You are on page 1of 22

C. A.

Bouman: Digital Image Processing - January 29, 2013 1


Tomography
Many medical imaging systems can only measure projec-
tions through an object with density f(x, y).
Projections must be collected at every angle and dis-
placement r.
Forward projections p

(r) are known as a Radon trans-


form.
p (r)

x
y
r
Objective: reverse this process to form the original image
f(x, y).
Fourier Slice Theorem is the basis of inverse
Inverse can be computed using convolution back pro-
jection (CBP)
C. A. Bouman: Digital Image Processing - January 29, 2013 2
Medical Imaging Modalities
Anatomical Imaging Modalities
Chest X-ray
Computed Tomography (CT)
Magnetic Resonance Imaging (MRI)
Functional Imaging Modalities
Signal Photon Emission Tomography (SPECT)
Positron Emission Tomography (PET)
Functional Magnetic Resonance Imaging (fMRI)
C. A. Bouman: Digital Image Processing - January 29, 2013 3
Multislice Helical Scan CT
Multislice CT has a cone-beam structure
X-ray Source
Detector Array
Path of Helical Scan
Plane of Desired
Image Reconstruction
C. A. Bouman: Digital Image Processing - January 29, 2013 4
Example: CT Scan
Gantry rotates under berglass
cover
3D helical/multislice/fan beam
scan
C. A. Bouman: Digital Image Processing - January 29, 2013 5
Photon Attenuation
XRay Source
x
0
Y
Material with density u(x)
T

Pin Hole Columnator


x
x - depth into material measured in cm
Y
x
- Number of photons at depth x

x
= E[Y
x
]
Number of photons is a Poisson random variable
P{Y
x
= k} =
e

k
x
k!
.
As photons pass through material, they are absorbed.
The rate of absorption is proportional to the number of
photons and the density of the material.
C. A. Bouman: Digital Image Processing - January 29, 2013 6
Differential Equation for Photon Attenuation
XRay Source
x
0
Y
Material with density u(x)
T

Pin Hole Columnator


x
The attenuation of photons obeys the following equation
d
x
dx
= (x)
x
where (x) is the density in units of cm
1
.
The solution to this equation is given by

x
=
0
e

_
x
0
(t)dt
So we see that
_
x
0
(t)dt = log
_

0
_
log
_
Y
x

0
_
C. A. Bouman: Digital Image Processing - January 29, 2013 7
Estimate of the Projection Integral
XRay Source
x
0
Y
Material with density u(x)
T

Pin Hole Columnator


x
A commonly used estimate of the projection integral is
_
T
0
(t)dt

= log
_
Y
T

0
_
where:

0
is the dosage
Y
T
is the photon count at the detector
C. A. Bouman: Digital Image Processing - January 29, 2013 8
Positron Emission Tomography (PET)
Detector i
Detector i
- detection rate A
ij
- emission rate
j
x
j
x
E[y
i
] =

j
A
ij
x
j
Subject is injected with radio-active tracer
Gamma rays travel in opposite directions
When two detectors detect a photon simultaneously, we
know that an event has occurred along the line connecting
detectors.
A ring of detectors can be used to measure all angles and
displacements
C. A. Bouman: Digital Image Processing - January 29, 2013 9
Example: PET/CT Scan
Generally low space/time resolu-
tion
Little anatomical detail couple
with CT
Can detect disease
C. A. Bouman: Digital Image Processing - January 29, 2013 10
Coordinate Rotation
Dene the counter-clockwise rotation matrix
A

=
_
cos() sin()
sin() cos()
_
Dene the new coordinate system (r, z)
_
x
y
_
= A

_
r
z
_
Geometric interpretation
y
x

r
z
Inverse transformation
_
r
z
_
= A

_
x
y
_
C. A. Bouman: Digital Image Processing - January 29, 2013 11
Integration Along Projections
Consider the function f(x, y).
x
y
f(x,y)
We compute projections by integrating along z for each r.
y
x

r
z
The projection integral for each r and is given by
p

(r) =
_

f
_
A

_
r
z
__
dz
=
_

f (r cos() z sin(), r sin() + z cos()) dz


C. A. Bouman: Digital Image Processing - January 29, 2013 12
The Radon Transform
The Radon transform of the function f(x, y) is dened as
p

(r) =
_

f (r cos() z sin(), r sin() + z cos()) dz


The geometric interpretation is
p (r)

x
y
r
f(x,y)
Notice that the projection corresponding to r = 0 goes
through the point (x, y) = (0, 0).
C. A. Bouman: Digital Image Processing - January 29, 2013 13
The Fourier Slice Theorem
Let
P

() = CTFT {p

(r)}
F(u, v) = CSFT {f(x, y)}
Then
P

() = F( cos(), sin())
P

() is F(u, v) in polar coordinates!

u
v
C. A. Bouman: Digital Image Processing - January 29, 2013 14
Proof of the Fourier Slice Theorem
By denition
p

(r) =
_

f
_
A

_
r
z
__
dz
The CTFT of p

(r) is then given by


P

() =
_

(r)e
j2r
dr
=
_

_
_

f
_
A

_
r
z
__
dz
_
e
j2r
dr
=
_

f
_
A

_
r
z
__
e
j2r
dzdr
We next make the change of variables
_
r
z
_
= A

_
x
y
_
.
Notice that the Jacobian is |A

| = 1, and that r = x cos()+


y sin(). This results in
P

() =
_

f (x, y) e
j2[x cos()+y sin()]
dxdy
=
_

f (x, y) e
j2[x cos()+y sin()]
dxdy
= F( cos(), sin())
C. A. Bouman: Digital Image Processing - January 29, 2013 15
Alternative Proof of the Fourier Slice Theorem
First let = 0, then
p
0
(r) =
_

f(r, y) dy
Then
P
0
() =
_

p
0
(r)e
2jr
dr
=
_

_
_

f(r, y) dy
_
e
2jr
dr
=
_

f(r, y)e
2j(r+y0)
dr dy
= F(, 0)
By rotation property of CSFT, it must hold for any .
C. A. Bouman: Digital Image Processing - January 29, 2013 16
Inverse Radon Transform
Physical systems measure p

(r).
From these, we compute P

() = CTFT{p

(r)}.

v
u
at each angle
Transformed Projection
Next we take an inverse CSFT to form f(x, y).
Problem: This requires polar to rectagular conversion.
Solution: Convolution backprojection
C. A. Bouman: Digital Image Processing - January 29, 2013 17
Convolution Back Projection (CBP)
Algorithm
In order to compute the inverse CSFT of F(u, v) in polar
coordinates, we must use the Jacobian of the polar coor-
dinate transformation.
du dv = ||d d
This results in the expression
f(x, y) =
_

F(u, v)e
2j(xu+yv)
dudv
=
_

_

0
P

()e
2j(x cos()+y sin())
||d d
=
_

0
_
_

||P

()e
2j(x cos()+y sin())
d
_
. .
g

(x cos()+y sin())
d
Then g(t) is given by
g

(t) =
_

||P

()e
2jt
d
= CTFT
1
{||P

()}
= h(t) p

(r)
C. A. Bouman: Digital Image Processing - January 29, 2013 18
where h(t) = CTFT
1
{||}, and
f(x, y) =
_

0
g

(x cos() + y sin()) d
C. A. Bouman: Digital Image Processing - January 29, 2013 19
Summary of CBP Algorithm
1. Measure projections p

(r).
2. Filter the projections g

(r) = h(r) p

(r).
3. Back project ltered projections
f(x, y) =
_

0
g

(x cos() + y sin()) d
C. A. Bouman: Digital Image Processing - January 29, 2013 20
A Closer Look at Projection Filter
1. At each angle, projections are ltered.
g

(r) = h(r) p

(r)
2. The frequency response of the lter is given by
H() = ||
3. But real lters must be bandlimited to || f
c
for some
cut-off frequency f
c
.
H( ) = | |
fc fc

So
H() = f
c
[rect (f/(2f
c
)) (f/f
c
)]
h(r) = f
2
c
_
2sinc(t2f
c
) sinc
2
(tf
c
)

C. A. Bouman: Digital Image Processing - January 29, 2013 21


A Closer Look at Back Projection
Back Projection function is
f(x, y) =
_

0
b

(x, y) d
where
b

(x, y) = g

(x cos() + y sin())
Consider the set of points (x, y) such that
r = x cos() + y sin()
This set looks like
x
setofpoints
alongasingle
backprojection
q
y
r
Along this line b

(x, y) = g

(r).
C. A. Bouman: Digital Image Processing - January 29, 2013 22
Back Projection Continued
For each angle back projection is constant along lines of
angle and takes on value g

(r).
x
setofpoints
alongasingle
backprojection
q
y
g
(
r
)
q
r
0
r
Complete back projection is formed by integrating (sum-
ming) back projections for angles ranging from 0 to .
f(x, y) =
_

0
b

(x, y) d


M
M1

m=0
b
m
M
(x, y)
Back projection smears values of g(r) back over image,
and then adds smeared images for each angle.