Module 14: Pre-Processing (continued)

Loading...

Из курса от партнера Johns Hopkins University

Principles of fMRI 1

340 оценки

Johns Hopkins University

Principles of fMRI 1

340 оценки

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

Из урока

Week 2

This week we will discuss the fMRI signal, experimental design and pre-processing.

Module 13: Pre-Processing of fMRI Data10:17

Module 14: Pre-Processing (continued)7:42

Познакомьтесь с преподавателями

Martin Lindquist, PhD, MSc
Professor, Biostatistics
Bloomberg School of Public Health | Johns Hopkins University
Tor Wager
PhD
Department of Psychology and Neuroscience, The Institute of Cognitive Science | University of Colorado at Boulder

Hi.

In this module we'll continue talking about pre-processing.

And this is a topic that we started in the previous module.

So here's the pre-processing pipeline, and so pre-processing is performed both on

the fMRI data and the structural scans collected prior to the experiment.

So in the last module we talked a little bit about things like looking for

spike artifacts.

We talked about slice time correction.

We talked about motion correction.

We also talked about co-registering the structural and

functional scan to one another.

In this module, we're going to talk about warping to atlas template and smoothing.

So in general, all brains are different.

And the brain size of two subjects can differ up to 30%.

And so there also may be substantial variation in the shapes of people's

brains.

And so what normalization does, is it allows one to stretch, squeeze, and

warp each brain so that it's the same as some standard brain.

And this is important when we do group analysis,

because we want to be able to compare different brains with one another, and

if we're looking at a single voxel we want to be able to look at that voxel

across the entire population of the subjects that we have.

In order to do that, we need all the brains to sort of live on the same

standard brain space..And so the structural MR image using

the coregistration procedures is now warped into the template space.

And so we take the structural image and we worked that into a normalized space.

So what are the pros and cons of doing this?

Well the pros are that special occasions can be reported and

interpreted in a consistent manner.

So when we talk about a certain coordinate,

if we have it on a standard space we'll always know where that lies.

Results can now be generalized with a larger population,

because we can compare different subjects with one another.

Also results can be compared across studies, if each study is normalized for

the same brain.

And results can also be averaged across subjects.

The cons are that it reduces spatial resolution and

introduces potential errors.

So what are these brain atlases, these standards places that we use?

Well there's two main ones that are often used.

The first is the Talairach atlas which is due to Talairach and

Tournoux and this is based on a single subject and so

this was a cadaver of a 60 year old female and it's based on a single hemisphere.

Here, the origin is set at the Anterior Commissure and

it's orient in the brain is oriented so that a line

drawing the Anterior Commissure with the Posterior Commissure is horizontal.

Recently, the MNI space, or

the Montreal Neurological Institute Brain Atlas has become increasingly popular.

This is combination of many MRI scans on normal controls.

There's 152 in the current standard.

All of these are right-handed subjects, and this brain

atlas tends to be more representative of the population than the atlas,

which is only based on the single subject.

So how do we do normalization?

How do we take a person's brain and stretch it and squeeze it and

warp it on to the standard thing?

Well there's lots of different ways of doing that.

There's landmark based methods.

Here we align anatomical features in the different brains,

we find different landmarks in the brain and we try to align them.

There's volume-based registration where we use a linear or

nonlinear transformations much like we talked about in the coregistration.

There's computational anatomy methods,

here we can use diffeomorphic transformations.

And there's also surface-based methods, so this work on cortical surfaces so

you sort of blow up the brain and do normalization that way.

Now, there's lots of different ways of doing this.

And we're not going to go into detail.

I just want to sort of give you the flavor for these methods.

So the next thing we will need to do after normalization is spacial filtering.

So in fMRI is common to spatially smooth

the acquired data prior to statistical analysis.

This can increase the signal to noise ration, but it can also

validate certain distributional assumptions and remove artifacts.

So for example, we often assume that data is normally distributed, and so

by smoothing we can increase that likelihood by averaging over lots of

the different voxels.

So what are the pros and cons of spatial filtering?

Well, one of the pros is that may overcome limitations in the normalization by

blurring and residual anatomical differences.

So the normalization procedure is not always perfect.

And so sometimes a little bit of smoothing is needed in order to sort of blur

differences between, in anatomical differences between subjects.

Also, the pro is that it can increase the signal-to-noise ratio.

So, if you have coherent region of activation, it's actually beneficial to

average over that region because you keep the same signal but

you decrease the noise, and thus you get higher signal to noise ration.

Again as we said, it may increase the validity of statistical analysis, and then

finally it's also required for Gaussian random field theory, which is often used

in multiple comparison, and which we'll talk about later in the course.

So what are some cons?

Well again we reduce the image resolution by spatially smoothing, and

so we lose information in space.

So when doing spatial filtering, we often use a Gaussian kernel.

And the size of the kernel is determined by something called the full width at

half maximum, and this measures the width of the kernel at 50% of its peak value.

And what does this mean in terms of the standard deviation?

Well, there is a relationship between he full-width half maximum and

the standard deviation.

So, the standard deviation is the full-width half maximum,

divided by two over the square-root of two, the natural logarithm of two.

>> So what should we use when performing spatial smoothing?

Well the matched filter theorem states that a filter that's matched to the signal

extent will give optimal signal to noise ratio.

So if we know that the spatial extent is of certain size, we should smooth over

those voxels, because then we retain the signal while reducing the noise.

But however, typically the amount of smoothing is chosen at priori before

we look at activations, and it's usually independent of the data.

So hence, the likelihood what we get

our smoothing kernel matched to the signal extent is very unlikely.

Furthermore, the same amount of smoothing is applied throughout the whole image.

And the spacial extent of activation is probably going to differ across the brain.

So in order to circumvent this, one could use adaptive smoothing methods.

And there are such methods including non-stationary

spatial Gaussian Markov random fields,

and this allows the smoothing to vary across both space and time.

And so, let's revisit the pre-processing pipeline now that we've talked about all

the different components.

So, again both the structural and the functional images undergo pre-processing.

So again the structural images which are often T1 weighted,

they're coregistered onto the functional data and

warped onto a stereotactic space such as the m and i space.

The functional data undergoes a little bit of denoising,

slice time correction and motion correction.

Then, the normalization parameters that were used, that were derived from

the structural images, are also applied to the functional images so that they

are also aligned to the MNI space, or whatever standard space that you're using.

Finally, often times the functional images are smooth, and

once you're done with that you're ready to start analyzing the data.

Okay, that's the end of this module.

Here we've talked about pre-processing, and

now we should be ready to start doing some statistical analysis.

Okay, I'll see you in the next module, bye.

Ознакомьтесь с нашим каталогом

Присоединяйтесь бесплатно и получайте персонализированные рекомендации, обновления и предложения.

Coursera делает лучшее в мире образование доступным каждому, предлагая онлайн-курсы от ведущих университетов и организаций.

© Coursera Inc., 2018 Все права защищены.

Загрузить из App Store

Загрузить в Google Play