EE290T: Advanced Reconstruction Methods for Magnetic...

86
EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging Martin Uecker

Transcript of EE290T: Advanced Reconstruction Methods for Magnetic...

Page 1: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

EE290T: Advanced Reconstruction Methods for MagneticResonance Imaging

Martin Uecker

Page 2: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Tentative Syllabus

I 01: Jan 27 Introduction

I 02: Feb 03 Parallel Imaging as Inverse Problem

I 03: Feb 10 Iterative Reconstruction Algorithms

I –: Feb 17 (holiday)

I 04: Feb 24 Non-Cartesian MRI

I –: Mar 03 (cancelled)

I 05: Mar 10 GRAPPA/SPIRiT

I 06: Mar 17 Nonlinear Inverse Reconstruction

I –: Mar 24 (spring recess)

I 08: Mar 31 SAKE/ESPIRiT

I 09: Apr 07 Model-based Reconstruction

I 10: Apr 14 Compressed Sensing

I 11: Apr 21 Compressed Sensing

I 12: Apr 28 Final Project: Presentations

Page 3: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Outline

I Review of last lecture

I Compressed Sensing (and Parallel Imaging)

I IEEE Eta Kappa Nu - Survey

Page 4: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nyquist-Shannon Sampling Theorem

Theorem 1: If a function f (t) contains no frequencies higher thanW cps, it is completely determined by giving its ordinates at aseries of points spaced 1/2W seconds apart.1

I Band-limited function

I Regular sampling

I Linear sinc-interpolation

1. CE Shannon. Communication in the presence of noise. Proc Institute of Radio Engineers; 37:10–21 (1949)

Page 5: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

A Puzzling Numerical Experiment1

Exact recovery of Shepp-Logan phantom from incomplete radialFourier samples:

(Figure: Block et al. 2007)

1. EJ Candes, J Romberg, T Tao. Robust Uncertainty Principles: Exact Signal Reconstruction From HighlyIncomplete Frequency Information. IEEE Trans Inform Theory; 52:489–509 (2006)

Page 6: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging
Page 7: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 8: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 9: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Sparsity

Definition:

I vector x ∈ Rn

I k-sparse: at most k non-zero entries

Example:

Notation: Number of non-zero entries ‖x‖0 (this is not a norm)

Page 10: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Sparsity

x ∈ R2

(0

1.3

) (1.51.3

)

Set of sparse vectors is a (non-convex) union of subspaces

Page 11: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Denoising

I Sparse vector

I Densoising by hard-thresholding

I Densoising by soft-thresholding (shrinkage)

Page 12: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Denoising

I Sparse vector and random noise

I Densoising by hard-thresholding

I Densoising by soft-thresholding (shrinkage)

Page 13: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Denoising

I Sparse vector and random noise

I Densoising by hard-thresholding

I Densoising by soft-thresholding (shrinkage)

Page 14: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Denoising

I Sparse vector and random noise

I Densoising by hard-thresholding

I Densoising by soft-thresholding (shrinkage)

Page 15: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 16: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 17: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Regular Under-Sampling: Point-Spread-Function

I Regular under-sampling in Fourier domain

I Coherent aliasing in the time domain

Point-Spread-Function

Page 18: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Coherent Aliasing

I Regular under-sampling in Fourier domain

I Coherent aliasing in the time domain

1 12 2 33

Signal

Page 19: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Random Sampling: Point-Spread-Function

I Random sampling in Fourier domain

I Incoherent aliasing in time domain

noise-like artifacts

Point-Spread-Function

Page 20: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Incoherent Aliasing

I Random sampling in Fourier domain

I Incoherent aliasing in time domain

Signal

Page 21: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Linear Measurements

Ax = y x ∈ Rn, y ∈ Rm

Measurements: m >= n

Reconstruction: x = A†y

Matrix A should be nearly orthogonal.Example: Fourier Matrix

A AHA

Page 22: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Incoherent Linear Measurements

Ax = y x ∈ Rn and k-sparse, y ∈ Rm

Measurements: k log(n) <= m <= n

Reconstruction: ?

Matrix A should be nearly orthogonal (restricted isometry property)Example: Fourier Matrix with some rows removed A = PF

A AHA

Page 23: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Restricted Isometry Property

A n × p matrix and 1 ≤ s ≤ p

s-restricted isometry property: There is a constant δs such thatfor every s-sparse vector y :

(1− δs)‖y‖22 ≤ ‖Ay‖2

2 ≤ (1 + δs)‖y‖22

Page 24: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 25: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing

Ingredients:

I Sparsity

I Incoherence

I Non-linear reconstruction

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc; 45:600–616 (1997) 2. EJ Candes, J Romberg, T Tao. Robust UncertaintyPrinciples: Exact Signal Reconstruction From Highly Incomplete Frequency Information. IEEE Trans InformTheory; 52:489–509 (2006) 3. DL Donoho. Compressed sensing. IEEE Trans Inform Theory; 52:1289-1306 (2006)

Page 26: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

L1-Norm and Sparsity

(01

)

Set of vectors with ‖x‖0 ≤ 1 not convex! ⇒ L1 instead of L0

Page 27: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

L1-Norm and Sparsity

(01

)

Set of vectors with ‖x‖0 ≤ 1 not convex! ⇒ L1 instead of L0

Page 28: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Linear Reconstruction

L2-regularization:

argminx‖Ax − y‖22 + α‖Wx‖2

2

Explicit Solution: (AHA + αWHW

)−1AHy

Page 29: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Reconstruction

L1-regularization:

argminx‖Ax − y‖22 + α‖Wx‖1

In general: no explicit solution!

Page 30: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

L2-Norm vs L1-Norm

−1

0

1

−1 0 1

‖x‖2 = 1

‖x‖1 = 1

Page 31: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

L1-Norm and Sparsity

Minimize ‖x‖pp subject to Ax = y

Ax = y

‖x‖22

Ax = y

‖x‖11

Page 32: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Inverse Problem with L1-Regularization

Minimize ‖x‖1 subject to ‖Ax − y‖2 ≤ ε

‖Ax − y‖2 = ε

‖x‖11

Page 33: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Linear Reconstruction

x = argminz |z − y |2 + λ|z |2

x =1

1 + λy

0

0−1

1

−11

Page 34: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Soft-Thresholding

x = argminz |z − y |2 + λ|z |

ηλ(x) =

x − λ x > λ0 |x | ≤ λx + λ x < −λ

0

0−1

1

−11

−λ

λ

Page 35: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Joint Thresholding

Shrink magnitude but keep phase/direction

I complex values:

ηλ(x) =

ηλ(|x |) x

|x | x 6= 0

0 x = 0

I vectors:

ηλ(x) =

ηλ(‖x‖2) x

||x ||2 x 6= 0

0 x = 0

Page 36: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Joint Thresholding

Shrink magnitude but keep phase/direction

Page 37: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Iterative Soft-Thresholding (IST)

Landweber1:

xn+1 = xn + µAH(y − Axn)

Iterative Soft-Thresholding2:

zn = xn + µAH(y − Axn)

xn+1 = ηλ(zn)

1. L Landweber. An iteration formula for Fredholm integral equations of the first kind. Amer J Math; 73:615–624(1951) 2. I Daubechies, M Defrise, C De Mol. An iterative thresholding algorithm for linear inverse problems witha sparsity constraint. Comm Pure Appl Math; 57:1413–1457 (2004)

Page 38: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Reconstruction: Iterative Soft-Thresholding

1. Data consistency: zn = xn + µAH(y − Axn)

2. Soft-thresholding: xn+1 = ηλ(zn)

iteration 0

Page 39: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Reconstruction: Iterative Soft-Thresholding

1. Data consistency: zn = xn + µAH(y − Axn)

2. Soft-thresholding: xn+1 = ηλ(zn)

iteration 1

Page 40: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Reconstruction: Iterative Soft-Thresholding

1. Data consistency: zn = xn + µAH(y − Axn)

2. Soft-thresholding: xn+1 = ηλ(zn)

iteration 2

Page 41: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Reconstruction: Iterative Soft-Thresholding

1. Data consistency: zn = xn + µAH(y − Axn)

2. Soft-thresholding: xn+1 = ηλ(zn)

iteration 9

Page 42: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Algorithms

I FOCUSS1

I Iterative Soft-Thresholding (IST)2

I Fast iterative Soft-Thresholding Algorithm3

I Split Bregman4

I Nonlinear Conjugate Gradients

I ...

1. IF Gorodnitsky, BD Rao. Sparse signal reconstruction from limited data using FOCUSS: a re-weighted minimumnorm algorithm. IEEE Trans Sig Proc 45:600–616 (1997) 2. I Daubechies, M Defrise, C De Mol. An iterativethresholding algorithm for linear inverse problems with a sparsity constraint. Comm Pure Appl Math; 57:1413–1457(2004) 3. A Beck, M Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAMJ Imaging Sci; 2:183–202 (2009) 4. T Goldstein and S Osher. The Split Bregman Method for L1-RegularizedProblems. SIAM J Imaging Sci; 2:323–343 (2009)

Page 43: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Statistical Model

Linear measurements contaminated by noise:

y = Ax + n

Gaussian white noise:

p(n) = N (0, σ2) with N (µ, σ2) =1

σ√

2πe−

(x−µ)2

2σ2

Probability of an outcome (measurement) given the image x :

p(y |A, x , λ) = N (Ax , σ2)

Page 44: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Bayesian Prior

I L2-Regularization: Ridge Regression

N (µ, σ2) =1

σ√

2πe−

(x−µ)2

2σ2 Gaussian prior

I L1-Regularization: LASSO

p(x |µ, b) =1

2be−|x−µ|

b Laplacian prior

I L2 and L1: Elastic net1

1. H Zou, T Hastie. Regularization and variable selection via the elastic net. J R Statist Soc B. 67:301–320 (2005)

Page 45: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing in Magnetic Resonance Imaging

I Sparsity in medical imaging

I Incoherent sampling

I Iterative reconstruction

I Combination with parallel imaging

M Lustig, D. Donoho, JM Pauly. Sparse MRI: The application of compressed sensing for rapid MR imaging. MagnReson Med; 58:1182–1195 (2007)

Page 46: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Sparsity Transform

I Medical images are usually not sparse

I Need to apply sparsity transform

Sparsity:

I Wavelet Transform

I Total Variation

I Temporal Constraints

I Prior images

I Adapted dictionaries

I Low rank

I . . .

Page 47: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Wavelet Transform

I Orthonormal basis (or almost)

I Multi-scale transform

I Localized in frequency and space

I Compresses many signals/images into few coefficients

I Efficient computation: O(N)

I But: not shift-invariant (cycle spinning)

L1-regularization term in wavelet domain;

R(x) = ‖Wx‖1

W wavelet transform

Page 48: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Wavelet Transform

brain image wavelet transform

Signal concentrated in few coefficients!

Page 49: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Cycle spinning

Problem: Not shift-invariantSolution: Cycle spinning (or random shifting)

Page 50: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Artifacts

blurring blocky artifacts good quality

Problem: Not shift-invariantSolution: Random shifting (cycle spinning)

Page 51: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation

Definition: For a function f ∈ L1(Ω) with Ω an open set Ω ⊂ Rn,the total variation of f is:

TV f = sup

∫dx f divφ : φ ∈ C 1

c (Ω,Rn), ‖φ‖L∞(Ω) ≤ 1

I Denoising1

I Image Reconstruction2

1. LI Rudin, S Osher, E Fatemi. Nonlinear total variation based noise removal algorithms. Physica D; 60:259–268(1992) 2. D Geman, C Yang. Nonlinear image recovery with half-quadratic regularization. IEEE T ImageProcessing; 4:932–946 (1995)

Page 52: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation

For differentiable function in one variable:

TV f =

∫dx |f ′(x)|

For differentiable function in many variables:

TV f =

∫dx‖∇f (x)‖2

Sparsity of the partial derivatives!

Page 53: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation

f (x)‖∇f (x)‖2 =√

|∂1f (x)|2 + |∂2f (x)|2

Page 54: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation: Approximations

Anisotropic Total Variation:

TV f =

∫dx√|∂1f (x)|2 + |∂2f (x)|2

≈∫

dx |∂1f (x)|+ |∂2f (x)|

Finite differences (backward):

∂1f (x1, x2) ≈ f (x1, x2)− f (x1 − h, x2)

|h|

∂2f (x1, x2) ≈ f (x1, x2)− f (x1, x2 − h)

|h|

h ≈ 1 Pixel

Page 55: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation: Staircase Artifacts

I Staircase artifacts (sparse differences)

I Solution: use of higher-order derivatives1

I Total Generalized Variation2,3

1. D Geman, C Yang. Nonlinear image recovery with half-quadratic regularization. IEEE T Image Processing;4:932–946 (1995) 2. K Bredies, K Kunisch, T Pock. Total generalized variation. SIAM J Imaging Sci; 3:492–526(2010) 3. F Knoll, K Bredies, T Pock, R Stollberger. Second order total generalized variation (TGV) for MRI.Magn Reson Med; 65:480–491 (2011)

Page 56: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Total Variation in Time Domain

f1 f2 f3 f4 f5 f6 f7 f8 f9 f10

t

TVt f ≈∫

dx∑l

|fl(x)− fl−1(x)|

HFL Chandarana, T Block, AB Rosenkrantz, R Lim, D Chu, DK Sodickson, R Otazo. Free-breathing dynamiccontrast-enhanced MRI of the liver with radial golden-angle sampling scheme and advanced compressed-sensingreconstruction. Proc. 20th ISMRM (2012)

Page 57: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

GRASP: Compressed-Sensing Reconstruction

Otazo et al, MRM 2010: 64

Narrow data window Few spokes Flickering streak artifacts

Prior knowledge: Contrast uptake occurs “smoothly” and “continuously”

CS approach: Find solution that

CG SENSE-type reconstruction with temporal Total Variation (TV) constraint

has lowest flickering

matches data in all windows

Ground truth Ground truth13 spokes 13 spokes

Dynamic image series Temporal differences

Courtesy of Tobias Block, NYU

Page 58: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Example: GRASP Liver Imaging

Chandarana et al, ISMRM 2012: 5529

Free-breathing scan over 5 min

Contrast injection after 20 s

Retrospective selection of temporal resolution

Example: 13 spokes 2 s resolution

Enables free-breathing liver perfusion imaging

Here: 384 x 384 x 30 matrix

Spatial resolution 1.0 x 1.0 x 3.0 mm3

Temporal resolution 1.5 s

Top: GriddingBottom: GRASP

Courtesy of Tobias Block, NYU

Page 59: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Difference to Prior Image

Sparse difference to reference image g

R(x) = ‖x − x0‖1

Prior image: x0 composite image, previous frame, ...

1. GH Chen, J Tang, S Leng. Prior image constrained compressed sensing (PICCS): a method to accuratelyreconstruct dynamic CT images from highly undersampled projection data sets. Med Phys; 35:660–663 (2008) 2.A Fischer, F Breuer, M Blaimer, N Seiberlich, PM Jakob. Accelerated dynamic imaging by reconstructing sparsedifferences using compressed sensing. Proc 16th ISMRM (2008)

Page 60: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Dictionary Learning

=

Patch Reconstruction

Dictionary Learning Example-based

...

=a +b +cCourtesy of Patrick Virtue, UC Berkeley

Page 61: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Spatio-temporal Dictionaries

Dictionary based reconstruction of dynamic complex MRI data. Jose Caballero, Anthony Price, Daniel Rueckert,and Joseph V. Hajnal. ISMRM 13

Courtesy of Jose Caballero, Imperial College London

Page 62: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Low-rank Approximation

Data matrix M ∈ Ct×s , e.g. time × space

Singular-Value-Decomposition: M = UΣVH

Low rank: rank M < K

Decomposition into K (temporal and spatial) basis functions:

M =∑K

ukσkvHk

1. Z Bo, JP Haldar. C Brinegarm ZP Liang. Low rank matrix recovery for real-time cardiac MRI. ISBI; 996-999(2010) 2. JP Haldar, ZP Liang. Spatiotemporal imaging with partially separable functions: A matrix recoveryapproach. ISBI 716–719 (2010) 3. SG Lingala, H Yue, E DiBella, M Jacob. Accelerated Dynamic MRI ExploitingSparsity and Low-Rank Structure: k-t SLR. IEEE Trans Med Imag; 30:1042–1054 (2011) 4. R Otazo, E Candes,DK Sodickson. Low-rank and sparse matrix decomposition for accelerated DCE-MRI with background and contrastseparation. ISMRM Workshop on Data Sampling and Image Reconstruction. Sedona (2013)

Page 63: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Low-rank + Sparse Reconstruction of Cardiac Cine

I 6-fold acceleration (ky-t random undersampling)

I Temporal resolution: 40 ms

I Spatial resolution: 1.3x1.3x3 mm3

I Std. CS with temporal FFT

CS L+S L S

Courtesy of Ricardo Otazo, NYU

Page 64: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Parallel MRI

Goal: Reduction of measurement time

I Subsampling of k-space

I Simultaneous acquisition with multiple receive coils

I Coil sensitivities provide spatial information

I Compensation for missing k-space data

1. DK Sodickson, WJ Manning. Simultaneous acquisition of spatial harmonics (SMASH): Fast imaging withradiofrequency coil arrays. Magn Reson Med; 38:591–603 (1997) 2. KP Pruessmann, M Weiger, MB Scheidegger,P Boesiger. SENSE: Sensitivity encoding for fast MRI. Magn Reson Med; 42:952–962 (1999) 3. MA Griswold, PMJakob, RM Heidemann, M Nittka, V Jellus, J Wang, B Kiefer, A Haase. Generalized autocalibrating partiallyparallel acquisitions (GRAPPA). Magn Reson Med; 47:1202–10 (2002)

Page 65: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Parallel MRI: Undersampling

Undersampling Aliasing

kread

kphase

kphase

kpartition

Page 66: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Parallel MRI as Inverse Problem

I Signal from multiple coils (image x , sensitivities cj):

sj(t) =

∫Vd~r x(~r)cj(~r)e−i~r ·

~k(t)

I Assumption: known sensitivities cj⇒ linear relation between image x and data y

I Image reconstruction is a linear inverse problem:

Ax = y

1. JB Ra and CY Rim, Magn Reson Med 30:142–145 (1993) 2. KP Pruessmann, M Weiger, MB Scheidegger, PBoesiger. Magn Reson Med 4:952–962 (1999)

Page 67: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Parallel MRI: Regularization

I General problem: bad condition

I Noise amplification during image reconstruction

I L2 regularization (Tikhonov):

argminx‖Ax − y‖22 + α‖x‖2

2 ⇔ (AHA + αI )x = AHy

I Influence of the regularization parameter α:

small medium large

Page 68: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Parallel MRI: Nonlinear Regularization

I Good noise suppression

I Edge-preserving

⇒ Sparsity, nonlinear regularization

argminx‖Ax − y‖22 + αR(x)

Regularization: R(x) = TV (x), R(x) = ‖Wx‖1, . . .

1. JV Velikina. VAMPIRE: variation minimizing parallel imaging reconstruction. Proc. 13th ISMRM; 2424 (2005)2. G Landi, EL Piccolomini. A total variation regularization strategy in dynamic MRI, Optimization Methods andSoftware; 20:545–558 (2005) 2. B Liu, L Ying, M Steckner, J Xie, J Sheng. Regularized SENSE reconstructionusing iteratively refined total variation method. ISBI; 121-123 (2007) 3. A Raj, G Singh, R Zabih, B Kressler, YWang, N Schuff, M Weiner. Bayesian parallel imaging with edge-preserving priors. Magn Reson Med; 57:8–21(2007) 4. M Uecker, KT Block, J Frahm. Nonlinear Inversion with L1-Wavelet Regularization - Application toAutocalibrated Parallel Imaging. ISMRM 1479 (2008) 5. . . .

Page 69: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Inversion with Non-Quadratic Regularization

Iteratively Regularized Gauss Newton Method (IRGNM)

xn+1 − xn = argminδx‖DFH(xn)δx + F (xn)− y‖22 + αnR(δx + xn)

Previously: Image regularized with L2-norm

R(x) = ‖ρ‖22 + ‖(1 + s|~k|2)lFTcj‖2

2

Now: Different regularization terms

R(x) = R(ρ) + ‖(1 + s|~k |2)lFTcj‖22

Knoll F, Clason C, Bredies K, Uecker M, Stollberger R, Magn Reson Med, 67:34-41 (2012).

Page 70: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Inversion

I Quality of the reconstructed images can be improved

I Acceleration: 3 x 2

I L1-Wavelet: Cohen-Daubechies-Feauveau 9/7

Page 71: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing and Parallel Imaging

I Parallel imaging

I Sparsity, nonlinear regularization

I Incoherent sampling

1. KT Block, M Uecker, J Frahm. Undersampled radial MRI with multiple coils. Iterative image reconstructionusing a total variation constraint. Magn Reson Med; 57:1086–1098 (2007) 2. C Zhao, T Lang, J Ji. CompressedSensing Parallel Imaging. Proc. 16th ISMRM; 1478 (2008) 3. B Wu, RP Millane, R Watts, P Bones. Applyingcompressed sensing in parallel MRI. Proc. 16th ISMRM; 1480 (2008) 4. KF King. Combining compressed sensingand parallel imaging. Proc. 16th ISMRM; 1488 (2008) 5. B Liu, FM Sebert, YM Zou, L Ying. SparseSENSE:randomly-sampled parallel imaging using compressed sensing. Proc. 16th ISMRM; 3154 (2008) 6. B Liu, K King,M Steckner, J Xie, J Sheng, L Ying. Regularized sensitivity encoding (SENSE) reconstruction using bregmaniterations. Magn Reson Med 61:145–152 (2009) 7. . . .

Page 72: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Sampling Schemes

uniform random Poisson-disc

Poisson-disc sampling:

I Minimum distance to exploit parallel imaging

I Incoherence for compressed sensing

M Murphy, K Keutzer, SS Vasanawala, M Lustig. Clinically feasible reconstruction time for L1-SPIRiT parallelimaging and compressed sensing MRI. Proc 18th ISMRM; 4854 (2010)

Page 73: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Variable-Density Sampling

variable-density Poisson-disc radial

Advantages:

I Auto-calibration for parallel imaging

I Graceful degradation

Page 74: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing and Parallel Imaging

Regularized SENSE:

argminx‖PFCx − y‖22 + R(x)

P projection onto samples, F Fourier transform, C coil sensitivities,

R(x) regularization

Simple (but slow): IST with R(x) = ‖Wx‖1

zn = xn + µCHFHP(y − PFCxn)

xn+1 = W−1ηλ(Wzn)

Page 75: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Compressed Sensing and Parallel Imaging

Linear reconstruction:

I R(x) = 0 ⇒ No regularization

I R(x) = ‖x − x0‖22 ⇒ Tikhonov

Nonlinear reconstruction:

I R(x) = TV (x) ⇒ Total Variation

I R(x) = ‖Wx‖1 ⇒ L1-Wavelet

I R(x) = ‖x − x0‖1 ⇒ Prior Image

I · · ·

Incoherent Sampling ⇒ Compressed Sensing

Page 76: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Example: Undersampled Radial with Total Variation

KT Block, M Uecker, J Frahm. Undersampled Radial MRI with Multiple Coils. Iterative Image ReconstructionUsing a Total Variation Constraint. Magn Reson Med 57:1086-1098 (2007)

Page 77: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Nonlinear Inverse Reconstruction with Variational Penalties

Experiments:I Siemens Tim Trio 3 T, 12-channel head coilI 3D FLASH, acceleration: R = 4 (pseudorandom sampling)

Knoll F, Clason C, Bredies K, Uecker M, Stollberger R, Magn Reson Med, 67:34-41 (2012).

Page 78: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

`1-SPIRiT

I Robust auto-calibrating parallel MRI

I Calibration of coil-by-coil operator G in k-space

Optimization problem:

arg minx

α‖Px − y‖22︸ ︷︷ ︸

data consistency

+ β‖(G − Id)x‖22︸ ︷︷ ︸

calibration consistency

+ γR(x)︸ ︷︷ ︸regularization

x estimated k-space, y data, P Projection onto samples, G SPIRiT

operator, R(x) regularization

M Lustig, JM Pauly. SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space. MagnReson Med; 64:457–471 (2010)

Page 79: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

`1-SPIRiT

“POCS-type” algorithm (α = β =∞):

yn = y + (Id − P)xn data consistency

zn = Gyn calibration consistency

xn+1 = W−1ηλ(Wzn) soft-thresholding

Page 80: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

`1-ESPIRiT

ESPIRiT:

I Flexibility and efficiency of SENSE

I Robustness of GRAPPA/SPIRiT

Algorithm:

I Calibration of coil-by-coil operator in k-space

I Sensitivity maps from eigendecomposition

I Extended (“soft”) SENSE reconstruction

M Uecker, P Lai, MJ Murphy, P Virtue, M Elad, JM Pauly, SS Vasanawala, M Lustig. ESPIRiT - An EigenvalueApproach to Autocalibrating Parallel MRI: Where SENSE meets GRAPPA. Magn Reson Med. Epub (2013)

Page 81: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

ESPIRiT: Reconstruction with Multiple Maps

Relaxed (“soft”) SENSE using multiple maps.

I Simultaneous reconstruction of multiple images mj

I Data consistency:

N∑i=1

‖yi − PFM∑j=1

S ji m

j‖22︸ ︷︷ ︸

“soft” SENSE

+ Q(m1, · · · ,mM)

mj images, S j multiplication with maps, F Fourier transform,

P sampling operator, y data, Q regularization

I Image combination (e.g. root of sum of squares)

Page 82: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Example: `1-ESPIRiT

M Uecker, P Lai, MJ Murphy, P Virtue, M Elad, JM Pauly, SS Vasanawala, M Lustig. ESPIRiT - An EigenvalueApproach to Autocalibrating Parallel MRI: Where SENSE meets GRAPPA. Magn Reson Med. Epub (2013)

Page 83: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Conclusion

Compressed Sensing: Theory

I Sparsity

I Incoherent sampling

I Nonlinear reconstruction

Application to Magnetic Resonance Imaging

I Sparsity transform (adapted to application)

I Incoherent sampling schemes

I Efficient iterative reconstruction

I Combination with parallel imaging

Page 84: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Software

http://www.eecs.berkeley.edu/~mlustig/Software.html

http://www.eecs.berkeley.edu/~uecker/toolbox.html

http://web.eecs.umich.edu/~fessler/code/index.html

http://www.imt.tugraz.at/index.php/research/

agile-gpu-image-reconstruction-library

http://gadgetron.sourceforge.net/

http://codeare.org/

http://impact.crhc.illinois.edu/mri.aspx

Page 85: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Acknowledgements

I Michael Lustig, University of California, Berkeley

I Tobias Block, New York University

I Ricardo Otazo, New York University

I Patrick Virtue, University of California, Berkeley

I Jose Caballero, Imperial College London

Page 86: EE290T: Advanced Reconstruction Methods for Magnetic …inst.eecs.berkeley.edu/~ee290t/sp14/lecture10.pdf · EE290T: Advanced Reconstruction Methods for Magnetic Resonance Imaging

Final Projects

I Presentation: ≈ 10 minutes, 28th April

I Report