This course covers the design, acquisition, and analysis of Functional Magnetic Resonance Imaging (fMRI) data. A book related to the class can be found here: https://leanpub.com/principlesoffmri
Module 9: fMRI Artifacts and Types of Noise
Loading...
來自 Johns Hopkins University 的課程
Principles of fMRI 1
392 個評分
從本節課中
Week 2
This week we will discuss the fMRI signal, experimental design and pre-processing.
與講師見面
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 to Principles of fMRI.
In this module, we're going to talk about fMRI artifacts and types of noise.
BOLD fMRI signal contains multiple sources of noise related to the hardware and
also to the participants themselves and what they do in the scanner.
Sources of noise include, thermal motion or
thermal noise caused by free electrons in the system, gradient and
magnetic field instability causing spikes, head movement and its interaction with
magnetic field which can have complicated effects on, noise properties.
And finally, physiological effects, including heartbeat and respiration, and
their effects on the movement of the chest wall, and
also on parameters that interact with the vasculature, like CO2 levels.
These artifacts appear in the data as high frequency spikes.
Image artifacts an distortions that affect all or part of the image for some or
all time points or slices, and low frequency or slow
drift across time, as well as periodic fluctuations at particular frequencies.
So let's talk a little bit about noise and artifact mitigation.
The first line of defense is acquisition, we need good,
quality control processes for the scanner to make sure that it's working correctly.
Things do change over time, hardware goes bad,
things come loose on the table and can create a ton of noise.
So, that's one step.
And the second one is to choose acquisition and
sequence parameters that are appropriate for your goals in neuroimaging.
And we'll talk about that more in the second part of the course,
the second course.
Third, the use of specialized sequences, like spin echo sequences which
are less artifact proned, simultaneous, multi slice imaging, that can avoid some
physiological artifacts for reasons we'll discuss later in this module.
And a specialized procedures like z shimming,
which can help assist help stability of artifacts in certain areas of the brain
like the orbital frontal cortex in the amydala.
And always, always minimize head movement in
the participants because head movement is a killer.
[LAUGH] Second line of defense is in the analysis.
So, when we're analyzing data, the most important first thing to think of is,
look at the data as it's acquired and check for
issues that you could fix before moving forward.
Secondly, we employ, and
many labs now employ, outlier or artifact identification and correction.
There's a number of software packages for doing that.
Standard preprocessing techniques are used to adjust for head movement and drift.
And we'll discuss some of those as we go along in preprocessing.
Number four there are helpful statistical procedures that are increasingly widely
used now with brain imaging data.
One of those is robust regression, which can provide some automated control of
outliers in the massive testing for framework that we are in.
A second one is hierarchical models which can down weight observations or
subjects with low precision, high error variance, due to motion or other things.
And finally, we can model low frequency drift and
periodic fluctuations over time in various ways.
Including with filters and with auto-correlation models, you know,
data analysis.
So here are some common kinds of acquisition artifacts and these are things
you can look for on your data and hopefully avoid if they're present.
So first thing, very basic, check and make sure that you don't have this situation.
The brain coverage that you're getting may not be what you intended.
So here, the top part of the brain has not been acquired, you cannot fix that later.
Number two, RF radio frequency noise and malformed images, this is
one example of a case based artifact with RF noise at particular frequencies.
It shows up like this in these spans across the image, but
there are many other ways in which images can be malformed or gone wrong.
Number three is transient gradient artifacts or spikes.
So what you see here is a plot from AFNI software with a time series of images, and
you see the actual images here corresponding to three points.
Two of them are identified as spikes or outliers.
You can see a huge deviation in the signal time series.
The third one looks okay.
And then what you can see when you look at those images is that there
are bright bands across the images.
And that means that certain slices of the brain have huge signal intensity and
not others.
So this is a transient acquisition artifact.
We all have some of those we have to try to minimize them and
deal with those in analysis to a degree that we can't prevent them.
Four is ghosting, so this is a spacial wrap-around effect,
usually in the Y direction for most images.
And you can see that here quite prominently.
There's a bit of ghosting that's present in all images, but
if it's visible like this, that's a problem.
And finally five susceptibility artifacts or dropout.
So what you can see here is an image where there's a big hole in
the front of the brain and the interior or the pre-frontal cortex.
This happens, the same contrast that creates the bold signal in the first place
also creates susceptibility artifacts.
So it's difficult to mitigate those entirely,
this one is not good for orbitofrontal or prefrontal cortical imaging.
So if that's the area of the brain you really care about, we have a problem.
And six, Task-correlated head movement.
And this is an example we'll look at a little bit later,
of a group analysis result where there are significant activation, in this case,
deactivation, in the ventricles which is not physiologically plausible.
And we think this happened because of task correlated head movement.
And this is something that we can asses as we go along as well
as trying to correct for it after the fact
All fMRI data contain some artifacts, so we have to live with some of them.
But it's very difficult to deal with bad artifacts during the analysis process, so
try to avoid them during acquisition by working with your physicist
to get the best acquisition possible.
>> When modeling fMRI time series data, it's important that we understand
certain non-signal-related components of the signal.
One of the main components is drift.
These are slow changes in voxel intensity over time.
This is also called low frequency noise which is always present in fMRI signal.
So originally people thought maybe it was due to physiological noise, but
now one of the primary reasons is thought to be scanner instabilities,
as drift is seen even in cadavers or in phantoms.
And it's important that we account for drift both in the pre-processing and
when we conduct statistical analysis.
Drift can have serious consequences and it's important that
experimental conditions that very slowly not be confused with drift.
So experimental designs should be used high frequencies,
so we should use more rapid alternations between stimulus on/off states.
Here's an example of a bad design.
Let's say that we do, we have two conditions let's say finger tapping and
rest, and we do two minutes of finger tapping followed by two minutes of rest,
well, this is what the typical drift pattern in magnitude might look like, and
this is what the typical signal magnitude will look like.
So, in this example,
you'll never be able to detect this little signal due to the drift.
However, if we do 20 seconds on, 20 seconds off,
20 seconds on, 20 seconds off.
Then we have a hope of being able to separate the signal from the drift, and so
this becomes an important component in designing our experiments,
which we'll talk about in a later module.
Subject motion during the experiment can also give rise to serious problems, and
typically motion correction is performed in the pre-processing
stages of the analysis.
We'll talk about that in coming modules, as well.
However, there remain some effects of motion even after we've done
motion correction.
This is due to so-called spin history artifacts.
Spin history artifacts is caused by through-plane motion and
complex interactions with a magnetic field.
So basically, even though we correct for
motion there still remains some motion related artifacts in the brain and
it's important we account for them in our statistical analysis.
Respiration and heart rate also gives rise to noise at a particular frequency.
It can potentially be incorporated in our statistical models but
if the TR is too low or the, the, the time resolution is too low
there might be a problem with something called aliasing.
So for standard TR values of roughly two seconds this type of noise is difficult
to remove and is often left in the data giving rise to temporal auto correlations.
So, periodic signals that occur more rapidly than the sampling rate will often
be aliased back to lower frequencies, and so here we see some examples of this.
For example, let's say that the blue sinusoidal curve here is the true signal,
and it's sampled near it's periodicity.
So we're sampling at the red dots here.
So basically, if we have these red dots we would model the component
as a straight line here, this is the fundamental frequency, a flat line.
For another example,
let's say that we sample at about half of the original frequency.
Well this still results in an alias periodicity here about one-forth
of the original frequency In fact, signals faster than one half the sampling rate,
called the Nyquist frequency, will always be aliased.
To avoid aliasing, we must sample at least twice
as the fastest frequency in the signal, and here's a movie illustrating this.
So as we sample quicker and
quicker, you'll see that once we reach two times the signal period,
we'll be able to kind of recapture the periodicity of the signal.
And that happens About there, and then you see that
though the amplitude is a little off, we still capture the signal frequency here.
And then around four times we have basically a perfect reconstruction here.
So, that's an argument for faster sampling of fMRI data.
So, some noise components can be removed prior to or during the analysis and
this includes things like low-frequency drift or
images that were identified as outliers.
However, it's impossible to remove all sources of noise and
therefore significant autocorrelation is usually present in the signal.
In fMRI, we typically use autoregressive AR or
autoregressive moving-averages ARMA processes to model the autocorrelation.
We'll revisit this in a later module.
In general, just what I want to show now is that
the spatiotemporal behavior of these auto-corelations is complex.
So, here I'm showing spatial maps of the model parameters from an AR(2) model,
estimate for each voxel's noise data.
And we see that the values differ across the brain and you kind of
make out structure in the brain and differences between different tissue.
Types, and so, interestingly enough, the noise won't be constant across the brain
but rather will depend on different tissue types and location.
And so this is something that will add increased complexity as we move along.
Okay, so that's the end of this module.
Here we've talked about different artifacts and
noise present in FMRI's signal.
Okay. I'll see you next lecture.
Bye.
