Summer school Clermont - Irène Buvat - july 2004 - 1
To m og ra ph ic re co nstru ctio n te ch niq ue s
Irène Buvat U494 INSERM, Paris
buvat@imed.jussieu.frhttp://www.guillemet.org/irene
New Technologies in Medical Imaging and Surgery New Technologies in Medical Imaging and Surgery
Summer school Clermont - Irène Buvat - july 2004 - 2
Outline
• Introduction: the issue of tomographic reconstruction
• Basics
• Tomographic reconstruction methods
• Regularization
• New challenges
• Conclusion
Summer school Clermont - Irène Buvat - july 2004 - 3
Introduction
Summer school Clermont - Irène Buvat - july 2004 - 4
What is tomography ?
! An indirect measurement of aparameter of interest, using a detectorsensitive to some sort of radiations À Algorithms to recover the cartography of the parameters the measurements
Summer school Clermont - Irène Buvat - july 2004 - 5
Direct (forward) problem
The tomographic system measures a set of “projection” data: integralsof a signal along certain specific directions.
unknown parameters measurements = projection
The mathematical formulation of the relationship between the unknownparameters and the measurements is the direct problem. direct problem
Summer school Clermont - Irène Buvat - july 2004 - 6
Inverse problem
Tomographic reconstruction is the inversion of the direct problem.
unknown parameters measurements = projections
inverse problem
Inverse problem: estimating the 3D cartography of unknownparameters from the measured data.
Summer school Clermont - Irène Buvat - july 2004 - 7
Ill-posed inverse problem
Limited spatial sampling
Ill-posed: several solutions compatible with the measurements 1 projection
projectiondirection 2 projections
Noisy Noisy measurements
No noise signal intensity
spatial direction
Summer school Clermont - Irène Buvat - july 2004 - 8
Different types of tomography
!Transmission tomography
- External radiation source- Measurement of radiationstransmitted through the patient- Parameters related to the interactionsof radiations within the body
e.g., Computed Tomography (CT)e.g., SPECT and PET - Internal radiation source- Measurement of radiation emittedfrom the patient- Parameters related to the radiationsources within the body À Emission tomography
Summer school Clermont - Irène Buvat - july 2004 - 9
Different types of tomography
PET
SPECT CT
Optical TomographySynchrotron Tomography
Summer school Clermont - Irène Buvat - july 2004 -
B asics
Summer school Clermont - Irène Buvat - july 2004 - 11
Factorization of the reconstruction problem
3Dimages from a set of 2D measurements
3D cartography of aparameter 2D projections
3D distribution
If real 3D: “Fully 3D reconstruction” 1D projections2D cartography of aparameter 2D images from a set of 1D measurements
transaxial slice
Summer school Clermont - Irène Buvat - july 2004 -
K ey notion 1: projection
p(u,q) = f(x,y)dv
-• +• Ú u projectionq Modelling the direct problem
u = x cosq + y sinqv = -x sinq + y cosq qx yvf(x,y)u
Summer school Clermont - Irène Buvat - july 2004 - 13
Projection: mathematical expression
u projectionq
p(u,q) = f(x,y)dv
-• +• Ú The 2D Radon transform
set of projections for q = [0, p ] = Radon transform of f(x,y)
R[ f(x,y) ] = p(u,
q)d
q0 pÚ
Summer school Clermont - Irène Buvat - july 2004 -
Key notion 2: sinogram
All detected signal concerning 1 slice
q q2q3 q1
t
Summer school Clermont - Irène Buvat - july 2004 - 15
Key notion 3: backprojection
f*(x,y) = p(u,q)dq0 p
Ú Tackling the inverse problem
u = x cosq + y sinqv = -x sinq + y cosq qx yvf(x,y)u u projectionq
Beware: backprojection is not the inverse of projection !
p(u,q)Summer school Clermont - Irène Buvat - july 2004 -
Methods of tomographic reconstruction
Summer school Clermont - Irène Buvat - july 2004 - 17
Two approaches
!Analytical approaches
- Continuous formulation- Explicit solution using inversionformulae or successive transformations- Direct calculation of the solution- Fast- Discretization for numericalimplementation only
f*(x,y) = p’(u, q ) d q
0 pÚ
À Discrete approaches- Discrete formulation- Resolution of a system of linearequations or probabilistic estimation- Iterative algorithms- Slow convergence- Intrinsic discretization
p
i= S r
ijf
jjSummer school Clermont - Irène Buvat - july 2004 -
Analytical approach: central slice theorem
p(u,q) = f(x,y) dv-• +•
Ú
-• P(r,q) = p(u,q)e -i2pru duÚ
+• Fourier transformqx yvu u = x cosq + y sinqv = -x sinq + y cosqrx =rcosqry =rsinqdu.dv= dx.dy
1D FT of p with respect to u = 2D FT of f in a specific direction P(r,q) = f(x,y) e -i2p(xrx +yr
-•-• +•+•
Ú Ú
p(u,q)P(r,q) = f(x,y) e -i2pru du.dv-• +•
-• +•
Ú Ú
P(r,q) = F(rx ,ry )
Summer school Clermont - Irène Buvat - july 2004 - 19
Analytical approach: filtered backprojection (FBP)
P(r,q) = F(rx ,ry ) f(x,y) = F(rx ,ry ) e i2p(xrx +yry ) drx .dry-• +•
-• +•
Ú Ú
= P(r,q) e i2p(xrx +yry ) drx .dry-• +•
-• +•
Ú Ú
= P(r,q)|r|e i2pru dr. dq0 p+•
Ú Ú
= p’(u,q)dq with p’(u,q)= P(r,q)|r|e i2pru dr -•
0 p
Ú
-• +•Ú
Ramp filter qx yvu u = x cosq + y sinqrx =rcosqry =rsinqr= (rx 2 + ry 2) 1/2rdrx dry = rdrdq
Summer school Clermont - Irène Buvat - july 2004 -
Filtered backprojection: algorithm
P(r,q) 1D-FT
p’(u,q) FT -1 f(x,y)backprojection reconstructed image f(x,y)= p’(u,q)dq with p’(u,q)= P(r,q)|r|e i2pru d0 p
Ú
sinogram
p(u,q)
|r| P(r,q) filtering
Ramp filter
Ú
-• +•Summer school Clermont - Irène Buvat - july 2004 - 21
Filtered backprojection: beyond the Ramp filter
f(x,y)= p’(u,q)dq with p’(u,q) = P(r,q)|r|e i2pru dr0 p
Ú Ú
-• +•r
0 0.8
| r |w (r)
0 1r | r |
0 0.8 10 |r|w(r)
high frequency amplificationnoisenoise control
loss of spatial resolution
Summer school Clermont - Irène Buvat - july 2004 -
Two approaches
!Analytical approaches
- Continuous formulation- Explicit solution using inversionformulae or successive transformations- Direct calculation of the solution- Fast- Discretization for numericalimplementation only
f*(x,y) = p’(u, q ) d q
0 pÚ
À Discrete approaches- Discrete formulation- Resolution of a system of linearequations or probabilistic estimation- Iterative algorithms- Slow convergence- Intrinsic discretization
p
i= S r
ijf
jjSummer school Clermont - Irène Buvat - july 2004 - 23
Discrete approach: model
p
if
jr11 r14
r41 r44 p1p2p3p4 f1f2f3f4
= p = R f f1 f2
f3 f4 p1 p2p3p4 p1 = r11 f1 + r12 f2 + r13 f3 + r14 f4p2 = r21 f1 + r22 f2 + r23 f3 + r24 f4p3 = r31 f1 + r32 f2 + r33 f3 + r34 f4p4 = r41 f1 + r42 f2 + r43 f3 + r44 f4
projector
Given p and R, estimate f
Summer school Clermont - Irène Buvat - july 2004 -
Discrete approach: calculation of R
• Geometric modelling- intersection between pixel and projection rays
f
jp = R f R
models the direct problem• Physics modelling- spatial resolution of the detector- physical interactions of radiations
r
ij=1 r
i’j=0 r
i=0 r
if
jf
j’p
ip
i’ parallelp
i fan beamspatial resolution
p
ir
ij≠ r
ij’=0, f
jf
j’Summer school Clermont - Irène Buvat - july 2004 - 25
Two classes of discrete methods
! Algebraic methods
- Generalized inverse methods À Statistical approaches
- Bayesian estimates- Optimization of functionals
- Account for noise properties
p
i= S r
ijf
jjSummer school Clermont - Irène Buvat - july 2004 -
Iterative algorithm used in discrete methods
f
0f
1correction p = R f
p = R f
p ^ comparison p ^ versus p
defines the iterative method:
additive if
f
n+1 =f
n +c
nmultiplicative if
f
n+1 =f
n .c
nSummer school Clermont - Irène Buvat - july 2004 - 27
A lgebraic m ethods
Minimisation of || p - R f ||
2p = R f
Several minimisation algorithms are possible to estimate a solution:
e.g., SIRT (Simultaneous Iterative Reconstruction Technique)Conjugate GradientART (Algrebraic Reconstruction Technique)
e.g., additive ART:
f
j n+1= f
j n i
+ ( p - p
k i nijik
) r / S r
2Summer school Clermont - Irène Buvat - july 2004 -
Statistical methods Probabilistic formulation (Bayes’ equation): proba ( f | p ) = proba ( p | f ) proba ( f ) / proba ( p ) p = R f
probability of obtaining fwhen p is measured likelihood of p Find a solution f maximizing proba(p|f) given a probabilistic model for prior on fprior on p
f
n+1= f
n .R
t( p /p
n)
e.g., if p follows a Poisson law: proba(p|f) =P
exp(-pk ).pk pk/pMLEM (Maximum Likelihood Expectation Maximisation):
and OSEM (accelerated version of MLEM) k
Summer school Clermont - Irène Buvat - july 2004 - 29
Regularization
Summer school Clermont - Irène Buvat - july 2004 -
Regularization
Set constraints on the solution f
Solution f: trade-off betw een the agreement with the observed data and the agreement with the constraints
Summer school Clermont - Irène Buvat - july 2004 - 31
Regularization for analytical methods
Filtering
f(x,y)= P(r,q)w(r)|r|e i2pru dr0 p
Ú Ú
-• +•Ramp filterButterworth filter
Summer school Clermont - Irène Buvat - july 2004 -
Regularization for discrete methods Minimisation of || p - R f ||
2+ l K( f )
l controls the trade-off betweenagreement with the projections and agreement with the constraints
proba ( f | p ) = proba ( p | f ) proba ( f ) / proba ( p )
prior on f, i.e. proba(f) non uniform
Examples of priors:f smoothf having discontinuities
Conjugate Gradient gives MAP-Conjugate Gradient (Maximum A Posteriori)MLEM gives MAP-EM
Summer school Clermont - Irène Buvat - july 2004 - 33
New challenges
Summer school Clermont - Irène Buvat - july 2004 -
Analytical reconstruction
• Fully 3D reconstruction for CT: - suited to the new designs of CT detectors (increased number of rows, increased gantry rotation speed, helical cone beam geometry) - real time
• Reconstruction algorithms managing source deformations (motions)
Summer school Clermont - Irène Buvat - july 2004 - 35
Iterative reconstruction: fully 3D Monte Carlo reconstruction p = R f
modelling R using numerical (Monte Carlo) simulationsof the imaging procedure in emission tomography
tissue densityand composition+cross-section tables ofradiation interaction
All propagation and detection physics can be accurately modelled3D propagation of radiation is taken into account stochastic modellingof physical interactions probability that an“event” in j be detectedin i j i
R
ijSummer school Clermont - Irène Buvat - july 2004 -
P ending issues
• Size of the fully 3D problem: 64 projections 64 x 64, R includes 64
6elements (256 Gigabytes)
• Time
Summer school Clermont - Irène Buvat - july 2004 - 37
Question 1
G iven 128 projections of 64 pixels by 64 pixels, you can easily create:
A. 64 sinograms with 128 rows and 128 columns
B. 128 sinogram s w ith 64 row s and 64 colum ns
C. 64 sinograms with 128 rows and 64 columns
D. 64 sinograms with 64 rows and 128 columns
E. 128 sinograms with 128 rows and 64 columns
Summer school Clermont - Irène Buvat - july 2004 -