Basic Processing for Structural Images Flashcards
What are the three main tools for image analysis?
SPM Central- Statistical Parametric Mapping
FSL - FMRIB’s Software Library (Oxford Centre for Functional Magnetic Resonance Imaging of the Brain
FreeSurfer
Why do most researchers tend to use the same types of software for image analysis?
Using the same tools are the wider research community increases our confidence that the what we do will be accepted by the wider community
What is FreeSurfer?
Designed to input a scan and the program itself will analyse the image, gives lots of outputs
It is the most “black box” approach
When given a T1 scan it delineates between grey and white matter and gives us 3D representations/ 3D maps of cortical strucutures, and metrics/measurements e.g. cortical grey matter thickness
What is FreeSurfer’s popular use for?
Cortical analysis
How are SPM Central and FSL similar?
Both come with pre-written pipelines for complex analyses
Both modular- you can tinker freely with your MRI scans
What are the three pre-processing steps for MRI scan analyses?
- File Conversion
- Ensuring correct image orientation data
- Intensity non-uniformity correction
What is the most popular tool for file conversion?
dcm2niix – its used to convert DICOM to nifti files
How do we ensure the correct image orientation?
Software tools like to know “which way is up”, and parts of a nifti header can modified / used to give this information.
- done by running the FSL tool “fslreorient2std” on a nifti image whihc will add correct orientation information
What is intensity non-uniformity?
The “coil” worn in an MRI scanner is an essential part of the scan, but creates distortions in the local magnetic field
A persons head automatically distorts the magnetic field too
- these translate into changes in pixel intensity in the images we aquire
-these result in collected images being generally brighter/darker towards the centre (considered to be noise in the scan)
How to we do intensity non-uniformity correction?
FSL will do this as part of its “FAST” pipeline, for tissue segmentation.
What are the 5 types of structural MRI analysis methods?
Brain extraction
Tissue segmentations
Masks
Image resolution
Pulling it all together
What is brain extraction/ skull stripping?
The removal (zeroing) of all voxels which are not in the brain (making all voxels 0 so they appear black in the scan)
-practically it makes the program run more efficiently as there is less to scan
How do we do brain extraction?
In FSL use BET (brain extraction tool)
What is a limitation of BET?
Can leave bits of brain tissue in the scan
Can also remove parts of the brain from the scan
- can never achieve perfect brain extraction but we can repeat the process until is is good enough
- often a trade off between including more or less things
What is tissue segmentation?
FSL reads a strucutural brain scan (ideally a T1 image) and produces additional images in the same space which denote which voxels contain white or grey matter or CSF
How do we do tissue segmentation?
Through it’s “FAST” pipeline.
What do we get out of tissue segmentation?
We esentially get new images in nifti files which we can load up in the same imaging software
New images are called TPMs (tissue probability maps)- these only contain data where the relevant tissue was present in the main scan
What happens to the value of voxels in the TPMs generated from tissue segmentation?
Voxel values are no longer arbitrary in TPMs
They now represent the % of the voxel which belongs to the type of tissue
e.g 0% CSF, 43% white matter and 57% grey matter
What can we do from tissue segmentations?
We can do useful measurements
e.g. tissue volume, how much brain there is
this can be done using fslstats
What do we need from tissue segmentation to work out the brain volume?
-The total volume of all voxels identified as having some grey matter/white matter
-The mean value of those voxels
Multiplying the volume of all voxels by the average amount of grey matter they have gives you how much grey matter this person has (in millilitres)
How useful is it to know the volume of someone’s brain?
Neurodegenerative conditions lead to brain atrophy - brain volume is a straightforward and intuitive way of measuring this, that structural MRI is perfect for.
What is a limitation of using brain volume as a comparative atrophy measurement between people?
People naturally have different sized brains- bigger brains doesn’t mean healthier
So comparing volume between people needs a normalising factor.
How do we normalise brain volume?
FSL has a pipeline (sienax) that will give you grey and white matter volumes, normalised to the size of a person’s skull
= tells us if a person’s brain has shrunk compared to their own healthy baseline
What is masking?
Concept of using image masks - once we have TPM we can see locationally where the grey matter is which means we can create a mask that gives just location information
Everything in mask tends to have values of 1s and 0s
What happens when you multply the original scan by the mask?
Leaves behind an image of just grey matter with its original values
This is a way of identifying and defining voxels of particular interest for further analysis- useful approach for getting rid of information we arent interested in
What is thresholding?
Making decisions about which voxels should be considered grey matter vs which ones should be considered white matter
Usually grey matter tends to be higher than 0.5
Sometimes researchers may only want pure grey matter voxels, e.g. value 1 only
What is image space?
Images are in the same space when they are in alignment. This means that the same co-ordinates in each image refer to the same bit of brain anatomy.
Overlaid images from one person share the same space e.g. scans layed over with white matter TPMs and grey mapper TPMs
Scans from different people do not share the same space
What is native space?
It is the raw image, unadjusted and how it came out of the scanner
each person has their own native space
How do we make images share the same space?
Via image registration
What is image registration?
Image registration is a processing technique that moves, stretches and squishes one image to match the alignment of another.
It always involves a target image, and the image you are wanting to register to the target.
What are the two broad approaches of image registration?
Linear
Non-linear
What is linear registration?
Used to register images from the same person
It will shift the brain around in the space of the image up/down, left/right, forward/back, roll it, spin it and re-angle it but without deforming its actual shape
What are the degrees of freedom (DOF) in linear and non-linear registration?
linear =With 6 DOF – “Rigid body” registration - this is simply moving the objects around in space without changing the data too much
-e.g. to align the many images that form the basis of an fMRI or DTI dataset before modelling
non-linear= With 12 DOF it’s more extreme - the number of ways the program can move, stretch and squish the image- this may result in a scan being upscaled or downscaled in resolution
What is a limitation with sharing image space?
It introduces noise into the data and into the results
Image registration can blur the image or things don’t end up quite where they should be
But its the best approximation and the experimental benefits it allows outweights the cost of the noise
What is non-linear registration?
Most extreme
Differentially warps section of the brain to make two or more different brains share the same space where the core anatomy is different
Where might non-linear registration be necessary?
To compare a before and after pf the same person’s brain if its had a lot of atrophy over 20 years
If someone’s had a stroke or a tumour and a chunk is now removed and other brain areas have morphed around that
Essentially when the brains have different anatomy to a significant degree then its non-linear we use
What does image registration use to change an image to match another?
Cost function
What is cost function?
A mathematical approach to determine how similar or dissimilar the images being registered are, and therefore when a good registration has been achieved e.g. when the two images are aligned sufficiently
What do we need to generate the newly-registered image?
A method in interpolation- having to change the data, modifying it- the way you go about this is the interpolation
What is nearest neighbour?
This method simply determines the “nearest” neighboring pixel and assumes its intensity value
-not changing any values, just assigning them to the closest pixel in the new image space
What is a caveat of nearest neighbour?
Whilst you preserve the data values in their entirety, its not neatly lined up anymore so the shape gets fairly distorted
What is bilinear interpolation?
Modify the pixel values, they become like a weighted average of the pixels that they are now closest too
What is a caveat of bilinear interpolation?
Whilst it does preserve the shape, it is at the expense of modifying the data
What can we do to make transformations from image A to image B easier?
Save it as a set of instructions on FSL so it can recreate the transformation whenever
It also lets us reverse the transformation too
What is a caveat of manual drawing e.g. drawing around a hippocampus?
Takes significant training to know where the hippocampus is
Human error in the drawings meaning it is not very repeatable
Many MRIs have hundreds of thousands of scans
Blinding- could be bias in the data if we know who is healthy vs disease
What is a template?
A template is not a brain scan of a single person, but the average brain scan of multiple people
It can be created by scanning many people, registering their T1 scans into the same space, and then taking a mathematical average
What are templates used for?
Can be used as a “target” image to register everyone’s scans in the same space
Can be used to pre-draw ROIs
How can a tempate be used with a scan from your participant?
Can take your scan from your participants, do a non-linear registration into the template space where we know where things are
Because we know how to move between the two spaces, we can reverse transform this atlas back into the participants native space to see where these structures are in their brain and how big they are
