MRI Image Formation
DARSHAN. B.S.
MEDICAL IMAGING TECHNOLOGIST.
DEPARTMENT OF RADIODIAGNOSIS.KMIO.BENGALORE
K-SPACE
PIXEL
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Spins precess at the Larmor rate:
 =  (B0 + B)
MR imaging is based on precession
]
field strength field offset
x
y
z
Magnetic Gradients
Gradient: Additional magnetic field which varies
over space
– Gradient adds to B0, so field depends on position
– Precessional frequency varies with position!
– “Pulse sequence” modulates size of gradient
High field
Low field
B0
• Spins at each position sing at different frequency
• RF coil hears all of the spins at once
• Differentiate material at a given position by selectively
listening to that frequency
Magnetic Gradients
Fast
precession
Slow
precession
B0
High field
Low field
Simple “imaging” experiment (1D)
increasing
field
Simple “imaging” experiment (1D)
Fourier transform
Signal
“Image”
Fourier Transform: determines amount of material at a
given location by selectively “listening” to the
corresponding frequency
position
2D Imaging via 2D Fourier Transform
2DFT
2D Image
x
y
2D Signal
kx
ky
1D Signal 1D “Image”
1DFT
Analogy: Weather Mapping
2D Fourier Transform
2DFT
Measured signal
(frequency-, or k-space)
Reconstructed
image
Fourier Transform can be applied in any number of
dimensions
MRI: signal acquired in 2D frequency space (k-space)
(Usually) reconstruct image with 2DFT
x
y
kx
ky
Gradients and image acquisition
• Magnetic field gradients encode spatial position in
precession frequency
• Signal is acquired in the frequency domain (k-space)
• To get an image, acquire spatial frequencies along
both x and y
• Image is recovered from k-space data using a Fourier
transform
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Sampling k-space
Perfect reconstruction of an object would require
measurement of all locations in k-space
(infinite!)
Data is acquired point-by-point in k-space
(sampling)
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
x x x x x x x x x x x x x
FT
Sampling k-space
1. What is the highest frequency we need to sample
in k-space (kmax
)?
2. How close should the samples be in k-space (k)?
kx
ky
kx
2 kx
max
Frequency spectrum
What is the
maximum
frequency we
need to measure?
Or, what is the
maximum k-
space value we
must sample
(kmax
)?
FT
kmax
-kmax
Frequency spectrum
Frequency spectrum
Frequency spectrum
Frequency spectrum
Frequency spectrum
Frequency spectrum
Higher frequencies
make the
reconstruction look
more like the original
object!
Large kmax
increases
resolution (allows us
to distinguish smaller
features)
2D Extension increasing
kmax
2 kx
max
kx
max
kx
max
ky
max
ky
max
kmax
determines image resolution
Large kmax
means high resolution !
Sampling k-space
1. What is the highest frequency we need to sample
in k-space (kmax
)?
2. How close should the samples be in k-space (k)?
kx
ky
kx
2 kx
max
Nyquist Sampling Theorem
A given frequency must be sampled at least twice per
cycle in order to reproduce it accurately
1 samp/cyc 2 samp/cyc
Cannot distinguish
between waveforms
Upper waveform is
resolved!
Insufficient sampling
forces us to interpret
that both samples are
at the same location:
aliasing
Nyquist Sampling Theorem
increasing field
Aliasing (ghosting): inability to differentiate between 2 frequencies
makes them appear to be at same location
Applied FOV Aliased image
max ive
frequency
max ive
frequency
x x
k-space relations:
FOV and Resolution
x = 1/(2*kx
max
)
FOV = 1/kx
kx
ky
kx
2 kx
max
k-space relations:
FOV and Resolution
2 kx
max
= 1/x
xmax
= 1/kx
kx
ky
kx
2 kx
max
k-space and image-space are inversely related:
resolution in one domain determines extent in other
k-space Image
Full sampling Full-FOV,
high-res
Full-FOV,
low-res:
blurred
Low-FOV,
high-res:
may be
aliased
Reduce kmax
Increase k
2DFT
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Visualizing k-space trajectories
k-space location is proportional to accumulated
area under gradient waveforms
Gradients move us along a trajectory through k-
space !
kx(t) =  Gx()
d
ky(t) =  Gy()
d
Raster-scan (2DFT) Acquisition
Acquire k-space line-by-line (usually called “2DFT”)
Gx causes frequency shift along x: “frequency encode” axis
G causes phase shift along y: “phase ecode” axis
Echo-planar Imaging (EPI) Acquisition
Single-shot (snap-shot): acquire all data at once
Many possible trajectories through k-space…
Trajectory considerations
• Longer readout = more image artifacts
– Single-shot (EPI & spiral) warping or blurring
– PR & 2DFT have very short readouts and few artifacts
• Cartesian (2DFT, EPI) vs radial (PR, spiral)
– 2DFT & EPI = ghosting & warping artifacts
– PR & spiral = blurring artifacts
• SNR for N shots with time per shot Tread :
SNR   Ttotal =  N  Tread
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Partial k-space
If object is entirely real, quadrants of k-space
contain redundant information
a+ib
aib
c+id
cid
ky
kx
1
2
4
3
Partial k-space
Idea: just acquire half of k-space and “fill in” missing data
Symmetry isn’t perfect, so must get slightly more than half
measured data
missing data
a+ib
aib
c+id
cid
ky
kx
1
Multiple approaches
kx
ky
Reduced phase
encode steps
Acquire half of each
frequency encode
kx
ky
Parallel imaging
(SENSE, SMASH, GRAPPA, iPAT, etc)
Object in
8-channel array
Single coil
sensitivity
Surface
coils
Multi-channel coils: Array of RF receive coils
Each coil is sensitive to a subset of the object
Object in
8-channel array
Single coil
sensitivity
Surface
coils
Coil sensitivity to encode additional information
Can “leave out” large parts of k-space (more than 1/2!)
Similar uses to partial k-space (faster imaging,
reduced distortion, etc), but can go farther
Parallel imaging
(SENSE, SMASH, GRAPPA, iPAT, etc)
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Slice Selection
0
frequency
Gz
gradient
RF
excited slice
2D Multi-slice Imaging
excited slice
All slices excited and acquired sequentially (separately)
Most scans acquired this way (including FMRI, DTI)
t1
t2
t3
t4
t5
t6
“True” 3D imaging
Repeatedly excite all slices simultaneously, k-space
acquisition extended from 2D to 3D
Higher SNR than multi-slice, but may take longer
Typically used in structural scans
excited volume
excited volume
Image Formation
• Gradients and spatial encoding
• Sampling k-space
• Trajectories and acquisition strategies
• Fast imaging
• Acquiring multiple slices
• Image reconstruction and artifacts
Motion Artifacts
Motion causes inconsistencies between readouts in
multi-shot data (structurals)
Usually looks like replication of object edges along
phase encode direction
PE
Gibbs Ringing (Truncation)
Abruptly truncating signal in k-space introduces “ringing”
to the image
EPI distortion (warping)
Field map EPI image
(uncorrected)
field offset
Magnetization precesses at a different rate than expected
Reconstruction places the signal at the wrong location
image distortion
EPI unwarping (FUGUE)
field map uncorrected
Field map tells us where there are problems
Estimate distortion from field map and remove it
corrected
EPI Trajectory Errors
Left-to-right lines offset from right-to-left lines
Many causes: timing errors, eddy currents…
EPI Ghosting
= +
undersampled
Shifted trajectory is sum of 2 shifted
undersampled trajectories
Causes aliasing (“ghosting”)
To fix: measure shifts with reference
scan, shift back in reconstruction
Image Formation
Matlab exercises (self-contained, simple!)
k-space sampling (FOV, resolution)
k-space trajectories
CENTRIC SAMPLING
OUTWARD IMAGE SAMPLING
SEQUENTIAL SAMPLING
K-SPACE FILLING TYPES
MRI IMAGE FORMATION -K-SPACE            .

MRI IMAGE FORMATION -K-SPACE .

  • 1.
    MRI Image Formation DARSHAN.B.S. MEDICAL IMAGING TECHNOLOGIST. DEPARTMENT OF RADIODIAGNOSIS.KMIO.BENGALORE
  • 2.
  • 3.
  • 4.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 5.
    Spins precess atthe Larmor rate:  =  (B0 + B) MR imaging is based on precession ] field strength field offset x y z
  • 6.
    Magnetic Gradients Gradient: Additionalmagnetic field which varies over space – Gradient adds to B0, so field depends on position – Precessional frequency varies with position! – “Pulse sequence” modulates size of gradient High field Low field B0
  • 7.
    • Spins ateach position sing at different frequency • RF coil hears all of the spins at once • Differentiate material at a given position by selectively listening to that frequency Magnetic Gradients Fast precession Slow precession B0 High field Low field
  • 8.
    Simple “imaging” experiment(1D) increasing field
  • 9.
    Simple “imaging” experiment(1D) Fourier transform Signal “Image” Fourier Transform: determines amount of material at a given location by selectively “listening” to the corresponding frequency position
  • 10.
    2D Imaging via2D Fourier Transform 2DFT 2D Image x y 2D Signal kx ky 1D Signal 1D “Image” 1DFT
  • 11.
  • 12.
    2D Fourier Transform 2DFT Measuredsignal (frequency-, or k-space) Reconstructed image Fourier Transform can be applied in any number of dimensions MRI: signal acquired in 2D frequency space (k-space) (Usually) reconstruct image with 2DFT x y kx ky
  • 13.
    Gradients and imageacquisition • Magnetic field gradients encode spatial position in precession frequency • Signal is acquired in the frequency domain (k-space) • To get an image, acquire spatial frequencies along both x and y • Image is recovered from k-space data using a Fourier transform
  • 14.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 15.
    Sampling k-space Perfect reconstructionof an object would require measurement of all locations in k-space (infinite!) Data is acquired point-by-point in k-space (sampling) x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x x FT
  • 16.
    Sampling k-space 1. Whatis the highest frequency we need to sample in k-space (kmax )? 2. How close should the samples be in k-space (k)? kx ky kx 2 kx max
  • 17.
    Frequency spectrum What isthe maximum frequency we need to measure? Or, what is the maximum k- space value we must sample (kmax )? FT kmax -kmax
  • 18.
  • 19.
  • 20.
  • 21.
  • 22.
  • 23.
    Frequency spectrum Higher frequencies makethe reconstruction look more like the original object! Large kmax increases resolution (allows us to distinguish smaller features)
  • 24.
    2D Extension increasing kmax 2kx max kx max kx max ky max ky max kmax determines image resolution Large kmax means high resolution !
  • 25.
    Sampling k-space 1. Whatis the highest frequency we need to sample in k-space (kmax )? 2. How close should the samples be in k-space (k)? kx ky kx 2 kx max
  • 26.
    Nyquist Sampling Theorem Agiven frequency must be sampled at least twice per cycle in order to reproduce it accurately 1 samp/cyc 2 samp/cyc Cannot distinguish between waveforms Upper waveform is resolved!
  • 27.
    Insufficient sampling forces usto interpret that both samples are at the same location: aliasing Nyquist Sampling Theorem increasing field
  • 28.
    Aliasing (ghosting): inabilityto differentiate between 2 frequencies makes them appear to be at same location Applied FOV Aliased image max ive frequency max ive frequency x x
  • 29.
    k-space relations: FOV andResolution x = 1/(2*kx max ) FOV = 1/kx kx ky kx 2 kx max
  • 30.
    k-space relations: FOV andResolution 2 kx max = 1/x xmax = 1/kx kx ky kx 2 kx max k-space and image-space are inversely related: resolution in one domain determines extent in other
  • 31.
    k-space Image Full samplingFull-FOV, high-res Full-FOV, low-res: blurred Low-FOV, high-res: may be aliased Reduce kmax Increase k 2DFT
  • 32.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 33.
    Visualizing k-space trajectories k-spacelocation is proportional to accumulated area under gradient waveforms Gradients move us along a trajectory through k- space ! kx(t) =  Gx() d ky(t) =  Gy() d
  • 34.
    Raster-scan (2DFT) Acquisition Acquirek-space line-by-line (usually called “2DFT”) Gx causes frequency shift along x: “frequency encode” axis G causes phase shift along y: “phase ecode” axis
  • 35.
    Echo-planar Imaging (EPI)Acquisition Single-shot (snap-shot): acquire all data at once
  • 36.
    Many possible trajectoriesthrough k-space…
  • 37.
    Trajectory considerations • Longerreadout = more image artifacts – Single-shot (EPI & spiral) warping or blurring – PR & 2DFT have very short readouts and few artifacts • Cartesian (2DFT, EPI) vs radial (PR, spiral) – 2DFT & EPI = ghosting & warping artifacts – PR & spiral = blurring artifacts • SNR for N shots with time per shot Tread : SNR   Ttotal =  N  Tread
  • 38.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 39.
    Partial k-space If objectis entirely real, quadrants of k-space contain redundant information a+ib aib c+id cid ky kx 1 2 4 3
  • 40.
    Partial k-space Idea: justacquire half of k-space and “fill in” missing data Symmetry isn’t perfect, so must get slightly more than half measured data missing data a+ib aib c+id cid ky kx 1
  • 41.
    Multiple approaches kx ky Reduced phase encodesteps Acquire half of each frequency encode kx ky
  • 42.
    Parallel imaging (SENSE, SMASH,GRAPPA, iPAT, etc) Object in 8-channel array Single coil sensitivity Surface coils Multi-channel coils: Array of RF receive coils Each coil is sensitive to a subset of the object
  • 43.
    Object in 8-channel array Singlecoil sensitivity Surface coils Coil sensitivity to encode additional information Can “leave out” large parts of k-space (more than 1/2!) Similar uses to partial k-space (faster imaging, reduced distortion, etc), but can go farther Parallel imaging (SENSE, SMASH, GRAPPA, iPAT, etc)
  • 44.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 45.
  • 46.
    2D Multi-slice Imaging excitedslice All slices excited and acquired sequentially (separately) Most scans acquired this way (including FMRI, DTI) t1 t2 t3 t4 t5 t6
  • 47.
    “True” 3D imaging Repeatedlyexcite all slices simultaneously, k-space acquisition extended from 2D to 3D Higher SNR than multi-slice, but may take longer Typically used in structural scans excited volume excited volume
  • 48.
    Image Formation • Gradientsand spatial encoding • Sampling k-space • Trajectories and acquisition strategies • Fast imaging • Acquiring multiple slices • Image reconstruction and artifacts
  • 49.
    Motion Artifacts Motion causesinconsistencies between readouts in multi-shot data (structurals) Usually looks like replication of object edges along phase encode direction PE
  • 50.
    Gibbs Ringing (Truncation) Abruptlytruncating signal in k-space introduces “ringing” to the image
  • 51.
    EPI distortion (warping) Fieldmap EPI image (uncorrected) field offset Magnetization precesses at a different rate than expected Reconstruction places the signal at the wrong location image distortion
  • 52.
    EPI unwarping (FUGUE) fieldmap uncorrected Field map tells us where there are problems Estimate distortion from field map and remove it corrected
  • 53.
    EPI Trajectory Errors Left-to-rightlines offset from right-to-left lines Many causes: timing errors, eddy currents…
  • 54.
    EPI Ghosting = + undersampled Shiftedtrajectory is sum of 2 shifted undersampled trajectories Causes aliasing (“ghosting”) To fix: measure shifts with reference scan, shift back in reconstruction
  • 55.
    Image Formation Matlab exercises(self-contained, simple!) k-space sampling (FOV, resolution) k-space trajectories
  • 56.
  • 57.
  • 58.
  • 59.

Editor's Notes

  • #6 Make color gradient on arrows & fast->slow precession?
  • #7 Make color gradient on arrows & fast->slow precession?