El 14 de noviembre de 2016, la Fundación Ramón Areces organizó un Simposio Internacional sobre tecnología aplicada al mundo de la medicina de la mano del Instituto Tecnológico de Massachusetts (MIT) y de la Fundación mVision. Este encuentro llevó por título 'La visión computacional se encuentra con la medicina'. Durante esta jornada, se analizó el impacto que están teniendo las nuevas técnicas de imagen en alta resolución para el diagnóstico de todo tipo de enfermedades.
Emanuele Schiavi-'La visión computacional se encuentra con la medicina'
1. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Advanced Mathematical Methods for MR Images
Processing and Analysis
Emanuele Schiavi
Computer Vision Image Processing Group (CVIP)
1Dpto de Matem´atica Aplicada, Ciencias de los Materiales y Tecnolog´ıa
Electr´onica.
Universidad Rey Juan Carlos, M´ostoles, Madrdid, Spain
In collaboration with the Research Laboratory of Electronics, Massachusetts
Institute of Technology and the A. A. Martinos Center for Biomedical Imaging,
Massachusetts General Hospital
Computer Vision meets Medicine:
Present and future of imaging modalities and biomarkers
Madrid, November 14th, 2016
Emanuele Schiavi URJC Advanced Mathematical Methods
2. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Collaborators
Emanuele Schiavi URJC Advanced Mathematical Methods
3. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Supporting Institutions
TEC2012-39095-C03-02 Biomarkers Based on Mathematical Models
Emanuele Schiavi URJC Advanced Mathematical Methods
4. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Projects and Contracts
2015 -TEC2012-39095-C03-02 Ref. Int. M1010 T´ıtulo del
proyecto: Biomarcadores Basados en Modelos Matem´aticos.
Proyecto de I+D. IP: E. Schiavi. (URJC). Equipo: 12
miembros Entidad Financiadora: Ministerio de Econom´ıa y
Competitividad. Desde 01/01/2013 hasta el 31/12/2015.
Programa Nacional. 3 a˜nos
2014 - Ref. Int. M1294 T´ıtulo del Proyecto: Reconstrucci´on de
Im´agenes de MRI. MRI Reconstruction. IP: E. Schiavi. (URJC)
Art´ıculo 83. Contrato de I+D. Equipo: E. Schiavi Entidad
Financiadora: Massachussets Institute of Technology.
Duraci´on: 6 meses. Desde 01/01/2015 hasta el 30/06/2015.
Emanuele Schiavi URJC Advanced Mathematical Methods
5. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Projects and Contracts
2014 - Ref. Int. M1088 T´ıtulo del Proyecto: Improving the
Safety and Efficiency of High-Fields MR Imaging. Subaward
no. 5710003517 IP: E. Schiavi. (URJC) Art´ıculo 83. Contrato
de I+D. Equipo: E. Schiavi Entidad Financiadora:
Massachussets Institute of Technology. Duraci´on: 1 a˜no y 6
meses. Desde 1/7/2013 hasta el 30/06/2014.
2013 - Ref. Int. M1019 T´ıtulo del Proyecto: Improving the
Safety and Efficiency of High-Fields MR Imaging. IP: E.
Schiavi. (URJC) Art´ıculo 83. Contrato de I+D. Equipo: E.
Schiavi Entidad Financiadora :Massachussets Institute of
Technology. Duraci´on: 1 a˜no y 6 meses. Desde 1/1/2012
hasta el 30/06/2013.
Emanuele Schiavi URJC Advanced Mathematical Methods
6. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Papers
Chatnuntawech, I., Martin, A., Bilgic, B., Setsompop, K.,
Adalsteinsson, E., Schiavi, E. Vectorial Total Generalized
Variation for Accelerated Multi-Channel Multi-Contrast MRI.
Magnetic Resonance Imaging, June 2016. Impact Factor:
1.980
Martin, A., Schiavi, E., Eryaman, Y., Herraiz, J. L., Gagoski,
B., Adalsteinsson, E.,Wald, L. L., Guerin, B. Parallel
Transmission Pulse Design with Explicit Control for the Specic
Absorption Rate in the Presence of Radiofrequency Errors.
Magnetic Resonance in Medicine, 73(5):1896-1903. Impact
Factor: 3.571
Emanuele Schiavi URJC Advanced Mathematical Methods
7. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Papers
Eryaman, Y., Guerin, B., Akgun, C., Herraiz, J. L., Martin, A.,
Torrado- Carvajal, A., Malpica, N., Hernandez-Tamames, J.
A., Schiavi, E., Adalsteinsson, E., Wald, L. L.: Parallel
transmit pulse design for patients with Deep brain stimulation
implants. Magnetic Resonance in Medicine, July 3.
doi:10.1002/mrm.25820 Impact Factor: 3.571
2015 Eryaman, Y., Guerin, B., Keil, B., Mareyam, A., Herraiz,
J. L., Kosior, R. K., Martin, A., Torrado-Carvajal, A., Malpica,
N., Hernandez-Tamames, J. A., Schiavi, E., Adalsteinsson, E.,
Wald, L. L.: SAR reduction in 7T C-spine imaging using a
”dark modes” transmit array strategy.
Magnetic Resonance in Medicine, 73(4):1533-1539. Impact
Factor: 3.571
Emanuele Schiavi URJC Advanced Mathematical Methods
8. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
PhD
Iv´an Ram´ırez. Deep Variational Models for HAR. PhD student.
Directores: J. Pantrigo, E. Schiavi. URJC, Madrid.
Eduardo Alca´ın. GPU accelerated Variational Models. PhD
student. Directores: A. Sanz, E. Schiavi. URJC, Madrid.
Adri´an Mart´ın, adrian.martin@uni-graz.at, Institute for
Mathematics and Scientific Computing University of Graz,
Austria. Nonlinear optimization Methods for Accelerating
Magnetic Resonance Imaging. Director: E. Schiavi. URJC,
Madrid, 2016
Juan Francisco Garamendi. A Unified Variational Framework
for Image Segmentation and Denoising. Directores: E. Schiavi,
N. Malpica. URJC, Madrid, 2011.
Emanuele Schiavi URJC Advanced Mathematical Methods
9. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
References
Gonzalo Galiano, Emanuele Schiavi, Juli´an Velasco
Well-posedness of a nonlinear integro-differential problem and
its rearranged formulation Nonlinear Analysis Pages 74-90.
2016
Adri´an Mart´ın, Emanuele Schiavi, Sergio Segura de Le´on. On
1-Laplacian Elliptic Equations Modeling Magnetic Resonance
Image Rician Denoising Journal of Mathematical Imaging and
Vision pp 1-23. 2016
A Martin, I Chatnuntawech, B Bilgic, K Setsompop, E
Adalsteinsson, E. Schiavi Total Generalized Variation Based
Joint Multi-Contrast, Parallel Imaging Reconstruction of
Undersampled k-space Data Proc. Intl. Soc. Mag. Reson. Med
23, 0080. 2015
Emanuele Schiavi URJC Advanced Mathematical Methods
10. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
References
Adri´an Mart´ın, Emanuele Schiavi Noise Modelling in Parallel
Magnetic Resonance Imaging: A Variational Approach Image
Analysis and Recognition Volume 8814. Lecture Notes in
Computer Science pp 121-128. 2014
Adri´an Mart´ın, Emanuele Schiavi Automatic Total Generalized
Variation-Based DTI Rician Denoising Image Analysis and
Recognition Volume 7950. Lecture Notes in Computer Science
pp 581-588. 2013
A Martin, A Marquina, JA Hernandez-Tamames, P
Garcia-Polo, E Schiavi MRI TGV based super-resolution
ISMRM 21st Annual Meeting, Salt Lake City, 20-26. 2013
Emanuele Schiavi URJC Advanced Mathematical Methods
11. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Magnetic Resonance Imaging
Figure: Fundaci´on CIEN-Hospital Fundaci´on Reina Sof´ıa, Lab. de
Neuroimagen, Madrid, Spain
Emanuele Schiavi URJC Advanced Mathematical Methods
12. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Motivation
MRI is one of the most widely used medical imaging techniques.
Powerful diagnosis properties without using ionizing radiation.
Slow compared with Ultrasonography or X-ray Computer
Tomography (CT).
Patient discomfort, motion during the acquisition...
Very expensive technique.
Emanuele Schiavi URJC Advanced Mathematical Methods
13. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Outline
MRI data acquisition .
Frequency Gaussian Noise and Rician Spatial Noise.
The Rician probability density function
Bayesian modelling for the Energy Functional
TGV based regularization and edge preserving
Super-resolution for MRI Rician Denoising.
VTGV Based Multi-Coils Multi-Contrast for pMRI
Reconstruction
Non-Smooth Non-Local Non-Convex Saliency detection for
Brain Tumor Segmentation
Emanuele Schiavi URJC Advanced Mathematical Methods
14. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MRI data acquisition
1 Acquisition in K-space (frequency domain) (C)
2 Inverse Fourier transform from K-space (C) to the spatial
domain (C)
3 Compute the magnitude image
Emanuele Schiavi URJC Advanced Mathematical Methods
15. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
The Rudin, Osher and Fatemi (ROF) Model
TV based Gaussian Denoising
J(u) =
Ω
| u|dx
Regularity of u
+
1
2λ Ω
|f − u|2
dx
Fidelity to f
L. Rudin, S. Osher and E. Fatemi. Nonlinear total variation
based noise removal algorithms. Physica D: Nonlinear
Phenomena. 60 (1992). pp 259-268.
A. Chambolle. An algorithm for total variation regularization
and denoising. Journal Math. Imaging. 20. (2004). pp 89-97.
F. Andreu-Vaillo, V. Caselles, J. M. Maz´on Parabolic
Quasilinear Equations Minimizing Linear Growth Functionals
(Progress in Mathematics) 2004. Birkh¨auser Basel
Emanuele Schiavi URJC Advanced Mathematical Methods
16. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
From Gaussian noise to Rician noise
1 K-space (C): MK = ( ˆDR + ˆN1(0, σ))
Real part
+i ( ˆDI + ˆN2(0, σ))
Imaginary part
2 Spatial domain (C):
MS = (DR + N1(0, σ)) + i(DI + N2(0, σ))
3 Magnitude image:
MR = ((DR + N1(0, σ))2 + (DI + N2(0, σ))2
H. Gudbjartsson, S. Patz. The Rician Distribution of Noisy
MRI Data. Magn. Reson. Med. ,34 (1995), pp. 910-914.
Emanuele Schiavi URJC Advanced Mathematical Methods
17. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Bayesian Modelling
Rician probability distribution function
p(u|f )p(f ) = p(f |u)p(u)
p(u|f ) ∝ p(f |u)p(u)
Maximizing p(u|f ) amounts to minimizing the (minus)
log-likelihood
arg min
u
{− log p(f |u) − log p(u)}
p(f |u) = exp(−H(u, f )) = exp −
Ω
h(u, f )dx ,
Emanuele Schiavi URJC Advanced Mathematical Methods
18. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Modelling Rician noise
Rician probability density function
p(f |u) =
f
σ2
exp −
u2 + f 2
2σ2
I0
uf
σ2
u : true image intensity
f : noisy image data
I0 : the modified zeroth-order Bessel function of the first kind
σ2: variance of the original noise
Basu, S., Fletcher, T., and Whitaker, R. Rician noise removal
in diffusion tensor MRI. (2006). MICCAI 2006 (pp. 117-125).
A. Mart´ın, J.F. Garamendi, E. Schiavi. Iterated Rician
Denoising. IPCV’11. Vol. I pp 959-963, 2011
Emanuele Schiavi URJC Advanced Mathematical Methods
19. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Bayesian Modelling
Rician likelihood
H(u, f ) =
Ω
h(u, f )dx =
Ω
u2
2σ2
− log I0
uf
σ2
dx
Energy minimization functional
arg min
u∈X
E(u) = arg min
u∈X
λJ(u) + H(u, f ) = (1)
= arg min
u∈X
λ
Ω
j(u)dx +
Ω
h(u, f )dx =
= arg min
u∈X
λ
Ω
j(u)dx +
Ω
u2
2σ2
− log I0
uf
σ2
dx
Emanuele Schiavi URJC Advanced Mathematical Methods
20. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Total Variation (TV)
Definition
TV [u] = |Du|(Ω) =
Ω
|Du| = sup
¯p∈P Ω
u · ¯pdx
with
¯p = (px1
, .., pxD
) ∈ C1
0 (Ω)D
, |¯p|L∞(Ω) ≤ 1
Notice that for u ∈ W 1,1(Ω):
|Du|(Ω) =
Ω
|Du| =
Ω
| u|dx
Rudin, L. I., Osher, S., Fatemi, E. Nonlinear total variation
based noise removal algorithms Physica D: Nonlinear
Phenomena, 60(1), 259-268.
Emanuele Schiavi URJC Advanced Mathematical Methods
21. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Total Generalized Variation (TGV)
Definition
Let Ω ⊂ Rd , u ∈ L1
loc(Ω, R), α = (α0, ..., αk−1) > 0 weights, the
TGV functional of order k ∈ N is defined as,
TGVk
α(u) = sup
Ω
u (divk
v) dx |v ∈ Ck
c (Ω, Symk
(Rd
)),
divl
v ∞ ≤ αl , l = 0, ..., k − 1
Bredies, K., Kunisch, K., Pock, T. Total generalized variation.
(2010). SIAM Journal on Imaging Sciences, 3(3), 492-526
TGVk
α is proper, convex, lower semi-continuous.
TGVk
α is translation and rotation invariant.
ker(TGVk
α) = Pk−1(Ω) polynomials of degree less than k.
Emanuele Schiavi URJC Advanced Mathematical Methods
22. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
TGV vs TV
Emanuele Schiavi URJC Advanced Mathematical Methods
24. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Noisy low-resolution MR modalities: FA-DTI...
Figure: Diffusion Weighted Images (DWI) and Fractional Anisotropy (FA)
Emanuele Schiavi URJC Advanced Mathematical Methods
25. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Edge-preserving Up (Down)-sampling operator
The Super-resolution model
p(f |Du) =
f
σ2
exp −
(Du)2 + f 2
2σ2
I0
(Du)f
σ2
Low-Resolution data f and High-Resolution image intensity u
Low-Resolution image intensity Du being D a down-sampling
operator and S its Up-sampling operator. D ◦ S = Id ,
S ◦ D = Id .
Edge preserving piecewise linear approximation
Sjk(x, y) = ujk + a(x − xj ) + b(y − yk)
Joshi, S. H., Marquina, A., Osher, S. MRI Resolution
Enhancement Using Total Variation Regularization. ISBI 2009,
pp. 161–164.
Emanuele Schiavi URJC Advanced Mathematical Methods
26. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Bayesian Modelling
Rician p.d.f.
p(u|f )p(f ) = p(f |Du)p(u)
p(u|f ) ∝ p(f |Du)p(u)
Maximizing p(u|f ) amounts to minimizing the (minus)
log-likelihood
arg min
u
{− log p(f |Du) − log p(u)}
p(f |Du) = exp(−λH(Du, f )) = exp −λ
Ω
h(Du, f )dx ,
Emanuele Schiavi URJC Advanced Mathematical Methods
27. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Bayesian Modelling
Rician likelihood
H(Du, f ) =
Ω
h(Du, f )dx =
Ω
(Du)2
2σ2
− log I0
(Du)f
σ2
dx
The TGV prior for Super-resolution
p(u) = exp(−J(u)) = exp −
Ω
j(u)dx ,
J(u) = − log p(u) = TGV2
α(u) = sup
Ω
u div2
v dx, v ∈ K
K = v ∈ C2
c (Ω, Sym2
(Rn
)), v ∞ ≤ α0, div v ∞ ≤ α1
BGV 2
α(Ω) = {u ∈ L1
(Ω) / TGV2
α(u) < ∞}
Emanuele Schiavi URJC Advanced Mathematical Methods
28. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
The Total Generalized Variation for MR-DTI
Super-resolution
Energy minimization functional
min
u∈X
E(u) = min
u∈X
J(u) + λH(Du, f ) = (2)
= min
u∈X Ω
j(u)dx + λ
Ω
h(Du, f )dx =
= min
u∈X Ω
j(u)dx + λ
Ω
(Du)2
2σ2
− log I0
(Du)f
σ2
dx
A Mart´ın, A Marquina, JA Hernandez-Tamames, P
Garcia-Polo, E Schiavi MRI TGV based super-resolution
ISMRM 21st Annual Meeting, Salt Lake City, 20-26. 2013
Emanuele Schiavi URJC Advanced Mathematical Methods
29. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
References for TGV, 2010
Bredies, K., Kunisch, K., Pock, T. Total generalized variation.
(2010). SIAM Journal on Imaging Sciences, 3(3), 492-526
Knoll, F., Bredies, K., Pock, T., Stollberger, R. Second order
total generalized variation (TGV) for MRI (2011). Magnetic
resonance in medicine 65(2), 480-491
Bredies, K., and Valkonen, T. Inverse problems with
second-order total generalized variation constraints. (2011).
Proceedings of SampTA, 1-4.
Emanuele Schiavi URJC Advanced Mathematical Methods
30. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Energy minimization
TGV prior: equivalent formulation
TGV2
α(u) = min
v∈BD(Ω)
α1|| u − v||M + α0||E(v)||M
BD(Ω) = {v ∈ L1
(Ω) / ||E(v)||M < ∞}
Energy Minimization Problem
min
v∈BD(Ω)
α1|| u − v||M + α0||E(v)||M+
+
Ω
(Ku)2
2σ2
− log I0
(Ku)f
σ2
dx
Emanuele Schiavi URJC Advanced Mathematical Methods
31. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Fractional Anisotropy images
Figure: High resolution FA images: NN interpolation Vs. Model Proposed
Emanuele Schiavi URJC Advanced Mathematical Methods
32. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Fractional Anisotropy images
Emanuele Schiavi URJC Advanced Mathematical Methods
33. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Fractional Anisotropy images
Emanuele Schiavi URJC Advanced Mathematical Methods
34. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Diffusion Tensor images
Emanuele Schiavi URJC Advanced Mathematical Methods
35. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Diffusion Tensor images
Emanuele Schiavi URJC Advanced Mathematical Methods
36. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Diffusion Tensor images
Emanuele Schiavi URJC Advanced Mathematical Methods
37. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Numerical Results with Diffusion Tensor images
Emanuele Schiavi URJC Advanced Mathematical Methods
38. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Emanuele Schiavi URJC Advanced Mathematical Methods
39. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Multi Contrast Imaging
Clinical MRI protocols typically include multiple acquisition of the
same ROI with different contrast settings.
Emanuele Schiavi URJC Advanced Mathematical Methods
40. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Fully-Sampled Acquistion
Fu = g
Emanuele Schiavi URJC Advanced Mathematical Methods
41. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Parallel Imaging: Using multiple receiver coils
Acquisition time can be reduced because of having multiple
samples of the same data.
FScu = gc c = 1, ..., Nc
Figure: S1u, S2u, ..., S8uEmanuele Schiavi URJC Advanced Mathematical Methods
42. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Parallel Imaging: Using multiple receiver coils
FScu = gc c = 1, ..., Nc
Figure: S1, S2, ..., S8
Emanuele Schiavi URJC Advanced Mathematical Methods
43. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Parallel Imaging: Using multiple receiver coils
FScu = gc c = 1, ..., Nc
Figure: g1, g2, ..., g8
Emanuele Schiavi URJC Advanced Mathematical Methods
44. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Parallel Imaging: Using multiple receiver coils
SENSE Reconstruction, 1999
u = min
u
Nc
c=1
FScu − gc
2
2
K.P. Pruessmann, M. Weiger, M. B. Scheidegger, P. Boesiger
SENSE: Sensitivity encoding for fast MRI Magn Reson Med.
1999;42(5):952–962
TV-SENSE Reconstruction, 2007
u = min
u
|Du|(Ω) +
Nc
c=1
FScu − gc
2
2
Liu, B., Ying, L., Steckner, M., Xie, J., Sheng, J. Regularized
SENSE reconstruction using iteratively refined total variation
method. 2007. 4th IEEE International Symposium on Biomedical
Imaging: From Nano to Macro (pp. 121-124). IEEE.Emanuele Schiavi URJC Advanced Mathematical Methods
45. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Sparse MRI: Compressed Sensing, 2007
Donoho DL. Compressed Sensing. IEEE Trans Inf Theory.
2006;52:1289–1306.
Candes E, Romberg J, Tao T. Robust uncertainty principles: Exact
signal reconstruction from highly incomplete frequency information.
IEEE Trans Inf Theory. 2006;52(2):489–509
Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of
compressed sensing for rapid MR imaging. Magn Reson Med.
2007;58(6):1182–1195.
Emanuele Schiavi URJC Advanced Mathematical Methods
46. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Sparse MRI: Undersampling below Nyquist rate
Undersampled MRI Reconstruction
u = min
u
Ψu 1 +
λ
2
MFu − g 2
2
Emanuele Schiavi URJC Advanced Mathematical Methods
47. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Multi-contrast MRI: Exploit structural similarities, 2011
Based in CS-concepts or/and PI, several methods have used shared
features across Multi-Contrast MRI to reduce the data acquired
while preserving the reconstruction quality
Bilgic B, Goyal, V, Adalsteinsson, E. Multi-contrast reconstruction
with Bayesian compressed sensing Magn Reson Med.
2011;66(6):1601–1615.
Huang, J., Chen, C., Axel, L. (2014). Fast multi-contrast MRI
reconstruction. Magnetic resonance imaging, 32(10), 1344-1352.
Gong E, Huang F, Ying K, Wu W, Wang S, Yuan C. PROMISE:
parallel-imaging and compressed-sensing reconstruction of
multicontrast imaging using SharablE information. Magn Reson
Med. 2015;73(2):523–35.
Emanuele Schiavi URJC Advanced Mathematical Methods
48. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Vectorial Total Generalized Variation (VTGV)- 2014
The TGV can be naturally extended for vector valued images
Definition
Let Ω ⊂ Rd , u ∈ L1
loc(Ω, RL), α = (α0, ..., αk−1) > 0 weights, the
VTGV functional of order k ∈ N is defined as,
VTGVk
α(u) = sup
Ω
L
l=1
ul (divk
vl ) dx |
v ∈ Ck
c (Ω, Symk
(Rd
)L
), divl
v l,∞ ≤ αl , l = 0, ..., k − 1
Bredies, K Recovering piecewise smooth multichannel images by
minimization of convex Functionals with Total generalized variation
penalty.. Glob Optim Methods LNCS. 2014;8293:44?77.
Emanuele Schiavi URJC Advanced Mathematical Methods
49. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MCMC-TGV-SENSE, 2016
Multi-contrast SENSE Reconstruction using VTGV
min
u∈U
VTGV2
α(u) +
λ
2
Nc
c=1
L
l=1
Ml FScul − gc,l
2
2 (3)
equivalent formulation
VTGV2
α(u) = min
v∈V
α1|| u − v||1V + α0||Ev||1W
A Martin, I Chatnuntawech, B Bilgic, K Setsompop, E
Adalsteinsson, E. Schiavi Total Generalized Variation Based
Joint Multi-Contrast, Parallel Imaging Reconstruction of
Undersampled k-space Data Proc. Intl. Soc. Mag. Reson. Med
23, 0080. 2015
Emanuele Schiavi URJC Advanced Mathematical Methods
51. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MC-TGV-SENSE compared to TV-SENSE: R=5, 8 coils
Emanuele Schiavi URJC Advanced Mathematical Methods
52. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MC-TGV-SENSE compared to TV-SENSE (zoom in)
Emanuele Schiavi URJC Advanced Mathematical Methods
53. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MC-TGV-SENSE compared to TV-SENSE (zoom in)
Emanuele Schiavi URJC Advanced Mathematical Methods
54. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
MC-TGV-SENSE compared to TV-SENSE (zoom in)
Emanuele Schiavi URJC Advanced Mathematical Methods
55. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Non-Smooth Non-Local Non-Convex Model for joint
Saliency Detection and Image Restoration
Non-local Energy Functional
E ,p(u) = αJNL
,p (u) +
1
α
H(u) + λF(u)
JNL
,p (u) =
1
4 Ω×Ω
w(x − y)φ ,p(u(y) − u(x))dxdy
φ ,p(s) =
2
p
s2
+ 2 p/2
−
2
p
p
H(u) = −
Ω
|1 − δu|2
dx, F(u) =
Ω
|u − f |2
dx
I. Ram´ırez, G. Galiano, N. Malpica, E. Schiavi. A Non-local Diffusion
saliency Model in Magnetic Resonance Imaging. Bioimaging, Oporto,
2017.
Emanuele Schiavi URJC Advanced Mathematical Methods
56. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Emanuele Schiavi URJC Advanced Mathematical Methods
57. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Figure: NL diffusion model for p = {0.1, 0.5, 1, 2, 3}.
Emanuele Schiavi URJC Advanced Mathematical Methods
58. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Figure: FLAIR subjects from BRATS2015 dataset. From left to right
(columns): original image, output image, perimeter of the segmentation
of the output image, output binary mask, perimeter of the ground truth,
ground truth segmentation, ground truth and output binary mask overlap
.
Emanuele Schiavi URJC Advanced Mathematical Methods
59. Collaborators Magnetic Resonance Imaging (MRI) Modelling Total Generalized Variation Model Proposed Numerical Resolu
Muchas gracias
Emanuele Schiavi URJC Advanced Mathematical Methods