AutoCalibrated Parallel Imaging Techniques

1
Auto-Calibrated Parallel Imaging Techniques

• Grant Elliott
• 6.556/HST.580
• 12/6/07

2
Parallel Imaging

• Undersampling in k-space leads reduced field of
  view and aliasing
• But many coils can sample the same k-space
  simultaneously
• Goal Reconstruct a full FOV image from
  simultaneous reduced FOV images

Full k-space
Subsampled k-space
3
Review of SMASH

• Decompose spatial harmonics into sum over coil
  sensitivities
• Recall that signal is fourier transform of spin
  density
• Formulate composite signal as weighted sum of
  coil signals

4
SMASH Example
Reference
SMASH
5
Disadvantages of SMASH

• SMASH is restricted to coil configurations that
  can be used to generate spatial harmonics
• Sensitivity profiles are not always available,
  particularly in biological imaging
• Need auto-calibration

6
AUTO-SMASH

• Capture R-1 additional autocalibration signals
• Use ACS to find mapping from to

Subsampled k-space with ACS
7
AUTO-SMASH Example
Reference
SMASH
AUTO-SMASH
8
VD-AUTO-SMASH

• Additional ACS captures improve weight estimation
• Increases robustness to noise, non-ideal coil
  behavior

Subsampled k-space with redundant ACS
9
VD-AUTO-SMASH Example
Reference
SMASH
AUTO-SMASH
VD-AUTO-SMASH
10
GRAPPA

• Fit separately for each coil instead of to
  composite ACS
• Fit to multiple phase encode rows
• Produces one image per coil, combined by sum of
  squares

Subsampled k-space with ACS
11
GRAPPA Example
Reference
SMASH
AUTO-SMASH
VD-AUTO-SMASH
GRAPPA
12
Implementations

• Demonstrate AUTO-SMASH, VD-AUTO-SMASH, and
  GRAPPA on four coil data set
• Coil sensitivities are not distributed in the
  phase encode direction
• Data represents complete sampling downsample to
  produce R2 data

13
AUTO-SMASH Implementation

• Determine weights by fitting 0 for all coils
  to composite ACS (blue)
• Solve the over-determined system with
  pseudoinverse
• Use coil weights to fill in missing k-space

14
AUTO-SMASH Reconstruction
Reference
AUTO-SMASH
15
VD-AUTO-SMASH Implementation

• Determine weights by fitting to all ACS
  differences of Solve the over-determined
  system with pseudoinverse
• Use coil weights to fill in missing k-space

16
VD-AUTO-SMASH Reconstruction
VD-AUTO-SMASH
Reference
17
GRAPPA Implementation

• Fit multiple rows of all coils to individual coil
  ACS
• Construct Images for each coil
• Combine by sum of squares to yield final image

18
GRAPPA Reconstruction
Reference
GRAPPA
19
GRAPPA Parameterization
1 Row Fit
2 Row Fit
3 Row Fit
4 Row Fit
20
Comparison of Results
Reference
AUTO-SMASH
VD-AUTO-SMASH
GRAPPA
21
Conclusions

• Generalized nature of GRAPPA makes it very
  powerful (de-facto standard)
• Difficult to reconstruct with Rgt2 for this
  dataset, due to coil configuration
• In practice RN reconstructions are possible, but
  at SNR cost

22
Further Reading

• SMASH, SENSE, PILS, GRAPPA How to Choose the
  Optimal MethodBlaimer, Martin Breuer, Felix
  Mueller, Matthias Heidemann, Robin M Griswold,
  Mark A Jakob, Peter M
• AUTO-SMASH A self-calibrating technique for
  SMASH imaging Jakob, Peter M Griswold, Mark A
  Edelman, Robert R Sodickson, Daniel K
• VD-AUTO-SMASH ImagingHeidemann, Robin M
  Griswold, Mark A Haase, Axel Jakob, Peter M
• Generalized Autocalibrating Partially Parallel
  Acquisitions (GRAPPA)Griswold, Mark A Jakob,
  Peter M Heidemann, Robin M Nittka, Mathias
  Jellus, Vladimir Wang, Jianmin Kiefer,
  Berthold Haase, Axel