final exam Flashcards
(124 cards)
Nuclear Magnetic Resonance
MRI
energy at specific radio frequencies is absorbed and reemitted by nuclei with non-zero spin
(electromagnetic energy but same concept as resonance)
w=yB
w=resonant (Larmor) frequency -> 85mHz typ.
y=gyromagnetic ratio
B=magnetic field strength
MRI signal
radio frequency pulse (at Larmor frequency to be absorbed by hydrogen in brain) goes into brain
excites atoms in brain and then energy is released
measures density of hydrogen atoms but pulse of certain freq. only absorbed by H atoms, measure how much is absorbed then released again
Gradients to encode space
- parts of brain have slightly different magnetic field strengths s
- resonance frequency for hydrogen will be different due to magnetic field
Encodes energy using frequency and phase (3 gradients to get separate measurements to create 3D image)
K space
represents density of hydrogen but changes timing to measure other things
records how much energy comes out of brain in each point in time during recording
convert k-space to real space
2D Fourier transformation (mathematical)
information needs to be decoded
different parts of brain represented by different frequencies and phases
Blood oxygenation level dependent (BOLD) contrast
key to fMRI
oxygenation of blood gives hint of how active areas of brain are
- cells require energy source when metabolically active so they use up oxygen and other resources provided by blood
Oxygenated hemoglobin
diamagnetic
zero magnetic moment
–> has iron in it, one of the more magnetically active elements
Deoxygenated hemoglobin
paramagnetic
significant magnetic moment
greater magnetic susceptibility
- causes faster T2 decay
timing from hemoglobin
spike of oxygenated blood rushes in to sweep away deoxygenated hemoglobin
(deoxygenated from active use)
this dip in deoxygenated blood (from replacement) takes about 6s and the response is much more spread out in time
*fMRI does not have very good temporal resolution (cannot see direct neural activity but slower smeared out response for circulatory system)
fMRI data
subjects –> run (sequence on trials and tasks) –> volume (allows to measure where changes happen) –> slices –> voxel
fMRI signal/data
in order to collect frames every 2 seconds, resolution is sacrificed
images are a lot blurrier for fMRI for quick image collection
doesn’t affect too much cause changes are what are being recorded (local areas serviced by certain blood vessels)
BOLD signal is really weak, most of signal is structure but BOLD changes sit on top of main structural signal
fMRI data preprocessing
1) slice timing correction
2) realignment
3) coregistration
4) normalization
5) smoothing
Slice timing correction
data collected slice by slice and it takes time to collect each slice so the time you collected data at the bottom is different from the top
takes 2 seconds to do whole brain
**data has to be interpolated
Realignment
cant assume a voxel equals same bit of brain across time (think small movements, few mms make a difference)
typical
- tries to correct for translations and rotations of the brain
uses 6 parameter rigid body transformation (3 translations 3 rotations)
Reslicing
after realignment, need to replace to match up voxels
Coregistration
align function images to structural image
- uses an affine transformation (translation, rotation, scale, shear)
collects functional and also clear structural images but using different techniques to make them comparable
Normalisation
warp each subject’s brain into a standard space using deformations by linear combination of smooth basis functions
- every brain is a snowflake but there is a standard brain other images are matched up with
segments into grey matter, white matter and CSF using tissue probability maps and Bayesian estimation
Smoothing
deals with remaining inconsistencies by smoothing/blurring images
Blocked design
trials are grouped into blocks
hemodynamic response is estimated per block
e.g. testing reactions to faces, first test response to fearful faces, then happy, then neutral
Event-related design
mix different trial types together and show them in random sequence
uses jittering between conditions
neural activity to each one of the trials will highly overlap with response to other trials (try to control for with jittering)
Mixed design
blocks of separated conditions within which there are separated temporally jittered events of interest
Mass-univariate general linear model (GLM) analysis
univariate analysis of each voxel but for every voxel
- data collected from each voxel in brain over time
- predicted time course for activity related to each event/condition of interest (try to explain data in terms of what’s going on in the experiment)
- look at what was recorded in brain for each voxel and look at pattern of activity for what voxels respond to
- estimates are made for how much each event contributes to signal in each voxel
Familywise error rate (FWE)
correction for multiple comparisons
family-wise error correction gets rid of false positives but also gets rid of most of the signal
False discovery rate (FDR)
correction for multiple comparisons
still has a pretty good job of detecting signal and getting rid of false positives (compromise)
e.g. 5% of active voxels are false positives instead of each test being 5% wrong