10/06/20041 Resolution Enhancement in MRI By: Eyal Carmi Joint work with: Siyuan Liu, Noga Alon, Amos Fiat & Daniel Fiat.

10/06/2004 1

Resolution Enhancementin MRI

By: Eyal Carmi

Joint work with:

Siyuan Liu, Noga Alon,

Amos Fiat & Daniel Fiat

2

Lecture Outline

Introduction to MRIThe SRR problem (Camera & MRI)Our Resolution Enhancement AlgorithmResultsOpen Problems

3

Introduction to MRI

Magnetic resonance imaging (MRI) is an imaging technique used primarily in medical settings to produce high quality images of the inside of the human body.

4

Introduction to MRI

The nucleus of an atom spins, or precesses, on an axis.

Hydrogen atoms – has a single proton and a large magnetic moment.

5

Magnetic Resonance Imaging

Uniform Static Magnetic Field –

Atoms will line up with the direction of the magnetic field.

0 0

0

0

w B

B MagneticField

GyromagneticRatio

w LarmorFrequency

6

Magnetic Resonance Imaging

Resonance – A state of phase coherence among the spins.

Applying RF pulse at

Larmor frequency When the RF is turned off

the excess energy is released

and picked up.

7

Magnetic Resonance Imaging

Gradient Magnetic Fields –

Time varying magnetic fields

(Used for signal localization)

x

y

z

8

Magnetic Resonance Imaging

Gradient Magnetic Fields: 1-D

X

Y

B0

B1

B2

B3

B4

0 0

0

0

w B

B MagneticField

GyromagneticRatio

w LarmorFrequency

9

Signal Localization slice selection

Gradient Magnetic Fields for slice selection

z

ω

z1 z2 z3

0B

Gz,1

Gz,2

z4

ω1

ω2FT

B1(t)

10

Signal Localizationfrequency encoding

Gradient Magnetic Fields for in-plane encoding

t

x

B0B=B0+Gx(t)x

t

Gx(t)

11

Signal Localizationphase encoding

Gradient Magnetic Fields for in-plane encoding

t

x

B0B=B0+Gx(t)x

t

Gx(t)

12

k-space interpretationy

x

Wx

Wy

1-D path

k-space

Δkx

Sampling Points

DFT

2i kr

Object

S k r e dr

13

Magnetic Resonance Imaging

Collected

Data

(k-space)

2-D DFT

14

The Super Resolution Problem

Definition:

SRR (Super Resolution Reconstruction): The process of combining several low resolution images to create a high-resolution image.

15

SRR – Imagery Model

The imagery process model:

Yk – K-th low resolution input image.

Gk – Geometric trans. operator for the k-th image.

Bk – Blur operator of the k-th image.

Dk – Decimation operator for the k-th image.

Ek – White Additive Noise.

NkEXGBDY kkkkk 1

16

SRR – Main Approaches

Frequency Domain techniques

Tsai & Huang [1984]

Kim [1990]

Frequency Domain

17

SRR – Main Approaches

Iterative Algorithms

Irani & Peleg [1993] : Iterative Back Projection

Current HR Best Guess

Back Projection

Back Projected LR images

Original LR images

Iterative Refinement

18

SRR – Main Approaches

Patti, Sezan & Teklap [1994]

POCS:

Elad & Feuer [1996 & 1997]

ML:

MAP:

POCS & ML

EXHY

2XHY

XYPXmaxarg

XPXYPX

maxarg

19

SRR – In MRI

Peled & Yeshurun [2000]

2-D SRR, IBP, single FOV,

problems with sub-pixel shifts.

Greenspan, Oz, Kiryati and Peled [2002] 3-D SRR (slice-select direction), IBP.

20

Resolution Enhancement Alg.

A Model for the problemReconstruction using boundary values1-D Algorithm2-D Algorithm

21

Modeling The Problem

Subject Area: 1x1 rectilinearly aligned square grid.

(0,0)

22

Modeling The Problem

True Image : A matrix of real values associated with a rectilinearly aligned grid of arbitrary high resolution.

23895

96644

53492

61476

88135(0,0)

23

Modeling The Problem

A Scan of the image:

Pixel resolution –

Offset –

),,(),( yxmFjiS mm

True Image

48

1632

1122

1122

4433

4433),( yx

),( yx

Image Scan, m=2

24

Modeling The Problem

Definitions:

Maximal resolution –

Pixel resolution =

Maximal Offset resolution – We can perform scans at offsets where

with pixel resolution

n1

n mmm ,

n

),( yx

3,2,1 ,, kkyx

Zccm ,1

25

Modeling The Problem

Goal: Compute an image of the subject area with pixel resolution while the maximal measured pixel resolution is

11 δn, nn 1

Errors:

1. Errors ~ Pixel Size & Coefficients

2. Immune to Local Errors =>

localized errors should have localized effect.

26

Multiple offsets of a single resolution scan?

2x2

27

Multiple offsets of a single resolution scan?

3x3

28

Multiple offsets of a single resolution scan?

4x2

29

Using boundary value conditions Assumption: or

Reconstruct using multiple scans with the same pixel resolution. Introduce a variable for each HR pixel of physical dimension . Algorithm: Perform c2 scans at all offsets & Solve linear equations (Gaussian

elimination).

mm

0000

000

CCC

CCC

0 C

30

Using boundary value conditionsExample: 4 Scans, PD=2x2

Second sample

31

Using boundary value conditionsExample: 4 Scans, PD=2x2

000

0

32

Using boundary value conditionsExample: 4 Scans, PD=2x2

33

Problems using boundary valuesExample: 4 Scans, PD=2x2

)0,0(

34

Problems using boundary valuesExample: 4 Scans, PD=2x2

-)1-,1(

35

Problems using boundary valuesExample: 4 Scans, PD=2x2

Add more information

•Solve using LS

•Propagation

problem

36

Demands On the algorithm No assumptions on the values

of the true image.

Over determined set of equations Use LS to reduce errors:

Error propagation will be localized.

????

????

???

???

2min bAx

lkl bklA &, where

A x b

37

The One dimensional algorithm Input: Pixel of dimension Notation:

gcd(x,y) – greatest common divisor of x & y.

(Extended Euclidean Algorithm)

, . gcd( , )a b ax by x y

Nyxyx , 1 & 1

ixiixv offset at pixel 1 theof value- ),,(

,a y b x

38

The One dimensional algorithm

(Extended Euclidean Algorithm)

, . gcd( , )a b ax by x y

Given all ( , ) & ( , )

We can compute all (gcd( , ), )

v x i v y i

v x y i

gcd( , ) 1 (1, )x y v i

39

One dimensional reconstruction

40

One dimensional reconstruction

41

The One dimensional algorithmAlgorithm: Given

w.l.g, let: a>0 & b<0

To compute , compute:

( , ),0

( , ),0

v x j j m x

v y i j m x

(gcd( , ), )v x y i

11

0 0

, ,ba

j j

v x i xj v y i yj

42

The One dimensional algorithmLocalized Reconstruction

1, 0, 0ax by a b

1 ( mod )ax y

Localized ReconstructionEffective Area: x+y high-resolution pixels

43

One dimensional reconstruction

44

Two and More Dimensions Given pixels of size:

Where x,y & z are relatively prime.Reconstruct 1x1 pixels.

Error Propagation is limited to an area of O(xyz) HR pixels.

, & x x y y z z

x x

y y

xy x

xy y

1xy

1-D

AlgorithmStack

45

Two and More Dimensions1

1

xy

xz

1x

1-D

Algorithm

gcd(xy,xz)=x

1

1

xy

zy

1y

gcd(xy,zy)=y

1

1

x

y

1 1

46

Example Two dimensional reconstruction

PD=5x5 PD=3x3 PD=15x1

47

Example Two dimensional reconstruction

PD=4x4 PD=3x3 PD=12x1

48

Example Two dimensional reconstruction

PD=5x5 PD=4x4 PD=20x1

49

Example Two dimensional reconstruction

50

Larger Dimensions

Generalize to Dimension k Using

k+1 relatively prime Low-Resolution pixels

51

Results

Model ResultsExperiment ResultsProblems…

52

Model Design

HR Scene Blur Sampling

LRImage

Noise

LRImage

LRImage

SRR

AlgorithmHR

Image

53

Results

54

Experiment

GE clinical 1.5T MRI

scanner was used. Phantom:

- plastic frames

- filled with water Three FOV: 230.4,

307.2 & 384 mm.

55

Experiment

56

Experiment Results

Modeled DataMRI Data

57

Experiment Results

Modeled DataMRI Data

58

Problems

Homogeneity of the phantomPhantom Orientation

59

Problems

Homogeneity of the phantomPhantom OrientationRectangular Blur Vs. Gaussian-like Blur

60

Problems

Homogeneity of the phantomPhantom OrientationRectangular Blur Vs. Gauss-like Blur

Truncated

61

Problems

Homogeneity of the phantomPhantom OrientationRectangular Blur Vs. Gauss-like Blur k-space and Fourier based MRI

62

k-space and Fourier based MRIy

x

Wx

Wy

1-D path

k-space

Δkx

Sampling Points

DFT

2i kr

Object

S k r e dr

Δkx

63

Problems

# SamplesNoiseProblem

InfiniteNoOne scan is enough

InfiniteYesSNR too low

FiniteYesNo perfect reconstruction

Apply manual shifts → Different experiment

64

Open Problems

Optimization problems: “What is the smallest number of scans we can do to reconstruct the high resolution image?“

65

Scan Selection Problem

66

Scan Selection ProblemS5x and…

67

Scan Selection ProblemS3x and…

68

Open Problems

Optimization problems: “What is the smallest number of scans we can do to reconstruct the high resolution image?“

Decision/Optimization problem: Given a set of scans, what can we reconstruct?

Design problem: Plan a set of scans for “good” error localization.

69

Questions