This course covers the analysis of Functional Magnetic Resonance Imaging (fMRI) data. It is a continuation of the course “Principles of fMRI, Part 1”
Module 5: Optimizing acquisition for your study goals
Loading...
来自 Johns Hopkins University 的课程
Principles of fMRI 2
85 个评分
从本节课中
Week 1
This week we will discuss psychological and behavioral inference, as well as advanced experimental design.
与讲师见面
Martin Lindquist, PhD, MScProfessor, Biostatistics
Bloomberg School of Public Health | Johns Hopkins University
Tor WagerPhD
Department of Psychology and Neuroscience, The Institute of Cognitive Science | University of Colorado at Boulder
Welcome back. In this module we're going to talk about
optimizing acquisition for your particular study goals.
There's no one best way to obtain fMRI images.
There's a sequence of tradeoffs among different choices.
And many acquisition choices are standardized across fMRI centers or
across studies within a center.
But otherwise, others can be customized to fit your particular study goals.
The goal of this module is to review acquisition parameters and
discuss some of the tradeoffs involved.
So that you can make smart choices for your study when possible.
One example of a standard parameter setting or
choice is which sequence to use.
So echo-planar imaging is the most commonly used.
Collect slice by slice data from the bottom to the top of the brain.
But there are many variants and alternatives for how to do this well, and
those relate to how the image space is sampled.
So there's also spiral acquisition or volumetric acquisition choices or
sequences.
How potential artifacts are dealt with, so these include things like spin echo
sequences or Z-shimming to reduce artifacts.
Multi-echo sequences for better neural specificity and
slice-specifc echo-time and flip angle sequences.
They also include other kinds of choices like accelerate the data
acquisition options.
And this includes parallel imaging and
multi band simultaneous motor slicing imaging as well.
So it might be difficult to have very deep expertise in all he possible techniques
and tradeoffs.
If you're not a physicist or an engineer working in the ares.
But you can learn how to talk to a collaborator or physicist or engineer and
how to make good choices for your study.
There's some basic limitations, the first one is the time.
Whatever sequence you use, there's a limitation in how much data collect how
fast, and that's a fundamental limitation that governs many of these tradeoffs.
Secondly, there's a signal to noise limitation.
FMRI data are inherently noisy, there's thermal noise related to the scanner.
There's physiological noise related to head movement and
head movement induced by other kinds of physiological artifacts, and
other kinds of vascular artifacts as well.
Artifacts are endemic in fMRI images.
All images have some artifacts.
In fact, the T2* weighting that produces sensitivity to the bold affect,
the blood flow and oxygenation level, depend on the effect of interest,
are also those that produce susceptibility artifacts.
Particularly at the boundaries between tissue types, and especially brain and
air spaces around the base of the skull.
So these vary across participants due to the head geometry,
how the sinuses are organized and
arranged and the topography of those spaces as well as other factors.
So they can't be eliminated, but we can choose sequences and
parameters in order to be smart about the basic tradeoffs.
So that's where we get to choose, and
there's a space of different points in a series of tradeoffs.
So on the top of the pyramid here we have coverage, how much brain you're covering.
Which tradesoff with the other parameters including spatial resolution,
temporal resolution, and susceptibility artifacts.
So let's look at coverage, that means how much of the brain you acquire.
Which always takes time, so
you can do things to increase the coverage very simply, acquire more slices.
But there are pros and cons to each listed here in green and red respectively.
So the coverage of getting,
the pros of getting better coverage include better registration of the brains,
better normalization or warping across people so there's better alignment.
And of course there's more brain areas to look at that might be of interest.
The cons are in general you have to go slower that reduces temporary resolution.
It reduces spatial resolution because you have to get bigger box holes if you're
acquiring more of the brain.
And it increases susceptibility artifacts as well to do those
things to sample more slowly and with bigger box holes.
So let's look now at spatial resolution.
In spatial resolution higher spatial resolution gives you sensitivity and
more localized signals.
And that's determined in part by how small your voxels are.
And also by the intrinsic spatial resolution of the scanner and sequence.
It's fundamentally limited by the underlying resolution, or the point spread
function of bolt, which depends in part on the vasculature and the field strength.
So some pros of increasing spatial resolution,
you can detect activity in smaller brain regions.
There's less partial voluming, which refers to averaging over regions
that do different things with different populations of neurons or even different
tissue types, grey matter and white matter, grey matter and fluid space.
In general, you can get better decoding and
prediction at 3T with voxel sizes down to about two millimeters.
And after that, there are additional issues.
At 7T, you can go to smaller voxels and that will likely improve
accuracy in decoding and lots of varied analysis as well.
And finally, there are reduced susceptibility artifacts.
Lot of pros, their cause in signal to noise,
as you get smaller in voxels and also in temporal resolution.
Let's now look at temporal resolution, the third corner of the pyramid.
In temporal resolution refers to how slow the signals changes need to be for
you to detect them.
In terms of underlying brain or physiological activity.
This is determined in part by the TR or the volume acquisition time.
There's a fundamental limitation here too, which is the slow human dynamic response,
which rises slowly and peaks at about five to six seconds, as we've seen.
However, artifacts can happen much more quickly, related to movement or
heartbeats, etc., in the brain.
And so there is some value in also sampling quickly to separate out
the artifacts from the real signals.
So some pros of increasing temporal resolution,
we get better sampling of the human dynamic response function and
connectivity, particularly across brain regions.
We have better ability to separate out and remove artifacts.
There's less aliasing of high frequency physiological artifacts into this
space of the task design or psychological pattern that you are interested in.
And there's lower susceptibility artifact primarily because there's reduced head
motion during the acquisition of a volume.
There's also large cost which is the cost in time and it's coverage and
in spatial resolution.
Now we'll turn to susceptible artifacts which are unwanted consequence
of bold imaging and susceptible artifacts take two basic forms.
Signal dropout or loss of signal and geometric distortion and
this is particularly treat the boundaries between tissue types like air and
sinuses with brain.
Moving structures,
like large blood vessels also produce susceptibility artifacts.
And as you can see from the movie here, this ping pong ball even though,
a spherical shape when you sample it in the scanner can often have distortion.
So that's an example of an artifact.
So some pros of sequences that reduce susceptibility include better signal
fidelity, better localization particularly in high susceptibility areas like
the orbitofrontal cortex, the amygdala, the ventral temporal cortices.
Some cons are reduced sensitivity to the true neurovascular signals of
interest because the signal strength is going down overall usually.
In some procedures can give you a win, but they cost in time, in coverage or
spatial resolution, z-shimming as an example of that.
Where there's a benefit in certain slices that are high susceptibility, but
there's also a cost because it takes longer to acquire those images.
So now let's look at some parameter choices and their effects in this section.
So let's review some parameters first.
We collect data slice by slice, there's one slice.
So one parameter is the number of slices, and
also the slice thickness in millimeters.
There's the head tilt angle, both of the person in the scanner, but
here we're referring to the prescription of the slices.
Where they're tilted or oriented along the head, that also matters.
There's a pre-scan period where certain procedures like
fat saturation that we'll talk about soon are set up, and
shimming is set up to homogenize the magnetic field.
And then we collect one volume per TR, as you see here.
When we sample one slice, the area that we're sampling in that slice is
the field of view, and there's also a grid that's the data sampled on.
And the matrix size refers to how many cubes you divide up that field of
view into.
So a typical matrix size is 64 for high res imaging, it can be greater.
And the voxel size is determined by the field of view divided by matrix size,
and also by the slicing this.
So 3x3x3 millimeters would be in terms of the x, y and z directions.
And isotropic voxels, mean ones that
are equal 92 to cross all dimensions.
So let's take a look at some of this parameters.
The TR is a volume of time between acquisitions and
the basic recommendation is to keep this under 2 seconds.
We found that 1.3 seconds has a good profile for
many people in terms of reduced aliasing of cardiac noise.
So 1.3 is a value that we like in many of our studies.
That also reduces head movement during the image more than a long TR.
You can also go much faster in some cases, we'll see that soon.
Voxel size is how big the voxels are in the three dimensions, and
one thing to know is that the signal to noise varies
inverse proportion to the volume of the voxel.
So if you have 3 millimeter voxels that's 27 cubic millimeters and
you compare that to 1 cubic millimeter, 1 x 1 x 1 voxels,
the 1 x 1 x 1 has 27 times less signal to noise.
Now that may be a rate limiting factor if thermal noise is a big issue.
Or it may not matter too much, because I might be dominated by
the physiological noise of moving your head and so forth.
So generally down to 1 or
2 millimeters is, you won't be at the thermal noise threshold usually.
But if you really tried to go smaller than that then thermal will become a big issue.
1 other point is that at 3 Tesla imaging, then the intrinsic
smoothness of the bold images are close to the 3 x 3 x 3 millimeters,
perhaps somewhat less, maybe 2.5 millimeters.
And so if you go smaller than that, you're really not getting anything in
terms of intrinsic resolution, but you're losing so much noise.
And that's why some studies, like Chen Et Al 2011 is a paper I like, and
they found that something close to 3 x 3 x 3 at 3T at 7 Tesla.
Smaller boxes are going to be helpful.
Now let's look at coverage, that's the amount of brain covered.
And I recommend whole brain or
near whole brain unless you have special reasons to do otherwise.
And that's primarily because of the issues with spatial registration and
normalization if you acquire a partial volume.
Finally, the field of view, and we like to set this,
this is our set based on the voxel size that we like.
So one thing to check for is aliasing or wrapping of the image in space, folding
over, if the field of view is too small and too tightly coupled around the brain.
And there are some ways to deal with that if that happens.
Another parameter is the flip angle and this is refers to,
these are parameters that are about how we require the images.
Flip angle is the amount of magnetization applied to the spins to
tip them over into this transfers frame.
90 degrees are a typical value for very long TR like three seconds or
more but it's TR dependent.
There's an angle called the Ernst angle,
which is a roughly optimal flip angle for a given TR.
So a rule of thumb is the Ernst angle minus about 10 degrees.
So for a 1.3 second TR the flip angle would end up being about 70 or so.
Echo time is the time from application of the RF pulse in the image
to the time that you read out.
And the optimal TE depends on the field strength, but
it also depends on the local susceptibility of artifacts.
The way to think about the tradeoff is, if there are high susceptibility artifacts,
you want to lose signal in order to also reuse the artifacts and
a shorter TE is going to be better.
This is true at the bottom of the brain.
At the top of the brain, you have lower susceptibility to artifacts and
you can go a little bit longer on the TE.
And still get really good signal, and then you get more signal strength too.
So at 1.5 T, 40 to 50 millisecond TEs are optimal at 3T,
about 22 to 20 milliseconds is a good range.
And on the lower end if you really care more about the bottom of the brain,
and on the higher end if you care more about the top of the brain.
There are optimized sequences which have slice specific TE values.
The bandwidth is the frequency at which the signal is digitized and sampled.
And if you've got higher bandwidth increases acquisition speed at some
cost to noise.
And this is often set close to the maximum value that the scanner can tolerate which
means you can get more images faster at higher temporal resolution at some
cost to SNR.
Finally, fat saturation refers to a pulse that's often applied
before each volume that's acquired every TR.
That essentially saturates the signal in fat with radio frequency energy and
this is recommended.
because it reduces facial distortion around the fat water boundaries at
the edges of the brain, so some time cost to doing that.
And finally, the tilt angle is something very simple that you can control.
And this is the angle of the slice prescription from roughly horizontal let's
say to tilted back.
And what we and many other groups like is something that's tilted 10 degrees
up from the anterior posterior common straight line.
Because that tends to reduce artifacts in the over frontal cortex and
the amygdala to some degree.
But this is also dependent on the scanner you choose and the sequence as well.
Now we've two more choices here, and the next one is a parallel imaging factor.
So this refers to acceleration of the imaging in a plane within a slice,
by skipping some portions of k-space and essentially undersampling the image and
then going fast.
It relies on the use of multiple coils,
are also called channels in the head coil to reconstruct the image and
unwrap these things which are aliased and folded on top of each other.
There are different names on different platforms.
On GE, it would be GRAPPA.
SENSE on Philips or iPAT settings on the Siemens scanners currently.
And 1 is no acceleration, or greater than 1 means more acceleration,
higher acceleration factor.
So some basic recommendations R,
some parallel imaging is probably okay and probably necessary if you want whole
brain coverage with the TR of 1.3 seconds in 3 millimeter voxels.
You'll have to do some parallel imaging and so that's fine.
There's only really a small cost often to SNR and
reduced artifacts due to faster imaging.
But you should check that on your images called some public data.
The second basic option is multibanner simultaneous multislice imaging.
And this refers to acceleration of the acquisition of a volume
by acquiring multiple slices at the same time.
So this is a relatively new innovation in 2015 and
again relies on the use of multiple coils to unwrap the image.
And this is recommended if its palliative at all at your center.
It can be quite a big win and a great thing but use it with caution,
because it's really in development now.
So there are benefits but the same benefits has led to parallel imaging but
greater speed up.
So right now we're using a multiband eight which is eight slices at once which
is pretty aggressive.
We've done a lot of tests, that's what we recommend for our set up at least.
Combining it with parallel imaging is not recommended, because that really
increases the artifacts especially if you're moving your head in the field.
So it multiplies the artifacts together.
So now here's a summary of these tradeoffs that we talked about.
And what we're looking at it the tradeoff between coverage, spatial resolution,
temporal resolution and susceptibility artifacts.
And here what you can see if things that you might want do on the rows there and
on the columns, choices that you can make.
So for example, if you scan at a higher field strength there are winds, and
contrast noise, and bold sensitivity.
Which is what you want at the end of the day, and special resolution.
But there are costs in susceptibility artifacts and
susceptibility to motion induced artifacts as well.
If you choose a shorter TE for
your field strength then you reduce susceptibility artifacts that's good.
But it cause you in contrast noise and both sensitivity to some degree.
If you choose smaller voxels then if it's in spatial resolution,
it's susceptibility artifact resistance.
But cause potentially in coverage,
in contrast noise in both sensitivity and temporal resolution.
If you choose a shorter TR, that's faster imaging,
there are gains and temporal resolution, susceptibility artifact reduction and
physiological artifact reductions.
And their costs, again in potentially brain coverage In contrast to noise and
in spatial resolution.
And finally,
if you choose accelerated imaging, you can get a lot of potential benefits.
They include increased coverage given other parameter settings,
increased spatial and temporal resolution, increased susceptibility reduction,
artifact reduction and reduced sensitivity to physiological artifacts.
But there are costs intrinsically in contrast and noise and
bold sensitivity because of the acceleration.
So this can help you think through some of the tradesoffs for your study.
That's the end of this module.
Tune in again next time.
