Reconstruction

The Problem

In imaging we measure g(l,q) and need to determine f(x,y)

 
g (l , )    f ( x, y)   ( x cos   y sin   l )dxdy
  
q y
p

??
g(q,l)
x

l f(x,y)
0
 Back Projection Method
A little trick that almost works!

Object Back Projection

 Back Projection Method
A little trick that almost works!
Back Projection

Object

We do this for every angle and then add

together all the back projected images
 Back Projection Method
Step #1: Generate a complete an image for each projection (e.g. for
each angle q)
b ( x, y )  g ( x cos( )  y sin( ), )
These are called back projected images

Step #2: Add all the back projected images together


f b ( x, y )   b ( x, y )d
0
 Back Projection Method
Original object Reconstructed object
Reconstructed Image

Kind of worked but we need to do better than

this. Need to come up with a better
reconstruction algorithm.
 Projection-Slice Theorem
This is a very important theorem in CT imaging
First take the 1D Fourier transform a projection g(l,q)


G (  , )  1D  g (l ,    g (l ,  ) e  j 2l
dl


g (l ,  )
 Projection-Slice Theorem
This is a very important theorem in CT imaging
First take the 1D Fourier transform a projection g(l,q)

G (  , )  1D  g (l ,    g (l ,  ) e  j 2l
dl


Next we substitute the Radon transform for g(l,q)

 
g (l , )    f ( x, y)   ( x cos   y sin   l )dxdy
  
  
G (  , )  
   
f ( x, y )   ( x cos   y sin   l )e  j 2l dxdydl
 Projection-Slice Theorem
This is a very important theorem in CT imaging
First take the 1D Fourier transform a projection g(l,q)

G (  , )  1D  g (l ,    g (l ,  ) e  j 2l
dl

Next we substitute the Radon transform for g(l,q)
 
g (l ,  )    f ( x, y)   ( x cos   y sin   l )dxdy
  
  
G (  , )  
   
f ( x, y )   ( x cos   y sin   l )e  j 2l dxdydl

Next we do a little rearranging

 
 
G (  ,  )    f ( x, y )   ( x cos   y sin   l )e  j 2l
dl dxdy
     
 Projection-Slice Theorem
This is a very important theorem in CT imaging
Next we do a little rearranging
 
 
G (  ,  )    f ( x, y )   ( x cos   y sin   l )e  j 2l
dl dxdy
     
Applying the properties of the delta function
 
G (  , )    
f ( x, y ) e  j 2  x cos  y sin   dxdy
  

 
This looks a lot like
F (u, v)   f ( x, y )e  j 2  xu  yv  dxdy
  

with u   cos( ), v   sin( )

 Projection-Slice Theorem
This is a very important theorem in CT imaging
 
G (  , )   f ( x, y )e  j 2  x cos  y sin  dxdy
  

G (  , )  F (  cos( ),  sin( ))

Question: So what does this mean?

Answer: If I take the 1D FT of a projection at an angle q the

result is the same as a slice of the 2D FT of the original
object f(x,y)
 Projection-Slice Theorem
If I take the 1D FT of a projection at an angle q the result is the same as a
slice of the 2D FT of the original object f(x,y)
u   cos( ), v   sin( )

2D FT

qo

f(x,y) F(u,v)
qo
g (l , o )


G (  ,  o )  1D  g (l ,  o )   g (l ,  o ) e  j 2l
dl

u   cos( ), v   sin( )
v
  u 2  v 2 ,   tan 1  
u
 The Fourier Reconstruction Method
2D IFT

f(x,y) F(u,v)
g (l ,  o ) qo
u   cos( ), v   sin( )


G (  , o )  1D  g (l , o )   g (l ,  o ) e  j 2l
dl


Take projections at all angles q.

Take 1D FT of each projection to build F(u,v) one slice at a time.
Take the 2D inverse FT to reconstruct the original object based on F(u,v)

f ( x, y )   1
2D  G (  , )
 Image Reconstruction
Using Filtered Backprojection

g ( , l )

Filter
t


Backprojection
 Filtered Back Projection
The Fourier method is not widely used in CT because of the
computational issues with creating the 2D FT from projections.
However, the method does lead to a popular technique called
filtered back projection.
 
j 2  xu  yv 
f ( x, y )   
  
F (u , v ) e dudv

In polar coordinates the inverse Fourier transform can be written as

2 
j 2  x cos  y sin 
f ( x, y )  
0 
 F (  cos  ,  sin  ) e dd

with u   cos( ), v   sin( )

 Filtered Back Projection
The Fourier method is not widely used in CT because of the
computational issues with creating the 2D FT from projections.
However, the method does lead to a popular technique called
filtered back projection.
In polar coordinates the inverse Fourier transform can be written as
2 
j 2  x cos  y sin 
f ( x, y )  
0 
 F (  cos  ,  sin  ) e dd

with u   cos( ), v   sin( )

From the projection theorem G (  ,  )  F (  cos( ),  sin( ))

2 
j 2  x cos  y sin 
We can write this as f ( x, y )   
0 
G (  ,  ) e dd
 Filtered Back Projection
The Fourier method is not widely used in CT because of the computational
issues with creating the 2D FT from projections. However, the method does
lead to a popular technique called filtered back projection.

2 
j 2  x cos  y sin  
We can write this as f ( x, y )  
0 
 G (  ,  ) e dd

Since g (l , )  g (l ,   ) you can show

 
f ( x, y )    G (  , )e j 2  x cos  y sin   dd
0 

which can be rewritten as




f ( x, y )     G (  ,  )e j 2 l
 d  d
0    l  x cos( )  y sin( )
 Filtered Back Projection
And Back Projection
A. Back Projection
b ( x, y )  g ( x cos( )  y sin( ), )

f b ( x, y )   b ( x, y )d
0

B. Filtered Back Projection


G (  , )  1D  g (l ,    g (l ,  ) e  j 2l
dl



 
f ( x, y )     G (  ,  )e j 2 l
 d  d
0   l  x cos( )  y sin( )
 Filtered Back Projection Method
Filtered Back Projection

This always works!

Object

Digital Filter
1) take 1D FFT of projection
2) multiply by ramp filter
3) take 1D inverse FFT
4) make a back projection
 Filtered Back Projection Method
Always works! Filtered Back Projection

Object

Digital Filter
1) take 1D FFT of projection
2) multiply by ramp filter
3) take 1D inverse FFT
4) make a back projection
 Filtered Back Projection Method
Always works! Filtered Back Projection

Object

Digital Filter
We do this for every angle 1) take 1D FFT of projection
and then add together all 2) multiply by ramp filter
the filtered back projected 3) take 1D inverse FFT
images 4) make a back projection
 Filtered Back Projection
verses Back Projection
A. Back Projection
b ( x, y )  g ( x cos( )  y sin( ), )

f b ( x, y )   b ( x, y )d
0

B. Filtered Back Projection


G (  , )  1D  g (l ,    g (l ,  ) e  j 2l
dl



 
f ( x, y )     G (  ,  )e j 2 l
 d  d
0   l  x cos( )  y sin( )
 Convolution Back Projection
From the filtered back projection algorithm we get

 
f ( x, y )     G (  ,  )e j 2 l
 d  d
0   l  x cos( )  y sin( )

It may be easier computationally to compute the inner 1D IFT

using a convolution

 
recall 1D F1 ( )  F2 ( )  f1 ( x)  f 2 ( x)
1



f ( x, y )   g (l ,  )  1D1 (  )  l  x cos( )  y sin( )
d
0
 Convolution Back Projection


f ( x, y )   g (l ,  )  1D1 (  )  l  x cos( )  y sin( )
d
0

c(l )  1D1 (  )
Let 
f ( x, y )    g (l , )  c(l ) l  x cos( )  y sin( ) d
0
 Convolution Back Projection
c(l )  1D1 (  )

f ( x, y )    g (l ,  )  c(l ) l  x cos( )  y sin( ) d
0

The problem is c(l )  1D (  )    e j 2l d does not exist

1


 Convolution Back Projection
c(l )  1D1 (  )

f ( x, y )    g (l ,  )  c(l ) l  x cos( )  y sin( ) d
0


The problem is c(l )  1D (  )    e j 2l d does not exist

1




The solution c~ (l )  1D (  W (  ))    W (  )e j 2l d

1



where W (  ) is called a weighting function


f ( x, y )    g (l ,  )  c~ (l ) d
0


c~ (l )  1D1 (   W (  ))    W (  )e j 2l d

 Convolution Back Projection

f ( x, y )    g (l ,  )  c~ (l ) d
0


c~ (l )  1D1 (   W (  ))    W (  )e j 2l d


Common window functions

 Hamming window
 Lanczos window (sinc function)
 Simple rectangular window
 Ram-Lak window
 Kaiser window
 Shepp-Logan window