0% found this document useful (0 votes)

16 views90 pages

VI. Reconstruction PET

Cours

Uploaded by

Makhlouf

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

16 views90 pages

VI. Reconstruction PET

Cours

Uploaded by

Makhlouf

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 90

Imaging using radiations

VI – image reconstruction in PET

Patrice Laquerriere
[email protected]

Cours M2 PRIDI, année 2021-2022

Data sampling
2RD

z Transverse Plane

Δdt
LD 2RD

fan-beam sampling
Δdt
Equal Angles: Δdt / 2RD

Δdz
Data sampling

Δxr
2
 x 
xr  d t 1   r 
 RD 

RD
xr d t  2
Nd Detector Number per ring
Data sampling
if Nd is even

Sinogram


++++++++++++++++ 0f 
++++++++++++++++
Nf  N d / 2
f N xr  N d / 2
f   / N f
0 x r  2 R / N x r  d t / 2
xr
Data sampling
In the axial direction

d z  LD / N R Ring Number

N R2 combinations possible for oblique sinograms

z  r1  r2  / 2 For the central plane

We label the ring with r1 and r2 or z, Δz with
r  r1  r2

Relation projection - sinogramme

p  xr , yr , f ,    s  xr , f , z, r 

yr  z cos 
with r
tan  
2 RD2  xr2
Data Reduction
Size of a complet set of data

2 Pixel size
N 
N T  N R2   d   N b
 2 

Number of sinogram Pixel Number in the sinogram

N R  24
Typically, we have N d  512 N T  72 Mo
Nb  2

Data Reduction needed

Data Reduction
In the transverse plane

Δxr = transverse intrinsic sampling

The Nyquist frequency give us the maximum spatial frequency that can be correctly estimated

1
N 
2 x r

In practice, c   N

Angular Reduction

Nf  N f  2  m avec f   f  2 m

The m factor is called the ”mashing” factor

m = 0, 1, 2 … m=2 reduction by 4
Data Reduction
Data Reduction
For the oblique sinograms

Introduction of the Michelogram

Δr = 21 Δr = 14 Δr = 7 Δr = 0
32

24 Δr = -7
In this example:
32 rings
Δr = -14 Span = 7
r2 16
Δrmax = 24
7 segments
Δr = -21
8

1
1 8 16 24 32

r1
Image Reconstruction
for a 3D acquisition

Two main différences majeures between 2D and 3D

Data Redondancy
Spatial Variation


p  xr , y r , f ,     f ( x, y, z )dz r

yr

xr
z
  x    sin f  cos f sin  cos f cos    xr 
zr  y    cos f  sin f sin  sin f cos    y r 
    
 z   0 cos  sin    zr 
 y
f
x
Data Redondancy

In the case of a 3D acquisition,

Enough data (segment 0) are recorded to reconstruct the image volume.
The other segments (oblique sinogram) must improved the SNR.

Spatial Variation
complete or incomplete projections

In order to develop the reconstruction algorithm, we must hypothise that

p  xr , y r , f ,   Is known for every  xr , y r 

In other words, we must record every LORs parallels to

zr  cos f cos  , sin f cos  , sin  

by intercepting the object Df for f(x,y,z) ≠ 0

Spatial Variation
complete or incomplete projections

If every projection are complete,

The detection efficiency of the system is independent for the position in the field of view.
The reconstruction problem is so stationary (« shift-invariant »)
So can be solved using classic deconvolution technics

The 3D acquisition is not stationary

Because the system field of view decrease when you are near from the edge

c D

The first step is to reconstruct the data for


2RD 2R zr  cos f cos  , sin f cos  , sin  
0 f ,   
L

LD
Reconstruction from complet
projections
Central slice Theorem in 3D
Generalisation of the 2D

  
F111  x , y , z  
 
f  x, y , z  e
 2 i x x  y y  z z

   
dxdydz

 
P  xr , yr , f ,   
 


  px , y ,f , e p  xr , y r , f ,   
 2 i xr x r  yr y r
r r dxr dyr  f ( x , y , z ) dz

r

  

    x   sin f  cos f sin  cos f cos    xr 

P  xr , yr , f ,   

 2 i xr x r  yr y r   
   f ( x, y , z ) e
    cos f sin f cos    yr 
dxr dyr dzr  y 
 sin f sin 

     z   0 cos  sin    zr 

Base Change

P  xr , yr , f ,    F  x , y , z 
xr   sin f x  cos f y
yr   cos f sin  x  sin f sin  y  cos  z
 zr
x x   y y z z
Example for  = 0
Introduction to the Orlov sphere

zr f , 
z

z 
r=1
y
x
y

x
z
z


 y
y
x
x 0

0  zr f ,   0  f   ,   0   zr f ,   0  f   ,   
 
The value    x , y , z  Can be calculated by using the central slice theorm


for every 2D if zr is orthogonal to n.

 
  zr f ,    x cos f cos    y sin f cos    z sin   0

generally, this equation has a continuous number of solutions is the set of measurement

  0  f   ,   

The data are then redundant

If we represent the frequency of the polar coordonnates  ,  , 

   x , y , z    cos  cos , sin  cos , sin 


cosf      tan tan  , pour   0

Solution for f only when sin   cos

F   Can be calculated by using the projections (f,) with

   lim , lim  And f defined by the previous equation

 , cos  sin 
lim  
 2   , cos  sin 
 z
 2 

 
  zr  0 If the data are consitents,
Every estimated F(n) on the circular arc are the same

When there is noise, the data are not consistents.
An optimal SNR is obtained by values average

A standard approach implies a weighted mean where the

weight factor is
cos 
cos2   sin 2 

In order to obtain a uniform angular sampling in (f,)

 lim
2 cos 
F    H ,   P  xr , yr , f , d
 lim cos   sin 
2 2

H eis the normalisation factor that allow to have F=1 if P=1

 lim
1 2 cos 
H , 
 
 lim cos2   sin 2 
d

 2
 cos  sin 
 4 arcsin  sin  
  cos  cos  sin 
 

z  z

 2 

 sin  
2  4 arcsin  
 cos 

  2   2
3D Algorithm of backprojection from
the filter projection
Projection Filtring (f,)

Calculate the FT of P  
P  xr , yr , f ,   
 
 p  xr , y r , f ,   e
 2 i xr x r  yr y r

  
dxr dyr

Multiply by the filter Hc and the apodisation window W

P F  xr , yr , f ,    P  xr , yr , f ,  H C  xr , yr ,  W  xr , yr 

Calculate the FT-1 in order to obtain the filtered projection

 
p  xr , yr , f ,     P  xr , yr , f , e

2 i xr x r  y r y r 
 d x r d y r
F F

  

Backprojection of the filtered projection

f  x, y , z   f  x, y , z   cosf  p F  xr , yr , f ,  

Repeat the previous steps for each f and 

Colsher filter

H C  xr , yr ,       H ,  2
xr
2
yr

Spatial componnent Angular Componnent

Similaire to the 2D ramp filter

with

 lim
1 2 cos 
H , 
 
 lim cos   sin 
2 2
d

 2
 cos  sin 
 4 arcsin  sin  
  cos  cos  sin 
 

and
 yr2 cos2 
tan   2
2

 xr   yr2 sin 2 
3DRP Algorithm
sinogram Rearrangement
goal: reduce the 3D problem into a simple 2D reconstruction problem
SSRB algorithm
(Single Slice ReBinning)
r1
Neglect the angles

z  r1  r2  / 2

r2
 rmax  z 
1
s2 D  xr , f , z    s  x , f , z, r dr
2 rmax z   
 rmax z
r

introduce distortion in the image

MSRB algorithm
(Multiple Slice ReBinning)
FORE Algorithm
(FOurier Rebinning)
Introduce by M. Defrise , this algorithm in an approximation of an exact inversion of the formula

He uses the relationship between frequency and distance,

A property of the Radon Transform introduce by par Edholm

2
S  xr , k , z , r     sx , f , z, r e
r
 2 ixr xr
dxr df
 0

The value of S  xr , k , z, r  Depends only on the value of f(x,y,z)

In the LOR point at a distance t

t  k /( 2 xr )

z  z  t tan 
FORE Algorithme
Initialize the 2D sinogramms to 0
S2 D  xr , k , z   0
For each oblique sinogramms

Calculate the 2D FT of the sinogramm

2
S  xr , k , z , r     s  x r , f , z ,  r e  2 ixr xr
dxr df
 0

For each frequency  xr , k 

Calculer z  z  t tan 
Ajouter S  xr , k , z, r  à S2 D  xr , k , z 

Normalize S2 D  xr , k , z 
Calculate the FT-1 of S to obtain s2 D  xr , f , z 
3DRP FORE
FOSA Algorithm
(FOurier Simple Averaging)
Iterative Algorithm

 Algorithms :

– ART : algebraic reconstruction technique

– MLEM : maximum likelihood expectation
maximization
– OSEM : ordered subsets expectation
maximization (gestion simultanée de
plusieurs sous ensembles de projections)
Iterative Algorithm

1 2 3

4 5 6

7 8 9
Iterative Algorithm

1 2 3

4 5 6

7 8 9

représentation
coupe de l’objet 
matricielle
Iterative Algorithm

1 2 3 1+2+3= 6
4 5 6 4 + 5 + 6 = 15
7 8 9 7 + 8 + 9 = 24
1 2 3
+4 +5 +6
+7 +8 +9
12 15 18
Iterative Algorithm

a b c

d e f

g h i
Iterative Algorithm

a b c 6
d e f 15
g h i 24
12 15 18
Iterative Algorithm : ART
(Algebraic Reconstruction Technique)

0 0 0

0 0 0
Iterative Algorithm : ART

0 0 0 6
0 0 0 15
0 0 0 24
Iterative Algorithm : ART

2 2 2 6
0 0 0 15
0 0 0 24

Je distribue la valeur 6 aux 3 cases de la 1ère ligne :

6/3=2
Iterative Algorithm : ART

2 2 2 6
0 0 0 15
0 0 0 24
Iterative Algorithm : ART

2 2 2 6
5 5 5 15
0 0 0 24
Iterative Algorithm : ART

2 2 2 6
5 5 5 15
8 8 8 24
Iterative Algorithm : ART

2 2 2

5 5 5

8 8 8
Iterative Algorithm : ART

15
2 2 2

5 5 5

8 8 8
Iterative Algorithm : ART

15 15 15
2 2 2

5 5 5

8 8 8
Iterative Algorithm : ART

15 15 15
2 2 2

5 5 5

8 8 8