advanced image-processing techniques in magnetic ... - univr
TRANSCRIPT
![Page 1: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/1.jpg)
1
Alessandro Daducci( work in collabora-on with Dr. Yves Wiaux )
Exploi'ng sparsityin diffusion MRI
30/03/2012University of Verona
![Page 2: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/2.jpg)
Outline
2
• Diffusion and why it is important
• Diffusion MRI
• Compressed Sensing (CS) framework
• Applica@on of CS-‐based techniques in diffusion MRI to reduce acquisi@on @mes
![Page 3: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/3.jpg)
What is diffusion?• Random movement of molecules from regions of high concentra@on to regions of lower concentra@on due to thermal agita@on
– EXAMPLE: in a glass of water, molecules diffuse randomly and freely, only constrained by the boundaries of the container
• First noted by Robert Brown in 1828– “…random mo-on without any apparent cause...”
3
![Page 4: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/4.jpg)
What is diffusion?• Random movement of molecules from regions of high concentra@on to regions of lower concentra@on due to thermal agita@on
– EXAMPLE: in a glass of water, molecules diffuse randomly and freely, only constrained by the boundaries of the container
• First noted by Robert Brown in 1828– “…random mo-on without any apparent cause...”
• Formally described by Albert Einstein in 1905mean displacement
of molecules coefficient of diffusionof the medium
observaBon Bme
3
![Page 5: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/5.jpg)
• Cerebrospinal Fluid (CSF)– Displacements follow a gaussian distribu'on in every direcBon (isotropic)
– Variance depends on the level of restricBon of the fluid
4
What happens in the brain?
![Page 6: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/6.jpg)
• Cerebrospinal Fluid (CSF)– Displacements follow a gaussian distribu'on in every direcBon (isotropic)
– Variance depends on the level of restricBon of the fluid
• White-‐maLer: neuronal 'ssues
4
What happens in the brain?
![Page 7: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/7.jpg)
• Cerebrospinal Fluid (CSF)– Displacements follow a gaussian distribu'on in every direcBon (isotropic)
– Variance depends on the level of restricBon of the fluid
• White-‐maLer: neuronal 'ssues
– Diffusion is more restricted in direcBons perpendicular to the main axis of the fiber than along the fiber itself
– Displacements sBll follow a gaussian distribuBon, but anisotropic
4
What happens in the brain?
![Page 8: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/8.jpg)
• Cerebrospinal Fluid (CSF)– Displacements follow a gaussian distribu'on in every direcBon (isotropic)
– Variance depends on the level of restricBon of the fluid
• White-‐maLer: neuronal 'ssues
– Diffusion is more restricted in direcBons perpendicular to the main axis of the fiber than along the fiber itself
– Displacements sBll follow a gaussian distribuBon, but anisotropic
• NOTE: regions with crossing fibers– This situaBon cannot be described by a gaussian distribuBon!
– SophisBcated techniques are required
4
What happens in the brain?
![Page 9: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/9.jpg)
Diffusion MRI in a nutshell• MRI sequences are made sensi@ve to diffusion by inser@ng two addi@onal magne@c field gradient pulses
- The goal is to dephase spins which move in a given direcBon
- Signal decays exponenBally as:
- b is called b-‐value and controls the diffusion weighBng (contrast) of images
5
![Page 10: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/10.jpg)
Diffusion MRI in a nutshell• MRI sequences are made sensi@ve to diffusion by inser@ng two addi@onal magne@c field gradient pulses
- The goal is to dephase spins which move in a given direcBon
- Signal decays exponenBally as:
- b is called b-‐value and controls the diffusion weighBng (contrast) of images
• Diffusion is strongly dependent on the gradient direc@on:
I-‐S L-‐R A-‐P5
![Page 11: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/11.jpg)
Main applica'ons• Fiber-‐tracking: algorithms which infer axonal trajectories inside the brain by exploi@ng the diffusion informa@on in each voxel
6
![Page 12: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/12.jpg)
Main applica'ons• Fiber-‐tracking: algorithms which infer axonal trajectories inside the brain by exploi@ng the diffusion informa@on in each voxel
• Connec'vity analysis: in-‐vivo and non-‐invasive assessment of structural wiring of the brain
6fiber-‐tracking corBcal segmentaBon adjacency matrix
![Page 13: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/13.jpg)
EAP, ODF and fODF• EAP (Ensemble Average Propagator)
– for every voxel, it’s the 3D PDF giving the probability of agiven displacement ➡ diffusion MRI is a 6D modality
– related to the signal aCenua-on E by a 3D FFT:
7
q-‐space
![Page 14: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/14.jpg)
EAP, ODF and fODF• EAP (Ensemble Average Propagator)
– for every voxel, it’s the 3D PDF giving the probability of agiven displacement ➡ diffusion MRI is a 6D modality
– related to the signal aCenua-on E by a 3D FFT:
• ODF (Orienta'on Distribu'on Func'on)– probability of diffusion along a given direc-on:
– funcBon on the unit sphere
7
q-‐space
![Page 15: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/15.jpg)
EAP, ODF and fODF• EAP (Ensemble Average Propagator)
– for every voxel, it’s the 3D PDF giving the probability of agiven displacement ➡ diffusion MRI is a 6D modality
– related to the signal aCenua-on E by a 3D FFT:
• ODF (Orienta'on Distribu'on Func'on)– probability of diffusion along a given direc-on:
– funcBon on the unit sphere
• fODF (fiber ODF)– sum of spikes idenBfying the fiber direcBons, with
amplitudes corresponding to the volume frac-ons
– funcBon on the unit sphere
7
q-‐space
![Page 16: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/16.jpg)
Diffusion TENSOR Imaging (DTI)•Assump'on: displacements follow a gaussian distribu-on
• Fully characterized by its covariance matrix– 3x3 symmetric posiBve semi-‐definite matrix, called Diffusion Tensor– plo^ed as an ellipsoid by means of its eigenvalues and eigenvectors – 6 degrees of freedom (3 rotaBons + 3 variances)
• The Diffusion Tensor can be es@mated by least-‐squares(at least 6 DWI images along different direc@ons are needed)
8
![Page 17: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/17.jpg)
Diffusion TENSOR Imaging (DTI)•Assump'on: displacements follow a gaussian distribu-on
• Fully characterized by its covariance matrix– 3x3 symmetric posiBve semi-‐definite matrix, called Diffusion Tensor– plo^ed as an ellipsoid by means of its eigenvalues and eigenvectors – 6 degrees of freedom (3 rotaBons + 3 variances)
• The Diffusion Tensor can be es@mated by least-‐squares(at least 6 DWI images along different direc@ons are needed)
•DTI scalar measures:Mean Diffusivity:high values = fast diffusion
Frac'onal Anisotropy:high values = fiber bundles
8
![Page 18: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/18.jpg)
Diffusion TENSOR Imaging (DTI)•Assump'on: displacements follow a gaussian distribu-on
• Fully characterized by its covariance matrix– 3x3 symmetric posiBve semi-‐definite matrix, called Diffusion Tensor– plo^ed as an ellipsoid by means of its eigenvalues and eigenvectors – 6 degrees of freedom (3 rotaBons + 3 variances)
• The Diffusion Tensor can be es@mated by least-‐squares(at least 6 DWI images along different direc@ons are needed)
•DTI scalar measures:
▲ Fast acquisi'ons: only 6 samples needed▼ Crossing fibers cannot be modeled
Mean Diffusivity:high values = fast diffusion
Frac'onal Anisotropy:high values = fiber bundles
8
![Page 19: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/19.jpg)
Diffusion SPECTRUM Imaging (DSI)• Measures directly the displacement of water molecules making no assump'ons about the underlying diffusion
– acquire data on a 3D cartesian grid
– FFT the data in q-‐space to obtain the EAP
– radial integra@on to obtain the ODF
– ODF maxima extrac@on to have the fODF
9
![Page 20: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/20.jpg)
Diffusion SPECTRUM Imaging (DSI)• Measures directly the displacement of water molecules making no assump'ons about the underlying diffusion
– acquire data on a 3D cartesian grid
– FFT the data in q-‐space to obtain the EAP
– radial integra@on to obtain the ODF
– ODF maxima extrac@on to have the fODF
• To recover the EAP, the q-‐space must be properly sampled:– 515 images usually acquired (acquisiBon Bme: DSI ≃ 60 min, DTI ≃ 3 min)
forming a 11x11x11 cartesian grid (all points inside a sphere of radius 5)
– high diffusion gradient strengths (b-‐value up to 8000 s/mm2, DTI ≃ 1000 s/mm2)
9
![Page 21: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/21.jpg)
Diffusion SPECTRUM Imaging (DSI)• Measures directly the displacement of water molecules making no assump'ons about the underlying diffusion
– acquire data on a 3D cartesian grid
– FFT the data in q-‐space to obtain the EAP
– radial integra@on to obtain the ODF
– ODF maxima extrac@on to have the fODF
• To recover the EAP, the q-‐space must be properly sampled:– 515 images usually acquired (acquisiBon Bme: DSI ≃ 60 min, DTI ≃ 3 min)
forming a 11x11x11 cartesian grid (all points inside a sphere of radius 5)
– high diffusion gradient strengths (b-‐value up to 8000 s/mm2, DTI ≃ 1000 s/mm2)
▲ Recovers the full EAP (full descrip@on of diffusion process)▼ Very long acquisi'ons, no clinically feasible▼ Powerful gradients are required
9
![Page 22: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/22.jpg)
Spherical deconvolu'on (CSD)
10
• The ODF can be seen as a convolu'on on the sphere
• The kernel characterizes the diffusion of a single fiber
= *ODF kernel fODF
![Page 23: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/23.jpg)
Spherical deconvolu'on (CSD)
10
• The ODF can be seen as a convolu'on on the sphere
• The kernel characterizes the diffusion of a single fiber
• It can be easily es@mated from the data1) iden@fy known areas of the brain with just one fiber;2) fit a tensor in each voxel and averaging
= *ODF kernel fODF
![Page 24: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/24.jpg)
Spherical deconvolu'on (CSD)
10
• The ODF can be seen as a convolu'on on the sphere
• The kernel characterizes the diffusion of a single fiber
• It can be easily es@mated from the data1) iden@fy known areas of the brain with just one fiber;2) fit a tensor in each voxel and averaging
▲ Reduces the number of samples down to 60▼ Sensi@ve to noise▼ Assumes the same diffusion proper@es across the brain
= *ODF kernel fODF
![Page 25: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/25.jpg)
Compressed Sensing in a nutshell• According to the CS theory, it is possible to recover a signal from less samples than those required by the Nyquist criterion, provided that the signal is sparse in some representa-on
11
![Page 26: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/26.jpg)
Compressed Sensing in a nutshell• According to the CS theory, it is possible to recover a signal from less samples than those required by the Nyquist criterion, provided that the signal is sparse in some representa-on
• Let be the signal to be recovered from the linear . measurements s.t. is sparse.If and obey some randomness and incoherence condiBons,then can be recovered by solving an inverse problem like:
11sparsity basis sensing basis
![Page 27: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/27.jpg)
• NOTE 1: l1 is a convex relaxa@on for l0 (“true definiBon” of sparsity)
![Page 28: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/28.jpg)
• NOTE 1: l1 is a convex relaxa@on for l0 (“true definiBon” of sparsity)
• NOTE 2: since and , then the system is under-‐determined and there are infinite solu@ons
If there is a unique sparse solu@on, CS will find it!n unknowns
m equ
aBon
s
![Page 29: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/29.jpg)
• NOTE 1: l1 is a convex relaxa@on for l0 (“true definiBon” of sparsity)
• NOTE 2: since and , then the system is under-‐determined and there are infinite solu@ons
If there is a unique sparse solu@on, CS will find it!
• NOTE 3: several variantsBPDN
l1 regularized Least Squares
LASSO
Total VariaBon minimizaBon
n unknowns
m equ
aBon
s
![Page 30: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/30.jpg)
Sparsity in diffusion MRI• The ODF is not sparse at all...
13
![Page 31: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/31.jpg)
Sparsity in diffusion MRI• The ODF is not sparse at all...
• ...but the fODF is very sparse!
13
![Page 32: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/32.jpg)
Sparsity in diffusion MRI• The ODF is not sparse at all...
• ...but the fODF is very sparse!
• Recast the reconstruc@on problem as:
where:– y is the acquired MRI data
– A is a dic'onary of atoms (as in Matching Pursuit) built from the single-‐fiber kerneles@mated from the data and rotated along each possible direc@on
– x is the fODF we want to recover (contribuBons, i.e. volume frac-ons, of each atom to the final ODF)
– ϵ can be sta@s@cally es@mated from data
13
is the iden'tyin this seing!
![Page 33: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/33.jpg)
• Simula'ons of voxel configura@ons with:– 2 fibers crossing at given angles, ranging from 30° to 90°
– different combina@ons of noise, b-‐value, volume frac-ons andorienta-on in space of the same configura@on
– 1000 repe--ons with different realiza@ons of the noise
14
![Page 34: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/34.jpg)
• Simula'ons of voxel configura@ons with:– 2 fibers crossing at given angles, ranging from 30° to 90°
– different combina@ons of noise, b-‐value, volume frac-ons andorienta-on in space of the same configura@on
– 1000 repe--ons with different realiza@ons of the noise
▲ CS-‐based reconstruc@on performs even be]er than DSI with only 10% of the samples!
14
![Page 35: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/35.jpg)
• Simula'ons of voxel configura@ons with:– 2 fibers crossing at given angles, ranging from 30° to 90°
– different combina@ons of noise, b-‐value, volume frac-ons andorienta-on in space of the same configura@on
– 1000 repe--ons with different realiza@ons of the noise
▲ CS-‐based reconstruc@on performs even be]er than DSI with only 10% of the samples!
▼ Unfortunately, already done...B.A. Landman et al. “Resolu-on of crossing fibers with constrained compressed sensing using diffusion tensor MRI”. NeuroImage (November 2011)
14
![Page 36: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/36.jpg)
State-‐of-‐the-‐art method• In [Landman, 2011] they formulated the problem as:
where λ is empirically set to .
15
![Page 37: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/37.jpg)
State-‐of-‐the-‐art method• In [Landman, 2011] they formulated the problem as:
where λ is empirically set to .
▼ Performs very well (30 samples)
15
![Page 38: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/38.jpg)
State-‐of-‐the-‐art method• In [Landman, 2011] they formulated the problem as:
where λ is empirically set to .
▼ Performs very well (30 samples)
▲ However, there’s s@ll room for improvements:(i) it turns out that sparsity is not properly exploited, since by defini@on:
(ii) they do not exploit any spa@al correla@on among voxels,
which instead is the case in real world
15
volume frac'ons
![Page 39: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/39.jpg)
Proposal 1: enhancing sparsity• Re-‐weighted l1 minimiza@on:
– ini@al weights
– at itera@on τ + 1, the problem is solved with weights computed from previous solu@on:
16
![Page 40: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/40.jpg)
Proposal 1: enhancing sparsity• Re-‐weighted l1 minimiza@on:
– ini@al weights
– at itera@on τ + 1, the problem is solved with weights computed from previous solu@on:
16
Wx
τ = 0
![Page 41: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/41.jpg)
Proposal 1: enhancing sparsity• Re-‐weighted l1 minimiza@on:
– ini@al weights
– at itera@on τ + 1, the problem is solved with weights computed from previous solu@on:
16
Wx
τ = 0 τ = 1
![Page 42: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/42.jpg)
Comparison se]ngs• 3 synthe@c fields
(1) 90° fiber crossing
(2) more realis@c fiber crossing
(3) single-‐voxel configura@onscrossing at given angles
• 3 quality measures
(1) MSE with the ground-‐truth ODFs
(2) average angular error in es@ma@ngthe orienta@ons of the fibers
(3) false fiber rate
17
![Page 43: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/43.jpg)
Results 1/3
18
![Page 44: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/44.jpg)
Results 2/3
19
![Page 45: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/45.jpg)
Results 3/3
20
Reconstruc@on with: 30 samples
![Page 46: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/46.jpg)
Results 3/3
21
Reconstruc@on with: 15 samples
![Page 47: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/47.jpg)
Proposal 2: spa'al correla'on• Re-‐arranging data:
22
voxels
direcBon
s
correla'onbetweenvoxels
in each voxeldirecBons are sparse
![Page 48: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/48.jpg)
Proposal 2: spa'al correla'on• Re-‐arranging data:
• Reconstruc@on problem re-‐formulated as:
– tries to enforce sparsity in each voxel
– tries to exploit the correla'on of direc@ons among voxels (as before, it is a convex relaxaBon for rk(A) )
22
voxels
direcBon
s
correla'onbetweenvoxels
in each voxeldirecBons are sparse
![Page 49: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/49.jpg)
Results 1/2
23
![Page 50: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/50.jpg)
Results 1/2
23
GROUND-‐TRUTHRe-‐weighted l1 LR
(simulaBons with 6 samples)
![Page 51: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/51.jpg)
Results 2/2
24
![Page 52: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/52.jpg)
LR minimiza'on: issues• With this formula@on we do not account for the conceptof “similar direc'ons”
25
![Page 53: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/53.jpg)
LR minimiza'on: issues• With this formula@on we do not account for the conceptof “similar direc'ons”
• “Sparsity term” and “low-‐rank term” are in compe''on!
25
voxels
direcBon
s
correla'onbetweenvoxels
in each voxeldirecBons are sparse
![Page 54: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/54.jpg)
Summary(1) Diffusion MRI is sparse!
(2) Re-‐weighted l1 improves the quality of reconstruc@ons-‐ State-‐of-‐the-‐art: DSI = 257 samples
L2L1 = 30 samplesWe achieve the same quality with 15 samples
(3) Is low-‐rank minimiza'on the key to go beyond?-‐ It is very effec've, but only in very simplis@c cases.
In more realis@c seings, it tends to over-‐simplify the structure
-‐ Sparsity is “compe'ng” with correla@on
-‐ Needs to be inves@gated further
26
![Page 55: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/55.jpg)
Summary(1) Diffusion MRI is sparse!
(2) Re-‐weighted l1 improves the quality of reconstruc@ons-‐ State-‐of-‐the-‐art: DSI = 257 samples
L2L1 = 30 samplesWe achieve the same quality with 15 samples
(3) Is low-‐rank minimiza'on the key to go beyond?-‐ It is very effec've, but only in very simplis@c cases.
In more realis@c seings, it tends to over-‐simplify the structure
-‐ Sparsity is “compe'ng” with correla@on
-‐ Needs to be inves@gated further
26
![Page 56: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/56.jpg)
Summary(1) Diffusion MRI is sparse!
(2) Re-‐weighted l1 improves the quality of reconstruc@ons-‐ State-‐of-‐the-‐art: DSI = 257 samples
L2L1 = 30 samplesWe achieve the same quality with 15 samples
26
DSI257 samples
CSD60 samples
L2L130 samples
L2L1 re-‐weighted15 samples
![Page 57: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/57.jpg)
Summary(1) Diffusion MRI is sparse!
(2) Re-‐weighted l1 improves the quality of reconstruc@ons-‐ State-‐of-‐the-‐art: DSI = 257 samples
L2L1 = 30 samplesWe achieve the same quality with 15 samples
(3) Is low-‐rank minimiza'on the key to go beyond?-‐ It is very effec've, but only in very simplis@c cases.
In more realis@c seings, it tends to over-‐simplify the structure
-‐ Sparsity is “compe'ng” with correla@on
-‐ Needs to be inves@gated further
26
DSI257 samples
CSD60 samples
L2L130 samples
L2L1 re-‐weighted15 samples
Low-‐rank? samples
![Page 58: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/58.jpg)
Data reconstrucBonYves Wiaux and me
Fiber-‐tracking and structural connecBvityAlia Lemkaddem
RelaBonship between structural-‐funcBonal networksAlessandra Griffa
StaBsBcal analysis of connecBvity graphsDjalel Meskaldji
Clinical studiesElda Fischi-‐Gomez
27
Alia Djalel Elda Alessandra Prof. Jean-‐Philippe Thiran(the boss)
Lausanne “diffusion group”
![Page 59: Advanced image-processing techniques in Magnetic ... - Univr](https://reader033.vdocuments.us/reader033/viewer/2022051403/627cf7aaea6e0a749806f4dc/html5/thumbnails/59.jpg)
Suggestions?
Questions?
Comments?