Diffusion Tensor Images
Anne Collard
1
, Silvère Bonnabel
2
, Christophe Phillips
3
, Rodolphe Sepulchre
1
1
Department of Electrical Engineering and Computer Science, University Of Liège, Belgium
2
RoboEcs Center, Mines ParisTech, Paris, France
3
Cyclotron Research Centre, University Of Liège, Belgium
In this work we propose to use the spectral decomposi7on of diffusion tensors in order to preserve all the diffusion informaEon. Let S1 and S2 be two diffusion tensors, whose spectral decomposiEons are , the proposed similarity measure is
Proposed similarity measure: based on the spectral decomposi7on
Interpola7on of diffusion tensors
[1] Le Bihan, D. (2001), ‘Diffusion tensor imaging: Concepts and applicaEons’, Journal of MagneEc Resonance Imaging, vol. 13, no. 4, pp. 534-‐546 [2] Arsigny V. (2006), ‘Log-‐Euclidean metrics for fast and simple calculus on diffusion tensors’, MagneEc Resonance in Medicine, vol. 56, no. 2, pp. 411-‐421 [3] André, E. , ‘IteraEve method for moEon correcEon of diffusion weighted images’, OHBM 2012, Poster #615 MT.
This paper presents research results of the Belgian Network DYSCO (Dynamical Systems, Control, and OpEmizaEon), funded by the Interuniversity AcracEon Poles Programme, iniEated by the Belgian State, Science Policy Office. The scienEfic responsibility rests with its authors.
ANISOTROPY PRESERVING INTERPOLATION OF DIFFUSION TENSORS
C
YCLOTRON
R
ESEARCH
C
ENTRE
|
hcp://www.cyclotron.ulg.ac.be
| Anne Collard|
anne.collard@ulg.ac.be
Each voxel of the image contains a symmetric posiEve definite matrix. The principal characterisEcs of tensors are [1]
-‐ mean diffusivity (characterizes the overall mean squared displacement of molecules) -‐ main direc7on of diffusiviEes (orientaEon in space of the structures of the brain) -‐ degree of anisotropy (describes the degree to which molecular displacements vary in space)
The processing of diffusion tensor images requires the development of methods adapted to ‘matrix-‐valued’ images. One of the most used method is the Log-‐Euclidean framework [2], for which
d
LE(S
1,S
2)
=|| log(S
1)
! log(S
2) ||
Fig: Zoom on a slice of a Diffusion Tensor Image
Si= Ui!iUi T
d
SD 2(S
1,S
2)
= k(!
1,
!
2) d
SO(3) 2(U
1,U
2)
+ d
D+(3) 2(!
1,!
2)
Spherical or isotropic tensors have low orientaEonal informaEon, so their orientaEon contribuEon to the metric should be minimal.
k is constructed to be small (around zero) when
any of the tensor is isotropic and close to 1 when both are highly anisotropic.
k
![0,1]
Distance between orientaEons is computed through the use of unit quaternions. This parametrizaEon of the rotaEons enables to speed up the computaEons.
Because of the non-‐uniqueness of the spectral decomposiEon, 4 different rotaEon matrices represent a single tensor. This implies 8 quaternions for a tensor. The distance between them is computed through
d(Q1,Q2)= minq2!Q2|| q1
r" q
2||
This is called the realignment of quaternions.
Eigenvalues of tensors are manipulated as geometric en77es (i.e. the logarithms of eigenvalues are used for computaEons).
The distance between the intensiEes of diffusion is thus given by d(!1,!2)= log 2 i
"
#1,i #2,i $ % & '()The mean of a group of N tensors is given by
Mean of a group of N tensors
S
µ= U
µ!
µU
Tµ!
µThe eigenvalues of the mean are given by the geometric mean of the eigenvalues.
!µ= exp( wi
i=1
N
"
log(!i)) Knowing the mean eigenvalues, the factor k can be computed for each tensor.ki= k(!µ,!i)
U
µThe orientaEon of the mean is given by the mean quaternion.
The most informa7ve tensor will be chosen as a reference for the computaEon of the mean. Each quaternion has to be realigned with the reference, because of the non uniqueness of the spectral decomposiEon.
Most informa7ve tensor: the one for which the product wiki is the maximum.
qm= wi i=1 N
!
kiqi,r ki i!
q
µ=
q
m|| q
m||
This method ensures the commuta7vity and the uniqueness of the mean.0 0,25 0,5 0,75 1 0 40 80 120 160 0 0 5 0.5 0.75 0 0.6 1.2 0 0 5 0.5 0.75 1 0.76 0.88 0.94 0 0.25 0.5 0 5 1 2.2 3 3.8 0,75 0,5 0,25 0 1 . .7 0,75 0,5 0,25 0 1 0.9 0 0,25 0,5 0,75 1 det(S) Φ FA HA 0 0,25 0,5 0,75 1 0 40 80 120 160 0 0.25 0.5 0.75 1 0 0.6 1.2 0 0.25 0.5 0.75 1 0.76 0.88 0.9 0.94 0 0.25 0.5 0.75 1 2.2 3 3.8 0,75 0,5 0,25 0 0 0 0,25 0,5 0,75 1 0.9 0 0,25 0,5 0,75 1 Log-‐Euclidean interpolaEon Proposed method
Fig: Geodesic interpolaEons between two tensors and variaEon of several characterisEcs (determinant, angle between the principal direcEons of diffusion, and fracEonal anisotropy) of tensors through these interpolaEons.
Clear differences can be seen between the two methods, namely the main tensor orientaEon and the degrees of anisotropy. These characterisEcs are well preserved by the developed framework, which is advantageous for any imaging or tractographic applicaEons.