8/1/2011
The Need for Speed: A
Technical and Clinical Primer
for Parallel MR Imaging
Nathan Yanasak, Ph.D.
Overview
• General Description of Parallel
Imaging
• Applications of Parallel Imaging
Chair, AAPM TG118
• pMRI Considerations
Assistant Professor
Department of Radiology
Director, Core Imaging Facility for Small Animals (CIFSA)
Georgia Health Sciences University
• Quality Control
What is pMRI?
Speed: less phase encodes = smaller
FOV (with same resolution)
• Uses spatial information obtained from
arrays of RF coils sampling data in parallel
• Information is used to perform some portion
of spatial encoding usually done by gradient
fields (typically phase-encoding gradient)
• Speeds up MRI acquisition times
– without needing faster-switching gradients
– without additional RF power deposited (key
for higher field MR)
aliasing
Smaller FOV
4
1
8/1/2011
Phased-Array of Coil Elements
Uneven SNR throughout
volume, but …
6
5
Multichannel Coils
12-channel head coil
Multi-element full body coil
SIGNAL-TO-NOISE RATIO (Arbitrary Units)
• Very high SNR at edge
• Lower SNR in middle
• SNR in middle is
generally better than
comparable volume
coil.
SNR: Surface Coil vs. Volume Coil
Surface Coil (single element)
Non-uniform SNR
Great SNR up close
a
5
4
3
(a)
(b)
(c)
(d)
Volume Coil
Uniform SNR
Average SNR
8 cm dia. Surface coil
10 cm dia. Surface coil
14 cm dia. Surface coil
Head coil
b
c
2
1
0
head
d
body
5
10
15
DEPTH (in cm)
6
Use of Phased Array Coil in
Parallel Imaging
Spatial sensitivity varies for each element  can
use this in conjunction with undersampling.
7
Conventional use of phased-array
(unaliased)
Parallel reconstruction of data
(aliased)
2
8/1/2011
Sensitivity Map
Coil Sensitivity Profiles
The spatial sensitivity of each coil element = sensitivity map.
A calibration scan is usually required to calculate this.
• Different approaches to solving the inverse
problem that recovers spatial information.
• The key information always required to solve this
problem is information on the spatial distribution
of the RF coils’ sensitivity.
• How you collect and use this information 
different methods.
Total
s1
s2
10
Using Coil Sensitivity to Un-alias
an Image: An Example
Coil Locations and Sensitivity
Maps
Object
being
imaged
11
12
3
8/1/2011
Using Coil Sensitivity to Un-alias an
Image
13
Two Parallel Approaches
• Image based: Reconstruct
images from each element,
then untangle (SENSE,
ASSET) (our demo)
• k-Space based: Untangle
data to create fully-filled kspace, then reconstruct image
(SMASH, GRAPPA)
14
The Encoding Matrix
S p   B pj  j
j
Sp: signal received by the coil, p.
j: proton density at the pixel index, j
Bpj: encoding function that connects the coil response
to the proton signal at a location.
In matrix notation: S = B
or inverting:
= B-1S
Thus if B-1 can be calculated,  can be determined.
4
8/1/2011
A Simplistic SENSE
Example
k-Space Based pMRI
Salias,1=B1,AIA + B1,BIB
A
• Assumes spatial harmonics of phaseencoding gradients can be omitted and
emulated by a linear combination of coil
sensitivities
IA
B
s1
I1
IB
A
B
s2
• Coil sensitivity still required (measured
in some manner, and complex).
I2
Salias,2=B2,AIA + B2,BIB
17
A Simplistic SMASH
Example
A Simplistic SMASH
Example
•
(from Sodickson, et al., MRM 41: 1009, 1999)
Weighted linear combination of element
responses are sensitive to different spatial
scales (but…no recon yet).
5
8/1/2011
A Simplistic SMASH
Example
A Simplistic SMASH Example
The black curves represent two of the effective sensitivities
using elements in this example.
The upper combination is sensitive
Fundamental: CA (add signal for all three)
to these spatial variations:
First Harmonic: CB (subtract middle signal)
…while the lower combination is
sensitive to these spatial variations:
•
Resultant combinations (spatial harmonics) allow for filling of all
lines in a composite k-space.
Auto-Calibration Methods
• Acquire reference lines (ACS lines) in k-space
rather than whole coil sensitivity images (data
from center of k-space acts like a sensitivity profile)
22
Parallel Imaging (Technique
Pros/Cons)
Image-based reconstruction: More artifacts, but
easier to implement the sequence.
Example: GRAPPA (k-space based)
• Missing k-space lines are synthesized by fitting
between reference data and nearest neighbor lines of
data
K-space based reconstruction: Depends more
strongly on coil design, less artifacts, but longer to
reconstruct.
• Fitting determines the weighting factors for
generating missing lines for each coil
Hybrids between both also exist...
24
6
8/1/2011
Parallel Imaging Flavors
Name
SENSitivity Encoding
Acronym
SENSE
Array Spatial Sensitivity Encoding
Technique
Auto-calibrating Reconstruction for
Cartesian imaging
ASSET
integrated Parallel Acquisition Techniques
iPAT
GeneRalized Auto-calibrating Partially
Parallel Acquisition
GRAPPA
modified SENSitivity Encoding
mSENSE
ARC
SPEEDER
Method
Image-based,
reference scan
Image-based,
reference scan
Hybrid (imageand k-space
based)
Used by all
pMRI
k-space based,
auto-calibrated
with reference
scan option
image based,
auto-calibrated
with reference
scan option
image-based,
reference scan
Manufacturer
Philips
General Electric
General Electric
Advantages/Uses of pMRI
Siemens
Siemens
Siemens
Toshiba
When Should You Use
Parallel MR Imaging?
• To reduce total scan time
• To speed up single-shot MRI methods
25
26
Use #1: Body Imaging
A: 22 sec acquisition w/ 15
sec breathhold
B: 11 sec acquisition w/ 11
sec breathhold + R=2
• To reduce TE on long echo-train methods
• To reduce total scan time
(or eliminate breath holds)
• To mitigate susceptibility, chemical shift and
other artifacts (may cause others)
• To decrease RF heating
(SAR) by minimizing
number of RF pulses
• To decrease RF heating (SAR) by minimizing
number of RF pulses (B2)
Margolis D et al. Top Magn
Reson Imag 2004; 15: 197-206
28
7
8/1/2011
Use #3: Reduce T2 Blurring
(FSE)
Use #2: Spinal Imaging
D: non-pMRI
E: R=2
• Image quality is
of similar quality
for ½ the scan
time
• Problem #1: Greater ETL
 lower SNR
• Problem #2: T2 relaxation
during acquisition of ETL
results in ―T2 blurring‖.
• Reductions in TEeff would
be useful.
Noebauer-Huhmann et al. Eur Radiol 2007; 17: 1147-1155
29
Use #3: Reduce T2 Blurring
Augustine Me et al. Top Magn Reson Imag 2004; 15:207
Glockner et all. RadioGraphics 2005; 25: 1279-97
Use #4: Susceptibility
Artifacts – Air Sinuses
• Regions of air/bone/soft tissue causes local gradients
due to differences in magnetic field susceptibility
8
8/1/2011
Susceptibility Artifact Reduction
with Parallel Imaging
―Turn Key‖ Parallel Imaging ?
R=1
R=2.0
R=2.8
R=3.2
R=4.0
• Shortening TE
helps (must have
less phase encodes
to do this).
• EPI-based
sequences gain
more in general
(e.g., DWI,
perfusion)
Top – normal
acquisition,
Bottom – R=2
acceleration
―Turn Key‖ Parallel Imaging ?
R=1
R=2.0
R=2.8
R=3.2
R=4.0
Use #5: Contrast-enhanced
MR (MRA)
Left: R~ 1.5; Right: non-pMRI with reduced FOV
• Improved spatial resolution for a given scan time.
Wilson, et al. Top Magn Reson Imag. 2004; 15: 169-185
9
8/1/2011
Use #6: Cardiac Imaging
Balanced FFE MRI
A&B: 11 sec breath
holds
C&D: 5 sec breath
holds + R=2
Van den Brink, et al. Eur. J. Rad. 2003; 46: 3-27.
Drawbacks/Consideration of
pMRI:
SNR Properties & Artifacts
37
SNR
38
Non-Uniformity of Noise
SNR is a concern with pMRI for three reasons:
• Non-uniformity of signal (array coils)
• Non-uniformity of noise (pMRI)
• Lower signal from acceleration (pMRI)
Larkman DJ et al. Magn Reson Med 2006; 55:153-160
10
8/1/2011
Key SNR Parameters
in Parallel Imaging
Key SNR Parameters
in Parallel Imaging
• SNR depends on number, size and orientation of
the coil elements
• SNR depends on number, size and orientation of
the coil elements
norm
SNR
PI
i , j ,k

SNR i , j ,k
g i , j ,k
norm
SNR

1
g (r , R)  R 2
PI 
SNR (r )
R
• R: acceleration factor
• g: coil-dependent noise amplification factor
(non-uniformity that we observed)
• R: acceleration factor
• g: coil-dependent noise amplification factor
(non-uniformity that we observed)
SNR vs. Acceleration
g-Factor Calculated Maps
32-channel coil, 1.5 T magnet
R=2
R=3
R=2
R=3
R=4
R=5
R=4
R=5
R=7
R=6
R=6
R-L Phase Encoding
Short-axis cardiac images – 32-channel coil – 1.5 T magnet
Reeder SB et al. MRM 54:748, 2005
• g-Factor changes with R
R=7
A-P Phase Encoding
Reeder SB et al. MRM 54:748, 2005
11
8/1/2011
Artifacts
Artifacts associated with pMRI may or may not
be subtle.
Artifact #1: Tissue Outside of FOV
(SENSE)—Wrap-around artifact
Unalias
What happens
Undersampling…
with pMRI
when the FOV is too
small?
Similarities to conventional MRI artifacts
(aliasing, ghosting).
Important to prescribe the acquisition
properly, and to avoid movement.
Center region in this example should be unaliased, for
Normal
FOV
acceleration
R=2.
Smaller
Treated as non-aliased tissue during FOV
reconstruction.
Examples: Phantom and
Patient
• With SENSE-based technique, tissue outside of the
FOV yields ―wrap-into‖ artifact
Artifact #2: Motion After Calibration
Scan (SENSE or GRAPPA)
Calibration scan must accurately represent tissue position.
Goldfarb, JMagn Reson Imag. 2004
Small FOV
SENSE
Normal
FOV
Small
displacement
Medium
displacement
Large
displacement
12
8/1/2011
Artifact #2: Motion After Calibration
Scan (SENSE or GRAPPA)
Clinical Artifact Examples
Affected by FOV choice as well.
Small FOV
Large FOV
Not aliasing, folks!
Clinical Artifact Examples
Pseudo-‖failure‖ of fat sat: Patient moved between reference and
3D artifact:
faintscans
ghost nearIAC
thestructure
middle ofghosting
FOV that resembles structures located at
SENSE
the edges of scanned volume (nose, ear).
pMRI and Traditional Artifacts
Appearance of traditional artifacts may be modified by pMRI
Susceptibility (artifact not perfectly represented on sensitivity
map)
Thin, bright structures in the periphery of sensitivity map—mismatch
between sensitivity and anatomy.
simulation
phantom
13
8/1/2011
ACR QC vs. pMRI QC
SNR=Signal
U=1|maxROILarge
-minROI
|/|maxROI
+min
ROI/Noise
Small
ROI ROI|
Quality Control
ACR approach to SNR, Uniformity
(using volume coil, signal and noise
are fairly uniform in many cases)
TG118
Preliminary
Results: SNR
ACR QC vs. pMRI QC
SNR=mean(I
NAAD=11/N
× (I
/[sqrt(2I
j,1+I
j,2)ROI
j - <I>)/<I>
j,1-Ij,2)ROI]
SNR
In pMRI, signal is not uniform (array
coils), and noise is not uniform (gfactor).
With
Update:
SNR TG118
ROI in has
center,
will
focused
artifacton
perturb
differences in
measurement?
SNR, uniformity as
acceleration changes.
14
8/1/2011
TG118 Preliminary
Results:
Uniformity
Thoughts from TG118 on
pMRI QA
• One must always remember that metrics and
system performance can be interconnected
(e.g., SNR vs. uniformity vs. artifacts)
• Not clear that any uniformity measure can
satisfy all three criteria below
• easy to measure
Uniformity
Update: TG118 has
on differences
Is focused
more sensitivity
good? in
SNR, uniformity as
acceleration changes.
• Sensitive to aberrant non-uniformity
• Insensitive to coil-inherent non-uniformity
2D SENSE (with 3DFT MRI)
Acceleration > 1D
2D SENSE reconstruction (2X in L-R and 2X in A-P) from an 8-channel
head array coil conjugated gradient iterative solver after 10 iterations .
http://www.nmr.mgh.harvard.edu/~fhlin/tool_sense.htm
15
8/1/2011
Undersampling in the Temporal
Domain
TSENSE
2D Cine TrueFISP
• Parallel imaging in the temporal
domain: Time Auto-Calibrated
GRAPPA (TGRAPPA); TSENSE
• Self-Calibrating Non-Cartesian SENSE
R=2
6 heartbeats
R= 3
4 heartbeats
R=4
3 heartbeats
36 msec true temporal resolution, 144 x 256 matrix
Peter Kellman, NHLBI, and Al Zhang, Siemens R&D, Chicago
Future Directions
Importance of pMRI
• Increases MR imaging speed
• pMRI with massively parallel arrays
• Transmit parallel imaging
• Is applicable to all MRI sequences
• Is complimentary to all existing MRI
acceleration methods
• Can often reduce artifacts
• Alters SNR in MR images
16
8/1/2011
Application of pMRI
• pMRI offers the promise of high resolution MR
imaging at speeds as fast as MSCT
• Applications of parallel imaging include FSE,
cardiac MR, diffusion and perfusion EPI brain
imaging methods, 3D FT MRI (and MRA).
Acknowledgments
• Current and Past Members of TG118
• Jason Stafford, Lisa Lemen, Max Amurao,
Geoff Clarke, Ron Price, Ishtiaq Bercha,
Michael Steckner
• Parallel imaging is tool for managing RF heating
in the body at 3T and higher field strengths
• Frank Goerner (UTHSCSA)
• Parallel imaging and dedicated RF coil design
are enabling technologies for high Bo MRI
• Ed Jackson (MDA), Lawrence Wald (MGH),
Jerry Allison (MCG)
The Future of the
Future?
• 64 strip
detectors used
to image a test
object with a
single readout
(MacDougal and
Wright, 2005).
17