Basic Principles of (P)MRI
or:
How I learned to stop worrying and love k-space
What is not in this talk:
Physics of MRI
(spin/magnetic field interaction: just cartoons ...)
Mathematics of MRI
(sampling theory, reconstruction algorithms)
Pretty pictures
(stolen from all over the internet)
What is in this talk:
What is really measured in a MRI scan?
How is the image generated?
Where does parallel imaging come in?
Principles of MRI 1:
Signal generation
or: The machine that goes ping
MRI is based on nuclear spin of hydrogen
Effect of external magnetic field:
spin axis aligns with
field ...
... and precesses
EM pulse at resonance frequency
(Larmor frequency)
energy is absorbed,
spin axis ...
... flips down
Energy is radiated at Larmor frequency
while spin axis realigns with field
rotating magnetic moment
induces current ...
...this is the
measured data
Three fundamental ideas
Signal strength proportional to hydrogen
density
Larmor frequency proportional to magnetic
field strength
Spin needs time to realign with magnetic
field (relaxation time)
Principles of MRI 2:
Spatial Encoding
or: Where did that ping come from?
Localization of signals
Goal: spatial map of hydrogen density
But: Coils measure composite signal from all
protons
Idea: Signal frequency = Larmor frequency
depends on magnetic field strength
Use gradient fields to map signal frequency
to spatial location
(e.g., Gx = x G0, Gy = y G0, Gz = z G0)
Localization of signals
Problem: Superposition of gradients not
unique: Gx+Gy+Gz=G0 at many different points
One possibility: use many gradient
orientations, filtered backprojection (slow)
In practice: Use gradients sequentially
Slice selection (z)
Use gradient field Gz
during excitation
Only thin slice has
resonance at
frequency of EM pulse
Signal comes from
single slice only
Frequency encoding (x)
Use gradient field Gx
during measurement
Resonance frequency
differs at each x
Signal is superposition
of frequencies,
FT gives contribution
of each x position
Phase encoding (y)
Use gradient field Gy to change phase
Field on: spins precess
at different speed
Field off: same speed,
but different phase
Phase encoding (y)
Problem: phase cannot be measured directly
But phase difference can!
Linear gradients: phase difference constant
Repeat measurement with different gradients
Gy = n y G0, n = -N,...,0,...,N
Phase difference is y-specific:
y larger
gradient change, phase change larger
What is k-space?
Matrix of measurements
kx: time signal
ky: diff. phase gradients
FT in kx: frequency in
time ( x-position)
FT in ky: frequency in
phase ( y-position)
like Fourier domain,
FT gives image
Image reconstruction in MRI
Time Dependent Signals
Fourier Transform
Frequency and Phase of Signal
Spatial Encoding
Image
Measured k-Space Data ...
ky ()
kx (=t)
... Fourier transformed
Basic steps in MRI scan
1. External field B0
7. Frequency encoding: Gx
2. Slice selection: Gz
8. Measure signal
3. Apply RF pulse
9. turn off Gx
4. turn off Gz
[Link] relaxation time
repeat steps 5-9 with
different Gy
5. Phase encoding: Gy
6. turn off Gy
Principles of MRI 3:
Parallel Imaging
or: I cant hold still that long!
Speed limitations
Need to turn on and off strong magnetic
fields repeatedly
(fun fact: thats why an MR scan is so loud)
Need to wait for the spin to realign with
mag. field before each phase encoding step
do less phase encoding steps
(e.g., half the number with double step size)
Aliasing artefacts
But: equivalent to
subsampling in y
2 periodic image
now 2/2 periodic
Wrap around,
size halved in y
What to do about it
Information is missing
get more!
Additional measurements from different coils
Coils should be independent, e.g., have
different spatial sensitivity
Modeling of Coil Sensitivities
Sensitivity S(x,y):
Magnitude of current through coil induced by
magnetic moment of unit strength at (x,y)
Biot-Savart law: Simplification of Maxwell eqs
Biot-Savart describes magnetic field around
current carrying conductor
By reciprocity: B.-S. describes current induced
by magnetic moment (rotating spin axis)
Examples of sensitivities
Examples of sensitivities
Reconstruction: SENSE
Measured pixel I(x,y) is superposition of two
modulated pixels from true image u
Sensitivities S1, S2 (approximately) known: solve
1 (, y) = S1 (, y)(, y) + S1 (, y + y)(, y + y)
2 (, y) = S2 (, y)(, y) + S2 (, y + y)(, y + y)
Overdetermined, model errors: use pseudoinverse
Center Lines
k-space like Fourier domain:
central frequencies most
important
do more phase encoding
steps for small n
weighted wrap around
Reconstruction: GRAPPA
Interpolation in k-space using center lines
Not acquired coefficients are linear combination
of measured coefficients
Compute interpolation weights from center lines
Enough center lines fully describe sensitivities:
GRAPPA equivalent to SENSE
What I didnt tell you
Different imaging types
(based on other properties such as relaxation time ...)
Pulse sequence design
(because timing matters ...)
Different sampling patterns
(radial, spiral, random ...)
Time dependent measurements
(and other advanced uses such as functional MRI)
If you want to learn more
Z.-P. Liang, E.M. Haacke,
Magnetic Resonance Imaging, Encyclopedia
of Electrical and Electronics Engineering
D. Weishaupt, V.D. Koechli, B. Marincek,
How does MRI work? An Introduction to the
Physics and Function of Magnetic Resonance
Imaging, 2nd edition, Springer, 2006
Mathematicians Summary
MRI scanner measures Fourier coefficients of image
Which coeffs: determined by magnetic fields, not
receiver (same for all coils!)
PMRI faster by leaving out lines of coeffs
Coil sensitivities are smooth modulations, depend on
coil geometry and magnetic properties of sample