Advanced Methods in Diffusion Weighted Imaging

1
Advanced Methods in Diffusion Weighted Imaging

• Sudhir K Pathak
• University of Pittsburgh

2
What we can do with DTI

• Simple and Easy to Understand
• Good for calculating Fractional Anisotropy
• You can find Major fiber bundle using Fiber
  Tractography algorithm ( Algorithm works well for
  FA gt 0.8 ).

3
Limitation of DT-MRI

• Major drawback of DT-MRI is that the Gaussian
  Model is often poor fit to the data. DT-MRI
  provide only one fiber-orientation estimate in
  each voxel.
• When Gaussian model is poor fit the two major
  selling-point of DT-MRI fails are
• Indices of anisotropy derived from diffusion
  tensor, such as Fractional anisotropy,
  underestimate the true directional variability of
  p. The Gaussian model for p smoothes out the
  ridges of p.
• Fiber-orientation estimate are incorrect

Actual p
Diffusion tensor estimation
Actual p
Principle direction of diffusion tensor
4
(No Transcript)
5
Modeling the ADC profile

• As explained earlier ADC profile gives estimate
  of diffusion within voxel. Modeling log(A(q))
  with higher order polynomials can be used to
  detect departure from the Gaussian model and to
  obtain more reliable indices of anisotropy.
• One can write ADC profile as linear combination
  of Spherical harmonics,
• The spherical harmonics Ylm, l 0, ., 8, m
  -l, , l form a basis for complex valued
  functions on the unit sphere in three dimensions
  S2.

6
Voxel Classification
Black show background, zero order tensor,
White/Grey show white matter, 2nd order tensor,
Yellow show higher order tensor
7
Q Space
8
Representation of displacementof water molecule
in six dimensional space
9
Terms Used in this Field

• HARDI Higher Angular Resolution Diffusion
  Imaging means Acquiring more Angular Direction
  Data.
• ODF Oriented Distribution Function
• FOD Fiber Oriented Distribution Function

Tournier 2007
10
Advance Reconstruction Algorithms

• Multi-tensor models
• Tuch et al MRM 48 2002, Parker and Alexander IPMI
  2003.
• Behrens et al MRM 2003
• Higher Order Tensor
• Diffusion Spectrum Imaging q-space
• Wedeen (ISMRM 1999)
• Qball (Tuch et al Neuron 40 2003)
• Spherical Deconvolution
• Tournier et al NeuroImage 23 2004, Anderson MRM
  54 2005
• PASMRI Jansons and Alexander, Inverse Problems 19
  2003
• MESD (Alexander IPMI 2005)

11
Multi-Tensor Diffusion Imaging

• Limitation
• Model Selection problem, hard to decide what n to
  choose.
• Non-linear Fitting makes solution unstable.

D C Alexander
12
Some Results
Sotiropoulos et al 2008
13
D C Alexander
14
Higher Order Tensor Model
15
(No Transcript)
16
Tensor Distribution Function
17
Examples
18
Discussion

• Multi-tensor
• - Model Selection with Multi-tensor Algorithms
  i.e., dont know how many fiber a voxels contain.
• Computationally Faster and Scan time is less
• Software exists CAMINO
• Higher Order Tensor
• - How many order tensor you want to use for a
  voxel
• Very intuitive way to extends Diffusion
  Tensor MRI
• Software doesnt exists but can request for some
  code
• Tensor Distribution

19
Diffusion Spectrum Imaging
20
Q Space
21
Theory
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
Discussion

• No model selection problem, This is best
  possible method theoretically
• - Long scan time which make this method
  impossible for clinical purpose.
• - Even 257 515 is not enough to cover all
  Q-Space.
• Software TrackVis and DSIStudio

26
Q-BallImaging
27
(No Transcript)
28
(No Transcript)
29
(No Transcript)
30
Results
31
Results
32
Results
33
Results
34
Results
35
Comparison
36
More Comparison
37
(No Transcript)
38
Discussion

• No model assumption
• Computationally FAST
• - Long Scan time, typically 256 gradient
  direction, it takes 38 min with PARALLEL Imaging.
• - We sampled Q-Space on single shell
• Software TrackVis, CAMINO, 3D Slicer and
  DSIStudio

39
SphericalDeconvolutionMethods
40
(No Transcript)
41
(No Transcript)
42
Examples showing Fiber Crossing
43
constrained super-resolvedspherical deconvolution
44
Results
Results from data set A (b3000 s/mm2, 60 DW
directions) in the centrum semiovale (location
shown on FA map in middle). The estimated FOD is
shown in each voxel as a surface rendered plot.
In each case, the main contributing fibre tracts
can be identified the corona radiata running in
the inferior-superior direction (in blue), the
superior longitudinal fasciculus (in green) and
the commissural projections from the corpus
callosum (in red).Top left filtered SD With
lmax8 top right CSD with lmax8 bottom left
super-CSD with lmax10 bottom right super-CSD
with lmax12.
45
Results
Results from data set C (b1000 s/mm2, 60 DW
directions) in the centrum semiovale (location
shown on FA map in middle). The estimated FOD is
shown in each voxel as a surface rendered plot.
In each case, the main contributing fibre tracts
can be identified the corona radiata running in
the inferior-superior direction (in blue), the
superior longitudinal fasciculus (in green) and
the commissural projections from the corpus
callosum (in red). Top left filtered SD with
lmax8 top right CSD with lmax8 bottom left
super-CSD with lmax10 bottom right super-CSD
with lmax12
46
Results
Results from data set D (b1000 s/mm2, 20 DW
directions) in the centrum semiovale (location
shown on FA map in middle). The estimated FOD is
shown in each voxel as a surface rendered plot.
Top left filtered SD with lmax4 top right
CSD with lmax4 bottom left super-CSD with
lmax6 bottom right super-CSD with lmax8.
47
Discussion

• With very less gradient direction you can
  resolve fiber crossings
• Computational time is reasonable
• - Sensitive to noise, new method have better
  way to handle it.
• - Works good if you have higher b-value
• Software MrTRIX

48
Persistent Angular StructurePASMRI
49
Persistent Angular Structure (PAS MRI)

• Jansons and Alexanders PASMRI algorithms
  computes another feature of p called the
  persistent angular structure (PAS). The PAS has
  sharper peaks than ODF an resolves the crossing
  fibers more consistently

50
(No Transcript)
51
(No Transcript)
52
Comparison of DTI and PAS-MRI of crossing fibers.
53
(No Transcript)
54
(No Transcript)
55
Discussion

• No model assumption, I think best method to
  resolve fiber crossing
• Very robust method w.r.t. gradient
  directions.
• - Computationally very very Expensive
• - Written in Java
• Compiled code is also available
• Software CAMINO

56
Suggestions

• Diffusion tensor imaging
• Minimum 7 gradient direction
• Scan time typically 1-3 minutes
• Measure FA, MD and Lattice Index, Fiber bundle
  with FA gt 0.8 is good

• Advanced Method
• 60 - 128 gradient direction
• Scan time typically 15 - 25 minutes
• Method like PASMRI, Spherical Deconvolution can
  be used.

• Advanced Method
• 128 - 256 gradient direction
• Scan time typically 25 - 35 minutes
• Method like QBall can be used.

• Diffusion tensor imaging
• 12 -24 gradient direction
• Scan time typically 5 - 10 minutes
• Probabilistic Tractography can be used, Tensor
  estimation will be more robust and reliable

• Advanced Method
• 256 gt gradient direction
• Scan time typically more than hour(s0
• Method like DSI can be used.

• Diffusion (Multi) - tensor imaging
• 30 - 60 gradient direction
• Scan time typically 10 - 15 minutes
• Probabilistic/Deterministic Tractography can be
  used, with Multi - Tensor algorithms

57
TOOL Reconstruction Method Fiber Tracking Comments
Camino Almost All PiCO, FACT Command Line tools, Advanced Algorithms
TrackVIs HARD, QBALL, TENSOR FACT Very Nice GUI and Rendering Capability
MEDInria Log-normal FACT Nice Rendering, you can import fMRI roi
FSL Using probability to estimate diffusion tensor ProbTrack Connectivity, Paracellation, connectivity based segmentation
GTract Tensor GTract, FACT, TEND, TensorLine Eddy current correction, B-spline registration, Graph theory based fiber Tracking
DTIQuery/AFNI Tensor FACT, TEND, Tensorline Nice Rendering, robust estimation of Tensor
58
TOOL Reconstruction Method Fiber Tracking Comments
SCIRun/TEEM/BioTensor Tensor FACT, TEND, Tensorline Nice programming capibility
Slicer/TEEM TENSOR FACT
FIberViewer/NeuroLib Tensor FACT Nice Clustering algorithms
BioImageSuite Tensor FACT Tensor rotation
mrVista Tensor FACT, TEND, TensorLine Matlab based
ExploreDTI Tensor FACT Matlab based
BrainVISA Tensor FACT, TEND, TensorLine Python based, Problem in instation
DPTOOL Tensor FACT Nice rendering
MrTRIX Sperical Decovolution FACT, Prob. Mixture of command line and GUI
doDTI Tensor FACT, TEND, Tensorline Clustering algorithms
Other SPM, CImg, R, BrainVoyager, DTVii