Ghost are the biggest limitation in high res fMRI. Similar to low resolution fMRI, ghosts in sub-millimeter EPI are arising from mismatch of k-space lines. This mismatch can be associated with (A) the actual readout itself or (B) inappropriate GRAPPA auto calibration data.
Here I try to make notes of strategies that I found helpful to minimize ghosts.
tSNR maps for various ACS schemes: FLASH is better than FLEET. FLEET is often better than single shot. And “segmented” is usually the worst.
FLASH-GRAPPA
FLASH GRAPPA can have a huge impact on signal tSNR. Here an example of segmented vs. FLASH GRAPPA.
In FLASH GRAPPA, the reference lines are acquired without most of the biases of EPI-based reference data. E.g. no distortions, no delays between lines etc.
There might, however be other inconsistencies. E.g. non-steady state with sub-optimal flip angles, stimulated echoes, reduced regularization of noisy ACS. Playing with the different parameters, usually can accounts for it.
You can change the kernel size and the “optimized” GRAPPA algorithm at the scanner as follows:
cmd → xbuilder → c:/MedCom/config/Ice/IceConfig.evp → ICE/CONFIG/iPAT/Improved Grappa set to 1
cmd → xbuilder → c:/MedCom/config/Ice/IceConfig.evp → ICE/CONFIG/iPAT/GrappaPSize set from 4 to 2
FLEET-GRAPPA
In FLEET, the reference lines are acquired in a different order.
Is supposed to work just as well as FLASH GRAPPA, may be even better. For me, however, it amplifies the ghosts super bad.
Ramp-Sampling
The SC72 (aka XR-AS82) body gradient coil in almost all the SIMENS 7Ts has pretty bad acoustic resonances in the range of 550 Hz and 1100Hz. For high-res acquisition below the the limit of PNS, I am almost always very close to the resonances.
The only thing that really helps, is to play with the bandwidth to change the echo spacing. Usually, the ghosts are minimal, with smaller ramp sampling ratios.
It’s the combination of being close to the acoustic resonances and having a large ramp sampling that makes the ghosting worse.educational animation to depict the entanglement of echo-spacing and bandwidth.
Note that in SIEMENS, you generally do not have a control on the amount of ramp sampling. See figure below on the gradient response to changing individual parameters.
The EPI-ghosting can be associated with the amount of ramp-sampling. For constant gradients during the readout, any delay is associated with a global phase offset of the k-space line. As soon as the gradient is not constant, a given delay has a variable phase offset across the position. This means that for ramp sampling, a “global” phase correction does not fully account for gradient delays anymore.
There are three phase correction algorithms implemented in SIEMENS.
global, which applies the same phase offset across the whole k-space line. (RetroRecon code = 7)
local, which has two parameters a global and a linear term. Ahn, C.B., and Cho, Z.H. (1987). New Phase Correction Method in NMR Imaging Based on Autocorrelation and Histogram Analysis. IEEE Trans. Med. Imaging MI, 32–36. (RetroRecon code = 68)
pixel by pixel. (RetroRecon code = 0)
If not implemented in the sequence special card, you can change that as follows:
cmd → twix → select dataset (left) → Book symbol with x bottom right → edit → iOnlinePhasCorrectionAlgo from 7 something else → save → in twixwindow “start”
2 thoughts on “How to deal with Ghosts in high-res EPI”
Thanks for the nice tutorial.
Something confused me, you first say
“the ghosts are minimal, with larger ramp sampling ratios”
and later
“As soon as the gradient is not constant, a given delay has a variable phase offset across the position. This means that for ramp sampling, a “global” phase correction does not fully account for gradient delays anymore”
your second statement means that ramp sampling leads to wrong phase correction, therefore, ghosting should me larger. Isn’t it?
Thanks for you comment Ali.
You are absolutely right.
I am sorry this was confusing. It was actually an error. I meant to say that the ghosts are minimal, when the ramp sampling ratios are smaller. For smaller ramp sampling ratios, the signal is acquired while the gradients have the same magnitude. This means that a gradient delay can be nicely approximated with a global phase offset. And Thus the Nyquist-ghost correction algorithms work more reliable.
layer fMRI blog 
Thanks for the nice tutorial.
Something confused me, you first say
“the ghosts are minimal, with larger ramp sampling ratios”
and later
“As soon as the gradient is not constant, a given delay has a variable phase offset across the position. This means that for ramp sampling, a “global” phase correction does not fully account for gradient delays anymore”
your second statement means that ramp sampling leads to wrong phase correction, therefore, ghosting should me larger. Isn’t it?
LikeLike
Thanks for you comment Ali.
You are absolutely right.
I am sorry this was confusing. It was actually an error. I meant to say that the ghosts are minimal, when the ramp sampling ratios are smaller. For smaller ramp sampling ratios, the signal is acquired while the gradients have the same magnitude. This means that a gradient delay can be nicely approximated with a global phase offset. And Thus the Nyquist-ghost correction algorithms work more reliable.
LikeLiked by 1 person