1. Receive Coil Arrays and Parallel Imaging for fMRI of the Human Brain
Jacco de Zwart
No longer accelerated. Accelerated to indicate that data were undersampled, but not all examples here are faster imaging
Advanced MRI section, LFMI, NINDS
National Institutes of Health
Bethesda, MD, USA

2. Outline Part I - Receive Coil Arrays Part II - Parallel Imaging
Part III - Parallel Imaging & fMRI
Nicely introduced by Kamil Ugurbil yesterday

3. Part I Receive Coil Arrays

4. Multi-Coil Imaging ≠ Parallel Imaging?!?
"Parallel Imaging" = multi-coil imaging where data are undersampled during acquisition
Will be discussed in Part II of this talk
Parallel Imaging (PI) requires a receive coil array
Typically results in a loss of SNR compared to the equivalent fully sampled case when using the same hardware
Receive coil arrays used without PI
Covered here in Part I of this talk
Generally yields an SNR increase compared to a conventional volume coil

5. Why Use a Receive Coil Array?
Smaller coils 'see' less noise  increased SNR close to the coil
BUT: small coils have a limited field-of-view
SOLUTION: cover the object with small coils
SNR in center same as with equally-sized volume coil (e.g. birdcage)
SNR everywhere else , highest gain close to coils
Signal originates from single voxel
Noise originates from all tissue "observed" by coil

6. What is the Limit? Sample noise should be the dominant noise source
Other noise sources: coil, preamplifier
Sample noise decreases with coil size (to the 3rd power!)
Sample noise increases with field strength (~linearly)
Optimal number of elements (our guesstimates for the human head):
~20 coil 1.5 T
~ T
~ T

7. Number of Coil Elements & Image SNR
average over entire brain
center of the brain
acquired 16-channel data
1, 2, 4 & 8 channel data derived from 16-channel data

8. 16-Element Array vs. Head Coil @ 3 T
image intensity scaling factor
SNR
SNR
GE head coil
128×96
192×144
rate-2 SENSE
128×96
16-element array
Performance gain: 2-fold in center, up to 6-fold in peripheral cortex!

9. Conclusion for Part I
Receive coil arrays outperform similarly-sized volume coils
Equal performance in center of object
Performance gain everywhere else, greatest in periphery

10. Part II Parallel Imaging

11. Undersampled MR Imaging
R-fold undersampling of MR-data
Yields R-fold reduction of acquisition time
BUT: Aliasing in the image
Information is lost due to this folding artifact
Signals from different object regions are superimposed and cannot be distinguished
Unless…
R=2

12. Undersampled MR Imaging
Unless… a receive coil array was used:
Sensitivity profile for each coil element different
 Relative contribution of superimposed signals different for each coil
Allows unaliasing images in post-processing
in image domain: "SENSE" [Pruesmann, Magn Reson Med 1999, 42:952]
in k-space: "SMASH" [Sodickson, Magn Reson Med 1997, 38:591]
Undersampled acquisition with receive coil array + Unaliasing during image reconstruction = Parallel Imaging

13. Parallel Imaging Penalty
With n coil elements, up to n-fold acceleration
BUT: SNR reduced due to reduced sampling
AND: additional noise introduced by reconstruction
generally referred to as g-factor (= spatially varying)
depends on coil configuration
acceleration rate
Parallel Imaging (PI) penalty increases with:
higher acceleration factors
lower number of coil elements

14. Example – human brain imaging w. PI+
Image obtained using SENSE reconstruction
1.5 T GE Signa LX
EPI w. 50% ramp sampling
64×48 / 64×24 matrix
220×165 / 220×83 mm2 FOV
2000 ms TR
40 ms TE
4 mm slice thickness
24.1 / 12.3 ms echo train
Full-FOV images for each individual coil  coil sensitivity information
(acquired only once!)
Undersampled images
4-element dome coil courtesy of Patrick Ledden, Nova Medical Inc, Wakefield, MA, USA

15. Conclusion for Part II
Parallel Imaging increases imaging speed at the cost of image SNR

16. Part III Parallel Imaging & fMRI

17. Does PI Make SENSE For fMRI?
Disadvantage: PI  reduced image SNR
~ g√R [Pruessman et al., MRM 1999, 42:952]
DUE TO: g-factor + R-fold reduction in sampling time
But: temporal stability determines fMRI sensitivity
temporal stability determined by sum of:
image SNR
scanner stability
physiologic noise
Therefore: PI penalty for fMRI typically less than reduction in image SNR
SNR loss is associated with most PI applications, which I call accelerated PI here, like acquisition train shortening or resolution increase
not affected by PI

18. PI-fMRI Sensitivity Penalty
gR
= full PI loss
PI noise increase 1
= no loss!
Illustrated by this figure / all voxels in brain of a normal volunteer / solid line is model / basically demonstrates that resolution can be optimized for given hardware+experiment such that experiment is not dominated by image SNR
intrinsic noise contribution
stability-limited
SNR-limited
All superior brain voxels; normal volunteer; 1.5 T; 4-element coil; 3.8×3.8×4.0 mm3 voxels; rate-2 SENSE; gradient-echo EPI [de Zwart et al., MRM 2002, 48:1011]

19. Does PI Make SENSE For fMRI?
But there are several advantages of PI use:
artifacts 
geometrical distortions 
signal loss in inhomogeneous areas 
temporal resolution 
gradient acoustic noise 
spatial resolution 
important for single-shot imaging at high field

20. Is PI Essential For fMRI At High Field?
When B0 
(A) NMR signal   CNR 
(B) More large vessel suppression  specificity 
BUT:
(C) T2  & T2* 
(D) T1 
(A)&(B)  allows higher spatial resolution
BUT: (C)  blurring 
Parallel Imaging can help:
 higher spatial resolution for given sampling window

21. Example: 3 T Results, R=2 @ 192×144
Mention first EPI time series image
Single-shot gradient echo 3.0 T, rate-2 SENSE
- 16-channel coil [de Zwart et al., Magn Reson Med 2004, 51:22]
- 16-channel receiver [Bodurka et al., Magn Reson Med 2004, 51:165]
- 1.1×1.1×1.5 mm3 resolution (192×144 matrix)
ms TR, 48 ms TE, 14 slices, 73.1 ms readout train
- 5-min scan; visual paradigm stimulates alternately peripheral (red/yellow) and foveal (blue) vision

22. 7 T Results: Single-Shot R=3 @ 192×120
single-shot EPI, rate-3 SENSE, ms readout, 5 min scan time
192×120 = 1.25×1.25×1.0 mm3
background = first EPI volume
finger tapping paradigm
zoom on next slide

23. 7 T Results: R=3 @ 192×120 EPI image from 1st time point
same slice with functional overlay
single-shot EPI, rate-3 SENSE, ms readout, 5 min scan time
192×120 = 1.25×1.25×1.0 mm3
finger tapping paradigm

24. Conclusion for Part III
PI: fMRI penalty < image SNR penalty
PI-fMRI benefits:
Reduce geometrical distortions
Reduce signal loss due to inhomogeneity
Increase spatial resolution
Increase temporal resolution
Reduce gradient acoustic noise
PI-fMRI is important (essential?) for BOLD-fMRI at high field

25. Further Information
Fifteen minutes is too short to cover it all, so if you have questions…
ask me! =
And/or:
The journal NMR in Biomedicine recently dedicated a special issue to Parallel Imaging (May 2006)

26. Acknowledgements National Institutes of Health, Bethesda, MD, USA
Jerzy Bodurka
Jeff Duyn
Peter van Gelderen
Martijn Jansma
Peter Kellman
Nova Medical, Wilmington, MA, USA
Patrick Ledden

27. Thanks a lot for your attention!
Advanced MRI section
LFMI/NINDS/NIH