Program on Quasi-Monte Carlo and High-Dimensional Sampling Methods for Applie...
DTI lecture 100710
1. Introduction to Diffusion
Tensor Imaging
Why DTI?
Diffusion – what it is, how it affects MR signal
Tensor – how we represent diffusion
Imaging – how we measure it in MRI
3. Why diffusion?
Diffusion is EXTREMELY SENSITIVE to
differences and changes in tissue microstructure
Myelination/Demyelination
Axon damage/loss
Inflammation/Edema
Necrosis
It is NOT a biomarker of white matter integrity
It is NOT just about white matter
Gray matter
Cardiac tissue
4. Example DTI image
“Fractional Anisotropy”
map
“map” is a computed
parameter, unlike an
“image” which is acquired
signal
Also called a “tractogram”
since it clearly shows
major white matter fiber
tracts
5. What is Diffusion?
stochastic movement of particles in a solvent,
driven by the thermal molecular motion of the
solvent…
… and also applies to motion of the solvent
itself (Einstein, 1905)
time τ∆
NOTE: In the limit N→∞, use the
Central Limit Theorem to assume
“step size” ∆ is fixed and equal to
the average of individual
displacements ∆i.
6. 1D Fick’s Law - what the flux?
t = t0 + τ
x
t = t0
0 2x − ∆ 0x − ∆ 0x + ∆0x 0 2x + ∆
t = t0 + τ
x
t = t0
0 2x − ∆ 0x − ∆ 0x + ∆0x 0 2x + ∆
What is the flux (J) through x0 after one time interval τ ?
C1(x)C2(x)
dx
dC
J
τ
2
2
1 ∆
−=
Adolf Fick, 1855: Flux is proportional to
the particle concentration gradient
(conservation of mass)
7. The Diffusion Coefficient
3D Fick’s Law
Note the minus sign: flux goes
from high to low concentration
del operator replaces partial
derivative
factor of 6, not 2 (why?)
D is the diffusion coefficient
This is the expression for
isotropic diffusion
τ62
∆=D
CDJ ∇−=
CJ ∇
∆
−=
τ6
2
9. Diffusion in Tissue (Anisotropic)
t
ink
r2
r3
r1
diffusion
ellipsoid
tDr 11 2=
tDr 22 2=
tDr 33 2=
x
y
z
laboratory
frame
DON’T
try this at
lab!!!!!
12. Tensor Invariants
Shape invariants:
analytical calculation
directly from tensor
coeffs
xx xy xz
xy yy yz
xz yz zz
D D D
D D D
D D D
( )1
3
av xx yy zzD D D D= + +
1
2
2 2 2
1
3
xx yy xx zz yy zz
surf
xy xz yz
D D D D D D
D
D D D
+ +
=
− − −
1
2 3
2 2
2
xx yy zz xx yz
vol
xy xz yz zz xy yy xz
D D D D D
D
D D D D D D D
−
=
+ − −
13. Scalar Anisotropy Indices
( )
avND
D
DADC
DDDADC
=
++= 321
3
1
( ) ( ) ( )
2
2
2
3
2
2
2
1
2
3
2
2
2
1
1
2
3
mag
surf
ND
D
D
D
FA
DDD
DDDDDD
FA
−=
++
−+−+−
=
14. FA vs. ADC
FA (x 10
-4
)
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Probability
0.0000
0.0002
0.0004
0.0006
0.0008
0.0010
WM
GM
CSF
MD (x 10 -3
mm2
/s)
0 500 1000 1500 2000 2500 3000 3500 4000
Probability
0.000
0.001
0.002
0.003
0.004
0.005
WM
GM
CSF
FA and ADC are very
useful clinically, but are
very different.
Tensor has a LOT of
information!
Q. Which metric would you use to detect brain cancer?
15. Vector anisotropy measures
We can use eigenvector
information from the
tensor as well
Represent direction of
primary eigenvector as
color on a scalar map
Or render the primary
eigenvectors as “fibers”
for astonishing* 3D
visualizations
Red = R/L
Green = A/P
Blue = S/I
*but how “real” is it? Many PhD
theses have asked….
18. Diffusion weighted MRI
∆ δ
G G
echoπα
( ) ( )( )2
0
δexpδ
3
M
G D
M
γ= − × ∆ − ×
∆ δ
( ) ( )2
δδ
3
b Gγ= × ∆ −
(boxcar gradients)
“b-value”
Consider
simplified diffusion
experiment…
19. MR Measurement of Diffusion Tensor
j
T
j j
j
p
G G q
r
= ×
ur
0
xx xy xz j
j j j xy yy yz j
xz yz zz j
D D D p
p q r D D D q
D D D r
j
b
S S e
− × × ×
=
( )
22 2
γ δ Δ δ 3b G= −
( )
1
6
j N
N
=
≥
Kjth
diffusion-
weighted
image
Diffusion
magnitude
Diffusion
direction
Gz
Gy
Gx
...
...
...
20. Solving for D
20
0
xx xy xz j
j j j xy yy yz j
xz yz zz j
D D D p
p q r D D D q
D D D r
j
b
S S e
− × × ×
=
1. Acquire T2W image (b = 0 s/mm2
)
3. Choose a diffusion gradient orientation2. Choose a b-value
4. Acquire image (Sj)
5. Repeat steps 1 – 3, j = 1 … N times
6. Solve for D…. How?
21. Let’s do some linear algebra…
[ ]
⋅
⋅⋅−=
zj
yj
xj
zzzyzx
yzyyyx
xzxyxx
zjyjxjj
DDD
DDD
DDD
bSS
α
α
α
αααexp0
( )
( )
( )
( )
( )
( )
⋅
⋅=
yz
xz
xy
xx
yy
xx
T
jxy
jxy
jxy
jzz
jyy
jxx
j
D
D
D
D
D
D
b
S
S
α
α
α
α
α
α
2
2
2
ln
2
2
2
0
1661 xNxNx xAY ⋅=
22. B-matrix formalism
22
( )yzyzxzxzxyxyzzzzyyyyxxxx DDDDDDb αααααα 222222
+++++⋅
( )yzyzxzxzxyxyzzzzyyyyxxxx DbDbDbDbDbDb 222 +++++=
∑∑= =
=
3
1
3
1i j
ijij Db
The “b-matrix”
The b-matrix formalism
summarizes total
attenuating effect of all
gradient waveforms in
all directions (including
imaging gradients)
23. T2W
(b = 0 s/mm2
)
Y, -ZY, Z-X, Y
X, Y-X, Z+X, Z
28. How low can you go?
High b-values mean more
attenuation, lower SNR
Lower b-values mean higher
SNR, room for more N
At very low b-values, imaging
gradients’ diffusion effects
are no longer negligible
Lower b-values also do not
probe same diffusion scale,
less clinically interesting
b=100 s/mm2
b=500 s/mm2
(N = 6, 8 NEX)
31. DT-MRI Alexander
Mapping Complex Diffusion
Based Upon Q-Space Theory – Model Independent
ODF – orientation density function (Tuch et al., Neuron 2003)
Diffusion Spectrum Imaging (DSI)
(Tuch et al. Neuron 2003, Wedeen et al. 2005)
High Angular Diffusion Imaging (HARDI), Q-Ball
(Frank 2002; Tuch et al. Neuron 2003)
Hinweis der Redaktion
Before I go over the diffusion enhancements, let me give a bit of background on diffusion imaging and why it is desirable
Heres a highly invasive technique for imaging the corpus callosum
And heres diffusion fiber tractography of the structure
The particles will move forward or backwards with equal probability, so half the particles undergo displacement -D and the other half undergo displacement +D .
After a single time interval tau we only need consider particles within distance D of a given point x0 to calculate the flux
half are displaced left and the other half are displaced right. Therefore, only half the particles between x0 and x0 D will cross x0 in one step.
THIS IS ONLY A CARTOON! demonstrate concept
injecting ink not part of methods and materials
if we repeat thought experiment in brain tissue, diffusing molecules will encounter obstructions from tissue microstructure
(unlike free water)
isotropic: degenerate case uynder this formalism where D1 = D2 = D3
asymmetric diffusion called “anisotropic”, Diffusion is strongly anisotropic in WMFT
NOTE: - ellipsoid true in limit of infinitesimal point
- oblique diffusion has components along three axes
we actually measure directly in DT-MRI : diffusion tensor coefficients, averaged over entire voxel. (volume element)
Diagonalize -> reduce to intrinsic reference frame of ellipsoid (lose orientation)
The shape invariants have the advantage that they can be constructed analytically from the tensor coefficients, unlike the eigenvalues which are usually obtained using an iterative computation.
The shape invariants have the advantage that they can be constructed analytically from the tensor coefficients, unlike the eigenvalues which are usually obtained using an iterative computation.
to visualize-> construct vector (2d) and scalar (1d) indices to characterize the anisotropy of tissue in that voxel
focus on scalar indices in this research
RA : VISUALIZATION gold standard, linear, 0-sqrt(2)
FA : more sensitive to low aniso, high aniso saturates (0-1)
ND/D definitions are equivalent - index VALUE is the same
--- EXTRA INFO --
- vector indices sensitive to noise, therefore formidable acquisition requirements (time, hardware).
- don’t need vectors : construct scalar indices from any set of tensor invariants
MR signal is attenuated by diffusion effects:
apply gradient
For every diffusion-weighted gradient we apply, we obtain one such equation.
6 unknowns in the diffusion tensor (due to symmetry),
need at least N = 6 diffusion weighted gradients.
more than 6 gradients -> use least-squares method like SVD to find the best fit
-- EXTRA info --
[ typically maximize Gd for shortest TE]
static spins see no net dephasing - 1800 pulse reverses polarity of first gradient. Diffusing spins, undergo net dephasing - experience different local magnetic susceptibility as
sensitize any MR pulse sequence by inserting diffusion-weighting magnetic field gradients
example of data set in Experiment 1 - N = 6, b = 500 s/mm2
diffusion-weighted images are attenuated from T2 baseline
signal varies with orientation of the gradient
fiber tracts’ orientation not known a-priori
equal weighting of anisotropy in all directions needed for max sensitivity
-> even spacing of diffusion-weighting gradients (“acquisition scheme”)
for this project. we developed acquisition schemes for N = 6, 13, 27, and 55
[ shown from same angle for consistency ]
N = 6: standard isocahedral encoding
[X, Z], [Y, Z], and [X, Y]
N = 27, 55: Perl algorithm
tiling unit hemisphere by equal solid angle disks
uses circular tiles as an approximation, cannot cover the surface precisely
algorithm performs poorly for N < 27
N = 13: modeled by electrostatic repulsion
each gradient treated as point charge on surface of hemisphere
iterative computation - solve for lowest energy configuration
all images are RA, gold standard commonly used
experiment 14 (N = 6)
exp. 15 - similar results
< 500 are worse, interference with imaging gradients