0

Iterative methods for image formation in MRI

Jeffrey A. Fessler
EECS Department
The University of Michigan
UM Applied Physics Seminar
Mar. 21, 2007
Acknowledgements: Doug Noll, Brad Sutton,
Chunyu Yip, Will Grissom
1
The Ends
X-ray CT MRI
www.gehealthcare.com www.cis.rit.edu
MRI: excellent soft tissue contrast, and no ionizing radiation.
(But, expensive, slow, big, small bone signal...)
2
Outline
MR imaging physics
MR image reconstruction introduction
(k-space, FFT, gridding, density compensation)
Model-based reconstruction overview
Iterations and computation (NUFFT etc.)
Regularization
Field inhomogeneity correction
Parallel (sensitivity encoded) imaging
Iterative methods for RF pulse design
3
Physics
4
Bloch Equation - Overview
Time evolution (phenomenological) of local magnetization MMM(rrr, t):
dMMM
dt
= MMMBBB
M
x
iii +M
y
jjj
T
2

(M
z
M
0
)kkk
T
1
Precession
Relaxation
Equilibrium
5
Bloch Equation and Imaging
dMMM(rrr, t)
dt
= MMM(rrr, t) BBB(rrr, t)
M
x
iii +M
y
jjj
T
2
(rrr)

(M
z
M
0
(rrr))kkk
T
1
(rrr)
Image properties depend on:
Spin (Hydrogen) density M
0
(rrr)
Longitudinal (spin-lattice) relaxation T
1
(rrr)
Transverse (spin-spin) relaxation T
2
(rrr)
Chemical shift
(resonant frequency of H is 3.5 ppm lower in fat than in water)
Applied eld BBB(rrr, t) includes three components we can control:
Main eld B
0
RF eld BBB
1
(t)
Field gradients rrr GGG(t) = xG
x
(t) +yG
y
(t) +zG
z
(t)
BBB(rrr, t) = B
0
+BBB
1
(t) +rrr GGG(t)kkk
6
Systems view of MRI
Applied
eld
BBB(rrr, t)
Patient
magnetization
pattern
MMM(rrr, t)

RF coil(s)
(Faraday
induction)

received
signal
s
r
(t)

demodulate
(Larmor
frequency)

baseband
signal
s(t)

sample
(A/D)

recorded
data
y
i
, i = 1, . . . , M

reconstruction
algorithm

displayed
image
f (r)
Research areas:
design of RF pulses / gradient waveforms (many possibilities!)
coil design
reconstruction algorithm development
7
Introduction to Reconstruction
8
Standard MR Image Reconstruction
MR kspace data Reconstructed Image
Cartesian sampling in k-space. An inverse FFT. End of story.
Commercial MR system quotes 400 FFTs (256
2
) per second.
9
K-space Sampling Strategies
Cartesian Truncated Partial
Undersampled Variable density NonCartesian
10
Non-Cartesian MR Image Reconstruction
k-space image
k
x
k
y
=
11
Example: Iterative Reconstruction under B
0
12
Example: Iterative RF Pulse Design
13
Textbook MRI Measurement Model
Ignoring lots of things:
y
i
= s(t
i
) +noise
i
, i = 1, . . . , M
s(t) =
Z
f (r)e
2

(t)r
dr,
wherer denotes spatial position, and

(t) denotes the k-space trajectory of the MR pulse sequence,

determined by user-controllable magnetic eld gradients.
e
2

(t)r
provides spatial information = Nobel Prize
MRI measurements are (roughly) samples of the Fourier
transform F(

) of the objects transverse magnetization f (r).

Basic image reconstruction problem:
recover f (r) from measurements y
i

M
i=1
.
Inherently under-determined (ill posed) problem
= no canonical solution.
14
Image Reconstruction Strategies
The unknown object f (r) is a continuous-space function,
but the recorded measurements yyy = (y
1
, . . . , y
M
) are nite.
Options?
Continuous-discrete formulation using many-to-one linear model:
yyy =A f +.
Minimum norm solution (cf. natural pixels):
min

f
_
_
f
_
_
subject to yyy =A

f

f =A

(AA

)
1
yyy =

M
i=1
c
i
e
2

i
r
, where AA

ccc = yyy.
Discrete-discrete formulation
Assume parametric model for object:
f (r) =
N

j=1
f
j
p
j
(r).
Continuous-continuous formulation
Pretend that a continuum of measurements are available:
F(

) =
Z
f (r)e
2

r
dr,
vs samples y
i
= F(

i
)+
i
, where

(t
i
).
The solution is an inverse Fourier transform:
f (r) =
Z
F(

)e
2

r
d

.
Now discretize the integral solution (two approximations!):

f (r) =
M

i=1
F(

i
)e
2

i
r
w
i

M

i=1
y
i
w
i
e
2

i
r
,
where w
i
values are sampling density compensation factors.
Numerous methods for choosing w
i
value in the literature.
For Cartesian sampling, using w
i
= 1/N sufces,
and the summation is an inverse FFT.
16
Conventional MR Image Reconstruction
1. Interpolate measurements onto rectilinear grid (gridding)
2. Apply inverse FFT to estimate samples of f (r)
0
0
k
x
k
y
Gridding from Polar to Cartesian
17
Gridding Approach 1: Pull from K nearest
0
0
k
x
k
y
Gridding by pulling from 10 nearest
18
Gridding Approach 2: Pull from neighborhood
0
0
k
x
k
y
Gridding by pulling from within neighborhood
19
Gridding Approach 3: Push to neighborhood
0
0
k
x
k
y
Gridding by pushing onto neighborhood
20
Gridding Approaches
Ignore noise: y
i
= F(

i
)
Pull:
for each Cartesian grid point, use weighted average
of nonuniform k-space samples within some neighborhood
Does not require density compensation
Requires cumbersome search/indexing to nd neighbors
Push:
each nonuniform k-space sample onto a Cartesian neighborhood

F(

) =
M

i=1
y
i
w
i
C(

i
)
C(

) denotes the gridding kernel, typically separable Kaiser-Bessel

Jackson et al., IEEE T-MI, 1991
denotes convolution.
() denotes the Dirac impulse
density compensation factors w
i
essential
21
Post-iFFT Gridding Correction
Gridding as convolution in k-space:

F(

) =
M

i=1
y
i
w
i
C(

i
) =C(

)
M

i=1
y
i
w
i
(

i
).
Inverse FT reconstruction:

f
initial
(r) =F
1
_

F(

)
_
= c(r)
M

i=1
y
i
w
i
e
2

i
r
.
Post-correction:

f
nal
(r) =

f
initial
(r)
c(r)
.
22
Gridding Kernels and Post-corrections
6 4 2 0 2 4 6
0.4
0.2
0
0.2
0.4
0.6
0.8
1
u
C
(
u
)
Convolution kernels


1.5 1 0.5 0 0.5 1 1.5
0.2
0
0.2
0.4
0.6
0.8
1
x/FOV
c
(
x
)
Postgridding correction functions
KaiserBessel, J=4 =18.5547
Sinc
23
Density Compensation
f (r) =
Z
F(

)e
2

r
d

i=1
y
i
e
2

i
r
w
i
.
Voronoi cell area
Bracewell, 1973, Astrophysical Journal; Rasche et al., IEEE T-MI, 1999
Jacobians Norton, IEEE T-MI, 1987
Jacksons area density Jackson, IEEE T-MI, 1991
Iterative methods Pipe and Menon, MRM, 2000
...
Tradeoffs between simplicity and accuracy.
24
Voronoi Cell Area
1 0 1
1
0
1

2
25
Limitations of Gridding Reconstruction
1. Artifacts/inaccuracies due to interpolation
2. Contention about sample density weighting
3. Oversimplications of Fourier transform signal model:
Magnetic eld inhomogeneity
Magnetization decay (T
2
)
Eddy currents
...
4. Sensitivity encoding ?
5. ...
(But it is faster than iterative methods...)
26
Model-Based Image Reconstruction: Overview
27
Model-Based Image Reconstruction
MR signal equation with more complete physics:
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
y
i
= s(t
i
) +noise
i
, i = 1, . . . , M
s
coil
(r) Receive-coil sensitivity pattern(s) (for SENSE)
(r) Off-resonance frequency map
(due to eld inhomogeneity / magnetic susceptibility)
R

2
(r) Relaxation map
Other physical factors (?)
Eddy current effects; in

(t)
Concomitant gradient terms
Chemical shift
Motion
Goal?
(it depends)
28
Field Inhomogeneity-Corrected Reconstruction
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
Goal: reconstruct f (r) given eld map (r).
(Assume all other terms are known or unimportant.)
Motivation
Essential for functional MRI of brain regions near sinus cavities!
(Sutton et al., ISMRM 2001; T-MI 2003)
29
Sensitivity-Encoded (SENSE) Reconstruction
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
Goal: reconstruct f (r) given sensitivity maps s
coil
(r).
(Assume all other terms are known or unimportant.)
Can combine SENSE with eld inhomogeneity correction easily.
(Sutton et al., ISMRM 2001, Olafsson et al., ISBI 2006)
30
Joint Estimation of Image and Field-Map
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
Goal: estimate both the image f (r) and the eld map (r)
(Assume all other terms are known or unimportant.)
Analogy:
joint estimation of emission image and attenuation map in PET.
(Sutton et al., ISMRM Workshop, 2001; ISBI 2002; ISMRM 2002;
ISMRM 2003; MRM 2004)
31
The Kitchen Sink
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
Goal: estimate image f (r), eld map (r), and relaxation map R

2
(r)
Requires suitable k-space trajectory.
(Sutton et al., ISMRM 2002; Twieg, MRM, 2003)
32
Estimation of Dynamic Maps
s(t) =
Z
f (r)s
coil
(r)e
(r)t
e
R

2
(r)t
e
2

(t)r
dr
Goal: estimate dynamic eld map (r) and BOLD effect R

2
(r)
given baseline image f (r) in fMRI.
Motion...
33
Model-Based Image Reconstruction: Details
34
Back to Basic Signal Model
s(t) =
Z
f (r)e
2

(t)r
dr
Goal: reconstruct f (r) from yyy = (y
1
, . . . , y
M
), where y
i
= s(t
i
) +
i
.
Series expansion of unknown object:
f (r)
N

j=1
f
j
p(r r
j
) usually 2D rect functions.
y
i

Z
_
N

j=1
f
j
p(r r
j
)
_
e
2

i
r
dr =
N

j=1
_
Z
p(r r
j
)e
2

i
r
dr
_
f
j
=
N

j=1
a
i j
f
j
, a
i j
= P(

i
)e
2

i
r
j
, p(r)
FT
P(

).
Discrete-discrete measurement model with system matrix AAA =a
i j
:
yyy = AAAfff +.
Goal: estimate coefcients (pixel values) fff = ( f
1
, . . . , f
N
) from yyy.
35
Small Pixel Size Need Not Matter
x true
128 0 127
128
0
127
0
2
N=32
128 0 127
128
0
127
0
2
N=64
128 0 127
128
0
127
0
2
N=128
128 0 127
128
0
127
0
2
N=256
128 0 127
128
0
127
0
2
N=512
128 0 127
128
0
127
0
2
36
Proles
100 50 0 50 100
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
horizontal position [mm]
|
f
(
x
,
0
)
|
true
N=32
N=64
N=128
N=256
N=512
37
Regularized Least-Squares Estimation

fff = argmin
fff C
N
( fff ), ( fff ) = |yyy AAAfff |
2
+R( fff )
data t term |yyy AAAfff |
2
corresponds to negative log-likelihood of Gaussian distribution
regularizing roughness penalty term R( fff ) controls noise
R( fff )
Z
| f |
2
dr
regularization parameter > 0
controls tradeoff between spatial resolution and noise
(Fessler & Rogers, IEEE T-IP, 1996)
Equivalent to Bayesian MAP estimation with prior e
R( fff )
Quadratic regularization R( fff ) =|CCCfff |
2
leads to closed-form solution:

fff =
_
AAA
/
AAA+CCC
/
CCC

1
AAA
/
yyy
(a formula of limited practical use)
38
Choosing the Regularization Parameter

fff =
_
AAA
/
AAA+CCC
/
CCC

1
AAA
/
yyy
E
_

fff
_
=
_
AAA
/
AAA+CCC
/
CCC

1
AAA
/
E[yyy]
E
_

fff
_
=
_
AAA
/
AAA+CCC
/
CCC

1
AAA
/
AAA
. .
blur
fff
AAA
/
AAA and CCC
/
CCC are Toeplitz = blur is approximately shift-invariant.
Frequency response of blur:
L() =
H()
H() +R()
where H = FFT(AAA
/
AAAe
j
) (lowpass) and R = FFT(CCC
/
CCCe
j
) (highpass)
Adjust to achieve desired spatial resolution.
39
Spatial Resolution Example
AA e
j
10 0 10
10
5
0
5
10
CC e
j
10 0 10
10
5
0
5
10
PSF
10 0 10
10
5
0
5
10
H()

Y
0

0

R()

Y
0

0

L=H/(H+R)

Y
0

0

Spiral k-space trajectory, FWHM of PSF is 1.2 pixels
40
Spatial Resolution Example: Proles
0
0
5
10
x 10
5
H
(

)
0
0
200
400
600
800
R
(

)
0
0.6
0.8
1
L
(

41
Iterative Minimization by Conjugate Gradients
Choose initial guess fff
(0)
(e.g., fast conjugate phase / gridding).
Iteration (unregularized):
ggg
(n)
=
_
fff
(n)
_
= AAA
/
(AAAfff
(n)
yyy) gradient
ppp
(n)
= PPPggg
(n)
precondition

n
=
_
_
_
0, n = 0
ggg
(n)
, ppp
(n)
)
ggg
(n1)
, ppp
(n1)
)
, n > 0
ddd
(n)
= ppp
(n)
+
n
ddd
(n1)
search direction
vvv
(n)
= AAAddd
(n)

n
= ddd
(n)
, ggg
(n)
)/AAAfff
(n)
, AAAfff
(n)
) step size
fff
(n+1)
= fff
(n)
+
n
ddd
(n)
update
Bottlenecks: computing AAAfff and AAA
/
yyy.
AAA is too large to store explicitly (not sparse)
Even if AAA were stored, directly computing AAAfff is O(MN)
per iteration, whereas FFT is only O(MlogM).
42
Computing AAAfff Rapidly
[AAAfff ]
i
=
N

j=1
a
i j
f
j
= P(

i
)
N

j=1
e
2

i
r
j
f
j
, i = 1, . . . , M
Pixel locations r
j
are uniformly spaced
k-space locations

i
are unequally spaced
= needs nonuniform fast Fourier transform (NUFFT)
43
NUFFT (Type 2)
Compute over-sampled FFT of equally-spaced signal samples
Interpolate onto desired unequally-spaced frequency locations
Dutt & Rokhlin, SIAM JSC, 1993, Gaussian bell interpolator
Fessler & Sutton, IEEE T-SP, 2003, min-max interpolator
and min-max optimized Kaiser-Bessel interpolator.
NUFFT toolbox: http://www.eecs.umich.edu/fessler/code
0
50
100
/2 /2
X
(

)
?
44
Worst-Case NUFFT Interpolation Error
2 4 6 8 10
10
10
10
8
10
6
10
4
10
2
J
E
m
a
x
Maximum error for K/N=2
MinMax (uniform)
Gaussian (best)
MinMax (best L=2)
KaiserBessel (best)
MinMax (L=13, =1 fit)
45
Further Acceleration using Toeplitz Matrices
Cost-function gradient:
ggg
(n)
= AAA
/
(AAAfff
(n)
yyy)
= TTT fff
(n)
bbb,
where
TTT AAA
/
AAA, bbb AAA
/
yyy.
In the absence of eld inhomogeneity, the Gram matrix TTT is Toeplitz:
[AAA
/
AAA]
jk
=
M

i=1
[P(

i
)[
2
e
2

i
(r
j
r
k
)
.
Computing TTT fff
(n)
requires an ordinary (2 over-sampled) FFT.
(Chan & Ng, SIAM Review, 1996)
In 2D: block Toeplitz with Toeplitz blocks (BTTB).
Precomputing the rst column of TTT and bbb requires a couple NUFFTs.
(Wajer, ISMRM 2001, Eggers ISMRM 2002, Liu ISMRM 2005)
This formulation seems ideal for hardware FFT systems.
46
NUFFT with Field Inhomogeneity?
Combine NUFFT with min-max temporal interpolator
(Sutton et al., IEEE T-MI, 2003)
(forward version of time segmentation, Noll, T-MI, 1991)
Recall signal model including eld inhomogeneity:
s(t) =
Z
f (r)e
(r)t
e
2

(t)r
dr .
Temporal interpolation approximation (aka time segmentation):
e
(r)t

l=1
a
l
(t)e
(r)
l
for min-max optimized temporal interpolation functions a
l
()
L
l=1
.
s(t)
L

l=1
a
l
(t)
Z
_
f (r)e
(r)
l
_
e
2

(t)r
dr
Linear combination of L NUFFT calls.
47
Field Corrected Reconstruction Example
Simulation using known eld map (r).
48
Simulation Quantitative Comparison
Computation time?
NRMSE between

fff and fff
true
?
Reconstruction Method Time (s) NRMSE NRMSE
complex magnitude
No Correction 0.06 1.35 0.22
Full Conjugate Phase 4.07 0.31 0.19
Fast Conjugate Phase 0.33 0.32 0.19
Fast Iterative (10 iters) 2.20 0.04 0.04
Exact Iterative (10 iters) 128.16 0.04 0.04
49
Human Data: Field Correction
50
Acceleration using Toeplitz Approximations
In the presence of eld inhomogeneity, the system matrix is:
a
i j
= P(

i
)e
(r
j
)t
i
e
2

i
r
j
The Gram matrix TTT = AAA
/
AAA is not Toeplitz:
[AAA
/
AAA]
jk
=
M

i=1
[P(

i
)[
2
e
2

i
(r
j
r
k
)
e

(
(r
j
)(r
k
)
)
t
i
.
Approximation (time segmentation):
e

(
(r
j
)(r
k
)
)
t
i

l=1
b
il
e

(
(r
j
)(r
k
)
)

l
TTT = AAA
/
AAA
L

l=1
DDD
/
l
TTT
l
DDD
l
,
DDD
l
diag
_
e
(r
j
)
l
_
[TTT
l
]
jk

M
i=1
[P(

i
)[
2
b
il
e
2

i
(r
j
r
k
)
.
Each TTT
l
is Toeplitz = TTT fff using L pairs of FFTs.
(Fessler et al., IEEE T-SP, Sep. 2005, brain imaging special issue)
51
Toeplitz Results
1 64
Image and support
1
64
0
2.5
1 64
Fieldmap: Brain
1
64
H
z
20
0
20
40
60
80
100
120
Uncorrected Conj. Phase, L=6
CGNUFFT
L=6
CGToeplitz
L=8
0
2.5
52
Toeplitz Acceleration
Precomputation NRMS % vs SNR
Method L BBB,CCC AAA
/
DDDyyy bbb = AAA
/
yyy TTT
l
15 iter Total Time 50 dB 40 dB 30 dB 20 dB
Conj. Phase 6 0.4 0.2 0.6 30.7 37.3 46.5 65.3 99.9
CG-NUFFT 6 0.4 5.0 5.4 5.6 16.7 26.5 43.0 70.4
CG-Toeplitz 8 0.4 0.2 0.6 1.3 2.5 5.5 16.7 26.4 42.9 70.4
Reduces CPU time by 2 on conventional workstation (Mac G5)
No SNR compromise
Eliminates k-space interpolations = ideal for FFT hardware
53
Joint Field-Map / Image Reconstruction
Signal model:
y
i
= s(t
i
) +
i
, s(t) =
Z
f (r)e
(r)t
e
2

(t)r
dr .
After discretization:
yyy = AAA() fff +, a
i j
() = P(

i
)e

j
t
i
e
2

i
r
j
.
Joint estimation via regularized (nonlinear) least-squares:
(

fff ,

) = argmin
fff C
N
,R
N
|yyy AAA() fff |
2
+
1
R
1
( fff ) +
2
R
2
().
Alternating minimization:
Using current estimate of eldmap

,
update

fff using CG algorithm.
Using current estimate

fff of image,
update eldmap

using gradient descent.
Use spiral-in / spiral-out sequence or racetrack EPI.
(Sutton et al., MRM, 2004)
54
Joint Estimation Example
(a) uncorr., (b) std. map, (c) joint map, (d) T1 ref, (e) using std, (f) using joint.
55
Activation Results: Static vs Dynamic Field Maps
Functional results for the two reconstructions for 3 human subjects.
Reconstruction using the standard eld map
for (a) subject 1, (b) subject 2, and (c) subject 3.
Reconstruction using the jointly estimated eld map
for (d) subject 1, (e) subject 2, and (f) subject 3.
Number of pixels with correlation coefcients higher than thresholds
for (g) subject 1, (h) subject 2, and (i) subject 3.
Take home message: dynamic eld mapping is possible, using iter-
ative reconstruction as an essential tool.
(Standard eld maps based on echo-time differences work poorly
for spiral-in / spiral-out sequences due to phase discrepancies.)
57
Tracking Respiration-Induced Field Changes
58
Regularization Variations
59
Regularization Revisited
Conventional regularization for MRI uses a roughness penalty
for the complex voxel values:
R( fff )
N

j=1
[ f
j
f
j1
[
2
(in 1D).
Regularizes the real and imaginary image components equally.
In some MR studies, including BOLD fMRI:
magnitude of f
j
carries the information of interest,
phase of f
j
should be spatially smooth.
This a priori information is ignored by R( fff ).
Alternatives to R( fff ):
Constrain fff to be real?
(Unrealistic: RF phase inhomogeneity, eddy currents, ...)
Determine phase of fff somehow, then estimate its magnitude.
Non-iteratively (Noll, Nishimura, Macovski, IEEE T-MI, 1991)
Iteratively (Lee, Pauly, Nishimura, ISMRM, 2003)
60
Separate Magnitude/Phase Regularization
Decompose fff into its magnitude mmm and phase xxx:
f
j
(mmm, xxx) = m
j
e
x
j
, m
j
R, x
j
R, j = 1, . . . , N.
(Allow magnitude m
j
to be negative.)
Proposed cost function with separate regularization of mmm and xxx:
(mmm, xxx) = |yyy AAAfff (mmm, xxx)|
2
+R
1
(mmm)+R
2
(xxx).
Choose to strongly smooth phase estimate.
Joint estimation of magnitude and phase via regularized LS:
( mmm, xxx) = argmin
mmmR
N
, xxxR
N
(mmm, xxx)
is not convex = need good initial estimates (mmm
(0)
, xxx
(0)
).
61
Alternating Minimization
Magnitude Update:
mmm
new
= argmin
mmmR
N

_
mmm, xxx
old
_
Phase Update:
xxx
new
= argmin
xxxR
N
(mmm
new
, xxx),
Since f
j
= m
j
e
x
j
is linear in m
j
, the magnitude update is easy.
Apply a few iterations of slightly modied CG algorithm
(constrain mmm to be real)
But f
j
= m
j
e
x
j
is highly nonlinear in xxx. Complicates argmin.
Steepest descent?
xxx
(n+1)
= xxx
(n)

xxx

_
mmm
old
, xxx
(n)
_
.
Choosing the stepsize is difcult.
62
Optimization Transfer
63
Surrogate Functions
To minimize a cost function (xxx), choose surrogate functions
(n)
(xxx)
that satisfy the following majorization conditions:

(n)
_
xxx
(n)
_
= (xxx
(n)
)

(n)
(xxx) (xxx), xxx R
N
.
Iteratively minimize the surrogates as follows:
xxx
(n+1)
= argmin
xxx
(n)
R
N

(n)
(xxx).
This will decrease monotonically; (xxx
(n+1)
) (xxx
(n)
).
The art is in the design of surrogates.
Tradeoffs:
complexity
computation per iteration
convergence rate / number of iterations.
64
Surrogate Functions for MR Phase
L(xxx) |yyy AAAfff (mmm, xxx)|
2
=
M

i=1
h
i
([AAAfff (mmm, xxx)]
i
),
where h
i
(t) [y
i
t[
2
is convex.
Extending De Pierro (IEEE T-MI, 1995), for
i j
0 and

N
j=1

i j
= 1:
[AAAfff (mmm, xxx)]
i
=
N

j=1
b
i j
e
x
j
=
N

j=1

i j
_
b
i j

i j
_
e
x
j
e
x
(n)
j
_
+ y
(n)
i
_
,
where b
i j
a
i j
m
j
, y
(n)
i
[AAAfff (mmm, xxx
(n)
)]
i
. Choose
i j
0 and

N
j=1

i j
= 1.
Since h
i
is convex:
h
i
([AAAfff (mmm, xxx)]
i
) = h
i
_
N

j=1

i j
_
b
i j

i j
_
e
x
j
e
x
(n)
j
_
+ y
(n)
i
_
_

j=1

i j
h
i
_
b
i j

i j
_
e
x
j
e
x
(n)
j
_
+ y
(n)
i
_
,
with equality when xxx = xxx
(n)
.
65
Separable Surrogate Function
L(xxx) =
M

i=1
h
i
([AAAfff (mmm, xxx)]
i
)
M

i=1
N

j=1

i j
h
i
_
b
i j

i j
_
e
x
j
e
x
(n)
j
_
+ y
(n)
i
_
=
N

j=1
M

i=1

i j
h
i
_
b
i j

i j
_
e
x
j
e
x
(n)
j
_
+ y
(n)
i
_
. .
Q
j
(x
j
; xxx
(n)
)
.
Construct similar surrogates S
j
for (convex) roughness penalty...
Surrogate:
(n)
(xxx) =
N

j=1
Q
j
(x
j
; xxx
(n)
) +S
j
(x
j
; xxx
(n)
).
Parallelizable (simultaneous) update, with 1D minimizations:
xxx
(n+1)
= argmin
xxx
(n)
R
N

(n)
(xxx) = x
(n+1)
j
= argmin
x
j
R
Q
j
(x
j
; xxx
(n)
) +S
j
(x
j
; xxx
(n)
).
Intrinsically guaranteed to monotonically decrease the cost function.
66
1D Minimization: cos + quadratic
... Q
j
(x
j
; xxx
(n)
)

r
(n)
j

cos
_
x
j
x
(n)
j
r
(n)
j
_
,
r
(n)
j
=
_
f
(n)
j
_

[AAA
/
(yyy AAAxxx
(n)
)]
j
+[m
j
[
2
N
M

i=1
[P(

i
)[
2
2 0 2
0
1
2
3
4
5
6
7
8
9
t
s
u
r
r
o
g
a
t
e
(
t
)
m(1cos(tp)) + bt + c/2 t
2
5 0 5 10 15
0
0.5
1
1.5
2
2.5
3
Surrogates for sinusoids
t
Simple 1D optimization transfer iterations...
67
Final Algorithm for Phase Update
Diagonally preconditioned gradient descent:
xxx
(n+1)
= xxx
(n)
DDD(xxx
(n)
)(xxx
(n)
)
where the diagonal matrix DDD has elements that ensure decreases
monotonically.
Alternate between magnitude and phase updates...
68
Preliminary Simulation Example
0

0

Undersampled Spiral

Y
1 32
|x| true
1
28
0
2
1 32
x true
1
28
0.1
0.6
1 32
|x| old
1
28
0
2
1 32
x old
1
28
0.1
0.6
1 32
|x| new
1
28
0
2
1 32
x new
1
28
0.1
0.6
69
Parallel Imaging
70
Sensitivity encoded (SENSE) imaging
Use multiple receive coils (requires multiple RF channels).
Exploit spatial localization of sensitivity pattern of each coil.
Note: at 1.5T, RF is about 60MHz.
= RF wavelength is about 3 10
8
m/s/60 10
6
Hz = 5 meters
RF coil sensitivity patterns
Array coil images
1 64
1
64
Regularized sensitivity maps
1 64
1
64
0
1.4
Pruessmann et al., MRM, 1999
71
SENSE Model
Multiple coil data:
y
li
= s
l
(t
i
) +
li
, s
l
(t) =
Z
f (r)s
coil
l
(r)e
2

(t)r
dr, l = 1, . . . , L = N
coil
Goal: reconstruct f (r) from coil data yyy
1
, . . . , yyy
L
given sensitivity maps
_
s
coil
l
(r)
_
L
l=1
.
Benet: reduced scan time.
Left: sum of squares; right: SENSE.
72
SENSE Reconstruction
Signal model:
s
l
(t) =
Z
f (r)s
coil
l
(r)e
2

(t)r
dr
Discretized form:
yyy
l
= AAADDD
l
fff +
l
, l = 1, . . . , L,
where AAA is the usual frequency/phase encoding matrix and
DDD
l
contains the sensitivity pattern of the lth coil: DDD
l
= diag
_
s
coil
l
(r
j
)
_
.
Regularized least-squares estimation:

fff = argmin
fff
L

l=1
|yyy
l
AAADDD
l
fff |
2
+R( fff ).
Can generalize to account for noise correlation due to coil coupling.
Easy to apply CG algorithm, including Toeplitz/NUFFT acceleration.
For Cartesian SENSE, iterations are not needed.
(Solve small system of linear equations for each voxel.)
73
RF Pulse Design
74
Example: Iterative RF Pulse Design
(3D tailored RF pulses for through-plane dephasing compensation)
75
Multiple-coil Transmit Imaging Pulses (Mc-TIP)
76
Summary
Iterative reconstruction: much potential in MRI
Even nonlinear problems involving phase terms e
x
are tractable
by using optimization transfer.
Computation: reduced by tools like NUFFT / Toeplitz
Optimization algorithm design remains important
(cf. Shepp and Vardi, 1982, PET)
Some current challenges
Sensitivity pattern mapping for SENSE
Through-voxel eld inhomogeneity gradients
Motion / dynamics / partial k-space data
Establishing diagnostic efcacy with clinical data...