Neurohacking describes how to use the R programming language (https://cran.r-project.org/) and its associated package to perform manipulation, processing, and analysis of neuroimaging data. We focus on publicly-available structural magnetic resonance imaging (MRI). We discuss concepts such as inhomogeneity correction, image registration, and image visualization. By the end of this course, you will be able to: Read/write images of the brain in the NIfTI (Neuroimaging Informatics Technology Initiative) format Visualize and explore these images Perform inhomogeneity correction, brain extraction, and image registration (within a subject and to a template).
Pre-Processing with ANTSR
Loading...
来自 Johns Hopkins University 的课程
Introduction to Neurohacking In R
133 个评分
从本节课中
Image Processing
In this section, we will discuss the steps done to process brain MRI data. We will discuss inhomogeneity correction, brain extraction or skull stripping, and various image registration techniques.
与讲师见面
Dr. Elizabeth SweeneyRice Academy Postdoctoral Fellow
Biostatistics
Ciprian M. CrainiceanuProfessor
Biostatstics
John Muschelli IIIAssistant Scientist
Biostatistics
[MUSIC]
This is Theme 3, Lecture 9, so
we're going to do some image pre-processing with the ANTsR package.
So two of the main things we will focus on
here is the bias field correction using the N4 algorithm, and image registration.
So, I have written the extra ANTsR package,
because I wanted to interface some of the ANTsR functionality
with tying it into some of the oro.nifti functionality, right.
So for the majority of the lectures, we've spoken about nifty objects in R,
implementing oro.nifti package,
and that's the main container for images I kind of use a lot.
So I wanted the ANTsR functionality to be able to work on these types of images.
So, I wrote a simple wrapper for the N3 bias correction and
the N4 bias correction.
Those are two functions in the ANTsR package.
And they're wrapped into a very simple function called bias_correct.
Right now the, again if you don't have the extra ANTsR package downloaded or
loaded up, you can install that through GitHub with
muschellij2, muschellij2/extrantsr.
And there's no double A, it's extrantsr or something like that.
I don't know exactly the name yet.
But the idea is you can use the bias_correct function.
And you can pass in either a final name or a nifti object, and then you just tell
what type of correction you would like to do, either using N4 or N3.
In general, we recommend using N4.
It is a very similar algorithm to the N3 algorithm that came around with sleds.
That's his last name, his dissertation.
And the N4 improves on that with respect to convergence property,
some speed, and just some problems that had occurred in
the N3 that they've tried to smooth out, no pun intended.
So, the function here, right, we pass in a nifti object, in this case nim.
There are a lot of instances where we previously discussed how to read in
a nifti object.
And on the left-hand side we just assigned it to a new object.
So on the left-hand side would be the bias corrected image for the N4 and the N3.
In this case we use the brainex data, the T1 image.
And on the left-hand side is the original image.
The original T1 from the brainex data on the right-hand side,
it's the N4 corrected Image.
Now, it's very hard to see any differences here.
Especially because when you do the plotting with the orthographic,
even though you might do an inhomogeneity correction,
if you don't keep the scales exactly the same, it can be hard to see.
So we wanted to just show you we took a difference between the images and
mapped those to some colors.
And you can see the color bar on the right-hand side of the image showing
where there are differences in the image between the original image and
the N4 corrected image.
So we took the original image, subtracted off the N4 image,
and we overlaid that with a color scheme here to
show you that there are differences in the top right-hand corner.
If you look at the axial sliced in the bottom left corner of the entire image,
you can see it's blue indicating that it was lower in the bias field correction.
Whereas, if you do the original minus the bias field,
that is negative which means the bias field corrected is higher.
And the red means the original image was higher so
just kind of visualize how the bias field correction did.
And the big things that we also implemented is, although there's
a function called antsRegistration in ANTsR package, and it's very good.
Just we wanted to add some additional functionality on top of that interface,
especially again to allow for oro.nifti nifti objects to be passed in.
So we wrote on the extrantsr package the ants_regwrite.
So this will allow you to do a registration.
So it will read in a file name or take in a nifti object, do a registration,
apply that transformation, and write out the files to disk if you want.
And it will also return a nifti object if that's what you would like.
So let's break down this function very simply.
The ants_regwrite, you put in a file name that can either be a character,
so it can be the actual path of the file name on your disk, or
it can be an oro.nifti object.
You say what you want to register that to.
So again this can be a file name or this can be a nifti object.
So if you've already read in a template object, so for example,
the MNI 152 template that we've given you in the data, that would be an example.
And then we want to transform the data using a rigid transformation.
So it'll take the data, translate it, rotate it, and put it to the image base,
and do an interpolation.
So the interpolation pretty much at the end of the day is if you transform one
image into a different space, right, the voxels don't necessarily line up.
So it's how do you want to get the original image into the new
image space right, and they don't exactly line up?
So what you have to do is usually some interpolation,
which essentially means local averaging.
So take all the voxels that kind of intersect, and average them in some way.
By default, it will do a linear interpolation, so just equally weighting.
And then there's other types of interpolations like sync windowing and
windowing, in a sense, they'll do Interpolation.
But they'll average things a little bit differently.
Things farther away from that voxel will get down weighted.
The remove.warp argument is not necessarily so
important if you're doing rigid transformations.
But when you do some of these advanced diffeomorphic registrations,
these nonlinear registrations,
some of these output fields that allow you to warp the image can be rather large.
And usually you want to get rid of them unless you want to use them later.
So, here's a very quick example we did.
Here's the MNI template on the left, and
that's actually what I registered it originally to.
On the right, we took the N4 bias field correction, or
bias field corrected image that we had gotten from the bias_correct function.
And then we rigidly aligned them.
You see this is the MNI template with the skull on.
You see the image on the right with the skull on.
And although they don't match up particularly well in some areas,
they are essentially in the same space.
We would use this for some rough estimates.
You wouldn't necessarily do this in a population and
say everybody's voxels mean same thing, but just kind of a reorientation and
realignment to a front facing image that all have the same resolution.
So, the different types of transformations,
the common ones I would say that you would use with this package,
there are many others that you can look in the documentation.
The Rigid, the Affine, and the SyN, so the symmetric normalization.
So this is one of those, it's called diffeomorphic.
It's a nonlinear registration.
It's one of the state of the art.
It's very good with respect to matching local areas of the brain.
And it's one we use for a lot of things when we really want to ensure that
hopefully voxel from person A and voxel from person B match up very well,
especially when we go to a template.
Thank you.
