1. Brain imaging using FSL
David Field
Thanks to….
Tom Johnstone, Jason Gledhill, FMRIB
Oxford Centre for Functional MRI of the Brain have given permission for re-use of their own course materials in this short course.
FSL stands for FMRIB Software Library
The brain is a small % of our body weight (about 2%), but it uses about 20% of the energy used by the body, in the form of glucose, so it is hungry….
2. This is an image of the vasculature of the human brain, which delivers all that nice glucose to the neurons. Of course, the blood carried in this vasculature is where the BOLD signal that we measure with fMRI takes place. We don’t measure anything in the neurons themselves.
3. And the result of all that is a nice big functional brain activation image! If you can get one like this I suggest you send it to Nature!

2. Overview Today’s practical session will cover
viewing brain images with FSLVIEW
brain extraction with BET
intra-subject registration with FLIRT
inter-subject registration with FLIRT
registration to standard space
This lecture aims to provide
the background needed for today's practical session
practical concepts
an overview of fMRI
Intra subject registration usually refers to the process of finding the point by point mapping between low spatial resolution function image series (4D) to a high resolution 3D structural image. But it can also refer to registration of any two images from the same subject, e.g. a PET scanner image with an MRI scanner image
Inter subject registration is achieved by registering the high resolution structural image to the standard space (MNI template). This registration uses 12 DOF, including expansion / contraction and shearing of the image to get a good fit to the template. Once the 12 DOF transformation matrix is known it can also be applied to the low-spatial res 4D functional series to convert them into the standard space

3. What is imaging? What is imaged in MRI What is imaged in fMRI
Coordinate system
Voxel
Physics
Image intensity values T1 T2
What is imaged in fMRI
Physics (T2*)
BOLD
Hemodynamic
response (HRF)
4D time series
Steps in the analysis of fMRI images
Motion correction
Registration
Affine transform
Brain extraction
Modelling the time course

4. Imaging
Quantity A is not measurable, but it is related to a measurable quantity, B.
requires a model of the relationship or coupling between A and B
As an example of an imaging system, imagine the sun and a wall, with a set of objects between them
Solid object, shirt, sunglasses, water vapour, pane of glass
Use a photometer to measure “directly” darkness of cast shadows
Infer opacity of objects to light
Could object size be inferred by measuring the size of the shadow?
In fMRI the coupling between the measured signal and neural activity is a multi-stage process
stimulus // neural response // vascular response // MR signal
Solid objects would have maximum signal (darkest shadow). A shirt on a clothes line or a cloud of water vapour would have lower signal. A pane of glass would have very low signal, possibly too small to measure.
Could the imaging system be used to infer other quantities? Object size would be hard because shadow size is determined by object size plus its distance from the wall.
The photometer measures photic energy (photons). Object size imaging would require knowledge of the distance of the object from the wall.
The simple model fo the coupling in fmri is experimental stimulus -> neural response -> vascular response -> MR signal. In fMRI there are many cases where the actual BOLD response to a stimulus does not fit the “canonical” model, but researchers often blithely ignore this.

5. What is imaged in MRI?
Images that are intended to provide information about anatomical structure (tissue contrast)
used in hospitals
no information about variation over the time course of the scan
In fMRI anatomical scans are acquired in addition to the functional scans
First, we will take a look at the end product, which is the inferred (imaged) quantity
Then, we will take a brief look at what is directly measured by MRI
In a BOLD fMRI study you acquire anatomical scans because they have higher spatial resolution, and less distorted information about the participants anatomy than the functional scans. You then register the two sets of images together.

6. MRI structural (anatomical) T1 image
Fluid appears dark (CSF)
Bone and fat tend to appear white
Cortical white matter has higher fat content than grey matter, so appears whiter
In a T2 image the grey / white matter contrast is reversed
sagittal
coronal
axial (transverse)
CSF = cerebro spinal fluid
The images show 3 different kinds of slice through the same structural scan. Such a scan is usually (but not always) acquired sagittally, but the other views are easily reconstructed.
The difference between how T1 and T2 images are acquired will be explored later. In medical imaging, T2 images are usually the ones that show pathology
(or transverse)

7. T1 exhibiting exceptionally good grey matter to white matter contrast
An image like this will perform really well if you want to segment grey matter from white matter to extract the cortical sheet, perhaps for subsequent computational unfolding and flattening of the cortical sheet. It’s the Roles Royce of structural scans

8. Directions and locations
dorsal
ventral
anterior
medial
lateral
Anterior also referred to as rostral. Posterior also referred to as caudal. These terms can be used as both locations and relative directions. So, the thalamus is rostral to the occipital lobe, and it is caudal to the prefrontal cortex
Dorsal also referred to as superior, and ventral as inferior
posterior

9. Stereotaxic (talairach) coordinate system
left
anterior
The talairach coordinate system is also referred to as the stereotaxic coordinate system. The origin is at the anterior commissure (more about that in a minute). All neuroimaging studies make use of this common coordinate system. It is important to realise that using the talairach coordinate system does not mean that the data has been registered to the talairach template brain. Most modern studies register their data to the MNI template brain rather than the unrepresentative talairach template brain, which is too small for one thing. But the coordinate system used is the same in both cases. For this reason it would be better if everyone used the term stereotaxic coordinate system to refer to this convention. However, the term talairach coordinate system persists because historically it was first used with the talairach brain.
Every voxel in the images you browse with FSLVIEW has a location defined by X,Y,Z values in mm in the stereotaxic space. You will also see the voxel coordinate, but this has a different origin (the most posterior, ventral, and rightmost voxel). Brain areas also have known X,Y,Z values in the stereotaxic space.
The Talairach coordinate system is defined by making two points, the anterior commissure and posterior commissure, lie on a straight horizontal line. Since these two points lie on the midsagittal plane, the coordinate system is completely defined by requiring this plane to be vertical. Distances in Talairach coordinates are measured from the anterior commissure as origin. Talairach coordinates are sometimes also known as stereotaxic coordinates.
posterior
right
This coordinate system is universal to all template brains. If you see “talairach coordinate system” written in a paper it does not follow that the data was registered to the talairach template brain

10. The origin of the coordinate system
Anterior commissure
Posterior commissure
The Talairach coordinate system is defined by making two points, the anterior commissure and posterior commissure, lie on a straight horizontal line. Since these two points lie on the midsagittal plane, the coordinate system is completely defined by requiring this plane to be vertical. Distances in Talairach coordinates are measured from the anterior commissure as origin. Talairach coordinates are sometimes also known as stereotaxic coordinates.
When I began running imaging studies I used SPM to analyze my data, which required me to manually set the origin of my images. FSL will find the origin automatically when you run registration to a template image. Before registration has taken place the origin will not be set to the anterior commissure and so you won’t be able to use atlases that rely on the origin being set.
Common origin & coordinate system enables comparison between studies

11. The origin of the coordinate system
Here is the anterior comissure shown on a t1 structural

12. Voxel
Each unique combination of X,Y, and Z coordinates defines the location of one voxel
“volumetric pixel”
Each voxel contains a single numerical value
“image intensity”
Grey matter tends to have image intensity values in a certain range, and white matter tends to have image intensity values in a different range
In structural images, voxels often have dimensions of 1*1*1 mm, while functional image voxels are usually larger
In the structural images shown here, a visual image is produced by mapping the numbers onto brightness values that can be shown on a PC screen
How does the scanner generate the number at each voxel location?
The key thing to understand is that one voxel always equal one number, and that number is all the information you have in the part of the brain equal to the area covered by that voxel.

13. The three most important components of an MRI scanner are the static magnet field, the gradient coils, and the RF coil. I’m going to say very briefly how each of these three components contribute to making images, that is values of image intensity at spatial locations defined by voxel coordinates.
Block diagram of an MRI scanner (reproduced from Jezzard, Mathews, & Smith)

14. 1 3T magnet is 50,000 times the earths magnetic field
1 3T magnet is 50,000 times the earths magnetic field. it is important that the region in the centre of the static magnetic field is homogenous.
homogenous field

15. When you put a material (like your subject) in an MRI scanner, some of the protons become oriented with the magnetic field.
Protons (hydrogen atoms) have “spins” (like tops). They have an orientation and a frequency.
Hydrogen is an atom with a single proton. It’s the hydrogen molecules in the water (H20) inside your body that are crucial in MRI. Protons have two properties defining their spin – orientation and temporal frequency. Putting the person in the static magnet field influences the orientation of spins but not their frequency. The protons in the magnet become oriented with the z direction (the long dimension of the bore) The RF coil and gradients are doing nothing at this point.
Slide borrowed from Culham fMRI for newbies

16. The radio frequency coil is usually both a transmitter (for excitation) and a receiver of energy (for reception). It’s called a radio frequency coil because it transmits energy in the radiofrequency portion of the electromagnetic spectrum, which is where most atomic nuclei (including hydrogen) have their RESONANT frequency. By transmitting energy at the correct frequency using the RF coil, you can get the protons in the hydrogen molecules to “resonate”, which means they absorb some of the energy that has been transmitted. The energy is only transmitted for brief periods called “pulses”
One familiar example of resonance is a playground swing, which acts as a pendulum. Pushing a person in a swing in time with the natural interval of the swing (its resonance frequency) will make the swing go higher and higher (maximum amplitude), while attempts to push the swing at a faster or slower tempo will result in smaller arcs. This is because the energy the swing absorbs is maximized when the pushes are 'in phase' with the swing's oscillations, while some of the swing's energy is actually extracted by the opposing force of the pushes when they are not.
It is said that an army has to break step when it marches over a bridge as a precaution to prevent the possibility of marching at the resonant frequency of the bridge, causing it to oscillate and then crack.
Generate electromagnetic field at the resonant frequency of hydrogen nuclei -RF pulse. Receive energy back from participant.

17. Resonance
Resonance is a fundamental physical phenomena that is exploited to allow the RF coil to selectively influence the orientation of the protons in the hydrogen atoms
It is easier to see a demonstration than listen to an explanation….
The key point is that in order for energy to be transmitted from object A to object B, A has to vibrate at the correct frequency, i.e. the fundamental frequency of B
The vibration frequency of the pulses produced by the RF coil can be varied, so that the pulse energy is absorbed by different atoms
In MRI, RF pulses are set to the fundamental frequency of hydrogen atoms, but importantly the fundamental frequency of the hydrogen atoms can vary (see later)
Every pendulum has a natural interval (resonant frequency) which is why you can use them to measure time in a clock. If you get a pendulum with the right length and weight its fundamental frequency will be one cycle per second. Bingo – the pacemaker for a clock!
Utube:
Science Geek
bridge:

18. After you turn off the RF pulse, as the protons return to their original orientations, they emit energy in the form of radio waves.
The emitted energy is received by the RF coil, becoming measurable current in the coil
Relaxation time is how long it takes the protons to return to their original alignment with the static magnetic field
When you apply radio waves (RF pulse) at the resonant frequency of hydrogen nuclei, you can change the orientation of the spins as the protons absorb energy.
The current measured in the receiver coil is the fundamental unit of measurement in MRI. The trick is to work out where in the subject the current is coming from, and then to understand how variation in the amount and duration of received electrical current is related to tissue type.
Electrical current is measured in amperes (amps)
The relaxation time is the time for the protons to become realigned with the static magnetic field Bo, and therefore stop emitting energy that the receiver coil can pick up.

19. Gradient coils allow spatial encoding of the MRI signal
If an MRI scanner had only a static magnetic field and an RF coil, you’d be able to measure the relaxation time of the nucleii you had excited, and you’d be able to use differerent pulse frequencies to excite nuclei other than hydrogen. However, you’d have no spatial information.
The gradient coils are used to introduce a spatial modulation into the static magnetic field, so that briefly it is no longer homogenous. Recall that the radio frequency pulse has to be at the correct frequency to pass its energy into the protons in the water molecules and make them “resonate”. What is the correct frequency to cause the resonance to occur is dependent upon magnetic field strength. By creating a spatial gradient across the magnetic field, different RF pulse frequencies will cause protons in different parts of space to resonate. This is a crude explanation of how the current received in the RF coil is related to spatial position.
Gradients can be created in the X, Y, and Z directions. Note that these are not independent magnetic fields, they are modulations of the strength of the static magnetic field (which is in the Z direction)
Gradient coils allow spatial encoding of the MRI signal

20. The effect of introducing a gradient
A lower frequency RF pulse will cause hydrogen nuclei here to resonate
A higher frequency RF pulse will cause hydrogen nuclei here to resonate
The grey wedge indicates the variation in the magnetic field introduced by the gradient coil. The fundamental resonant frequency of hydrogen varies depending upon the strength of the magnetic field it is in. So, by imposing a spatial gradient on the field strength, you are able to tie received signal from the RF pulse to a particular location along the gradient.
Gradients can be created in the X, Y, and Z directions. Note that these are not independent magnetic fields, they are modulations of the strength of the static magnetic field (which is in the Z direction)

21. Recap
The magnet causes protons in hydrogen atoms to become aligned with the magnetic field
The RF coil transmits energy into the subject and receives energy that is returned as the raw MR signal
variation in electrical current
The gradient coils allow the signal to be assigned to spatial locations
a voxel
So, the image intensity value at each voxel is derived from the amount of electrical current received by the RF coil for that voxel location
The amount of current received by the RF coil at each voxel varies with tissue type
Why?

22. Relaxation time of hydrogen nuclei varies depending on the surroundings
The time for relaxation to occur is governed by a rate constant, called T1
T1 is longer for H20 in CSF than it is for H20 in tissue
remember that CSF appears black in the structural image (the examples I showed were “T1 structural”)
remember that tissue was brighter (higher voxel intensity value), reflecting it’s shorter T1 relaxation time
Applying a single RF pulse does not generate tissue contrast. Why?
T1 tissue contrast is realized by applying multiple excitation pulses in quick succession using the RF coil
tissues with short T1 rate constant have time for substantial relaxation to occur between pulses and generate high signal in the receiver (because they emit a lot of energy)
tissues with long T1 rate constant give lower signal, because most of their relaxation process does not have time to occur before the next RF pulse is applied (so they emit little energy)
Relaxation is an exponential process governed by the T1 rate constant
A single RF pulse does not generate T1 tissue contrast because all the tissue types have time for full relaxation and therefore full energy emission. It is only by using multiple RF pulses to prevent the tissues with long relaxation times having time to relax and emit its energy that T1 tissue contrast is generated. Time between RF pulses is called TR.

24. Definitions
TR = the time interval between successive RF pulses (time to repetition)
TE = how much time elapses after the RF pulse before the receiver coil is switched on to measure the returned energy (time to echo)
short TE will only allow tissue with short t1 constant to generate strong signal
long TE will allow all tissue types to release the energy fully, resulting in the same t1 signal value for all tissue types

25. T1 Weighted image T2 Weighted image TR = 14ms TE = 5ms
On the left is an axial slice from a t1 weighted anatomical image. Notice that the TR value is very low, so as only to allow tissue with a short t1 constant to generate signal. Signal is low in the ventricles (containing CSF). On the right is a T2 image. This has a considerably longer TR. The TR is long enough for full relaxation of hydrogen nuclei in all tissue types. Therefore all tissue types return their full energy to the receiver, and there is no difference in t1 signal between the different tissue types.
But in the T2 image you lengthen the TE in order to give time for T2 signal to build up. What is the TE? It stands for time to echo.
In the T2 image the grey matter is whitish and the CSF is very bright.
TR = 14ms TE = 5ms
TR = 4000ms TE = 100ms

26. T2 signal and the TE
If you wait longer before turning on the receiver part of the RF coil then….
just after the RF pulse the billions of hydrogen nuclei begin emitting energy in temporal phase
but they gradually drift out of phase with each other
this arises because of very small local variations in the magnetic field
going out of phase causes an exponential loss of the summed signal intensity as a function of time
the time constant (T2) of the exponential decay of signal is different for the water in different tissue types so as for T1, the current measured varies
T2 images have opposite contrast from T1 images
The longer the TE the less the signal will be from structures that have a short T2 constant.
CSF has a long T2 constant and is white in the T2 image
As the T2 constant of a tissue tpe gets shorter, its signal gets less and it is darker in the image. The initial dip in brain activation makes the image darker, but the late positive response makes it brighter.

27. T1 Weighted image T2 Weighted image TR = 14ms TE = 5ms
T2 images are good for showing tissue abnormalities
On the left is an axial slice from a t1 weighted anatomical image. Notice that the TR value is very low, so as only to allow tissue with a short t1 constant to generate signal. Signal is low in the ventricles (containing CSF). On the right is a T2 image. This has a considerably longer TR. The TR is long enough for full relaxation of hydrogen nuclei in all tissue types. Therefore all tissue types return their full energy to the receiver, and there is no difference in t1 signal between the different tissue types.
But in the T2 image you lengthen the TE in order to give time for T2 signal to build up. What is the TE? It stands for time to echo.
In the T2 image the grey matter is whitish and the CSF is very bright.
TR = 14ms TE = 5ms
TR = 4000ms TE = 100ms

28. What is imaged in fMRI – BOLD signal
Red blood cells containing oxygenated hemoglobin are diamagnetic
diamagnetic materials are attracted to the magnetic field and don’t distort it or induce gradients
Red blood cells containing deoxygenated hemoglobin are paramagnetic
paramagnetic materials repel and distort the applied magnetic field
this in turn causes the spins to go out phase with each other faster and so the local T2 signal strength falls faster
Therefore, when the local proportion of deoxygenated blood increases the recorded image intensity falls
Hence BOLD imaging (Blood Oxygen Level Dependent)
This effect is described by the T2* relaxation time, which is less than the T2 relaxation time
In a nutshell, variation in the BOLD signal is determined by the ratio of oxygenated to deoxygenated blood
The simple story is to say that the BOLD signal is related to the ratio of oxygenated blood to deoxygenated blood in each voxel. This would imply that the BOLD signal is not influenced by such factors as blood volume, blood pressure, blood flow, etc. Of course, the story is actually more complex, but lets stick with the simple story for now.

29. Magnetic susceptibility artifact
This scan (for "TIA-like symptoms, stroke?") was stopped and the patient removed from the scanner after the localizer and T1 sequences showed this strong magnetic susceptibility artefact, apparently centred over the left orbit. He was feeling some heat and pain in and around his left eye, which was tearing and showed conjunctival injection (redness) on examination. There was no visual acuity or field impairment. On direct questioning, the patient finally recalled an incident at age 5, when "something got into my eye while playing outside". His mother had taken him to the Eye and Ear Hospital, but they had left without being seen after waiting for 2 hours in the ED ("I didn't look that bad and my mother got sick of waiting"). He had never had any symptoms since and had completely forgotten about it until this episode.
The patient had been screened as per normal safety protocol and had reported no previous eye injury, surgery or exposure to metallic foreign bodies. We then checked a recent CT which had been done for his "TIA".
The report of the brain CT did not mention any findings in the orbits. The images were then retrieved from archive and reviewed. On routine brain window images (top row) it is indeed not easy to spot the abnormality in the left orbit, especially if they are reviewed in thick (5mm) reconstructed slices on film alone. On bone window, however, a sizable metallic foreign body was clearly identifiable in the left orbit, lodged in the recess between the eye and the lateral wall. It was obviously asymptomatic and was therefore presumably subconjunctival.
The signal dropout observable here occurred because a ferrous object distorted the magnetic field producing a large gradient – the same effect as that caused by increased deoxyhemoglobin on a larger scale

30. The BOLD signal and the physiology of the hemodynamic response
The initial effect of an increase in neural activity within a voxel is an increase in the proportion of deoxygenated hemoglobin in the blood
reduced image intensity due to the shortening of the T2* relaxation time produced by paramagnetism (“initial dip”)
But the brain responds to the fall in the oxygenation level of the blood by flooding the tissue in the voxel with fresh oxygenated blood
the proportion of deoxygenated hemoglobin in the voxel now falls below the “baseline” (baseline = resting state neural activity?)
therefore, image intensity begins to increase
the “late positive” response is much larger than the “initial dip”
The flooding with oxygenated blood (late positive response) is usually interpreted as an “overkill” response, but there are alternative theories in which it is not seen as overkill. These theories focus on resupply of glucose to the tissue rather than resupply of oxygen.
The late positive response may spread beyond the area of activated neurons (BOLD spread). It is thought that the initial signal dip caused by the increase of deoxyhaemoglobin is more spatially restricted to the site of active neurons. So, in principle we should be imaging the initial dip rather than the late positive response to achieve better spatial resolution in functional scans, but signal to noise ratio is just too low for this to be practical in most cases.

31. The initial dip and the late positive response
BOLD response % is basically image intensity. The stimulus duration shown here is about 16 seconds of a stimulus that does not produce adaptation (boredom, habituation) in the neurons that process it. For example, a flickering checkerboard in V1.
The initial dip occurs about a second after stimulus presentation. The peak of the late positive response occurs 5-6 seconds after stimulus presentation.
Undershoot usually observed, initial dip more rarely. in FMRI you are mostly modeling the overshoot (late positive response). Note that the haemodynamic response (HRF) is assumed to be linear and additive, and this more or less holds up empirically. So, two events that stimulate the same voxel, are both brief, and are separated in time by about 4 seconds will produce two separate HRF’s that summate with each other.
Note that the time to peak of the late positive, and its dispersion, DOES vary between brain areas and between individuals. The canonical model is based on the response in primary visual cortex – so beware!.

32. BOLD imaging caveats
The simple story is that the BOLD signal is determined by the ratio of oxygenated blood to deoxygenated blood in each voxel
And therefore provides a good index of the brains response to the metabolic needs of neurons
There are some important caveats…
but let’s save them for another time
Also caveats on the physiology of the hemodynamic response

33. Relative spatial resolutions of T1 structural and single shot functional EPI images
This slide illustrates why you need to acquire a structural image and register the EPI to it. EPI voxel size here is 3*3*3 mm, versus 1*1*1 for the structural.
The TR for the single shot EPI is about 2 seconds. you have to acquire it that quickly or you fail to capture the temporal dynamics of the HRF. But that then places a constraint on voxel size.

34. Temporal sampling of the hemodynamic response
TR of a single shot whole brain functional EPI is typically about 2.5 sec
fMRi data is 4D or “time series”, whereas MRI data is 3D
We can only measure the BOLD response once every few seconds

35. Steps in the analysis of fMRI data
fMRI data consists of a series of consecutively acquired volumes (3D plus time = 4D)
Each volume is made up of X*Y*Z voxels
Each voxel position is sampled once per unit time equal to the TR (approx 1-4 sec)
What happens if the participant
moves during the experiment?

36. Head motion is always a problem
It can produce false task related activations if head motion is temporally correlated with an experimental condition
The top middle voxel in the grid changes its value from 89 to 507 due to a small head motion. This is a massive change in intensity compared to the BOLD signal. To understand this remember that not all variation in the t2* epi images is BOLD signal, most of it is still anatomical or just noise or other factors
Head motion violates the basic assumption of time series analysis that what you have sampled at each time point is the same thing in terms of space. Voxels remain in fixed locations in the scanner, not relative to the brains position.
The worst case scenario is when head motion patters become temporally correlated with experimental tasks, e.g. a participant who tends to nod their head slightly each time they press a button…
Obviously it is impossible for the subject to lie completely still. Small high frequency movements are caused by heart beat and respiration, though generally at a frequency you cannot resolve with fmri (unless you have a TR of around 1 sec). Type “pulsatility fmri” into google….

37. Motion correction Select one volume as a reference
first or middle volume of series
Realign all other volumes in the series to the reference volume
rigid body registration with 6 DOF
(more on registration in a minute)
This reduces the problems caused by head motion, but does not remove them
because moving produces moving gradients in the magnetic field / interacts with existing inhomogeneities in the static magnetic field, and this changes the measured image intensities as well as their voxel locations
More problematic if head motion is correlated with the experimental time course
If a person moves a lot, you can’t use the data
The potential artefacts introduced by head motion are quite complex, and I might cover them in more detail in a later session. There are also new and more sophisticated solutions to the problem of head motion correction, which may be a slight improvement.
Motion correction does not fix changes caused by head motion * bias field interaction, or interpolation effects.
One technological solution is to reposition the slices between each volume acquisition (which our scanner can do, PACE), but then you loose your record of how much the subject moved, so you end up just having to assume that they kept still….not good. We may be able to get Siemens to provide a record of the adjustments made during scanning (watch this space)

38. Spatial transformations for registration
These are used in
motion correction
registration of structural to template images
registration of T1 structural to T2 and/or low resolution functional images of the same participant
registration of a PET (or other modality) scan of a participant to an MRI scan of the same participant
They can be characterised by the number of degrees of freedom of the transformation
rigid body (assumes both brains same size and shape)
affine (can change size and shape of brain)
The transform parameters are found by iteratively minimizing a cost function
The cost function can be least squared difference between the two images to be registered, or maximum magnitude of correlation (better as invariant to t1 /t2 differences in polarity of image contrast). Probably the best const funciton to use now that computers are getting faster is mutual information.

39. Rigid body (6 DOF)
Used for intra subject registration, including motion correction
3 rotations (pitch, roll, yaw)
3 translations (X,Y,Z)
translation Y
translation X
rotation (yaw)
If you are registering a low res functional EPI volume to a high res t1 structural of the same person you might think you need a 7th degree of freedom to take into account the different sizes of the images in terms of voxel dimensions. No, because the registration is done in mm this is unnecessary

40. Affine linear (12 DOF)
Allows registration of two brains differing in size and shape
registration of participant to template brain
registration of participant 1 to participant 2
As rigid body plus
3 scalings (stretch or compress X, Y, or Z)
3 skews / shears
scale Y
shear
Note that nonlinear registration (FNIRT) will do a better job of inter subject and subject to template registrations, but is more complex to set up and takes longer.

41. Registration in FSL (versus SPM)
There are 2 steps in registration
estimating the transformation
resampling
resampling is applying the transform to produce a third image that you write to the hard disk
The third image is “in the space of” the target image
SPM usually performs resampling immediately after estimation
once for motion correction, once for registration to template image, once for slice timing correction etc
produces lots of intermediate images on hard disk
some loss of image quality inherent in resampling is transmitted between processing stages
FSL delays resampling until after modelling stage
all the individual linear transforms can be added together
By delaying resampling until after the GLM has been estimated and contrasts performed FSL avoids loss of image quality and minimizes disk use. You only ever apply the transformation matrix to resample the necessary images for viewing your results (e.g. contrast images of t statistics overlaid on anatomy). But if you are used to SPM this can be a bit confusing at first.

42. Brain Extraction (BET)
Skull can have similar images intensity values to white matter in some cases
could confuse some registration processes
Templates brains are skull free
Skull and CSF is source of individual variation it is best to get rid of before you do anything else
Also reduces number of voxels that have to be processed in later steps
BET makes use of the large high contrast boundary produced by the CSF / Brain interface to remove both CSF & skull
By default FSL assumes that the structural image you use for registration to the standard space template has had the skull removed, so you need to do this as a first step.

43. fMRI: modelling the voxel time course
Modelling the response
modelling the changes in voxel image intensity over time as measured in the 4D functional data
Often, the model is just the time course of the experimental stimuli
After fitting the model
search for individual voxels where a statistically significant proportion of the variance over time is explained by the model
there are a very large number of voxels, which results in a serious multiple comparisons problem
This topic will be covered in more detail in later weeks

50. Output of the modelling stage
Begin with a 4D image series
Model change over time
The output of the model can be thought of as a single 3D volume
Each voxel in the 3D volume has a value representing how well the temporal model fits at that X,Y,Z spatial location
Modelling is actually a way of compressing the data by “removing” the time dimension
If the model is good then you can send an interested party the model instead of the data
the model can be used to recreate the data
If you were collaborating with a scientist in the USA, and your internet access was limited to a 56K modem, then there would be genuine value in sending the model instead of the data, because it would take a very long time to send fMRI data over a 56k modem!

51. Thresholding
At each voxel you have to decide whether the fit between the time course and the model time course is statistically significant
probability of a fit that good occurring through random sampling from a null distribution where the true fit is zero, and all variation across the time course is random
“false positive”
Convention suggests that a 5% risk of a false positive is acceptable
But using 5% with approximately voxels there will be 5000 false positive results…
Bonferoni correction is too conservative
more on this next week
This topic will be covered in more detail in later weeks

52. The Problem of Multiple Comparisons (64000 voxels)
P < (1682 voxels)
P < (364 voxels)
P < (32 voxels)
What’s the spatial structure of the false positives?
This image shows a single slice of the brain with 64,000 functional voxels projected onto it. The data are just two sets of random noise, and then one set of noise was subtracted from the other on a voxel by voxel basis. The numbers show the Type 1 error rates for this t test.
At p < 0.05 you expect one in 20 voxels to be classified as active by chance.
Observation: these Type I error voxels have a random spatial distribution across the slice. It is rare to see two adjacent ones. Real activated voxels tend to cluster spatially, which is helpful in finding a solution to the multiple comparison problem

53. Some advantages of FSL over SPM
Origin automatically set to anterior commisure by registration
SPM requires manual setting
Easier to obtain atlas information for activations
Can view single voxel time course and fitted model in FSLVIEW
not possible to view time course in SPM unless you are a MATLAB programmer!
View 4D as movie
FSL is very well documented….

54. Old material beyond this point

55. Resonance – the R in MRI Resonance is a fundamental concept in physics
a playground swing is a pendulum with a resonant frequency
if you push the swing in time with its resonant frequency the energy from the push is transferred to the swing and it gains height
pushing at other frequencies is unsuccessful….
All atomic nuclei have a resonant frequency
hydrogen will absorb energy from electromagnetic waves that match its resonant frequency (just like the swing)
The resonant frequency of hydrogen is in the radio frequency range, hence the name RF coil
Every pendulum has a natural interval (resonant frequency) which is why you can use them to measure time in a clock.
Utube:
Science Geek
bridge:

56. TE and T2 weighted images
With a long TR (e.g. 2 seconds) all tissue types have time for full T1 relaxation between RF pulses, so no T1 signal is generated
T2 signal builds up when the TE is longer
TE is the amount of time you wait after transmitting the RF pulse before making the measurement of received energy from relaxation of hydrogen nuclei
A short TE is necessary for good T1 contrast

57. What is imaged in fMRI?
fMRI functional images (usually EPI – echo planar imaging) are T2 weighted images
The T2 time constant describing the rate of signal loss becomes much shorter near local gradients in the magnetic field
Water molecules diffuse through the gradients, causing their resonance frequencies to alter, reducing the coherance of the spins
sending them out of phase with each other, thereby speeding up T2 decay
An extreme case of this is caused by the gradient in the magnetic field caused by the presence of a ferromagnetic object in or near the imaged volume
Note that the alteration of the frequency of oscillation of the hydrogen nuclei by the local gradient is basically the same mechanism as that the experimenter induces when using the gradient coils to influence the frequency of spins by making it dependent on spatial location.