = P (k ) . y ) are known.4. y ) for each point in the plane by calculating the inverse Fourier transform of P(k . θ Practically this is the FFT of the projection function in the θ direction. which makes the resultant estimate of f ( x.4 Fourier Reconstruction The computational problem with the Central Section Theorem is that a 2D inverse transform is required. re-sampling from polar to rectangular coordinates involves considerable interpolation.

1 Back Projection For reference. The second method is to use the filtered backprojection method. 2. both of which can lead to round-off and truncation errors.4.4. Unfortunately.5.5 Filtered Backprojection (FBP) 2.5 is reproduced below: Figure 2. Figure 2. This is illustrated in Figure 2. The first is to use smaller values for ∆θ and ∆s .Dove – Physics of Medical Imaging 10/8/2004 ways around this problem. it will cause many more calculations to be performed. Thus the measured values are “smeared” across the unknown density line as if a line of wet ink is drawn across the line. this will also increase the hardware requirements of the machines. which will decrease the noise introduced by the spatial interpolation scheme.5 (repeated for reference) In back projection the measurements obtained at each projection are projected back along the same line (or equivalently same θ ).11 for the case of an object consisting of a single point. Page 18 .

Essentially we are stating that we know that the point of density is somewhere along the line so that a “crude’ reconstruction will result if we assign the measured value along the entire line. If these delta functions are back-projected. we obtain the laminogram: fb ( x. Each projection of a delta function at the origin is also a delta function. We will now correct this distortion.Dove – Physics of Medical Imaging 10/8/2004 Figure 2. each projection is identical.17) Adding up these densities at all angles. Mathematically the back projection of a single measured projection along an unknown density is given by: bθ ( x. The δ (L) function is the Dirac delta function that has the following properties: δ (r ) =   1 if r = 0 0 otherwise (2. y )dθ 0 π =∫ π ∞ (2. y ) = ∫ pθ (r )δ ( x cosθ + y sin θ − r )dr (2. the resulting response (known as the impulse response) is simply: Page 19 . In this illustration. y ) = ∫ bθ ( x.11 Projections of a point at the origin back projected.18) 0 −∞ ∫ pθ (r )δ ( x cosθ + y sin θ − r )drdθ The laminogram represents a distorted picture of the anatomy through which the X-ray beam passes. y ) is the back-projected density due to the projection pθ (r ) .16) where bθ ( x.

” For example. weight the result with the 2D Fourier transform of the reciprocal of 1 . There have been two methods for accomplishing this r task.Dove – Physics of Medical Imaging 10/8/2004 hb (r ) = ∫ π π 0 −∞ ∫ δ (r )δ (r cos(θ − φ ) − r )drdθ 1 = r (2. and then perform the r inverse 2D Fourier transform of the product. The computational requirements and potential for noise introduction (due to re-sampling. F ( k .20) The double asterisk (**) signifies convolution. then the following relationship holds: r and assuming linear system theory fb ( x.20). The early reconstructions were based on Equation (2.20). Operations on generalized functions can be “esoteric. The first is to calculate the 2D Fourier transform of Equation (2.21) Page 20 . etc. however.θ ) = since the 2D Fourier transform of 1 is 1 .19) The result is derived after knowing something about generalized functions. with marginal results at best. δ ( x − xn ) f ′( xn ) Knowing that the response to an impulse is 1 applies.θ ) k (2. The Fourier transform of Equation (2.) are enormous. y ) = f ( x. the lack of knowledge of generalized functions does not invalidate the result. which is correct by all accounts. y ) is blurred by the 1 term. Thus. r k The 1 blurring must be removed. The second method is to use the filtered backprojection algorithm. Practically this means that the actual image f ( x. generalized functions predict that δ [ f ( x)] = ∑ where xn are the roots of f ( x) . the image that one calculates with the r backprojection method (the laminogram) is a blurred version of the actual image. y ) ** 1 r (2.20) is Fb (k .

26) It is necessary to use k instead of k because the integration includes negative values. We begin by restating the back-projection relationship for the laminogram fb ( x.26) with Equation (2. the Central Section Theorem. This is accomplished with our powerful r friend.24) yielding f ( x.θ )δ ( x cosθ + y sin θ − r )drdθ (2. y ) = ∫ ∫  ∫ F (k .4. y ) = ∫ π π 0 −∞ ∫ p(r.24).25) This can be modified to conform to the form of Equation (2. we obtain f b ( x.θ ) exp  j 2π k ( x cosθ + y sin θ ) k dkdθ (2. as expected.22) Restructuring this into a Fourier transform mode and applying the CST to substitute the inverse transform of P (k . y ) = ∫ π ∞ 0 −∞   ∫ F (k . y ) as f b ( x. y ) = ∫ π ∞ 0 −∞ ∫ F (k . Substituting ℑ{ pθ (r )} for F (k . we can rewrite the 2D Fourier transform in polar form.24) This equation is actually quite important.24).θ ) ∞  fb ( x. yielding 2π ∞ f ( x.θ ) exp [ j 2π k ( x cosθ + y sin θ )]dkdθ (2.2 Filtered backprojection method It is desirable to be able to use the elegant simplicity of back projection and to undo the 1 blur without requiring 2D transforms.θ ) in Equation (2.5.θ ) exp  j 2π k ( x cosθ + y sin θ ) kdkdθ 0 0 (2.θ ) exp ( i 2π kr ) dk δ ( x cosθ + y sin θ − r )drdθ 0 −∞  −∞  π ∞ (2. we see that they differ only by the k weighting term. and multiplying and dividing by k .23) Performing the integration over r. we obtain Page 21 . To appreciate this statement.Dove – Physics of Medical Imaging 10/8/2004 2. Comparing Equation (2. y ) =   ∫ ∫ F (k .

This term is the ramp filter. and 4. which is bad if the goal is to derive an image that faithfully reproduces the anatomy. the CT image “blows up” for each θ .28) This reconstruction is known as the filtered back-projection system. y ) = ∫ π ∞ ℑ{ gθ (r )} exp  j 2π k ( x cosθ + y sin θ )  k dk   k 0 −∞ ∫ (2. dθ becomes ∆θ . This accounts for the blurred reconstruction of fb ( x. The scheme is: 1. which is not stable (there are values of k for which things are not well behaved). The result is back projected.5. 3. The “fix” for this blurring is to pre-weight the transform of the projection prior to back projecting. 2. and dk becomes ∆k .29) =− Note that the inverse Fourier transform of the filter has a singularity at r=0.27) This equation states that the transform of each projection has been weighted by k along each radial line in Figure 2.28).3 Discrete FBP Of course digital computers require discrete representations of Equation (2. Each projection is individually transformed. y ) = ∫ π ∞ ℑ{ pθ (r )} k exp  j 2π k ( x cos θ + y sin θ )  k dk   k 0 −∞ ∫ (2. The transformed projection is weighted with k .4. y ) . Conversion would be straightforward if it were not for the pesky k term. that is f ( x. Is this important? Unfortunately it is.11). The result is inverse transformed. Page 22 .Dove – Physics of Medical Imaging 10/8/2004 f b ( x. consider the inverse Fourier transform of the filter: F {k }= −1 ∞ −∞ ∫ k exp ( j 2π kr ) dk 1 2π 2 r 2 (2. and is the current standard algorithm for calculating a tomographic image. The instability makes this filter unusable in practice. To see this.5 (or Figure 2. This system involves only a one-dimensional Fourier transform that is fast (because the FFT algorithm is actually used). 2. Going from continuous space to discrete space is usually straightforward: ∫ become ∑ .

The Ram-Lak filter can be written as the difference of a block and a triangle (Figure 2. substitute rn = n∆r and kn = 1 . To make the filter discrete.12(a). named after its inventors Ramachandran and Lakshiminarayanan.12: Decomposition of Ram-Lak filter (a) into the difference between a block (b) and a triangle (c).32) Page 23 . Applying a smoothing filter (usually the Shepp-Logan filter.30) becomes: (2∆r )  1  2  4(∆r )  q(rn ) = 0  1 − 2  (π n(∆r ) )  if n =0 if n is even if n is odd (2. Hamming. or other) suppresses the highest spatial frequency and reduces noise. A generalized Hamming window is given by α + (1 − α ) cos π k kN for k ≤ k N  H (k ) =  for k ≥ k N 0  (2. Thus Equation (2.12(a)) is known as the Ram-Lak filter. or other low-pass filters such as Hann. The resulting filter (graphed in Figure 2.Dove – Physics of Medical Imaging 10/8/2004 What to do? We know from our discussion on sampling that the measured data are limited to frequencies below the Nyquist frequency k N as shown in Figure 2.31) Usually frequencies below k N are corrupted by aliasing and noise. The inverse Fourier transform then gives the following: q (r ) = k N sin(2π k N r ) 1 − cos(2π k N r ) − 2π 2 r 2 πr (2.30) The resulting filter function has no singularity at r=0.12(b) and (c)). Figure 2.

spline. Back-project. the complete formula of discrete FBP can be written by combining a discrete convolution using the kernel in Equation (2. 6. y j ) = ∑  ∑ p (θ m . 5. this involves either a desired size.4. linear. Figure 2. N = 2*floor(size(P. this may involve interpolation if the size of the image is different from the size of the projection matrix. The interpolation scheme could be nearest neighbor. or other) in the Fourier (frequency) domain.33) 2. we have the Hann (or Hanning) window.1)/(2*sqrt(2))) where P is the projection matrix. Hann. Hamming.22) as M −1 N −1   f ( xi .13 shows the results of applying the Hann smoothing filter to the Ram-Lak filter. we have the Hamming window. 2.5 . Determine image size. or if unspecified. Filter the projections by multiplying the filter with the Fourier transformed projections (remember: convolution in the space domain is multiplication in the frequency domain). Calculate 1D FFT of the projections for each θ . and often defines the individual characteristics of the particular CT machine. 3. rn ) ⋅ q ( xi cos θ m + y j sin θ m − rn ) ∆r ∆θ m=0  n=0  (2. Page 24 .13 Product of a smoothing filter and the Ram-Lak filter. 4. or a calculation of the diagonal similar to diag=2*ceil(N/sqrt(2))+1. Scale the image elements (pixels) if necessary. Figure 2. Design a smoothing filter (Shepp-Logan.4 Implementation Actual implementation of the FBP algorithm might follow this outline (assuming parallel projections): 1.Dove – Physics of Medical Imaging 10/8/2004 For α = 0. or other.5. 7. and for α = 0. where N is the specified number rows. Add ad hoc image processing to prepare the image for ultimate clinical use.31) and a discrete back projection given in Equation (2. This ad hoc processing is usually vender specific. After all of this development.54 .

Page 25 . The second generation consisted of a small fan beam source. This generation of machines acquired data with rotation only.Dove – Physics of Medical Imaging 10/8/2004 2.15.14. and translation and rotation about the patient.6 Fan beam FBP The CT equipment has undergone various “generations” of development as illustrated in Figure 2.4. Figure 2. linear detector. This is illustrated in Figure 2. no translation was necessary. and a stationary detector array. and then linear translation across the patient. The third generation consisted of a wider fan beam (about 30°) where the fan beam could cover the entire object under study. This electron-beam CT (EBCT) shows high spatial and temporal resolution – there are no moving parts.14 Generations of X-ray CT machines (from Cho) The first generation was Hounsfield’s configuration and consisted of a pencil beam source and detector. There is a “fifth generation” machine that consists of an electronically steered electron beam swept across a target ring of tungsten. The fourth generation of machine consisted of a rotating fan beam.

pediatric. trauma and geriatric patients where speed of acquisition is essential. In Multi-Slice sequence scanning. The images from the EBCT machine are impressive. In a SVS sequence timed scans are acquired while the patient moves incrementally through the scanner. along with real time ECG cardiac triggering provides cine and flow capabilities.15 EBCT geometry From Imatron’s literature: Single Slice mode scanning utilizes 100 millisecond sweeps along one of the four target rings. without skipping a beat.Dove – Physics of Medical Imaging 10/8/2004 Figure 2. the Left Anterior Descending (LAD) coronary artery looks like: Page 26 . In a CVS sequence data is acquired by moving the patient continuously through the x-ray beam path. provides extremely high temporal resolution and high patient compliance without exposure to unnecessary dose. the electron beam sweeps along one to four target rings in rapid succession covering up to 8 centimeters of the body without moving the patient through the scanner. For example. This sequence provides two distinct acquisition features: Continuous Volume Scanning (CVS) and Step Volume Scanning (SVS). In SVS diagnosticians can utilize real time ECG triggering to provide a volume acquisition of the heart at a prescribed phase of the cardiac cycle. The ability of the ECG to trigger the Imatron EBT scanner at high heart rates. The exceptional scan speed of the Imatron EBT scanner provides the diagnostician with critical diagnostic information when imaging cardiac. airway physiology and blood flow analysis. This unique functionality. joint motion. This very fast scanning speed and temporal resolution is ideal for evaluating ventricular function.

from EBCT. Page 27 .17.16 LAD artery with lesion. but rather in fans as shown in Figure 2. The reconstruction algorithms that we have developed are all based on the assumption that the data have been acquired with parallel-beam geometry.17 Fan-beam geometry (a) and required measurement range (b). Data from modern machines are actually not ordered in parallel subsets. The fan angle is the formed by the fan. How can the mathematical models developed in the previous section be used in this new geometry? Three are two approaches.Dove – Physics of Medical Imaging 10/8/2004 Figure 2. Figure 2.

2) Unfortunately.Dove – Physics of Medical Imaging 10/8/2004 1. This nonlinearity is removed if we use the log of the measured transmission where ln I0 = µ ( z )dz I ∫ (3. Most machines compensate for this distortion on an ad hoc basis by using a nonlinear function of the log of the measurements. These data then require the same algorithm as developed above. These line integrals are the sum of the function in different directions. and the back projection involves a quadratic weighting factor rather than the uniform weighting of the parallel rays. All machines report derived linear attenuation coefficients using a CT number. and interpolate into parallel data. Adopt the FBP algorithm. and the beam is infinitesimally small. 2.). Record the fan beam data. In the case of X-ray attenuation. this is not going to work. Here the convolution kernel is slightly different.3) If the scale factor K=1000. The CT number of a pixel with attenuation coefficient µ x is: CT Number = K µ x − µ water µ water (3. we measure the exponent of the desired line integral as − µ ( z ) dz I = I 0e ∫ (3. then the units of the CT Number is Hounsfield Units (HU).1) Note that there is a nonlinear relationship between the measured projections and the desired line integral. Page 28 . The following table shows data for various materials.5 CT number The reconstruction methods studied so far indicate how a 2D function can be reconstructed from its line integrals. however. This process is known as re-binning. This approach is widely used in existing scanners. 2.1) was derived with the assumption that the source is mono-energetic. Equation (3. This approach uses a modified convolution backprojection system of Gullberg and Denton (see refs.

used in the earliest CT machines. The photo-diode is based on a pn-junction. Gas ionization chambers have a low quantum efficiency and slow response time. giving rise to a measurable electric current. consist of a scintillation crystal (NaI. Material Bone Water Muscle Fat Air CT Number (HU) µ (cm-1) 0. giving rise to measurable electric signals. The newest scanners typically use semiconductor detectors. this is called a sequential scan.21 0.7 Imaging in three dimensions 2. Page 29 . Gas ionization chambers consist of a pressurized gas chamber (Xe or Kr) with two electrodes. The resulting free electrons and ions drift towards the anode and cathode. Some of the incident X-ray photons cause ionization of the gas (by the photoelectric absorption). a good stability. electron-hole pairs are produced. high quantum efficiency.6 X-ray detectors There are three kinds of X-ray detectors used in modern CT equipment. and good reliability. This is illustrated in Figure 2.20 0.38 0. gas ionization chambers. The disadvantage is the low packing density. they are susceptible to scattered radiation and thus require external collimation. Scintillation detectors.0 808 0 -48 -142 -1000 2.1 CT Number and µ for various materials at 60 keV. The advantage is high packing density.Dove – Physics of Medical Imaging 10/8/2004 Table 2. A solid-state conductor consists of a scintillation crystal (CsI) and a photo-diode. These are scintillation detectors.18 0.7.1 Multiple 2D acquisitions The most straightforward way to image an entire volume is to scan number of consecutive slices. On the other hand. 2. These detectors have a very fast response time. When a scintillation photon passes through the depletion layer (boundary layer containing no mobile carriers). The advantages of this technique are high quantum efficiency. and solid-state (semiconductor) detectors. and fast time response.18(a) for a 360˚ acquisition. CaF2) in combination with a photo-multiplier tube.

but simultaneously the patient is slowly translated through the gantry. Let’s assume a rectangular slice thickness (or profile) of ∆z . the axial sampling density must obey the Nyquist criterion. Figure 2.18 Acquisition schemes: (a) multiple 2D acquisition. The pitch ratio is the ratio between the table feed and the slice thickness. Page 30 . In order to acquire a complete 3D data set containing all information needed to calculate a reconstruction. or in words. we need to acquire at least two slices per slice thickness. has several advantages: lower dose. The table feed (TF in the figure) is the axial distance over which the table translates during the time needed by the source to complete a 360˚ orbit. β represents the angular position of the X-ray tube.5).0 is used. 2. Data are acquired for discrete axial positions zi and for angular tube positions β ranging from 0 to 2π .7. This technique could use some smart biomedical engineer to continue its development. (c) 360° linear interpolation. and in order to not miss any lesions or other anatomical details. In practice.18(b) illustrates this type of data collection scheme. a pitch ratio between 1. We come to the conclusion that the maximum distance between slices is ∆z . A higher pitch.0 and 2. this technique has not lived up to the original expectations. Unfortunately. This means that the real data is convolved with an axial block of width ∆z . and less motion artifacts. for example. In this figure. and thus faster scanning. less tube load. 2 2. The tube describes a helical orbit relative to the patient.3 Cone-beam CT Some experimental machines have a 2D detector array in order to measure the entire volume on one single orbit of the X-ray source. data are simultaneously collected in β and in z. a slice thickness of 2 mm is combined with a table feed of 3 mm (pitch=1.Dove – Physics of Medical Imaging 10/8/2004 Figure 2. (b) 360° nearest neighbor. and z its axial position relative to the patient. The X-ray source rotates continuously around the patient.2 Spiral CT A technique that is widely used is spiral (or helical) CT.7.

and up to 1 mm axial resolution. Size of the detector elements. Reconstruction kernel (or convolution filter). c. Spectrum of the X-ray tube. A typical modern CT machine has sub-mm in-plane resolution. which determine the slice thickness. the major factors being: 1. which can be tuned to enhance high frequencies (for determining edges). Size of the focal spot. at the expense of more noise.2 Spatial resolution The spatial resolution of a CT image depends on several factors. 4. Increasing the number of detectors. The narrower the spectrum (more like mono-energetic photons) the better the contrast.and post-patient (axial) collimators. b. 2. Better resolution can be obtained by: a.1 Image contrast Image contrast depends on several factors. the size of the detector elements and the size of the focal spot determine the beam width. The effective linear attenuation coefficient of the tissue depends on the incident spectrum and therefore the contrast depends on the incident spectrum.8. Using a reconstruction kernel with less smoothing. Pre. The dynamic range of the hardware limits the contrast. at the expense of more noise. 2. which is the area on which the electrons hit the anode. Tube current (mAs) per slice (the more current supplied to the tube. the most important being: 1. limit the image contrast. but also the more radiation delivered to the patient). 3. 4. the better the contrast. or low frequencies (for attenuate noise and oth er artifacts) 5.8. 2. Page 31 . 3. etc. Nonlinearities such as scatter.Dove – Physics of Medical Imaging 10/8/2004 2.8 Image quality 2. Using a smaller pixel (or voxel) size.

Noise in the projections in CT exhibits a characteristic structure. Increasing the energy of the X-ray beam can reduce noise. The variance of the number of photons is proportional to the actual number of photons. (f) beam hardening. (e) view under-sampling. and can be described by the Poisson distribution. (g) scatter. and (i) object motion. (b) second phantom. (d) detector under-sampling.Dove – Physics of Medical Imaging 10/8/2004 Figure 2. Page 32 . (h) nonlinear partial volume effect. A high noise level results in alternating dark and light thin streaks radiating from locations with high attenuation (bright areas) as shown in Figure 2.19(c).19 Example of image artifacts: (a) test phantom. (c) noise. (unpublished results) 2.3 Noise The main source of noise is quantum noise. The number of photons passing through the patient and striking the detectors is susceptible to quantum fluctuations.8. therefore an increase in the number of photons used to improve contrast will result in more noise.

Page 33 .1 Under-sampling Taking too few samples results in two distinct phenomena (shown in Figure 2.8.8. If the number of views is too small.20 in which the white arrows point to the streak artifacts due to the partial volume effect. 2. Another example is shown in Figure 2.2 Beam hardening Low energy photons are preferentially absorbed.19(f)). The larger the attenuation gradients perpendicular to the beam. 2. Aliasing due to the number of detectors being too small. As a consequence. then streaks appear. The large white “blob” is highly absorbent contrast material in the patient’s stomach. 2.4. This results in streaks tangent to edges. This phenomenon is illustrated in Figure 2.8. every measurement represents an intensity that is averaged over the beam aperture. such as “cupping” and streaks (Figure 2.4. hardening artifacts appear.19 (h).Dove – Physics of Medical Imaging 10/8/2004 2.8.19): 1. This causes the X-ray beam too “harden” as it passes through the body. the larger the underestimate.4 Artifacts 2. This value is always an underestimate of the actual attenuation.4.3 Partial volume effects Due to a finite beam aperture. This average intensity is log-converted to give an effective beam attenuation.

Dove – Physics of Medical Imaging 10/8/2004 Figure 20. Page 34 . Image with artifactual streaks due to the partial volume effect.

time-invariant. In 1807 Fourier submitted a manuscript for publication describing the Fourier Series and Fourier Transform. and mathematics. Fourier was interested in diffusing the heat generated in cannons used by Napoleon Bonaparte. In 1753 Bernoulli criticized the use of trigonometric series to represent vibrations on strings. lumped (LTIL) system is a sinusoid. The manuscript was never published. Blood pressure transducers are designed based on Fourier analysis. The Babylonians used "trigonometric series" to predict astronomical events. Let's let u(t)=exp(st).Dove – Physics of Medical Imaging 10/8/2004 3.2) Page 35 . the list grows daily. Jean Baptiste Joseph Fourier entered this controversy around 1800. Circuits are design based on Fourier analysis. discontinuous slopes) using trigonometric series. Only Lagrange objected to the theory that signals can be represented as trigonometric series. but scaled in magnitude and shifted in phase. Appendix I – Fourier analysis 3. In 1748 L. including Laplace and Lagrange. this becomes: y (t ) = H ( s )e sτ (4. Euler preformed the same calculations that we will perform in the next section. and the magnitude of the vibrations was sinusoidal. then the output will also be a sinusoid of the same frequency.e. His criticism was based on his own belief that it was impossible to represent signals with corners (i.2 Response of LTIL system to eigenfunctions It is easy to show that if the input to a linear. The manuscript went to four reviewers. Thus we have: ∞ y (t ) = −∞ s ( t −τ ) st ∫ h (τ ) e d τ = e ∞ −∞ ∫ h (τ ) e − sτ dτ (4. Euler investigated the motion of a vibrating violin string and found that the vibrations along the length of the string were periodic. The ideas offered by Fourier have had an enormous impact in engineering. We know that a LTIL system is always characterized by the convolution of its impulse response h(t) and the input u(t). 3. The theory was finally given the "light of day" in Fourier's book The Analytical Theory of Heat published in 1822. If the input to a LTIL system is a complex exponential. A signal for which the system output is a (possibly complex) constant times the input is referred to as an eigenfunction of the system.1) By the definition of the Laplace transform. The list of engineering systems designed based on Fourier analysis is never-ending.1 Introduction: The development of Fourier analysis has a long history involving many individuals. and the amplitude factor is referred to as the system's eigenvalue. the output is also a complex exponential with only a change in (possibly complex) amplitude.. science.

4) (4.Dove – Physics of Medical Imaging 10/8/2004 The function H(s) is a complex constant the value of which depends on a complex number s=σ+jω. Let u (t ) = c1e s1t + c2 e s2t + c3e s3t Then by superposition we know that y (t ) = c1 H ( s1 )e s1t + c 2 H ( s 2 )e s2t + c3 H ( s3 )e s3t (4. 3. the constant H(s) for a specific value of s is the eigenvalue associated with the eigenfunction exp(st). If the input is written as u (t ) = ∑ c n e snt n (4. We can do an example to illustrate this concept. The value of H(s) and the system impulse response are a Laplace Transform pair.7) . We know however that complex exponentials are related to real functions. Furthermore.3) Extending this concept leads to a convenient expression for the response of a LTIL system.6) Thus if the input can be represented as a series of complex exponentials. In Fourier analysis the value of s is restricted to purely imaginary numbers. s= jω. that is. Euler proved that e jω 0t = cos(ω 0 t ) + j sin(ω 0 t ) Thus. each coefficient in this representation of the output is obtained as the product of the corresponding coefficient of the input cn and the system's eigenvalue H(sn).5) then the output will be written as y (t ) = ∑ cn H ( sn )e snt n (4.3 Fourier Analysis of Periodic Signals The value of the variable s in the above treatment is an arbitrary complex number. Thus in Fourier analysis only complex exponentials exp(jωt) are considered. complex exponential are periodic with period T0 given by Page 36 (4. Thus. then the output can also be represented as a series of complex exponentials.

To generalize this result. c2=c-2=1/2.10) Thus we see that the summation of a series of complex exponentials is equivalent to the summation of real periodic signals. In this case. Substituting these values and collecting terms.13) This can be further simplified by writing the coefficient cn in rectangular form as cn = An + jBn where An and Bn are real. Let’s express the signal as x(t ) = k = −3 ∑c e k 3 jk 2πt (4.Dove – Physics of Medical Imaging 10/8/2004 T = 2π = ω0 0 1 f (4. Fourier showed that x(t) can be written as x(t ) = n = −∞ ∑c e n ∞ − jnw0t (4.8) 0 Let’s look at an example of what this means. 7. in general. Eq. Eq.11) (4. complex. 10 can be written as x(t ) = 1 + 1 cos 2πt + cos 4πt + 2 cos 6πt 2 3 (4. and c3=c-3=1/3.9) where c0=1.12) The constant c is. c-1=c1=1/4. This can be rewritten as x(t ) = c0 + ∑ [cn e jnω0t + c−n e − jnω0t ] n =1 ∞ (4. 13 can be rewritten as: Page 37 . we get x(t ) = 1 + 1 (e j 2πt + e − j 2πt ) + 1 (e j 4πt + e − j 4πt ) + 1 (e j 6πt + e − j 6πt ) 4 2 3 Using Euler's relation in Eq. Consider a periodic signal x(t) with fundamental frequency ω0 of 2π.

In any finite interval of time. there are only a finite number of discontinuities. Parseval showed that for a continuous-time periodic signal x(t). then ∞ 1 | x(t ) | 2 dt = ∑ | a k | 2 ∫ TT k = −∞ (4.14) This is usually rewritten in the more familiar form x(t ) = a0 + ∑ an cos nω 0t + ∑ bn sin nω 0t n =1 n =1 ∞ ∞ (4. x(t) must be absolutely integrable.15) Equations 13 and 15 are equivalent if the following holds cn = an 2 −j bn 2 (4. The Dirichlet conditions for the existence of the Fourier Series are: 1. 3. the infinite series converges to the average value on either side of the discontinuity. Stated in another way. 2. except at isolated values of t for which x(t) is discontinuous. The frequency nωo rads/sec is said to be the n-th harmonic frequency. The Fourier coefficients an and bn are sometimes called the spectral coefficients of x(t). Thus. The fundamental frequency is ω0 rads/sec. the integral of x(t) over one period must be finite.17) The term on the left-hand side of this equation is the average power (energy per unit of time) in one period of the signal x(t). that is. x(t) must have a finite number of maxima and minima. Dirichlet proved that under certain conditions. These coefficients are a measure of the portion of the signal x(t) that is at each harmonic of the fundamental component.Dove – Physics of Medical Imaging 10/8/2004 ∞  ∞ x(t ) = A0 + 2∑ An cos nω 0t − ∑ Bn sin nω 0t  n =1  n=1  (4. In any finite interval of time.16) This is the Fourier Series in the trigonometric form. the average power of a periodic signal is Page 38 . Parseval's theorem proved that the total average power in a periodic signal equals the sum of the average powers in all of its harmonic components. Over any period. x(t) equals its Fourier Series representation. At these values.

There are at least two methods that can be used to find the coefficients.22) Thus the first term in the Fourier Series is simply the average value of the waveform.18)  0 if m ≠ n  cos nωt cos mωtdt =  T otherwise 2  T ∫ (4. T ∫ cos nωt sin mωtdt = 0 (4. but in practice.Dove – Physics of Medical Imaging 10/8/2004 conserved across the Fourier Series.23) Page 39 .21) If we multiply both sides of this equation by x(t) and integrate over T0 (one period) and take advantage of the fact that integrating a sine or cosine over one period is zero. the unknown coefficients an and bn must be found.19)  0 if m ≠ n  sin nωt sin mωtdt =  T otherwise 2  T ∫ (4. If we multiply both sides of the Fourier Series by cos(mω0t) and sin(mω0t).20) The Fourier Series is x(t ) = a 0 + ∑ a n cos nω 0 t + ∑ bn sin nω 0 t n =1 n −1 ∞ ∞ (4. Two functions are said to be orthogonal if their inner product is zero. then we get that 1 a0 = T0 T0 ∫ x(t )dt 0 (4. 3. For example. and take advantage of the orthogonality of sinusoids. however. we get: 2 an = T0 T0 ∫ x(t ) cos(nω t )dt 0 0 (4. Thus. and power will not be lost. The original method of Fourier.4 Calculating the Fourier Coefficients In order to take advantage of the power of Fourier analysis. was based on the orthogonality of trigonometric functions. one can work in the time domain or the frequency domain. one uses the Fast Fourier Transform (more later).

If the signal is periodic with period T0 seconds.” Prentice-Hall.25) The cn coefficients are found from solving this equation. O. Remember that Euler said that cos x = exp( jx) + exp(− jx) exp( jx) − exp(− jx) and sin x = 2 2j Using the definition of cn in Eq. this is the reason that the algorithm is fast. performing this integration for each harmonic can be time consuming. One can buy an FFT on a chip for just a few dollars. First most biomedical signals are not periodic. Second. 1974). and then altering the signal in the new domain. 16. then f(t) can be written as a Fourier Series. The FFT algorithm was proposed by Cooley and Tukey in the early 1960's (see Brigham. This algorithm calculates the complex coefficients cn at a blazingly fast speed. Euler's relations for the cosine and sine can be used to develop a practical approach for calculating the Fourier Series coefficients cn. Page 40 .Dove – Physics of Medical Imaging 10/8/2004 bn = 2 T0 T0 ∫ x(t ) sin(nω t )dt 0 0 (4.5 Fourier Analysis of Long and Possibly Non-periodic Waveforms Given a time function f(t). E.24) Thus we have an easy way to calculate the Fourier Series coefficients for any periodic waveform x(t). one can predict the response of a linear system (filter) by first converting the time function to the frequency domain. The algorithm is also easily implemented in hardware. the Fourier Series coefficients can be written as 1 cn = T0 T0 ∫ f (τ ) exp(− jnω τ )dτ 0 0 (4. or from the Fast Fourier Transform (FFT) algorithm. This problem grows quadratically as T0 increases. Unfortunately there are two problems with this formulation. 3. “The Fast Fourier Transform. The algorithm is easily implemented in software using only a few multiplications and many shifts. or if a computer is used to sample the function with a sampling window of T0 seconds in width.

The coefficients from the FFT algorithm (assuming that scaling has already been applied) are thus arranged as a vector C of complex numbers given by  c0  c  1  M   cN p 2   c− N p 2   M  c  −1             C= (4. Thus the data array will be  data 0 + j 0   data + j 0  2     M   data N p −2 + j 0  data N −1 + j 0  p   D= (4. the number Np must be a power of two (2). Thus the total number of samples is Np = T0 fs. data1. etc. 128. The real parts will be the sampled data. or equivalently. The array will usually be complex with the imaginary parts all zero. Once the C coefficients (or equivalently the an and bn coefficients) are obtained. the points in the vector of data passed to the FFT algorithm must be scaled by Np. The results from the FFT will be arranged as n=0 to Np/2. cycle in this sampling window. and for performing Fourier analysis of non-periodic signals. 1024. where Np is the number of points sampled in T0 seconds. one must scale the results (output) of the FFT by 1/Np. 256.26) 3. Assemble the Np samples data0. these are followed in reverse order as n=-(Np/2) to -1. Calculate the values of cn using the FFT algorithm. Thus the cn coefficients from the FFT must be multiplied by 1/Np. 20). The fundamental frequency is that frequency that produces one complete = 1 .27) Thus. Page 41 . is as follows: 1. Sample the waveform for T0 seconds at a sampling rate of fs samples per second. one can reconstruct the time function from the Fourier Series equation (Eq. The value of T0 is the sampling window.Dove – Physics of Medical Imaging 10/8/2004 3.6 Practical Procedure for Performing Fourier Analysis The procedure for handling large amounts of data. such as 32. only the first Np/2 points in C are useful. and in most implementations. In the conventional FFT algorithm. By convention. 64. Thus T = 2π 0 ω0 f 0 2. … into an array.

This graph is sometimes called the line spectrum. If x(t)=sin(ωt). FN/2. and the variable N is the total number of samples. one can either pad the data sequence with zeros. If this is not the case. in seconds. or truncate the sequence to a power of two. Often the values of cn are displayed as a graph with the magnitude of cn on the ordinate. or use the integration formula for calculating the a and b coefficients directly. Page 42 . There are many FFT algorithms available. The MATLAB package has a discrete FT (DFT) algorithm that calculates the coefficients for sampled data. The first two columns are digital frequencies. The corresponding frequencies of the DFT are summarized in the following table. Frequencies of DFT Components in Different Units Hz-sec Symbol Frequency of F1 Frequency of Fm Frequency of FN/2 Sampling Frequency N Rad ω 2π/N 2πm/N Hz f 1/NT m/NT 1/2T 1/T rad/sec W 2π/NT 2πm/NT 1/N m/N 0. and the last two columns are analog frequencies.Dove – Physics of Medical Imaging 10/8/2004 In sum. or C++. The frequency Hz-sec is obtained by multiplying the frequency in Hz by the time Tk. then x[k]=sin(WTk). of the DFT of any real series is exactly onehalf the sampling frequency in accordance with Shannon's Sampling Theorem. The IMSL library has an easy-to-use algorithm that is callable from FORTRAN. the FFT yields a scaled version of the cn coefficients. From these. The two are related by ω=WTk where ω is the analog frequency and W is the digital frequency.5 1. One can then use the Fourier Series representation to reconstruct the time waveform after filtering.0 π 2π π/T 2π/T The frequency of the final component. one can filter the time signal f(t) by altering the cn coefficients in the frequency domain. C. Note that in the Table the variable Tk is the time step (in seconds) between data samples. The period of the Fm component is just the reciprocal of the frequency in all cases. Thus. All of the available FFT routines require that the number of points (samples) passed to the routine be a power of 2. the inter-sample time. and frequency on the abscissa. the period of Fm is N/m samples or NT/m seconds.

It turns out that Rxx(0) is the variance of the Page 43 . 1988) Pxx ( f ) = FT {R xx (t )} (4. The units of power density are therefore power per hertz. The function Rxx{t} is the autocorrelation function given by R xx (t ) = E f (τ ) ∗ f (τ + t ) { } (4. The area of the periodogram for positive frequencies gives the total power in the random signal.28) FT{} represents the Fourier transform.Dove – Physics of Medical Imaging 10/8/2004 3. The units of power are squared units of the sampled data xn. For finite data lengths. which assumes that M. the power spectral density (PSD).” Prentice-Hall.29) where E{} represents the expected value (expectation) operator. In contrast to the methods for estimating the line spectrum. of a wide-sense stationary random process xn is defined as (Kay.7 Frequency Analysis of Random Signals The above analysis is not the same as finding the power spectrum. The superscript * represents the complex conjugate operator. Spectral estimation is the determination of the spectral content of a random process based on a finite set of observations from the process. Formally. Englewood Cliffs. the number of samples. The PSD represents the distribution of power with frequency of the random process. For this case. This is called the periodogram and is similar to a probability distribution function that describes how the power of a random signal is distributed over frequency. This is not the same as the line spectrum that describes how the time signal is decomposed into its sinusoidal components.M.30) This new definition says to take the magnitude squared of the Fourier transform of the data (calculated with the FFT) and then divide by the record length. no one method is universally accepted for estimating the PSD. the autocorrelation function cannot be found exactly . “Modern Spectral Estimation. There are many other methods for estimating the PSD. is large. which is denoted by Pxx(f).only estimates are available. S. Often the values of Pxx(f) are displayed on a graph with the magnitude (in dB) of the PSD on the ordinate and frequency (in Hz) on the abscissa. the definition given for the autocorrelation function is changed to the following: Pxx ( f ) = 1 FT {x (t )}2 M [ ] (4. This approach is the one used by MATLAB.

In sum. “Signal Processing Algorithms.32). then Fourier analysis states: f ( x) = a0 + ∑ an cos ( nω 0 x ) + ∑ nn sin ( nω 0 x ) n =1 n =1 ∞ ∞ (4. (Stearns. and an infinite number of cosine waves. If the signal is discrete. if the signal f ( x) is periodic. then the Fourier Transform becomes the Discrete Fourier Transform (DFT) that is given below (copied from MATLAB help files) Page 44 .31) or Equation (4. David. Englewood Cliffs. is also the total power.Dove – Physics of Medical Imaging 10/8/2004 random signal. 1988).. Fourier analysis is based on the idea that one can decompose a signal (either a function of time.” Prentice-Hall. or space) into the summation of sine and cosine waves. S. an infinite number of weighted sine waves. Figure A – Decomposition of signal into summation of sine and cosine functions. R. For example.A. if the signal is Gaussian. which.32) The notion of Fourier transform is graphically illustrated below in Figure A in which the complex signal is decomposed into its constituent sinusoids of different frequencies by applying either Equation (4. Fourier theory supports the statement that any continuous signal can be exactly written as the sum of a constant.31) If f ( x) is not periodic.D. then one can write that the Fourier Transform of f ( x) is: F (ω ) = ∞ −∞ ∫ f ( x) exp ( − jω x ) (4.

as ours will most likely be. from MATLAB): where There is also a two-dimensional Fourier transform that is given by the following discrete equation (again copied from MATLAB): As an example. then the above equation becomes (again.Dove – Physics of Medical Imaging 10/8/2004 If the signal is real. consider the 2D figure below. Figure B – Test figure from MATLAB Page 45 .

grid on xlabel('cycles/year') title('Periodogram') period = 1. freq = (1:N/2)/(N/2)*nyquist. respectively.1). nyquist = 1/2.Dove – Physics of Medical Imaging 10/8/2004 The result of applying the 2D DFT to the test figure is shown below. plot(year. plot(freq.2).power). The actual calculation of the DFT is performed with an algorithm called the Fast Fourier Transform (FFT) invented by Tukey and Cooley.power). N = length(Y). grid on ylabel('Power') xlabel('Period(Years/Cycle)') [mp index] = max(power). Figure C – 2D DFT of object in Figure B. wolfer = sunspot(:.dat year = sunspot(:./freq. Try the following MATLAB code: load sunspot. figure.wolfer) title('Sunspot Data') Y = fft(wolfer). plot(period. power = abs(Y(1:N/2)). The MATLAB function fft() and ifft() calculate the forward FFT and inverse FFT. axis([0 40 0 2e7]). figure. Y(1) = []. period(index) Page 46 .^2.

Hose. (ed) The Physics of Medical Imaging. Shung. Smallwood. Medical Imaging Systems. D. New York. 1883. This reference is complete. A. the book is out of print. Englewood Cliffs. so sad. but useful. San Diego. barber. Page 47 . B. Berger.. Principles of Medical Imaging. Too bad. this text is out pf print. and E.) Introduction to Bioengineering. S. Brown. R. but we could not get more than 30 from the publisher. Institute of Physics Publishing. and D. Unfortunately. Lewis (eds. 1999. This is a complete text covering all aspects of medical imaging. This is a wonderful brief text from which many of the images are drawn. This was to be our textbook for the semester.Dove – Physics of Medical Imaging 10/8/2004 References 1. 5. Unfortunately. but it is written at a beginning graduate level. 1996. Tsui. Unfortunately. 4. 2. Webb. There are about 30 copies in the IMU bookstore. 1992. The section on medical imaging is brief. S. Institute of Physics Publishing. M. 1992. K. W. Prentice-Hall.. Academic Press. Macovski. Bristol. Goldsmith. Lawford. This small text contains the mathematical treatment of imaging that we all need to know. Smith. B. P. Oxford. Bristol. Medical Physics and Biomedical Engineering. this text is out of print. 3.