structured sparsity through reweighting and …structured sparsity through reweighting and...
TRANSCRIPT
![Page 1: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/1.jpg)
Structured Sparsity through reweighting and Application to
diffusion MRIMarch 25th, 2015
EPFL-Idiap-ETH Sparsity Workshop 2015
Anna Auria
Signal Processing Laboratory (LTS5)École Polytechnique Fédérale de Lausanne (EPFL)
Tuesday, March 24, 15
![Page 2: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/2.jpg)
2
Introduction and Outline
Outline:
✓ Diffusion MRI and problem formulation
✓ Structured sparsity through reweighting
✓ Results
✓ Discussion and future work
Problem: Recovery of multiple correlated sparse signals
Tuesday, March 24, 15
![Page 3: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/3.jpg)
3
Diffusion MRI (dMRI)✤ What is it?
✓ Diffusion MRI measures the Brownian motion of water molecules in a fluid due to thermal energy.✓ In ordered tissues, water does not diffuse equally in all directions (anisotropic diffusion).
Study the spatial order in living organs in a non-invasive way.
Tuesday, March 24, 15
![Page 4: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/4.jpg)
3
Diffusion MRI (dMRI)✤ What is it?
✓ Diffusion MRI measures the Brownian motion of water molecules in a fluid due to thermal energy.✓ In ordered tissues, water does not diffuse equally in all directions (anisotropic diffusion).
Study the spatial order in living organs in a non-invasive way.
Tuesday, March 24, 15
![Page 5: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/5.jpg)
3
Diffusion MRI (dMRI)✤ What is it?
✓ Diffusion MRI measures the Brownian motion of water molecules in a fluid due to thermal energy.✓ In ordered tissues, water does not diffuse equally in all directions (anisotropic diffusion).
Study the spatial order in living organs in a non-invasive way.
✤ STRUCTURAL NEURAL CONNECTIVITY
Tuesday, March 24, 15
![Page 6: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/6.jpg)
3
Diffusion MRI (dMRI)✤ What is it?
✓ Diffusion MRI measures the Brownian motion of water molecules in a fluid due to thermal energy.✓ In ordered tissues, water does not diffuse equally in all directions (anisotropic diffusion).
Study the spatial order in living organs in a non-invasive way.
✓Why? Neuroscience / Clinical applications✓How? Fiber tracking (tractography)
✤ STRUCTURAL NEURAL CONNECTIVITY
Tuesday, March 24, 15
![Page 7: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/7.jpg)
4
dMRI: Local Reconstruction problem
Recover the fiber orientation in every voxel of the brain.
✓ Fiber Orientation Distribution (FOD)
Probability of having a fiber along a given direction (function on )S2
Function of interest:
Tuesday, March 24, 15
![Page 8: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/8.jpg)
5
dMRI: FOD recovery via sparse reconstructionAssumptions:
1. Diffusion characteristics of all fiber in the brain are identical.
2. No exchange between spatially distinct fiber bundles.f1S1(✓,�) + f2S2(✓,�) = S(✓,�) = R(✓) ⌦ F (✓,�)
Signal attenuation Fiber ✓ KERNEL: Response generated by a single fiber estimated from the data.
Tuesday, March 24, 15
![Page 9: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/9.jpg)
5
dMRI: FOD recovery via sparse reconstructionAssumptions:
1. Diffusion characteristics of all fiber in the brain are identical.
2. No exchange between spatially distinct fiber bundles.f1S1(✓,�) + f2S2(✓,�) = S(✓,�) = R(✓) ⌦ F (✓,�)
Signal attenuation Fiber ✓ KERNEL: Response generated by a single fiber estimated from the data.
Spherical Deconvolution methods assume the signal can be expressed as the convolution of a kernel with the Fiber Orientation Distribution:
*=
SIGNAL FODKERNEL
Tuesday, March 24, 15
![Page 10: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/10.jpg)
5
dMRI: FOD recovery via sparse reconstructionAssumptions:
1. Diffusion characteristics of all fiber in the brain are identical.
2. No exchange between spatially distinct fiber bundles.f1S1(✓,�) + f2S2(✓,�) = S(✓,�) = R(✓) ⌦ F (✓,�)
Signal attenuation Fiber ✓ KERNEL: Response generated by a single fiber estimated from the data.
Spherical Deconvolution methods assume the signal can be expressed as the convolution of a kernel with the Fiber Orientation Distribution:
*=
SIGNAL FODKERNEL
✓ non-negative
✓ sparse
Tuesday, March 24, 15
![Page 11: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/11.jpg)
6
y = �x + �✓ is the acquired diffusion MRI data and is the FOD (of a single voxel).
✓ is the sensing basis or dictionary.
✓ represents de acquisition noise.
y��
x
The intra-voxel recovery problem can be expressed voxelwise in terms of the following linear formulation: (Jian and Vemuri, 2007)
Each atom of the dictionary is associated to a discrete direction on the sphere
dMRI: FOD recovery via sparse reconstruction
Reweighted constrained minimization
sparsity term
whereminx�0
k�x� yk22 s.t. kxkw,1 k
�1
kxkw,1 =X
i
wi|xi|
(Candes et al, 2008)
Tuesday, March 24, 15
![Page 12: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/12.jpg)
6
y = �x + �✓ is the acquired diffusion MRI data and is the FOD (of a single voxel).
✓ is the sensing basis or dictionary.
✓ represents de acquisition noise.
y��
x
The intra-voxel recovery problem can be expressed voxelwise in terms of the following linear formulation: (Jian and Vemuri, 2007)
Each atom of the dictionary is associated to a discrete direction on the sphere
dMRI: FOD recovery via sparse reconstruction
Reweighted constrained minimization
sparsity term
w(t)i � 1/x(t�1)
iSolving a sequence of these weighted problems with
whereminx�0
k�x� yk22 s.t. kxkw,1 k
�1
kxkw,1 =X
i
wi|xi|
(Candes et al, 2008)
Tuesday, March 24, 15
![Page 13: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/13.jpg)
7
sparsity and structurespatial regularisation
✓IDEA: Solve the FOD field for all voxels simultaneously to exploit spatial coherence between neighboring voxels:
minX2Rn⇥N
+
k�X� Yk22 s.t. kXkW,1 K.
Structured Sparsity through reweighting in dMRI
kXkW,1 =X
d,v
Wdv|Xdv|
voxels
dire
ctio
ns { {X 2 Rn⇥N+
Proposed formulation:
with
✓Assumption: neighbor voxels should present the same/neighbor directions.
Tuesday, March 24, 15
![Page 14: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/14.jpg)
8
✓Assumption: neighbor voxels should present the same/neighbor directions.
N (d)
N (v)
directions {d ... ... ...
X
angular neighbourhood
voxels
{v...
... Xspatial neighbourhood
...
...
Structured Sparsity through reweighting in dMRI
Definition of neighborhood:
Tuesday, March 24, 15
![Page 15: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/15.jpg)
9Spherical deconvolution: global problem
Xdvvoxels
directions { {d
v...
...
... ...
...
...
W(t+1)dv =
1
⌧ (t) +P
d0v02N(dv) |X(t)
d0v0 ||N (v)|
✓Assumption: neighbor voxels should present the same/neighbor directions.Definition of neighborhood:
Definition of the weights:
Tuesday, March 24, 15
![Page 16: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/16.jpg)
10
Simulations and Results
Exploiting spatial coherence Undersampling regimes Speed up acquisition
Tuesday, March 24, 15
![Page 17: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/17.jpg)
11
Simulations and Results
Exploiting spatial coherence Undersampling regimes Speed up acquisition
Tuesday, March 24, 15
![Page 18: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/18.jpg)
12
Conclusions
CONCLUSIONS:
✓ Spatially structured sparsity guaranties robustness to noisy and ability to go to higher undersampling regimes.
✓The method is versatile and can be generalised to recover multiple correlated sparse signals
FUTURE WORK:
✓ in dMRI: application to recovery of microstructure properties of the tissue, tractography methods,...
Tuesday, March 24, 15
![Page 19: Structured Sparsity through reweighting and …Structured Sparsity through reweighting and Application to di!usion MRI March 25th, 2015 EPFL-Idiap-ETH Sparsity Workshop 2015 Anna Auria](https://reader036.vdocuments.site/reader036/viewer/2022081222/5f7880b83f93b22ab9481b57/html5/thumbnails/19.jpg)
13
THANK YOU
Tuesday, March 24, 15