compressed sensing magnetic resonance imaging usingfourier … · 2020. 8. 13. · compressed...

151
Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non-Fourier Based Bunched Phase Encoding This dissertation is submitted for the degree of Doctor of Philosophy KAZI RAFIQUL ISLAM Principal Supervisor:Jingxin Zhang Co-supervisor:Cishen Zhang School of Software and Electrical Engineering Faculty of Science Engineering & Technology Swinburne University of Technology, John St, Hawthorn, VIC-3122, Australia July 2020

Upload: others

Post on 05-Mar-2021

2 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non-Fourier

Based Bunched Phase Encoding

This dissertation is submitted for the degree of Doctor of Philosophy

KAZI RAFIQUL ISLAM

Principal Supervisor:Jingxin Zhang

Co-supervisor:Cishen Zhang

School of Software and Electrical Engineering

Faculty of Science Engineering & Technology

Swinburne University of Technology, John St, Hawthorn, VIC-3122, Australia

July 2020

Page 2: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

ii

Declaration

I hereby declare that the contents of this thesis are original and have not been submitted in

whole or in part for the award of any other degree, qualification or any other university

except where specific reference is made to the work of others.

Kazi Rafiqul Islam

Page 3: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

iii

Acknowledgements

PhD study has been one of the valuable parts of my education life. During this period, I have

acknowledged help and encouragement from many people and institute. Here, I would like

to remember their name.

First of all, I would like to praise my creator and sustainer Allah Ta’ala for his bounty and

Rahmah. Then, I would like to express my wholehearted thanks to my supervisor Prof.

Jingxin Zhang. Without his proper supervision, I could not finish my PhD study. During my

PhD study, He has always been helpful, resourceful and patient. His continuous

encouragement, kindness and sympathetic helped me a lot throughout my research. Also, I

would like to thanks Prof. Cishen Zhang for his guidance and advice towards me.

IwouldalsoliketothankSwinburneUniversityofTechnologyforprovidingscholarships, research facilities to me so that I could enjoy a peaceful life, and concentrate my time on research. Staff from this institute are always willing to help. I am grateful for their kind support and willingattitude.

I would like to thank my colleagues from Swinburne University and Technology and

Monash University, who helped and cooperate me during my research, especially Kamlesh

Powar.

Finally and most importantly, I would like to thank my family, who encouraged and supported me to obtain my PhD degree. I could not have achieved my degree without their support.

Page 4: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

iv

Abstract

Magnetic resonance imaging (MRI) technology is one of the most significant biomedical

innovations. However, the MRI process speed is still too slow to meet the demand of

biomedical research and clinical diagnosis. Its applications, affordability, and accessibility

remain limited by the time required to sample adequate information of k-space for the

desired field of view (FOV), resolution, and signal-to-noise ratio (SNR). A large number of

methods have been developed over the years to accelerate MR imaging speed.

However, the ultimate targeted acquisition speed with quality is still a big issue that has

inspired active research in recent years on accelerated MRI. Compressed sensing (CS) is an

emerging technique to accelerate conventional MRI by reducing the number of acquired

data. Bunched phase encoding (BPE) is a technique that uses oscillating trajectories to

minimise phase encoding steps, and hence to accelerate data acquisition. These two

techniques have proven their respective advantages standalone, and their combination may

result in further acceleration of MRI data acquisition, but has never been exploited in the

literature.

This research work presents a novel method that combines CS with BPEusing Fourier

encoding to further accelerate data acquisition and preserve image quality. Physically BPE

trajectories are never the same as the theoretically calculated ones due to gradient

imprecision and field inhomogeneity, which results in poor image quality. A fix to this

problem is to measure the coordinates of the trajectories and use them in BPE image

reconstruction. Such measurements generally require prescan calibration that complicates the

process and increases operation cost.

To overcome this difficulty, we present a simple and effective method to estimate the

coordinates of BPE trajectories from normal scan dataand demonstrate its efficacy in image

reconstruction of in vivo scan data acquired from ZIGZAG trajectory.To smooth and remove

the aliasing from the image, a modulation map estimation method has been incorporated in

compressed sensing part of our proposed method.

Page 5: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

v

Non-Fourier encoding like spread spectrum instead of Fourier encoding in CS increases the

level of incoherence between the sparsifying transform matrix and the measurement

(sensing) matrix and hence the quality of reconstruction image. Therefore, a non-

Fourierencoding like Chirp modulation encoding method is used in this workto further

improve the quality of reconstructed image. It is realized by controlling the complex

orthogonality of the quadratic phase encoding. The experiments on the simulated and real-

life MRI data demonstrate that the proposed methods outperform the existing CS MRI and

BPE MRI methods in the literature.

Page 6: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

i

Table of Contents Chapter 1 Introduction .......................................................................................................... 1

1.1 Overview ................................................................................................................ 1

1.2 The Time Limitation of MRI .................................................................................. 3

1.3 Research Objective and Technical Approaches ....................................................... 7

1.4 Main Contributions ................................................................................................. 8

1.5 Thesis Outline ........................................................................................................ 9

Chapter 2 Fundamentals of MRI Process ............................................................................ 11

2.1 Introduction .......................................................................................................... 11

2.2 Theory of MRI ..................................................................................................... 11

2.2.1 Physics of NMR ............................................................................................ 11

2.2.2 Physical Behaviour of Tissue Particle ............................................................ 16

2.3 NMR Signal Excitation......................................................................................... 17

2.3.1 Slice Selection and Gradient Selection ........................................................... 17

2.3.2 Frequency and Phase Encoding ..................................................................... 18

2.3.3 Image Formulation ........................................................................................ 19

2.3.4 The K-space Signal ........................................................................................ 21

2.3.5 2D Spatial Encoding ...................................................................................... 22

2.3.6 Field of View (FOV) and Spatial Resolution ................................................. 23

2.3.7 MR Imaging Pulse Sequence ......................................................................... 25

2.4 Image Reconstruction Technique of MRI ............................................................. 26

2.4.1 Reconstruction Complexity ........................................................................... 26

2.4.2 Cartesian Trajectory ...................................................................................... 27

2.4.3 Non-Cartesian Trajectory .............................................................................. 28

2.4.4 Noise and System Uncertainty in Image Reconstruction ................................ 29

2.5 Image Reconstruction from Subsampled Data in Cartesian Coordinates................ 31

2.5.1 SENSE .......................................................................................................... 31

Page 7: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

ii

2.5.2 Extensions of SENSE .................................................................................... 34

2.5.3 GRAPPA ....................................................................................................... 35

2.5.4 Extensions of GRAPPA ................................................................................. 37

2.5.5 Compressive Sensing based Image Reconstruction Method ........................... 40

2.5.6 Extension of CS-based Image Reconstruction Method ................................... 42

2.6 Novel Sequence Design ........................................................................................ 43

2.6.1 Bunched Phase Encoding .............................................................................. 44

2.6.2 Rotating RF Coil ........................................................................................... 45

2.6.3 Wave-CAIPI ................................................................................................. 45

2.7 Some Limitations of the Present Research ............................................................ 45

2.8 Conclusion ........................................................................................................... 46

Chapter 3 Fundamentals of Sparse Sampling ...................................................................... 48

3.1 Introduction .......................................................................................................... 48

3.2 Theory of Compressive Sensing ........................................................................... 49

3.3 Restricted Isometry in Compression Detection...................................................... 52

3.4 Incoherence in Compressive Sensing .................................................................... 53

3.5 Compressed Sensing in Magnetic Resonance Imaging .......................................... 54

3.6 Application of Compressive Sensing .................................................................... 55

3.6.1 Sparse Error Correction ................................................................................. 55

3.6.2 Linear Regression and Model Selection ......................................................... 56

3.6.3 Group Testing and Data Stream Algorithms .................................................. 57

3.6.4 Coronary Heart Imaging ................................................................................ 58

3.6.5 Brain Imaging ............................................................................................... 58

3.6.6 Rapid 3-D Angiography ................................................................................ 58

3.7 Some Limitations of CSMRI ................................................................................ 59

3.8 Conclusion ........................................................................................................... 60

Chapter 4 Compressive Sensing MRI using Fourier Based Bunched Phase Encoding ......... 61

4.1 Introduction .......................................................................................................... 61

Page 8: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

iii

4.2 BPE and CS Techniques ....................................................................................... 64

4.3 Multichannel CSBPFE ......................................................................................... 74

4.4 Aliasing Coefficient Correction ............................................................................ 76

4.5 Smoothness Enhancing ......................................................................................... 80

4.6 Methods ............................................................................................................... 81

4.7 Simulations and Experimental Results .................................................................. 91

4.7.1 Multichannel CSBPE Simulation ................................................................... 91

4.7.2 Multichannel CSBPE Experiment .................................................................. 97

4.7.3 Phantom Scan Data Experiment................................................................... 100

4.7.4 Brain in vivo Scan Data Experiment ............................................................ 103

4.8 Computation Complexity of CSBPE as Compared with CSMRI. ........................ 103

4.9 Conclusion ......................................................................................................... 104

Chapter 5 Compressive Sensing MRI using Non Fourier Based Bunched Phase Encoding 105

5.1 Introduction ........................................................................................................ 105

5.2 Chirp Modulation based CSBP Encoding ........................................................... 106

5.3 Chirp Modulation of Multichannel CSBPE ......................................................... 110

5.4 Chirp Modulated CSBPE Sampling Pattern ........................................................ 111

5.5 Simulation and Experimental Results ................................................................. 112

5.5.1 Simulation Results ....................................................................................... 114

5.5.2 Experimental Results ................................................................................... 115

5.6 Conclusion ......................................................................................................... 122

Chapter 6 Conclusion and Future work ............................................................................. 123

6.1 Conclusion ......................................................................................................... 123

6.2 Future Work ....................................................................................................... 124

REFERENCES .................................................................................................................. 126

Page 9: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

iv

List of Figures Figure 2.1: Nuclear spin in different condition (a) Spins without magnetic field (b) Spins with the magnetic field ....................................................................................................... 12

Figure 2.2: Net magnetization as a function of time ............................................................ 14

Figure 2.3: Transverse relaxation Mxy as a function of time ................................................ 15

Figure 2.4: FID process in the x-y plan after 900RF pulse ................................................... 16

Figure 2.5: The process of the slice selection by the gradient pulse ..................................... 18

Figure 2.6: The image resolution and FOV ......................................................................... 24

Figure 2.7: The sample spacing ∆kx in k-space and the extent kx,max, relate to the FOV, FOVx and voxel size ∆x, respectively, of the reconstructed image in the x-direction. Similarly, ∆ky and ky,max are connected to FOVy and ∆y. ........................................................................... 25

Figure 2.8: Typical pulse sequence for MRI data acquisition .............................................. 26

Figure 2.9: the trajectory pulse sequence in the non-Cartesian coordinates .......................... 29

Figure 2.10: Comparison between the original and the subsampling images ........................ 32

Figure 2.11: Schematic diagram of GRAPPA reconstruction technique .............................. 36

Figure 2.12: The image reconstruction method of TGRAPPA ............................................. 38

Figure 2.13:The image rebuilding method of k-t GRAPPA ................................................. 39

Figure 2.14:schematic diagram of gradient and BPE encoding technique ............................ 44

Figure 4.1: Sequential steps of CSBPE image reconstruction process.................................. 64

Figure 4.2: Schematic diagram of k-space trajectories of bunched phase encoding. (a) Rectilinear sampling data. (b) zigzag sampling data in the BPE scheme. ............................. 65

Figure 4.3: Generalized 1D bunched phase encoding technique in k-space.......................... 67

Figure 4.4: Schematic diagram of compressive sensing BPE ............................................... 73

Figure 4.5: Baseline data and shifted data in the Zigzag trajectory of Bunched phase encoding. The image from all PE baseline data on left and rest of all images from shifted PE line data and RO line data for all four sets of image. The whole 256 × 1028 BPE data image is showing on below. .......................................................................................................... 78

Figure 4.6: K-space lines on or near the k-space centre for eight individual k-space volume of the matrix represents the same image with the entire k-space shifted by the step size dk(n) and n = [0:7], top: n = 0 and bottom: n = 7. ......................................................................... 80

Figure 4.7: A sample pulse sequence design of bunched phase encoding; (Gse) Gradient in slice direction; (Gpe) Gradient in phase encoding direction; (Gro) Gradient in readout direction. A zigzag gradient is incorporated in the phase encoding direction during readout. The amplitude and period of the oscillatory gradient in this figure are not to scale. ............. 82

Figure 4.8:(a) Inverse Fourier transform of the acquired all 32 channel data gives the aliased image; (b) Inverse Fourier transform of the acquired channel 4 data gives the aliased image; (c) k-space of channel 4. ..................................................................................................... 83

Figure 4.9: Phase encoded data sampling technique; (a) Random sampling; (b) Variable density Random sampling. .................................................................................................. 83

Figure 4.10: BPE k-space data acquisition. Ccalculated (simulated) trajectory which shown in red marks (----); Original (measured) trajectory where data positions are marked as (×’s,

Page 10: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

v

×’s, ×’s, ×’s), baseline (zero shift) data points are marked as ×’s, measured shifted data points are marked as ×’s, ×’s, ×’s. ....................................................................................... 85

Figure 4.11: The sequential steps of Aliasing Coefficient Correction by cross correlation technique ............................................................................................................................ 87

Figure 4.12: Calculated (red) and measured (blue) BPE k-space positions. The data were shifted to be symmetrical around zero PE shift.................................................................... 88

Figure 4.13: BPE image reconstruction (channel 4), phase images, using the synthetic BPE k-space PE shifts, Ks (left) and using the measured and corrected BPE k-space PE shifts, Kc (right). ................................................................................................................................ 89

Figure 4.14: BPE image reconstruction, all channels combined, R = 2, matrix 128 x 128, using the synthetic BPE k-space PE shifts, Ks (left), and using the measured and corrected BPE k-space PE shifts, Kc (right). ....................................................................................... 90

Figure 4.15: The coil Sensitivity maps used in CSBPE simulation. The sensitivity maps are estimated from data acquired on MR scanner ...................................................................... 92

Figure 4.16: The mean relative error versus the acceleration factor in CS-MRI, CSBPE for simulation data. .................................................................................................................. 93

Figure 4.17 and Figure 4.18 show the simulation results of Fourier based compressed sensing reconstruction using BPE data. Figure 4.17 is magnitude image and Figure 4.18 is phase image respectively. The BPE acquired data images are all aliased images which has obtained before the CS reconstruction and the middle column show the aliased images after CS reconstruction at reduction factor R = 4, 6, 8 and 10 respectively. Right column shows the error images of Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from BPE data can preserve image resolution. ................................................................................................................ 93

Figure 4.17: Simulation results of Fourier based compressed sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the BPE acquired data image and CS reconstructed images at reduction factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from BPE data can preserve image resolution. .................................................................................................. 94

Figure 4.18: Simulation results (phase image)of Fourier based compressed sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the BPE acquired data phase image and CS reconstructed phase images at reduction factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from BPE data can preserve image resolution. ........................................................ 95

Figure 4.19: Simulation results for comparing the conventional CS and Fourier based CSBP encoding technique. (a) reference image; (b)-(d) reconstructed images of Fourier based CS for acceleration factors of 4, 6, and 8 respectively; (e)-(g) error images of Fourier based CS for acceleration factors of 4, 6, and 8 respectively; (h)-(j) reconstructed images of Fourier based CSBP encoding for acceleration factors of 4, 6, and 8 respectively; (k)-(m) error images of Fourier based CSBP encoding for acceleration factors of 4, 6, and 8 respectively; The result of reconstructed images of CSBPE outperforms the CS for preserving image resolution. ........................................................................................................................... 96

Figure 4.20: Comparison between with and without aliasing coefficient correction of reconstructed MRI images from BPE acquired in vivo k-space data; (a) reference image; (b) Reconstructed MRI image using calculated (simulated) trajectories; (c) Reconstructed MRI

Page 11: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

vi

image after aliased coefficient correction using cross-correlation technique; (d) and (e) are error images. The maximum error in the calculated trajectory is 0.2095 and in corrected trajectory is 0.0423. ............................................................................................................ 98

Figure 4.21 : Comparison between CSBPE image reconstruction using estimated modulation maps and non-estimated modulation maps; (a) Original image (b) CSBPE image reconstruction without an estimated modulation map at reduction factor R = 4, (c) CSBPE image reconstruction using the estimated modulation map at reduction factor R = 4. (d) and (e) are error images. ............................................................................................................ 99

Figure 4.22: Experimental results of Fourier based CSBPE phantom data. The left column top image represents the reference image and Fig. Rest of the left column shows reconstruction images using Fourier based CSBPE for down sampling R = 2, 4, 6, 8 respectively and the right column show the error images. The error for the reduction factor 2, 4, 6 and 8 are 0, 0.021, 0.3637, and 0.3867 respectively. ................................................... 101

Figure 4.23: Experimental results of Fourier based CSBP encoding. The left column represents the reference image and Fig. (a) show image reconstruction using Fourier based CSBPE for down sampling R = 4 and Fig. (b) – (c) show reconstruction images of Fourier based CSBP encoding for acceleration factors of R = 6, 8 respectively and Fig.(d) - (f) (right Column): show the error images. The error for the reduction factor 4, 6 and 8 are 0.021, 0.478 and 0.523 respectively. ............................................................................................ 102

Figure 5.1: Sequential steps of Chirp modulated CSBPE image reconstruction process ..... 107

Figure 5.2: (a) Fourier modulated k-space (Mesh view), (b) Chirp modulated k-space (Mesh view), (c) Chirp modulated k-space .................................................................................. 109

Figure 5.3: (a) shows that the Fourier encoded variable density sampling pattern and (b) shows the chirp modulated random sampling pattern which energy is spread along the phase direction. .......................................................................................................................... 112

Figure 5.4 The mean relative error versus the acceleration factor in CS-MRI, CSBPE Fourier encoding and CSBPE chirp encoding for simulation data. ................................................. 113

Figure 5.10 shows the comparing real phantom results between the Fourier based CSBPE and chirp modulation based CSBP encoding schemes. Figure 5.10 (b)-(d) show Chirp modulated Fourier based CSBPE reconstructed images at different acceleration factors. Figure 5.10 (e)-(g) show error images with Chirp modulated Fourier based CSBPE for acceleration factors of 4, 6, and 8 respectively. Figure 5.10 (h)-(j) show images reconstructed with Fourier encoding based CSBPE using for acceleration factors of 4, 6, and 8 respectively. Figure 5.10 (k)-(m): show error images with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively. But, for the real phantom data case as shown in Figure 5.10, The Chirp modulated CSBPE results are not significantly high quality compare to Fourier based CSBPE methods. ...................................................................... 115

Figure 5.5: Simulation results for performance evaluation of chirp based compressed sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the chirp based BPE acquired data image and CS reconstructed images at reduction factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from chirp based BPE data can preserve image resolution. .................................... 116

Figure 5.6: Simulation results for performance evaluation of chirp based compressed sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the chirp based BPE acquired data image and CS reconstructed images at reduction factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS reconstruction for

Page 12: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

vii

acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from chirp based BPE data can preserve image resolution. .................................... 117

Figure 5.7: Simulation results for Chirp modulated Fourier encoding based CSBPE. Left column represents the reference image 256 × 256 (up/down: phase encodes, left/right: frequency encode), Fig. (a) - (d) (middle column): shows reconstructed images by Chirp modulated Fourier basedCSBPE with random down sampling patterns for acceleration factors of R= 4, 5.7, 6, 8 respectively and Fig.(a)-(d) (right Column): shows the error images. ............................................................................................................................. 118

Figure 5.8: Comparing results between the Fourier based CSBPE and chirp modulation based CSBPE encoding schemes (up/down: phase encodes, left/right: frequency encode). (b)-(d): show images reconstructed with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; (h)-(j): show images reconstructed with Fourier encoding based CSBPE using for acceleration factors of 4, 6, and 8 respectively; (k)-(m): show error images with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively; ................................................................. 119

Figure 5.9: Comparing results between the CS-MRI and chirp modulation based CSBPE encoding schemes (up/down: phase encodes, left/right: frequency encode). (b)-(d): show images reconstructed with Fourier encoding based CSBPE using for acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively; (h)-(j): show images reconstructed with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; (k)-(m): show error images with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; .................................................................................................................. 120

Figure 5.10: Compare the results between the Fourier based CSBPE and chirp modulation based CSBPE encoding schemes (up/down: phase encodes, left/right: frequency encode). (b)-(d): show images reconstructed with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8 respectively; (h)-(j): show images reconstructed with Fourier encoding based CSBPE using for acceleration factors of 4, 6, and 8 respectively; (k)-(m): show error images with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively; ................................................................. 121

Page 13: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

viii

List of Tables Table 1: Calculated trajectory vs corrected trajectory of all 32 channel. .............................. 88

Table 2: Comparison of the artifact power (AP) between proposed CSBPE-MRI methods and the reference CS-MRI in a simulated brain image at different acceleration factor (R). .. 93

Table 3: Compare the simulation results of Fourier based CSBPE and chirp modulated CSBPE for single channel data. ........................................................................................ 115

Page 14: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

ix

Nomenclature

B radius of weak-Apball

B0 external magnetic field

B1 magnetic field caused by the RF pulseaconstant

C complex set

d negative gradient of objective function

E energy

∆E energy differencefrequency

f(x) regularized functiongradient of objectivefunctionGramMatrix

Gx x component of magnetic gradient

Gy y component of magnetic gradient

Gz z component of magnetic gradient

Gs slice selectiongradient

GΦ phase encoding gradient

H hard thresholdingoperation

h the Planck’sconstant

I image/identity matrix

i index

j index

K sparsity of the originalsignal

k Boltzmannconstant

A00norm

A11norm

A22norm

Ap pnorm

A∞infinitynorm

M length of the measureddata

M0 net magnetization

Mx x component of the netmagnetization

My y component of the netmagnetization

Mxy transverse component of the net magnetization due to T2

M∗xytransversecomponentofthenetmagnetizationduetoT2

Mxy0 Mxyat the start of the dephasingprocess

Page 15: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

x

Mz z component of the netmagnetization

N length of the originalsignal

Nc number ofcoils

[N] a set (1,2,….. ,N)

card(N) cardinality of set[N ]

PD protondensity

p aconstant

r residual

R reductionfactor

Rnet net reduction factor

Rnom nominal reduction factor

R realset

s transformed compressiblesignal

si(j) acquired black data points in coil i at locationj

s4,ACS the gray data in the ACS line of coil4

s(kx,ky) acquired signal in Fourierdomain

S NMRsignal

t time

tp a timeduration

T1 spin-lattice relaxationtime

T2 spin-spin relaxationtime

T2,inhomorelaxation time caused byinhomogeneity

T2∗ combined relaxation time

TE echotime

TR repetitiontime

T measurementmatrix

TF partial Fourier encodingoperator

T̃F fullFourierencodingoperator

T̂F Fourierencodingoperatorwithwavelettransform

TN partial noiselet encodingoperator

T̃N fullnoiseletencodingoperator

T̂N noiseletencodingoperatorwithwavelettransform

Ts selected encoding technique with sub-samplingapplied

Page 16: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

xi

T̂s wholesub-samplingandencodingprocess

U transformmatrix

W wavelet transformoperator

wi(j) GRAPPAweightsincoiliatlocationj x adirection

x original signal orimage

xi the ithentry ofx

x̂1 overlappedpixel

x̂2 overlappedpixel

x(1) pixel in aimage

x(2) pixel in aimage

x̂ reconstructedsignalorimage

y adirection

y measureddata

yi sub-sampled data from coili

z adirection

α aconstant

β aconstant

Γ1(1),Γ1(2) sensitivity in the first coil

Γ2(1),Γ2(2) sensitivity in the second coil

Γi sensitivity profile of coili

Γ̂i transformedΓi

γ the gyromagneticratio

δ ratio of M toN

δK,δ2K RIPconstant

s parameters with relationship tonoise

θ degree of RFpulse

Λ supportset

λ regularizationparameter

µ coherencevalue

ν frequency ofproton

ν0 Larmorfrequency

ξ positive smoothingparameter

π aconstant

ρ ratio of K toM

Page 17: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

xii

ρ(x,y) spin density at location (x, y)

σ stopcriterion

σmin minimum singularvalues

σmax maximum singularvalues

τ parameters with relationship tonoise

Φ sensingmatrix

Φ̂ asquaresensingmatrix

φi ithrow or column of sensing matrix Φ

ΦK a matrix consisting of K columns fromΦ

Φsub(K) sub-matrix formed from K distinct columns ofΦ

χ(x) mother bases function of noiselet

Ψ sparsifyingmatrix

ψi ithrow or column of sparsifying matrixΨ

ψi(x) ithentry of transformed vector x under Ψ

Ψ̃ Ψwitheachcolumnnormalized

Ψ̂ dictionary

ψ̂i ithcolumnofΨ̂

ω frequency of sinewave

∅ emptyset

|·| magnitude

||·||0 number of non-zeroelements

||· ||1 sumofthemagnitudeofallelements

||·||2 sumofthesquareofallelements

||·||p pthrootofsumofpthpowerofthemagnitudeofallelements

||·||∞ maximum magnitude of avector

<·,·> innerproduct

∞ infinity

Ⓢ element-wisemultiplication

Page 18: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

Chapter 1 Introduction

Introduction

1.1 Overview

Visual observation is one of the essential keys of almost all human activities. Hence,

the most useful information source, such as video and image, plays an indispensable

role in our daily lives. Along with science and technology, a significant number of

imaging techniques have been developed and applied in almost all of our life events.

The impact of imaging in medicine on clinical diagnosis and research has been

unprecedented. Among all of the clinical imaging techniques, magnetic resonance

imaging (MRI) is the most popular and advanced imaging technique.

The main concept of MRI is based upon the practical application of Nuclear Magnetic

Resonance (NMR) which was first explained independently by Bloch and Purcell, in

1946, who was awarded Nobel Prize in Physics in 1952 [1]. After establishing the

theory, scientists started to realise that imaging could be a great application of NMR.

For the first time imaging by NMR was used for a military purpose. At the time of cold

war, aircraft navigation in the water project was established using NMR signals in the

former USSR in the 1960s. During this project, Lieutenant Vladislav Ivanov noted that

this technique could be used in the human body because living species are made up of

water [2].

However, his proposal was denied due to political reasons. A couple of years later, the

real era of MRI had literally begun. Researcher Lauterbur published a paper unfolding a

new technique to form images of biological structure using NMR signal of protons[3].

After establish this concept, Mansfield published a series of paper that introduced the

Page 19: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

2

fundamental concept of spatial encoding in MR imaging. Due to the major

contributionin medicine, Researcher Lauterbur and Mansfield were awarded the Nobel

Prizein Medicine in 2003[4].

Now, MRI [5] is a widely used medical imaging modality in contemporary biomedical

research and diagnosis.The importance of MRI is grown dramatically to radiologists,

clinicians and researchers for its competency to produce high-quality images without

the side effects of harmful radiation since its invention in the 1970s.MRI can construct

an image from soft tissue throughout the whole body [6] and can efficiently differentiate

grey and white matters in imaging the brain. It provides excellent contrast in imaging

the different parts of the body such as heart, muscles, blood and cancer cell compared

with other modalities such as CT or X-rays. It can distinguish blood flow and can be

used to measure necessary quantities diagnostically such as cortical thickness [7-9].

In addition, MRI has an extra feature to generate various contrast–weighted images such

as T1, T2, and T2*. This flexibility in image contrast is advantageous in clinical diagnosis

and medical research. Furthermore, thenon-invasive technique has made MRI harmless

for the patient. Hence, MRI has excellent versatility and numerous applications in both

medical research and clinical diagnostic and preoperative surgical imaging.

MRI is suitable for application to any part of the body, such as the heart, brain, neck,

spine, kidney, abdomen and feet. Besides that, there are many applications of MRI for

different types of diagnosis and research. For example, functional MRI (fMRI) can be

used to measure the signal activity of neuron in the brain; dynamic contrast-enhanced

MRI (DCE-MRI) is useful to evaluate specific contrast agents of the muscle blood flow;

3D MRI can evaluate 3D volumetric images of different body parts, etc. These are some

applications of the MRI. It is understood that with further development in the software

and hardware, MRI will be even more broadly applied in different medical and non-

medical fields.

Page 20: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

3

1.2 The Time Limitation of MRI

Despite the ever increasing importance and wide application of MRI in biomedical

research and clinical diagnosis, it still suffers from some major limitations.For example,

under the normal imaging conditions, a conventional MR imaging of a brain may take

around ten minutes, and inother cases, it mayprolonguptoan hour. This constraint is a

hardship for some people, such as children,aged patients, and patients suffering from

chronic pain. Moreover, most MRI machine bores are narrow bounded spaces, which

may cause claustrophobia for some patients.Also, multiple images are typically required

for various applications. This may extend ascan sessionto more than one hour in

duration, resulting in increasedexpenses and reducedaccessibility of MRI scanner [10].

MRI data acquisition speed is restricted by biological limitations associated with the

properties of spatially fluctuating magnetic fields in the body. Varying applied magnetic

fields can induce fluxes at high rates indifferent parts of the nervous system and may

stimulate the nerves causing discomfort andeven damage in the subject [11].As the

electromagnetic fields used to encode the data for Fourier coefficients, this constraint

limits the rate that can collect MRI data.Hence, in some time-critical applications, such

as fMRI (functional MRI), the spatial image resolution has to be compromised for

higher temporal resolution to accelerate MRI process, and researchershave focused on

modifying the sampling pattern and acquiring multiple samples concurrently.

Different kinds of sampling patterns or multiple simultaneous samples have some

advantages and disadvantages. Modifying the data sampling pattern may

reduceimageresolution, lose phase informationand also may require more complicated

reconstruction methods [12, 13]. Moreover, fast MR image acquisition methods can use

multiple echoes to acquire an image in shorter time while decreasing contrast or

increasing susceptibility to field inhomogeneity [14-16].

Several image processing approaches have been developed for accelerating MRIoverthe

last two decades, e.g., parallel MR imaging process and post-processing technique to

Page 21: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

4

recover images from fewer data samples. Most popular parallel imaging methods are

SENSE (SENSitivity Encoding), GRAPPA (Generalized Auto Calibrating Partially

Parallel Acquisition) [17-20]. These techniques use multiple receive channels with

different non-uniform magnetic sensitivities, whereas conventional MRI has no such

capacity to accelerate image processing due to using a single channel with uniform

sensitivity to magnetisation. Parallel MR imaging (pMRI) techniques are able to

accelerate different types of imaging which have already been implemented in

commercial scanners. However, those pMRI schemes alone are insufficient to achieve

the level of acceleration we would like to achieve due to the number of difficulties

associated with pMRI.These difficulties are responsible for degrading the quality of the

reconstructed image if these are not appropriately addressed and solved during the

reconstruction process.

These limitations are summarized below:

Inadequate sampling data for reconstruction: For example, keyhole imaging

technique is applied to reconstruct the full-size image using only a subset of the

acquired k-space data.

Noise in acquired data: It is nearly impossible to acquire data without noise in

all MRI systems[21-23]. Moreover, it has more effects on accelerated imaging

because fewer amounts of data are acquired in k-space. Hence, we have

observed significant SNR loss in accelerated imaging [19].

Process Noise: The procedure of encoding is responsible for imposing noise

during the acquisition, and it affects the reconstruction results, especially when

the others condition are not fulfilled.

Anonymous imaging scheme:This kind of problem can arise from unknown

sensitivity functions of receiver coils. As the imaging procedure is unidentified,

it requires to be estimated. Butestimation process may create many difficulties,

such as process estimation error and system structural constraints.

Page 22: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

5

Process uncertainty: The main reasons for the uncertainty in the imaging

technique are process noise and process estimation error. It can also affect image

quality, such as field’s inhomogeneity, improper pulse sequence design and

some other machine related issues.

Anewly invented methodof image reconstruction from undersampled k-space data is

known as compressed sensing (CS) [24-27].CS has a unique feature of constructability

to recover data from fewer randomly selected measurements than the original size of the

signal by using a suitable recovery technique. However, two prerequisites of CS

theoryshould be fulfilled in order to obtain the maximum benefit from CS as mention

below:

(1) Sparsity:The image signal should be sparse or sparse after a specific

transformation.

(2) Incoherency and RIP: The acquired signal should be sensed randomly.

Moreover, the random sensing matrix should satisfy a particularcondition

such as the restricted isometric property (RIP)[24, 28, 29] and

incoherence[30, 31]to ensure successful reconstruction of the image signal.

So, MRI is one of the appropriate candidates for CS because MRI data are generally

sparse in the wavelet transformed domain, and can be acquired with some random

sampling patterns[32-36]. Hence, MRI has the ability to fulfil the condition of RIP and

coherency to reconstruct the images from sub-sampled k-space data successfully.

Over the last two decades, the researchers have endeavoured to achieve the highest goal

of the MRI process time by combining different methods. The combination of

compressed sensing with some other existed popular pMRI methods such as SENSE,

and GRAPPA was already established in recent years [19, 20]. Some of the researchers

havetried to use different encoding techniques such as Fourier, Noiselet, and chirp

modulation. Some other researchers have also developed different encoding trajectories

such as Cartesian and non-Cartesian to enhance the processing speed.

Page 23: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

6

The various under-sampled data acquisition techniquesand compressive sensing

reconstruction methods have the capability to reduce the total scan time and running

costto some extent. This can give the opportunityformore patients or allow more

imaging to be completed per unit time. That is very important for systemic metabolic

and genetic diseases, where multiple organ systems are involved. Additionally, other

scan types can be made feasible. However, there are still some limitations in the present

work as listed below:

Problem 1.Parallel MRI techniques such as SENSE, GRAPPA require a large

number of coils when few number of data are acquire to reconstruct the

image.Also due to inconsistency in coil geometries and sensitivities, severe

residual aliasing artifacts and amplified noise may occur when high reduction

factor is chosen.

Problem 2.As a non-array coil method like BPE also has some problem when

reduction factor is increased. That means signal to noise ratio of BPE approach

becomes considerably reduced as the reduction factor is increased and we observe

that if reduction factor is greater than two, the noise of reconstructed image is

gradually increased even through aliasing artefact are not exist.

Problem 3. Similarly, the reduction factor is also crucial for high-quality

reconstruction in compressed sensing even though it is used with pMRI. That

means if it is increased to a certain level, the artifacts of the image and SNR also

increased simultaneously.

Problem 4. In the conventional CS-MRI, Fourier matrix as a sensing matrix and

Wavelet matrix as a sparsifying transform matrix respectively are not optimally

incoherent. Moreover, Fourier encoding usually concentrates energy in the centre

of the k-space known as low frequencies region. This imposes restriction on the

subsampling pattern to adequately sample the low frequency region and

insufficiently sample the high frequency region at high acceleration factors,

Page 24: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

7

resulting in degraded image quality. Such limitation can cause a considerable loss

in image resolution.

The purpose of the research is to address above mentioned limitations to accelerate

imaging speed and improve the imaging quality within the constraints of present

commercial MRI scanners. Currently, the scan time of MRI is still too slow to meet the

demand of clinical diagnosis and biomedical research. Therefore, the accelerated

acquisition and reconstruction techniques need to be employed to reduce encoding and

hence scan time. Inthis research, we develop a method to improve and accelerate the

conventional signal acquisition and high-resolution image reconstruction using a hybrid

method.These methods will provide economic benefits for patients and doctors.

1.3 Research Objective and Technical Approaches

The objective of this research is to advance further the present state of reconstruction

speed with quality in MRI. The fundamental concept of this research is to find a new

optimal way to reconstructhigh-quality images using less amount of acquired data by

implementing different approaches. These Novel techniques offer optimal solutions in

accelerated imaging and hence, increase the image quality. In this thesis, the enhanced

imaging methods are modelled, analysed and solved from a system point view by using

the technique extensively from the theories of system modelling and parameter

estimation, system inversion and system optimization.

The technical approaches of the research work areto achievea higher acceleration factor

with a quality image using both compress sensing and Fourier and non-Fourier

encoding of non-Cartesian zigzag trajectories. Non-Cartesian trajectory such as

Bunched Phase encoding (BPE) can be directly combined with the compressed sensing

reconstruction framework. Some existing methods such as Sparse SENSE [27] and L1

SPIRiT [37] follow this approach, yielding a sparsity promoting standardized

reconstruction technique that can reconstruct high-quality images from moderate

accelerations with random under sampling.Initially, general simulations will be used to

validate these ideas. After validation, the MR pulse sequence will beused to acquire data

Page 25: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

8

from the MRI scanner for physical experiment and verification of the proposed

methods.

The novel results of this thesis are two new accelerated MR imaging techniques. The

first one is a technique to reduce scan time in 2D static imaging and is called

compressed sensing MRI using Fourier based bunched phase encoding.This technique

accelerates the conventional MR signal acquisition and improves high-resolution image

reconstruction using a hybrid method which combines multiple coils compressed

sensing MRI with Fourier or non-Fourier based bunched phase encoding.

The second one is the non-Fourier chirp modulatedBPE to further reduce scan time in

MR imaging and is called compressed sensing MRI using non-Fourier based bunched

phase encoding. The motivation of this work is that chirpmodulated encoding schemes

outperformFourier encoding in compressed sensing reconstruction because it spreads

signal energy in the measurement domain that improves sampling incoherence.

1.4 Main Contributions

The major contributions of this thesis are summarized as follows:

1. In order to improve the quality of image reconstruction of compressive sensing

for high reduction factor, this thesis has proposed the first compressive sensing

bunched phase encoding (CSBPE) technique. The simulation results of Fourier

based CSBPE have shown that the quality of the reconstructed image is higher

than that of the conventional compressive sensing at the same reduction factor.

These results are presented in Chapter 4.

2. Due to magnetic field inhomogeneity and imprecise field gradients, the phase

of the zigzag trajectories is different from theoretically calculated ones, which

lowers image quality. To overcome this problem, we have introduced a new

technique called cross-correlation to remove the phase deviation from the

zigzag data position. The experiment results show that the CSBPE with zigzag

Page 26: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

9

data-position corrected by cross-correlation outperforms the conventional BPE

without the correction at high reduction factor. This part of the study is

presented in Chapter 4.

3. To smooth and remove the aliasing from the image, a modulation map

estimation method has been incorporated in compress sensing part of our

proposed method.This part of the study is presented in Chapter 4.

4. In order to further improve the image quality of CSBPE, we have introduced a

non-Fourier encoding scheme called chirp modulated Fourier encoding. The

simulation result shows that the chirp modulated CSBPE outperform the

Fourier method. This research result is given in chapter 5.

1.5 Thesis Outline

The thesis is organized as follows:

• Chapter 1: Presents the limitation of CSMRI, motivation and objectives of this

thesis. A general background of the study is explained to highlight the

importance of the research. The objectives and technical approaches are then

provided. Finally, the thesis outline is given.

• Chapter2:Presents the fundamentals of the MRI process. First, the theory of

NMR is reviewed to introduce the physics of NMR, It then explains the NMR

signal excitation, image reconstruction technique of MRI, the non-cartesian

trajectory, different types of sub-sampling and reconstruction in the Cartesian

coordinates such as SENSE, and GRAPPA. Finally,it explains some novel

sequence designs, including bunched phase encoding, rotating RF coil, Wave–

CAIPI.

Page 27: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

10

• Chapter 3: Summarizes the fundamental of sparse sampling. First, the theory

of compressed sensing is reviewed to introduce restricted isometry

propertyand, incoherence in compressive sensing. It then explains the

application of CS in magnetic resonance imaging and other fields.

• Chapter 4: Discusses compressive sensing MRI using bunched phase

encoding scheme. It first explains the theory of CSMRI using BPFE and

multichannel CSBPE, then explainshow to correct the reconstruction error of

the reconstruction techniques. It further explains the method of CSBPE and

describes the simulation and experimental results. Finally, the discussion and

conclusion of this chapter are presented.

• Chapter 5:Introduces compressive sensing MRI using bunched phase non-

Fourier encoding scheme. It first describes the theory of CSBPE with chirp

modulated Fourier encoding and explains the multichannel CSBPE with chirp

encoding. It then presents the simulation results with qualitative and

quantitative analysis. Finally, the discussion and conclusion of this chapter is

presented.

• Chapter 6:Concludes the thesis with discussions of the presented results and

recommendations for future study.

Page 28: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

11

Chapter 2 Fundamentals of MRI Process

Fundamentals of MRI Process

2.1 Introduction

One of the purposes of this chapter is to identify the limitationsof present work in the

literature review. At first, the fundamentals of MRI will be discussed then the open

problems in the existing literature will be identified.

2.2 Theory of MRI

Magnetic resonance imaging (MRI) concept works by using the principle of nuclear

magnetic resonance. The spatial gradient of the applied magnetic field provides spatial

information to create images. In MRI, the spatial frequency domain is known as k-

space. It transforms the time-domain signal to the spatial frequency domain. The

sampling trajectory of MRI in the k-space domain is determined by the time of the

applied magnetic field gradient. The appropriately designing the magnetic field is the

most important engineering aspect of MRI to induce nuclear magnetic resonance to

capture the spatial distribution of the hydrogen protons in the body.

2.2.1 Physics of NMR

The physical basis of Nuclear Magnetic Resonance (NMR) phenomenonis the concept

of Nuclear Spin, which is first invented by Bloch and Purcell in 1946[38-40]. NMR

signals can generate by flipping the nuclear protons from high energy state to low

energy state. Nuclear spin is a natural form of intrinsic angular momentum carried by

atomic and sub-atomic and other elementary particles. It is a hypothesis that protons of

Page 29: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

12

Hydrogen Atom get aligned in parallel or anti-parallel in the direction of the magnetic

field [41].Same directional spin particle repeal but opposite directional particle make

pair with each other.

In NMR, only unpaired spins are played most of the role. The rotate nuclei act as a

magnetic dipole with random orientations. However, in the natural environment, the net

magnetization of nuclei which donated as M0 is untraceable due to random orientations

and the magnetic effect iscancelled by each other.

When a proton spins on its axis, it generates a magnetic field. Hence, the nucleus can be

considered to be a tiny bar,and these small bar magnets are randomly oriented in space.

However, in the presence of a magnetic field B0, they are oriented with or against this

applied field. Therefore, the difference in numbers of dipoles orientation with and

without this field leads to the NMR signal.

Figure 2.1: Nuclear spin in different condition (a) Spins without magnetic field (b) Spins with the magnetic field

The energy difference between the two states isminimal,and it is defined as,

∆E = hγ|𝐁𝐁𝟎𝟎|2π

(2.1)

Page 30: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

13

Where,ℎ, 𝛾𝛾, |𝐁𝐁𝟎𝟎|is the Planck’s constant, gyromagnetic ratio, and magnitude of the static

magnetic field.Spins can switch between the two energy states at a particular frequency

by absorbing a photon. The energy of the photon E is related to its frequency f, therefore

𝐸𝐸 = ℎ𝑓𝑓 (2.2)

The photon only can move from one state to another state, when the energy difference

between the two energy states is the same as the photon energy, which means

𝐸𝐸 = ∆𝐸𝐸 (2.3)

Hence, we can rewrite the equation (2.1) as

ℎ𝑓𝑓 = hγ|𝐁𝐁𝟎𝟎|2π

(2.4)

Therefore the frequency of the photon is then given as

𝑓𝑓 = γ|𝐁𝐁𝟎𝟎|2π

(2.5)

The frequency fis called Larmor frequency,and the magnetization is resonant at this

frequency.

2.2.1.1 Spin-Lattice Relaxation The net magnetization of nucleiM0reaches an equilibrium state after placed in a

magnetic fieldB0 for a specific time, and the direction of M0isparallel to the magnetic

field B0. In the coordinate system, the direction of the field B0is generally considered

parallel to as the z-direction, whereas perpendicular to the x-direction and y-direction.

At the equilibrium state, the net magnetization M0 is same as the z component

magnetization Mz, while x and y component of the magnetization is zero. The linear

condition of the net magnetization with B0 can be changed by a sequence at Larmor

Frequency. The net magnetization can only reach its equilibrium condition after time

and the time constant can explain how it reaches its equilibrium state is called the spin-

lattice relaxation time or longitudinal constant which is denoted as T1time constant. The

net magnetization comes to its equilibrium state can be express by the following

equation,

Page 31: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

14

𝑀𝑀𝑧𝑧 = 𝑀𝑀0 �1− 𝑒𝑒−𝑡𝑡𝑇𝑇1�. (2.6)

Figure 2.2 shows that the net magnetization Mz as a function of time t when M0= 1 and

T1 = 256. Due to natural phenomenon, spin tends to come low energy state, when the

energy of the magnetization comes to its equilibrium state, the energy of the spin goes

to the lowenergy state. The energy of the spin is deteriorated as a transfer of heat

through the electromagnetic interaction, collisions and rotations.

Figure 2.2: Net magnetization as a function of time

2.2.1.2 Spin-spin Relaxation When a radiofrequency (RF) is applied to the magnetization field, it is switch into the

transverse plane which is perpendicular to the z-direction. The applied RF pulse should

be equal to Larmor frequency due to modifying the net magnetization. The modified

magnetic field due to RF pulse known as B1 magnetic field. After applying RF pulse,

the flip angle α will be

𝛼𝛼 = 𝑓𝑓𝑡𝑡𝑝𝑝 = 𝛾𝛾𝐵𝐵1𝑡𝑡𝑝𝑝. (2.7)

Page 32: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

15

Wheretp is the time duration, f is the RF frequency,and B1 is the modified magnetic

field. The flip angle can be varied from 0 to 1800. However, when the flip angle is 1800,

the net magnetization is the same as B0 but opposite direction. At any other angle except

1800, the rotating magnetization appears around z-axis which is known as precession.

The rotation speed fallows the Larmor frequency. After the external RF pulse cut off,

the net magnetization of the z component will experience T1 relaxation, and on the other

hand, the net magnetization Mxy will start, which expose a different relaxation process.

Suppose after 900RF pulse excitation, the transverse component of the magnetization

can be represented as

𝑑𝑑𝑀𝑀𝑥𝑥𝑥𝑥

𝑑𝑑𝑑𝑑= −𝑀𝑀𝑥𝑥𝑥𝑥

𝑇𝑇2 (2.8)

Since, after a 900excitation 𝑀𝑀𝑥𝑥𝑥𝑥(0) = 𝑀𝑀0, so

𝑀𝑀𝑥𝑥𝑥𝑥 = 𝑀𝑀0𝑒𝑒(−𝑑𝑑/𝑇𝑇2) (2.9)

Figure 2.3: Transverse relaxation Mxy as afunction of time

The characteristics of the decay of transverse magnetization are called Transverse

relaxation or spin-spin relaxation,and the time constant T2 is called spin-spin time

constant. Due to the intermolecular interaction of the spins,the transverse magnetization

Mxy dephases in the transverse plane resulting inthe time constant T2 decay. The T2

decay curve is shown in Figure 2.3. However, transverse relaxation process also

Page 33: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

16

depends on the z-component magnetic field oscillation. The longitudinal fluctuation

dominates transverse relaxation process, therefore T2 < T1. Due to the domination of the

z-component fluctuations,a new relaxation is evolved that is called T2*. In the presence

of transverse component oscillation, the z-component israpidlydephased, which is

characterized as T2*decay, which also called free induction decay (FID).Figure2.4 is

indicated in the FID process.

Figure 2.4:FID process in the x-y plan after 900RF pulse

2.2.2 Physical Behaviour of Tissue Particle One of the common element in the human and animal bodyiswater and hydrocarbons,

which are based on the hydrogen (1H) atom. This hydrogen (1H) atom is highly

sensitive to applied magnetic fields. When the gradient magnetic field is applied to the

body, the hydrogen atom of the body tissue gets excited to spin in one direction. This

spin happens because hydrogen Atom has only one proton, and it could be aligned

easily with MRI magnet.

The proton possesses the property of spin, which causes the nucleus to generate an

NMR signal that further catches by the radio frequency (RF) coils to produce k-space

data. Since bones have not had any water; it does not generate any data for the image.

Page 34: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

17

Bones leave a blank area in the images; that is why MRI scanners are best for scanning

soft tissues. The Radio Frequency coils detect the NMR signals generated by the

nuclear spin.

2.3 NMR Signal Excitation

2.3.1 Slice Selection and Gradient Selection The core magnetic field B0 of the MRI scanner is considered as homogeneous. When an

RF signal is applied to excite, the net magnetization is switched into the transverse

plane,and the FID will be recorded from the spins in the magnetic field since both the

spins and RF pulse have the Larmor frequency. A specific magnetic field gradient helps

us to select a particular position. The gradient field is a variation of the magnetic field

familiarized locations,and this gradient field is generated by gradient coils in an MRI

scanner.

The MRI scanner has three types of gradient coils,and those gradient coils generate

fluctuating magnetic field so that spins at different location precess at frequencies

unique to their location, allowing us to reconstruct 2D or 3D images. The x, y, and z

directional gradient known as Gx, Gy, and Gz, respectively.

MR images are drawn a slice of the object with a specific thickness. A one dimensional

(1D) selected magnetic gradient pulse is applied along z-direction to accomplish slice

selection during the period of the RF signal. During processing, the applied exciting

magnetic field B1 with a linear gradient field Gz tips the net magnetization B0 into the

transverse plane in a specified period. Spins of the net magnetization field outside the

slice selection gradient are not influenced. After thatacquired signal from the selected

slice and both the RF pulse and slice selection gradient pulse are turned off, and the

main magnetic field B0back to the initial homogeneous state. Figure 2.5 shows a

schematic diagram of the slice selection process.

Page 35: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

18

Figure 2.5: The process of the slice selection by the gradient pulse

2.3.2 Frequency and Phase Encoding During the image data acquisition period, the slice selection gradient along the z-

direction is used to select a specific slice with thickness.The frequency encoding

gradient and the phase encoding gradient are used to define the x and y directional value

respectively to select a specific location in the transverse plane. The midpoint of the

transverse plane, the magnetic field is homogeneousB0 and frequency is Larmor

frequencyf which point is called isocenter. Anx-directionallinear magnetic gradient is

applied to fluctuate linearly as a function of position. The magnetic field along this

direction can be represented as

𝐵𝐵(𝑥𝑥) = 𝐵𝐵0 + 𝑥𝑥𝐺𝐺𝑥𝑥 . (2.10)

And the corresponding Larmor frequency is

𝑓𝑓(𝑥𝑥) = 𝛾𝛾𝐵𝐵(𝑥𝑥) = 𝛾𝛾𝐵𝐵0 + 𝛾𝛾𝑥𝑥𝐺𝐺𝑥𝑥 = 𝑓𝑓 + 𝛾𝛾𝑥𝑥𝐺𝐺𝑥𝑥 (2.11)

The equation (2.11) shows that the Larmor frequency f also varies linearly as a function

of gradient position. Therefore, the encoded NMR signal has a different frequency at a

Page 36: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

19

different location. We can determine the x-directional location by measuring those

frequencies of the acquired MRI signal.This is known as frequency encoding technique.

For the 2D image, frequency encoding can determine the x-directional data position that

is called frequency encoding direction. By the same way, to determine y-directional

position value, phase encoding technique is used. A one-dimensional linear magnetic

field gradient along the phase encoding direction is used to accomplish phase encoding.

The equation of the magnetic field along the y-direction represent as

B(y) = 𝐵𝐵0 + yG𝑥𝑥. (2.12)

And the corresponding Larmor frequency is

𝑓𝑓(𝑦𝑦) = 𝛾𝛾𝐵𝐵(𝑦𝑦) = 𝛾𝛾𝐵𝐵0 + 𝛾𝛾𝑥𝑥𝐺𝐺𝑥𝑥 = 𝑓𝑓 + 𝛾𝛾𝑦𝑦𝐺𝐺𝑥𝑥 (2.13)

Therefore, we can define a three-dimensional gradient where slice excitation is one of

the first dimension,and frequency encoding provides the second dimension. The third

dimension is provided by phase encoding, which is done by applying a gradient in the

y-direction parallel to both slice selection and frequency encoding. From equation

(2.13), we conclude that the phase can be different for the same frequency, and the

phase encoding direction can be identified by measuring the phase.

2.3.3 Image Formulation The time-varying magnetization due to free relaxation is called the free induction decay.

The process of relaxation needs to transform into voltage or current in order to be

detected and further processed.The transformation of FID into voltage signal can be

done by using Faraday’s law of electromagnetic induction,the equation is given by

𝑉𝑉(𝑡𝑡) = 𝜕𝜕𝜕𝜕𝑥𝑥𝛷𝛷(𝑡𝑡) = 𝜕𝜕

𝜕𝜕𝑥𝑥 ∫ 𝐶𝐶(𝑟𝑟).𝑀𝑀��⃗ (𝑟𝑟, 𝑡𝑡)𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.14)

Page 37: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

20

where Φ(𝑡𝑡) denoted as the magnetic flux of the RF coil, VOI represent as volume of

interests of the subject being imaged, and 𝐶𝐶(𝑟𝑟) denoted as the reciprocity of the coil

calculated by the Biot–Savart Law

𝐶𝐶(𝑟𝑟) = 𝜇𝜇4𝜋𝜋 ∮

𝑑𝑑𝑑𝑑×𝑅𝑅�⃗

𝑅𝑅3, (2.15)

where 𝜇𝜇 represent as the material permeability, 𝑙𝑙 denoted as the path of the coil, 𝑅𝑅�⃗

represent the vector between 𝑑𝑑𝑙𝑙 and the spatial position𝑟𝑟. Equation (2.15) represents the

spatial sensitivity of the receive coil. Visually assuming, the applied transverse RF field

𝐵𝐵�⃗ 𝑟𝑟𝑟𝑟 is uniformly applied to sample, the image will show brighter where the coil’s

reciprocity is strong and vice versa. The coil’s reciprocity 𝐶𝐶(𝑟𝑟)value in single receive

coil scanners is uniform compare to multireceiver coil scanner. In multireceiver coil

scanners, a set of 𝐶𝐶(𝑟𝑟)s is used, each coil value can be designed to have a different and

localized sensitivity region for the purpose of acceleration imaging.

The Eq. (2.14) can be written for the multiple receiver coils as

𝑉𝑉(𝑡𝑡) = ∫ �𝐶𝐶𝑥𝑥(𝑟𝑟) 𝜕𝜕𝑀𝑀𝑥𝑥(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

+ 𝐶𝐶𝑥𝑥(𝑟𝑟) 𝜕𝜕𝑀𝑀𝑥𝑥(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

+ 𝐶𝐶𝑧𝑧(𝑟𝑟) 𝜕𝜕𝑀𝑀𝑧𝑧(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

� 𝑑𝑑𝑟𝑟.𝑉𝑉𝑉𝑉𝑉𝑉 (2.16)

Here, 𝜕𝜕𝑀𝑀𝑧𝑧(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

is considered as negligible component compare to other stronger

transverse components, so the MR signal will be

𝑠𝑠(𝑡𝑡) = ∫ �𝐶𝐶𝑥𝑥(𝑟𝑟) 𝜕𝜕𝑀𝑀𝑥𝑥(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

+ 𝐶𝐶𝑥𝑥(𝑟𝑟) 𝜕𝜕𝑀𝑀𝑥𝑥(𝑟𝑟,𝑑𝑑)𝜕𝜕𝑑𝑑

� 𝑑𝑑𝑟𝑟.𝑉𝑉𝑉𝑉𝑉𝑉 (2.17)

Finally, we can represent MR signal in the following,

𝑠𝑠(𝑡𝑡) = 𝜔𝜔0 ∫ 𝑒𝑒−𝑡𝑡

𝑇𝑇2(𝑟𝑟��⃗ ) 𝐶𝐶𝑥𝑥,𝑥𝑥(𝑟𝑟)𝑀𝑀0(𝑟𝑟, 0) 𝑒𝑒−𝑗𝑗(∆𝜔𝜔(𝑟𝑟)𝑑𝑑+∅(𝑟𝑟,𝑑𝑑)−𝜗𝜗(𝑟𝑟) 𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.18)

Where 𝑠𝑠(𝑡𝑡) is the MR signal, ∆𝜔𝜔(𝑟𝑟) is denoted as demodulated frequency, ∅(𝑟𝑟, 𝑡𝑡) is

the accumulated phase of the procession up to time t and 𝜗𝜗(𝑟𝑟) is represented as the

Page 38: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

21

receive field directional angle with x-axis. However, the global signal 𝑠𝑠(𝑡𝑡) cannot be

used to produce image signal from the spatial distribution of hydrogen nuclei. It needs

to establish a new technique called spatial encoding to obtain the image successfully.

2.3.4 The K-space Signal The MRI encoding system generates and acquires data in k-space by applying varying

magnetic field gradients, and this system is called spatial encoding of MRI. The

spatially encoded MRI signal are used to reconstruct the digital image through spatial

algorithm process. To understand the whole MRI encoding idea, the mathematical

explanation of k-space is first introduced.

We can represent the Equation (2.18) to obtain the MR signal in terms of the transform

of spatially weighted spin density function

𝑠𝑠(𝑡𝑡) = ∫ 𝐶𝐶 (𝑟𝑟) 𝑃𝑃(𝑟𝑟) 𝑒𝑒−𝑗𝑗(∆𝜔𝜔(𝑟𝑟)𝑑𝑑+∅(𝑟𝑟,𝑑𝑑)) 𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.19)

Where all the constants and the relaxation terms are dropped and 𝐶𝐶 (𝑟𝑟)is represent in

place of 𝐶𝐶𝑥𝑥,𝑥𝑥(𝑟𝑟) and 𝑒𝑒−𝑗𝑗𝜗𝜗(𝑟𝑟) is absorbed into 𝑃𝑃(𝑟𝑟) which is called

𝑃𝑃(𝑟𝑟) ≔ 𝑃𝑃0(𝑟𝑟)𝛾𝛾2ℏ2𝐵𝐵04𝑘𝑘𝑇𝑇

𝑒𝑒−𝑗𝑗𝜗𝜗(𝑟𝑟). (2.20)

Further, Introducing ∅𝐺𝐺(𝑟𝑟, 𝑡𝑡) = ∆𝜔𝜔(𝑟𝑟)𝑡𝑡 + ∅(𝑟𝑟, 𝑡𝑡), the Eq. (2.19) can be simplified to

𝑠𝑠(𝑡𝑡) = ∫ 𝐶𝐶 (𝑟𝑟) 𝑃𝑃(𝑟𝑟) 𝑒𝑒−𝑗𝑗∅𝐺𝐺(𝑟𝑟,𝑑𝑑)) 𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.21)

Now, ∅𝐺𝐺(𝑟𝑟, 𝑡𝑡) can be replaced by an encoding field gradient �⃗�𝐺 (𝑡𝑡) to give

𝑠𝑠(𝑡𝑡) = ∫ 𝐶𝐶 (𝑟𝑟) 𝑃𝑃(𝑟𝑟) 𝑒𝑒𝑗𝑗 𝛾𝛾 𝑟𝑟 ���⃗ ∫ 𝑑𝑑�́�𝑑𝑡𝑡0 �⃗�𝐺(�́�𝑑) 𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.22)

Now, the k-space can be represented as follows

Page 39: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

22

𝑘𝑘�⃗ = 12𝜋𝜋𝛾𝛾 ∫ �⃗�𝐺(�́�𝑡)𝑑𝑑

0 𝑑𝑑�́�𝑡 (2.23)

Therefore, the encoded k-space signal can be represented as

𝑠𝑠�𝑘𝑘�⃗ � = ∫ 𝐶𝐶 (𝑟𝑟) 𝑃𝑃(𝑟𝑟) 𝑒𝑒𝑗𝑗 2𝜋𝜋 𝑘𝑘�⃗ 𝑟𝑟 ���⃗ 𝑑𝑑𝑟𝑟,𝑉𝑉𝑉𝑉𝑉𝑉 (2.24)

2.3.5 2D Spatial Encoding

The 2D spatial encoding method is the process of encoding to generate the k-space

signal𝑠𝑠�𝑘𝑘�⃗ �. As observed from Eq. (2.23), the full k-space signal can be produced by

changing the encoding gradient or the time duration. In 2D slice selective encoding, az-

directional gradient set as a readout gradient to produce

𝑃𝑃(𝑥𝑥, 𝑦𝑦) = ∫ 𝑃𝑃(𝑥𝑥, 𝑦𝑦, 𝑧𝑧) 𝑑𝑑𝑧𝑧𝑎𝑎2− 𝑎𝑎2

(2.25)

Where 𝑎𝑎 is denoted as slice selection thickness. 2D encoding techniques can be applied

after a slice is selected. Here, The Fourier encoding technique is explained to

demonstrate the principle of spatial 2D encoding.

𝑘𝑘𝑥𝑥 = 𝛾𝛾𝐺𝐺𝑥𝑥(𝑡𝑡 − 𝑇𝑇𝑒𝑒) (2.26)

𝑘𝑘𝑥𝑥 = 𝛾𝛾𝛾𝛾∆𝐺𝐺𝑥𝑥𝑇𝑇𝑝𝑝𝑒𝑒 (2.27)

Where 𝐺𝐺𝑥𝑥 is denoted as frequency gradient, 𝑇𝑇𝑒𝑒 is frequency encoding time, ∆𝐺𝐺𝑥𝑥 is thei-

th phase encoding cycle and 𝑇𝑇𝑝𝑝𝑒𝑒 represent the phase encoding time. Hence, x-direction

is used as frequency encoding,whereas y-direction performs as phase encoding. Thus,

the frequency and phase-encodedk-space can be illustrated as

Page 40: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

23

𝑠𝑠�𝑘𝑘𝑥𝑥 ,𝑘𝑘𝑥𝑥� = ∬ 𝐶𝐶 (𝑥𝑥, 𝑦𝑦) 𝑃𝑃(𝑥𝑥, 𝑦𝑦) 𝑒𝑒𝑗𝑗 2𝜋𝜋 (𝑘𝑘𝑥𝑥𝑥𝑥+𝑘𝑘𝑥𝑥𝑥𝑥) 𝑑𝑑𝑥𝑥 𝑑𝑑𝑦𝑦,𝐹𝐹𝑉𝑉𝑉𝑉 (2.28)

Where FOV denoted as the 2D field of view of the receiver coil. In the single-channel

receiver coil, the area of FOV has to be equal to the region of interest (ROI). However,

when multiple channels are used for imaging, the encoded k-space equation will be

⎩⎪⎪⎪⎪⎨

⎪⎪⎪⎪⎧ 𝑠𝑠0�𝑘𝑘𝑥𝑥 ,𝑘𝑘𝑥𝑥� = ∬ 𝐶𝐶0 (𝑥𝑥, 𝑦𝑦) 𝑃𝑃(𝑥𝑥, 𝑦𝑦) 𝑒𝑒𝑗𝑗 2𝜋𝜋 �𝑘𝑘𝑥𝑥𝑥𝑥+𝑘𝑘𝑥𝑥𝑥𝑥�𝑑𝑑𝑥𝑥 𝑑𝑑𝑦𝑦,

𝐹𝐹𝑉𝑉𝑉𝑉

𝑠𝑠𝑑𝑑�𝑘𝑘𝑥𝑥 ,𝑘𝑘𝑥𝑥� = ∬ 𝐶𝐶𝑑𝑑 (𝑥𝑥, 𝑦𝑦) 𝑃𝑃(𝑥𝑥, 𝑦𝑦) 𝑒𝑒𝑗𝑗 2𝜋𝜋 �𝑘𝑘𝑥𝑥𝑥𝑥+𝑘𝑘𝑥𝑥𝑥𝑥�𝑑𝑑𝑥𝑥 𝑑𝑑𝑦𝑦,𝐹𝐹𝑉𝑉𝑉𝑉

𝑠𝑠𝐿𝐿−1�𝑘𝑘𝑥𝑥 ,𝑘𝑘𝑥𝑥� = ∬ 𝐶𝐶𝐿𝐿−1 (𝑥𝑥, 𝑦𝑦) 𝑃𝑃(𝑥𝑥,𝑦𝑦) 𝑒𝑒𝑗𝑗 2𝜋𝜋 (𝑘𝑘𝑥𝑥𝑥𝑥+𝑘𝑘𝑥𝑥𝑥𝑥) 𝑑𝑑𝑥𝑥 𝑑𝑑𝑦𝑦,𝐹𝐹𝑉𝑉𝑉𝑉

(2.29)

Where L is the number of receiver channel, 𝑠𝑠𝑑𝑑�𝑘𝑘𝑥𝑥 ,𝑘𝑘𝑥𝑥� is the received signal at the l-th

channel in the array, and 𝐶𝐶𝑑𝑑 (𝑥𝑥, 𝑦𝑦) is the sensitivity function of the l-th channel and this

value are not same for all channel. This multiple coils system are used in parallel MRI

to accelerate image processing.

2.3.6 Field of View (FOV) and Spatial Resolution

The resolution and the FOV of the final image are determined by the highest spatial

frequency sampled (kmax) and sample rate (Δk)respectively,which shown in Figure

2.6.The sampling rate in the frequency encodes direction (x) is determined by the ADC

receiver, where the aliasing problem rarely exists. But, the sampling rate in the phase

encoding direction (y) is determined by the magnitude of the k-space shift executed by

the phase encoding gradient. The phase encoding process takes a long time to acquire

phase directional data. To minimize the number of steps, a fixed extent of k-space are

used to reduce the scan time. There is two way to reduce the scanning time; we have to

take as large step as possible or minimize the FOV.

Page 41: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

24

Figure 2.6: The image resolution and FOV

However, if the FOV has no ability to contain the whole image, it is represented as an

aliasing image. And using different reconstruction technique, the full-size image can be

reconstructed. New decade are emphasising on the area to accelerate the MRI

processing.

The appropriate selection of the field of view (FOV) and the spatial resolution is very

important in the MRI data acquisition. The FOV should be sufficiently larger than the

dedicated size of the image to remove the aliasing artifacts as well as the voxel size

should be enough small to find the smallest features we are intent to observe.

If we consider the frequency domain spacing is Δk between two samples in k-space,

then the FOV will be FOV = 1/Δk and the voxel size (denoted as Δ) is inversely

proportional to the range of k-space (-kmax to kmax) of the sampled object. That mean ∆ =

1/(2kmax). The physical illustration of those parameters are presented in Figure 2.7. We

need more scan line and length to improve the quality of image because a larger FOV or

smaller voxels in the phase encode direction require more scan lines, and smaller voxels

in the frequency encode direction increases the length of those scan lines.

Also, image quality is affected by the signal-to-noise ratio (SNR), if SNR is too low,

tissue contrast and abnormal regions are troublesome to differentiate from observation

noise. SNR is roughly proportional to acquisition time; therefore, reducing the

Page 42: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

25

acquisition time is in the midst of the same reduction in SNR. The degradation in SNR

is going to be a significant concern in reconstructing quality pictures from accelerated

MRI information.

Figure 2.7: The sample spacing ∆kxin k-space and the extent kx,max, relate to the FOV, FOVxand voxel size ∆x, respectively, of the reconstructed image in the x-direction. Similarly,∆ky and ky,maxare connected to FOVyand ∆y.

2.3.7 MR Imaging Pulse Sequence Figure 2.8shows a typical timing diagram of gradient-echo (GRE) MRI pulse sequence

to generatek-space data from objects. This timing diagram design with an RF pulse, a

slice selection gradient signal, a phase gradient pulse, a frequency gradient pulse and, an

output signal. At first slice selection, z-gradient is applied in a specific part of the

objectto create a linear variation of the field along the z-direction in the presence of the

uniform magnetic field B0,and at the same time, RF pulse also applied. And this applied

RF pulse which is usually 0 to 900flips the net magnetization of a particularslice in the

transverse (x-y) plane.

After applying RF pulse, a y-directional gradient pulse is applied for some time to

generate phase variation along the y-direction. In this stage,the frequency encoding

signal is applied which is called a readout signal. As a result, the net magnetization

precesses along x-direction with different frequency. The total procedure is repeated

with different phase gradient pulse each time,and k-space is filled up correspondingly

Page 43: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

26

through the analogue to digital converter (ADC). For example, 256 phase encoding line

and 256 sampling points acquired during the readout process for a 256×256 dimensional

image.

Figure 2.8: Typical pulse sequence for MRI data acquisition

2.4 Image Reconstruction Technique of MRI

2.4.1 Reconstruction Complexity MRI is one of the leading age technology in biomedical diagnosis field. It has many

advantages,such as high resolution, radiation-free , flexible slice orientation [5].

However, MRI has somecritical problem; long data acquisition time is one of those.

Suppose, If the length of TRcycle is one second,then a total of 256 seconds or 4 minutes

and 16 seconds require for 256 phase encoding steps. Some application use repeated

phase encoding steps to improve the signal to noise ratio (SNR). Therefore, the total

scanning time will be double, when each phase encoding steps are repeated twice.

Besides, in multi-slice or volume imaging, the multi-sliceacquisition is essential to

acquire k-space data which extend the total scanning time up to one hour.Generally,

Page 44: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

27

contemporarytime reducing scanning methods of multi-coil scheme take advantages in

modern MRI process.

The advanced MRI scanner is equipped with phased array coils,and these phased array

coils can collectdata from multiple coils simultaneously.Each phase array coils has an

individual sensitivity profile which is used to emphasizes some part of the proton

density. A full image from all coils can obtain by the sum of square (SOS) of individual

channel image,and the SOS of the image is displayed the proton density with uniform

distribution weights. There are several methods have been proposed to reduce the data

acquisition time such as data sub-sampling of Cartesian trajectory, using non-Cartesian

trajectory, novel pulse sequence, and combinations of those methods.

2.4.2 Cartesian Trajectory The Cartesiantrajectoryis the spatial transform of the magnetization by acquiring a

Cartesian grid of samples. In this technique, the frequency encoding lines are acquired

one by one that means parallel to each other. Also, the distance between the linesis the

same. However, the phase encoding line is perpendicular to the frequency encoding

line. This method is widespread for generating images from the object, and many pulse

sequences techniquesare now present to implement Cartesian trajectory. The critical

part of the Cartesian and other Fourier sampling methods is the use of spatial gradient

fields during the relaxation of excited spins. As we describe above, the spacing and

extent of the Cartesian grid of samples both affect the acquisition time and the field of

view and a voxel size of the resulting image.

In the Cartesian coordinates, the brightness and the main structure of the image are

defined by the centre part of the frequency encoding lines of the k-space, whereas, the

details of the images are described by the peripheral of the k-space of the frequency

encoding line. In a nutshell, the different frequency encoding line in the k-space has a

different importanceto reconstruct the image. The quality of the reconstructed image

will not be profoundly affected due to the dropped out of some frequency encoding

lines.However, the image quality will be significantly changed due to discarding the

centre part of the frequency encoding lines of the k-space.

Page 45: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

28

The image reconstruction technique can be implemented easily and fast by applying 2D

Fourier transform. However, apparently, the Cartesian coordinates of the MRI data

acquisition techniquesuffer from long scan time for multiple TR and the phase encoding

steps. To increase the data acquisition speed, some frequency encoding steps are

dropped out deliberately.

2.4.3 Non-Cartesian Trajectory

The systems that assemble data in the Cartesian coordinates are Multi-slice imaging,

The Gradient Echo Sequenceand Volume Imaging. The frequency encoding lines are

parallel to each other, and the distances between adjacent frequencies encoding line are

identical. The phase encoding direction is perpendicular to the frequency encoding

direction. Images are reconstructed using the 2D Fourier transform from data acquired

with the trajectory in the Cartesian coordinate, and this is fast and not very hard to

implement.

However, trajectories within the Cartesian coordinates experience the lengthy scan time

equal to the multiplication of the repetition time (TR) and the total phase encoding steps.

The frequencyencoding lines at the middle of the Fourier domain decides the main

structure and brightness of the image, while the frequency encoding lines at the

periphery of the Fourier domain describes the facts of the image. In other words, the

importance of each frequency encoding lines is different in the k-space. Therefore, the

significance of varying frequency encoding lines across the Fourier transform is not

identical.

For example, certain frequency encoding lines are discarded at the periphery of the

Fourier transform, the resolution of the image will not be heavily influenced. However,

if some frequency encoding lines are removed at the centre part of k-space, the quality

of the image will be severely affected. Figure 2.9 below indicates a radial trajectory of a

signal structure in the non-Cartesian coordinates.

Page 46: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

29

Even though the time used for the non-Cartesian trajectories is not very high, the signal

reconstruction procedure is not direct. Supplementary stages are essential toreconstruct

images from acquired data using trajectories in non-Cartesian coordinates. Therefore,

the use of trajectories in non-Cartesian coordinates is constrained to certain processes. A

large chunk of scans is accomplished using trajectories in the Cartesian coordinates [42-

45]

Figure 2.9: the trajectory pulse sequence in the non-Cartesian coordinates

2.4.4 Noise and System Uncertainty in Image Reconstruction Noise in reconstructed image in MRI has been analysed in different works [21-23]. The

noise in the acquired data sample is generally exhibited as white additive noise with

variance

𝓋𝓋2 ∝ 4𝑘𝑘𝑇𝑇 × 𝑅𝑅 × 𝐵𝐵𝐵𝐵 (2.30)

Where 𝓋𝓋 is denoted as noise, R is the resistance loaded into the coil, coil resistance and

other electronic resistance. The BW is the bandwidth of the given imaging scheme

directed by the imaging parameters.

The discrepancy between mathematical estimated model and reality are described as

imaging scheme uncertainty. It is denoted as∆𝒯𝒯. For example, in the conventional two

dimensional Fourier encoding have many possible imperfect imaging conditions that

Page 47: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

30

can deviate the real encoding process 𝒯𝒯 away from the perfect Fourier basis. There are

some reasons behind this imperfection, such as field inhomogeneity of B0, the motion of

the imaged substance and selective excitation offset. In the model, these issues need to

resolve precisely in order to obtain the perfect reconstruction.

In exercise, these type of effects can reduce significantly by correctly choosing pulse

sequence in the current state of the art single receive coil scanners. Generally, the

inverse discrete Fourier transform (DFT) and Singular Value Decomposition (SVD) can

be used for image reconstruction without significantly affecting image quality.

However, due to the unknown coil sensitivity in pMRI, it needs to be estimated. And,

the estimated error is generally significant compare to a single receive coil scanners.

Therefore, in the applied parallel MRI system, both the noise and the system uncertainty

have to analyse explicitly. The right imaging function is thus given by

𝒯𝒯 = 𝒯𝒯0 + ∆𝒯𝒯, (2.31)

Where 𝒯𝒯0 is the estimated function, and the image reconstruction function is specified

by

𝒥𝒥 = 𝒯𝒯0−1. (2.32)

The encoded data measurement can be written as

𝒮𝒮 = 𝒯𝒯{𝑃𝑃} + 𝓋𝓋

= 𝒯𝒯0{P} + ∆𝒯𝒯{P} + 𝓋𝓋. (2.33)

Therefore, the reconstructed image function is given by

P� = 𝒥𝒥{𝒮𝒮}

= 𝒥𝒥𝒯𝒯0{P} + 𝒥𝒥∆𝒯𝒯{P} + 𝒥𝒥𝓋𝓋.

Page 48: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

31

= P + 𝒥𝒥[𝓋𝓋 + ∆𝒯𝒯P], (2.34)

Also, the corresponding reconstruction error is given by

ℰ = 𝒥𝒥[𝓋𝓋 + ∆𝒯𝒯P]. (2.35)

As can be observed from Equation (2.35),the error ℰ is not only dependent on the size

of noise 𝓋𝓋 and uncertainty function ∆𝒯𝒯 but also the gain of the reconstruction system𝒥𝒥.

The errorℰ can be significant if the gain of reconstruction system 𝒥𝒥 is substantial.

However, in the single receiver case, the error ℰ is not an important issue because the

hardware setup for both 𝓋𝓋 and ∆𝒯𝒯 are minimal, and the gain of the reconstruction

system is 1. Hence, there is no amplification.

2.5 Image Reconstruction from Subsampled Data in

Cartesian Coordinates

There is about 128 or 256 phase encoding line normally utilized to form an image in the

Cartesian coordinates. In order to decrease the time for acquiring the image signal, some

frequency encodinglines are intentionally escaped. The factor by which the number of

frequency encodinglines is reduced is referred to as the reduction factor R, which is

defined by dividingthe number of total frequency encoding lines by the number of

actual frequencyencoding lines in the actual scan.Reconstructedimages will have

artifacts by directly applying two-dimensional inverse Fourier transform to the under-

sampled k-space data. Specifically designed image reconstruction procedures are

required to reconstruct images with satisfactoryquality from under sampled k-space

data. SENSE [19], GRAPPA [20], CS-based reconstruction[35] and their some

extensions are the main focus for image reconstruction methods in MRI.

2.5.1 SENSE SENSitivity Encoding is a concept of reconstructingfull FOV images from sub-sampled

k-space data of multiple coils by using the sensitivity profile of coils[19]. If the k-space

Page 49: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

32

signal is under-sampled, the reconstructed signal has aliasing. For example, the images

are shown in Figure 2.10 (a) is initially transformed to k-space, and the k-space data are

under-sampled by assembling one for every two-frequency encoding lines. The image

are reconstructed by using 2-Dimensional inverse Fourier transform is depicted in

Figure 2.10 (b). The two bright pixels in the original image are vertically

symmetrical. The reconstructed image in (b) portrays an image that shows that these

two pixels have overlapto the point that they can no longer be differentiated. Actually,

the top and bottom halves are combined to form a 1/2 FOV image.

(a) The original image (b) The reconstructed image

Figure 2.10: Comparison between the original and the subsampling images

SENSE has two steps to unfold the reconstructed image with aliasing. In the first step of

this technique, the aliased image of each coil is reconstructed by applying 2-

Dimensional inverse Fourier transform to the under-sample k-space data. The second

step involves the reconstruction of the whole FOV image from this intermediate aliasing

images.

The SENSEprocess is demonstrated in the following example. Assume that the two

bright pixels in the figure above are represented by x(1) and x(2). Let’s also assume that

the scanning process is done using two coils for which their sensitivities in these bright

regions are r1(1), r1(2) and r2(1), r2(2) corresponding to the first and second coil

respectively. Also, suppose that 𝑥𝑥�1and 𝑥𝑥� 2represents the overlapped pixel in the two

Page 50: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

33

intermediate aliased images. The process of aliasing is illustrated using the equation

below:

r1(1)x(1)+ r1(2)x(2)= 𝑥𝑥�1, (2.36)

r2(1)x(1)+ r2(2)x(2) = 𝑥𝑥�2. (2.37)

In the matrix format, the above expressions are represented as

�𝑟𝑟1(1) 𝑟𝑟1(2)𝑟𝑟2(1) 𝑟𝑟2(2)� �

𝑥𝑥(1)𝑥𝑥(2)� = �𝑥𝑥�1

𝑥𝑥�2� (2.38)

In the above matrix expression, two linear expressions and two unknown variables

exist. This issue can be easily solved by multiplying the inverse matrix of the first

matrix expression on the LHS to both sides of the equation. The number of the coils, Nc,

must not be lower than R (Reduction factor). IfNc ≥ R, there are an equal number of

equations or more exists as the unknown variable. Using a pseudo-inverse methodology,

the unknown variables can be obtained. However, ifNc< R, the image reconstruction

issue may not be certain. To solve the equation, the pseudo-inverse-matrix or inverse

matrix is required to be present.

The precondition of the SENSE technique for successfulimage reconstruction process is

the precise sensitivity profile of coils. Assume that the sensitivity profiles for the coils

are not known, the process inversion cannot be carried out. A prescan can be exploited

to establish the sensitivity profiles of coils such that signals with a lower FOV are

obtained[19]. But then, in the course of the real scan, sensitivity profiles could be

dissimilar from the prescan to some extent due to many factors, such as the body

movements of the patients. Hence, the author suggestscreating a sensitivity profile

during the actual scan, by adopting variable density sub-sampling pattern, which means

collecting more frequency encoding lines the low frequency part of k-space, instead of

sub-sampling them the same way as in the high frequency part[46]. Sincethe calibration

data are acquired simultaneously with the data to be reconstructed,errors due to

sensitivity miscalibration are eliminated.

Page 51: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

34

Although these two methods differ in the time when the scanning for

sensitivityinformation is conducted, both share a similar sensitivity estimation

procedure. First,a low resolution image for each coil is reconstructed from k-space data

obtained bythe prescan or the low frequency part. Then the reference image is obtained

by theSOS of the sensitivity encoded images from all coils. The sensitivity profile of

eachcoil is calculated by dividing the sensitivity encoded image by the reference

image.An interpolation step is finally adopted to increase the size of the sensitivity

profiles.

One of the successful application of SENSE is real-time imaging of cardiac

diagnosis[47] and also, some other clinical practice[48, 49]. However, the sensitivity

profile is difficult to estimate in some application to fulfil the requirement of accurate

sensitivity.

2.5.2 Extensions of SENSE A lot of extensions have been utilized with the conceptualization of the SENSE

technique. Some of these extensions are introduced in this thesis.

2.5.2.1 2-Dimensional SENSE In [50], the authors propose to extend SENSing Encoding to 2-Dimensional under-

sampling, called 2D SENSE, which is an under-sampling technique used in 2-

Dimensions.Since there are two phase encoding directions in volume imaging, under-

sampling can be done in either direction. Hence, aliasing happens in both directions.

There are two successive steps in 2-Dimensional SENSE. And each of these steps

unfolds unidirectional aliasing[50].

2.5.2.2 Auto-SENSE Auto-SENSE was proposed in [51] and has been applied in cardiac imaging, which is

known as dynamic MRI. Sequential frames for dynamic MRI are acquired in a time

Page 52: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

35

series.A frequency encoding line is shifted when under-sampling k-space in adjacent

even or odd frames. Just like noted before, a higher frequency region of the k-space

decides the details of the image while the lower frequency region of the k-space is the

deciding factor of the main shape of the image. In dynamic MRI, it is understood that

the modifications between two neighbouring frames are significantly low.

Consequently, the lower frequency regions of k-space from neighbouring frames are

thought to be the same or very similar, while only the high-frequency region of Fourier

transform data alter seriously[51]. After estimated sensitivity profile of the coil, the

SENSE reconstruction technique is applied in both the odd and even frame to generate

images. Auto-SENSE can save data acquisition time for avoiding prescan, and

sensitivityprofiles can be dynamically updated, which leads to fewer artifacts and less

noise.

2.5.2.3 Tikhonov Regularization This is an image reconstruction algorithm that effectively decreases the loss of Signal-

to-Noise-Ratio (SNR) as a result of certain geometric connections in the spatial

information from the coil elements. In this algorithm, a reference image is utilized as

preceding information about the reconstruction image in order to provide regularized

approximates for image reconstruction. Pre-scanning is a method of getting this

reference image. It is frequently understood that this technique can decrease g-factors

and raise image resolutions in SENSE image reconstruction. In a characteristic SENSE

method, precise sensitivity profiles should be supplied before the image reconstruction

occurs[52].

2.5.3 GRAPPA GeneRalized Autocalibrating Partially Parallel Acquisitions (GRAPPA) is a widely

used image reconstructionmethods from the under-sampled k-space data [20]. Contrary

to SENSing Encoding, GRAPPA does not need precise data on sensitivity profiles. As a

substitute, GRAPPA needs that the lower frequency region of the k-space is fully

sampled. An additional milestone that differentiates the SENSE image reconstruction

Page 53: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

36

scheme and the GRAPPA method is that SENSE reconstruction is majorly completed in

the image domain while GRAPPA is majorly finished in the Fourier domain.

The process of reconstruction in GRAPPA is demonstrated below. Phase encoding

direction is shown by the horizontal region while the various coils are shown in the

vertical region. Perpendicular to the plane is the frequency encoding direction. Those

circles with holes in them denote the frequency encoding lines that were unsampled.

Those circles with a totally black solid grey or black colour represent sampled

frequency encoding lines. The solid black circle follows the same sampling module

used in SENSE meaning that an R line is obtained. Those circles that are grey represent

the additional calibration signal lines (ACS) which are deliberately obtained in

GRAPPA.

Figure 2.11: Schematic diagram of GRAPPA reconstruction technique

On a broad basis, GRAPPA image reconstructionmethod has two vital steps. The

GRAPPA weights are defined by an interpolation procedure in the very first step. Just

as showed in the diagram above, the solid black circles obtained from all coils in the

neighbouring phase encoding steps are utilized to interpolate the grey circles in the ACS

line of coil 4.The method of interpolation is represented by

∑ ∑ 𝑠𝑠𝑖𝑖4𝑗𝑗=𝑑𝑑

4𝑖𝑖=𝑑𝑑 (𝑗𝑗)𝑤𝑤𝑖𝑖(𝑗𝑗) = 𝑠𝑠4,𝐴𝐴𝐴𝐴𝐴𝐴, (2.39)

Page 54: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

37

i and j respectively denote the coil and phase encoding steps. si(j) represents the points

in the i of j domain obtained black data in the coil and 𝑠𝑠4,𝐴𝐴𝐴𝐴𝐴𝐴 ,stands for the grey data in

the ACS line of coil 4 and 𝑤𝑤i(j)stands for the weights of GRAPPA.

Each of the solid black circle data points is assigned a weight to stabilize the right- and

left-hand part of the equation of interpolation. For the scenario, a total of about 16

weights are to be determined, but just one equation is needed. There are about 15 free

variables that mean the equation has endless solutions. The course of interpolation is

iterated for all data points in the frequency encoding direction. Suppose there are 256

data points in the frequency encoding direction, meaning that there are 256 equations

but only 16 unknown variables, which makes the problem of finding GRAPPA weights

over-determined. In GRAPPA, it is assumed that the GRAPPA weights across the k-

space are the same. In addition to moving along the frequency encoding direction, the

same interpolation process is repeated along the phase encoding direction in the fully-

sampled low-frequency part of k-space. Since there are more equations than unknown

variables, the GRAPPA weights canbe obtained by a pseudo-inverse step, as in SENSE.

With the same procedure, theGRAPPA weights for interpolating the gray data points in

the other 3 coils areobtained.

After when the interpolation step has been finished, the points where the data are

missing are rebuilt. Those points are kept on the right part of the equation (2.17) in

place of the ACS lines, which contains the grey data points. Because the weights and

black data points are recognized, the points where the data are missing are simply

computed by using simple and straight forward mathematical process. Finally, 2-

Dimensionalinverse Fourier transform is applied to transform the reconstructed image

using the data in the k-space. The greatest benefit of GRAPPA in contrast to SENSE is

that GRAPPA does not require the coil’s precise sensitivity profiles. Rather, it uses a

variable density under-sampling module that acquires the ACS lines in the lower

frequency region of k-space. It is advised that in circumstances where it is possible to

approximate the sensitivity profiles, then it is also advised first to consider SENSE.

2.5.4 Extensions of GRAPPA

Page 55: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

38

From the time when GRAPPA was introduced, a lot of other extensions for

reconstruction images have also been introduced to make the ease of image

reconstruction a tangible reality. This thesis also strives to point out those extensions.

TGRAPPA[53] is an example of such an extension. This extension was introduced for

the purpose of applying GRAPPA in dynamic Magnetic Resonance Imaging. K-space

data are obtained at different times in dynamic MRI, which is normally being obtained

in ahigher dimensional k-t space. There are no extra ACS lines obtained in TGRAPPA.

The combination of the obtained frequency encoding lines in the lower frequency region

of the k-space in every R frames is one way of obtaining the ACS lines. Upon the

scanning of a new frame, the oldest frame is discarded. Hence, as time advances, the

ACS lines are upgraded. When the ACS lines are upgraded, GRAPPA weights can be

estimated and applied in every frame. The reconstruction principle of TGRAPPA is

demonstrated in the diagram below (Figure 2.12).

Figure 2.12: The image reconstruction method of TGRAPPA

The extension, k-t GRAPPA is introduced in a journal [54]. In contrast to the

TGRAPPA technique, k-t GRAPPA obtains ACS lines in each frame. This method

utilizes data from neighbouring time frames, in conjunction with data from

neighbouring blocks, for the interpolation of GRAPPA weights. The image

reconstruction technique of k-t GRAPPA is demonstrated in Figure 2.13.

Page 56: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

39

Figure 2.13:The image rebuilding method of k-t GRAPPA

The extensions k-t GRAPPA and TGRAPPA may be applied in either multiple or one

coil because since they also explore information in time-space. For the GRAPPA

extension, data obtained in neighbouring phase encoding steps in all coils are employed

to interpolate one data point in the ACS line of a coil. In a journal [55], they suggest not

only obtained data in neighbouring phase encoding steps of every coil are to be utilized,

but they also suggest that data obtained in the same frequency encoding lines are also to

be utilized, because close to the ACS line is the data point. By bringing in additional

data points during the course of interpolation, GRAPPA weights become increasingly

precise. The weights in the GRAPPA extension are understood to be uniform at every

region. When the weights are estimated, they are used across the whole k-space. In a

journal [56]titled “[55],” Magnetic Resonance in Medicine, the writers assert that for

higher and lower frequency regions of k-space, the differences should be reflected upon,

meaning that a plethora of varying GRAPPA weights is to be estimated and used in

these areas.

In order to get a better image reconstruction outcome, there should be a segmentation of

ACS lines along the frequency encoding direction. The GRAPPA weights in different

regions are to be computed individually. Furthermore, the Fourier transform data are be

sampled out of the ACS area, using varying factors of reduction in accordance with

locations.

According to the journal [57]titled “2D-GRAPPA-Operator for faster 3D parallel

MRI,” Magnetic Resonance in Medicine”, authored by M. Blaimer et al and published

in 2006, GRAPPA is used in 3-Dimensional Fourier transform data, and called 2D

Page 57: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

40

GRAPPA Operator or 2-Dimensional GRAPPA, in relation to the way of reconstruction

image from under-sampledFourier transform data in two-phase encoding directions. For

3-Dimensional Fourier transform data, asnoted before, under-sampling in two-phase

encoding directions is utilized to advance the gross rate of reduction. Just like the first

GRAPPA extension, 2-Dimensional GRAPPA weights are obtained using the method of

interpolating the data points in ACS lines using the data obtained in every direction.

This procedure is called 2-Dimensional GRAPPA because it is fairlyrelated to the

reconstruction process from under-sampled unidirectional Fourier transform.On the

other hand, the reconstruction process is divided into two individual successive 1-

Dimensional GRAPPA operations. For each of the GRAPPA operation, it reforms the

data missing in a unidirectional fashion. This is known as 2-Dimensional GRAPPA

Operator.

So far, no extension arising from GRAPPA has probed into this issue of noise.

Realistically data obtained by scanning, are usually damaged by noise, leading to

incorrect GRAPPAweights. In order to determine the off points, the authors of the

journal “Robust GRAPPA reconstruction and its evaluation with the perceptual

difference model,” devised a methodology named RobustGRAPPA[58]. Using this

methodology, the above points are multiplied by zero or a small constant so the weights

can then be estimated with higher precision because GRAPPA weights are supposed to

be uniform in the whole k-space, it is feasible to approximate GRAPPA weights from

reconstructed lines that are missing. GRAPPA weights can be calculated with

theconsideration of reconstructed missing data. GRAPPA weights are initially estimated

from a small number of ACS lines, and then the missing frequency encoding lines are

reconstructed[59].

2.5.5 Compressive Sensing based Image Reconstruction

Method In accordance with the current proposition of the compressive sensing theory, a signal

can be reconstructed with an overwhelming probabilityfrom data which are randomly

sampled if the original signal is sparse. Images obtained by the method of MRI are

intrinsicallysparse, such as angiography, or compactable following some types of

Page 58: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

41

change, such as brain images. Hence, it is straightforward to use compressive sensing in

the problem of image reconstruction from under-sampled k-space data in Magnetic

Resonance Imaging.

We can reconstruct images by minimizing the ℓ1 norm of the transformed image, subject

to data reliability restraints. In the compressive sensing based image reconstruction

process, the k-space data is under-sampled by a variable density under-sampling

procedure[32]. The equation for the image reconstruction problem can be represented as

𝑚𝑚𝛾𝛾𝑚𝑚||Ψx||1 s.t.||TFx−𝑦𝑦||2< ε (2.40)

x represents the objective image, while Ψ represents the wavelet transform operator. TF

is the partial Fourier transform operator obtainedby sub-sampling the fulldiscrete

Fourier matrix using variable density under-sampling method, ε is that factor that relates

to noise and 𝑦𝑦 is the measured data.In order to adequately address the equation above, it

is paramount first to restrain the problem to the unrestrained form:

𝑚𝑚𝛾𝛾𝑚𝑚λ||Ψx||1 + || TFx−𝑦𝑦||22 (2.41)

where λ is the factor in ascertaining the exchange between data sparsity and data

fidelity. λ is ascertained by working out the equation above with different numbers, and

then selecting λ so that the reliability of the data in Eq. (2.18) is addressed. The authors

propose to solve Eq. (2.19) by using a conjugate gradient descent algorithm with

backtracking line search. The function f(x) is characterized as in the equation above.

The conjugate gradient ∇f (x) is

∇f (x) = λ∇||Ψx||1+ 2𝑇𝑇𝐹𝐹∗(TFx−𝑦𝑦) (2.42)

The ℓ1 norm is described as the sum of absolute values of entries, not smooth for every

x value. Hence, the ℓ1 norm is estimated using a smooth function expressed as |𝑥𝑥| =

�𝑥𝑥 ∗ 𝑥𝑥 + ξ. ξ is termed the positive smoothing parameter. Using this estimation, the

conjugate gradient of the ℓ1 norm can be written as

Page 59: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

42

∇||x||1 = 𝑥𝑥�𝑥𝑥∗𝑥𝑥+ξ

(2.43)

Hence, ∇f (x) of function x can be represented as

∇f (x)= λ Ψ∗Ψx�(Ψx)∗(Ψx)+ξ

+ 2𝑇𝑇𝐹𝐹∗(TFx−𝑦𝑦) (2.44)

In [32], it is demonstrated that the enhanced spatial resolution and accelerated

acquisition for multi-slice fast spin echo brain imaging and 3D contrast enhanced

angiography are attained. The conjugate gradient descent algorithm with backtracking

line search is briefly described in Algorithm 3 [32].

2.5.6 Extension of CS-based Image Reconstruction Method A lot of extensions for the Compressive Sensing based image reconstruction scheme

have been reported over the decade. In[35], the authors explain the requirements for

successful CS reconstruction.They describe the natural fit of CS to MRI, and give an

intuitive understanding ofCS reconstruction by describing it as a process of interference

cancellation. Some example and applications of Compressive Sensing in MRI and

certain parameters influencing the use of Compressive Sensing in MRI are introduced

here. Some of the restrictions experienced here are forced by a malfunctioning of the

MR hardware, the characteristics of different types of images, and clinical concerns.

Similar to GRAPPA and SENSE, it is upfront to use Compressive Sensing in dynamic

MRI. Dynamic MRI provides additional sparsity than traditional 2-Dimensional

imaging schemes, by making use of the k-space data in the temporal dimension, if only

that part of FOV gets altered at temporal rates which are high, while other regions stay

immobile or alter gradually. Random under-sampling for Compressive Sensing can be

appreciated by arbitrarily dodging frequency encoding lines in every dynamic

frame[60].

Page 60: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

43

On a broader basis, CS-based image reconstructionmethod from under-sampled k-space

data in MRI is computed by efficient algorithms. A primary issue with these algorithms

is the expensive and lengthy computational time[61]. In[62], the use of multicore CPUs

can resolve this computational cost and efficiency by forwarding multichannel data to

multicore CPUs. Also, GPUs can use to accelerate computation[63]. Parallel computing

is a promising method to reduce the reconstruction time of CS based reconstruction

methods when multiple coils are used in MRI.

2.6 Novel Sequence Design

This section introduces some novel sequence designs that are different from the

conventional Fourier transform scheme of data encoding.

Page 61: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

44

2.6.1 Bunched Phase Encoding In conventional 2-Dimensional sequences such as GRE, methodically to form an image

of size 256×256, a total of about 256 phase encoding steps are implemented. Also, 256

data points are acquired in the frequency encoding direction. In [63], the authors

propose a high rated data acquisition technique in the frequency encoding direction and

along a zigzag trajectory. Basically, data are acquired in a bunch of phase encoding

steps simultaneously, and hence, it is called BPE (bunched phase encoding)[63]. The

difference between an oscillating frequency encoding line in BPE and a straight

frequency encoding line in GRE is shown in the figure below. It is observed that the

solid black circles in this figure stand for individual data points in its place of frequency

encoding lines.

Figure 2.14: Schematic diagram of gradient and BPE encoding technique

In bunched phase encoding, additional data are acquired in the frequency encoding

direction while TR is unchanged. Hence, phase encoding steps is decreased, leading to

reduce data acquisition time. Even though BPE needs just a coil, it may combine with a

phased array of coils to provide an even higher factor of reduction. A combination of

BPE plus SENSE takes advantage of many coils in the system[64]. In the initial step,

the aliased image is reformed for each coil by BPE reconstruction. Then the aliased free

image is formed by SENSE by making use of the sensitivity profiles of the coils. In

BPE, tailored sequences and quickly varying gradients are needed to obtain bunched

data. These necessities restrict its use in regular MR scanners. It has been suggested by

some authors that the GRAPPA operator gridding (GROG) technique to create a cluster

Page 62: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

45

of data, that may be achieved using MR scanners[65]. More details explanation about

BPE are given in chapter 4.

2.6.2 Rotating RF Coil The RF pulse that is required to flip the net magnetization to the transverse surface is

formed by an RF coil, which is generally stationary in the scanner. The authors

recommend a rotating RF coil system in papers [66-68]. The RF coil spins around the

object being scanned for spin excitation and signal acquisition. The rotating RF coil

system may prevent the requirement for multiple channels and complex RF decoupling

of several coils[66-68]. The signal reconstruction from data with a rotating RF coil is

similar to the images reconstruction from data with the conventional RF system.

2.6.3 Wave-CAIPI A high reduction factor is hard to attain in conventionalpMRI. According to the journal

[69], partition encoding gradients and sinusoidal phase encoding occur concurrently in

wave- CAIPI during the readout of frequency encoding lines. Hence, inter-shifts are

formed by modifying the partition encoding and the phase encodingstrategy. By using

this scheme, highly efficient k-space data sampling can be achieved, which extends the

aliasing consistently in all three dimensions. Highly accelerated volume imaging with

low artifacts and lowSNRpenalties is gainedbecause wave-CAIPI completely utilizes

the spatial variation in coil sensitivities.

2.7 Some Limitations of the Present Research The use of undersampled data acquisition has the potential to reduce the total scan time

and cost of MR acquisitions and to reduce the burden of imaging to patients. This can

permit either greater patient throughput, or permit more imaging to be completed per

unit time. However still there are some limitations in the present workas listed below:

Page 63: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

46

o Parallel MRI, such as SENSE and GRAPPA, requires large number of coils to

achieve a high reduction factor. Due to inconsistency in coil geometries and

sensitivities, severe residual aliasing artefacts and amplified noise may occur when

high reduction factor is chosen.

o The non-array coil method such as BPE also has some crucial problem when

reduction factor is increased. The signal to noise ratio of BPE may become

considerably reduced as the reduction factor is increased.It has been observed that

asreduction factor R>2, the noise of reconstructed image gradually increases even if

the aliasing artefactis absent.

2.8 Conclusion The MRI scheme is an extensively imaging scheme in the context of recent medical

practice. It is proficient in revealing images of the human body with no invasive

procedure being undergone. This procedure has a challenge that is comparatively

extended data obtaining time as a result of physical restraints. Hence as to decrease the

time for obtaining data, under-sampling is used in the phase encoding direction.

SENSE, GRAPPA, and Compressive Sensing have been productively used to the MRI

image reconstruction from under-sampled k-space data. In traditional MRI with Fourier

encoding, the majority of the energy of the calculated data is situated in the lower

frequency region of k-space.

Hence, variable density under-sampling is usually employed in CS-based

reconstruction, which gets more data from the lower frequency region of k-space and

fewer data from the high-frequency region of k-space. 2-Dimensional image

reconstruction outcomes show that a reformed image from erratically under-sampled

BPE data is of higher resolution than that of variable densely under-sampled Fourier

transform-encoded data. This thesis has addressed the aspects of MRI analysis and CS-

based MRI image reconstruction scheme so as to expand in the perception of the feat of

Page 64: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

47

Compressive Sensing-based MRI image reconstruction applying under-sampled data

from Fourier encoding and BPE.

In this chapter, the fundamentals of MRI process has reviewed briefly. Here, we have

explained the MR physics from the generation of bulk magnetization to the observable

MR signal in the time domain. In addition, the spatial domain data is introduced, which

is known as k-space data of MRI. The sampling of k-space and data acquisition

procedure is also discussed where we have mentioned the limitation for image

acceleration. In addition, different kind of system uncertainty and noise related issue in

MRI processing has been discussed here.

Page 65: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

48

Chapter 3 Fundamentals of Sparse Sampling

Fundamentals of Sparse Sampling

3.1 Introduction

The purposes of this chapter are to identify the limitations in the current research work

of popular compressive sensing MRI schemes. At first, the fundamentals of

compressive sensing will be discussed, then the open problems in the existing literature

are analysed results and identified.

Compressive sensing (or compressed sensing, sparse sampling, or compression

sampling) (CS) is a relatively new methodology that has been used in many research

areas where MR imaging is one of them. One of the major advantages of CS

techniquethat permits the faithful reconstruction of the signal from the data acquired

below the Nyquist sampling rateif the signal domains are sparse also it can accelerate

the speed of data acquisition in MRI.

Magnetic Resonance Imaging (MRI) is one of the most useful imaging modalities that

have been extensively used in medical science for the acquisition of biomedical data

over the years[70]. MRI has capability to analyse the detail structure of a human body

as well as the metabolic processes in the body. So, the application of CSMRI

significantly reduces the time of image reconstruction as well as improved the quality of

images.Compressive sensing has been applied to MRI to accelerate the acquisition

process and has been demonstrated in various MRI procedures. CSMRI has been

applied in clinical settings for capturing the infant’s bio-data, where it has effectively

reduced the overall time needed for data diagnosis [71].

Page 66: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

49

3.2 Theory of Compressive Sensing

The present popular compression technology Compression Sensing(CS) was first

introduced inthe journal [26]titled “Compressed sensing” breaks the traditional twice–

bandwidth boundary set by the Nyquist-Shannon sampling theorem.Compressed

sensing has the special capacity to reform the data in compressed form and to

reconstruct the actual data from fewer measurements than usually required by the

Nyquist–Shannon sampling theorem.

There are several differences between the latest compression technique CS with the

Nyquist-Shannon Sampling Theorem. Such as, CS mainly emphases on data sensing

and recovering discrete signals with finite length, whereas the Nyquist-Shannon

Sampling Theorem typically considers continuous signals with infinite length. In

addition, CS measures data in the form of inner products between the original signal and

a sensing matrix, while in the Nyquist-Shannon Sampling Theorem the continuous

signal is sampled at equally spaced time points. In CS, the randomness of the sensing

matrix plays a vital role, which will be analyzed in this thesis. Finally, the two

frameworks differ in how the original signal is reconstructed. In CS, signal

reconstruction is achieved by solving nonlinear convex optimization problems, which

may be hardware-demanding and time-consuming. In the Nyquist-Shannon framework,

signal recovery is achieved through sinc interpolation, which is a linear process that

requires little computation and has a simple interpretation. Although the reconstruction

process of the original signal by CS appears to be a problem, it is not a major concern in

CS, since the aim of CS focuses on acquiring less data.

Articles published by Terence Tao, Emmanuel Candes and Justin Romberg on MRI

examined the random sub-sampling of the k-spaces[24-26, 35]. For a compressive

sensing technique to thrive, there have to be at least three significantrequirements:

•Sparse transformation: The intended signal should possess a partitioned

expression in a recognized domain of transformation (that is to say that the signal

should have the ability to undergo compression by transformation encoding)[30].

Page 67: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

50

•Inconsistency of sub-sampling artifacts: Linear reconstruction of data as a result of sub-sampling the Fourier coefficients should be inconsistent in terms of diffusion transformation field [30].

•Non-linear image reconstruction method: The signal must be reconstructed

using a non-linear approach which applies to a lower density of image expression

plus the coherence of the recreation of the acquired signals. The initial

circumstance is vividly fulfilled using MRI data just as clarified above. The

awareness of the importance of inconsistency, the awareness that the data

obtained from MRI can be designed in order to carry out inconsistent sub-

sampling and the fact which implies the existence of sophisticated and realistic

algorithms for data reconstruction may not be made available in this thesis.

Hence, we are turning to a very simple example. Compressed sensing does not

need a preceding idea of the signal or even any suppositions to take into account

to initiate the reconstruction process. A simple set of measures, from a technical

point of view, is all that is needed to begin the reconstruction process.

The signal must contain more than pure noise. Even if the process is more efficient

when parsimony rates make it possible to define elements like zero, it is still possible

that this theory and its applications remain applicable when the principal coefficients are

close to zero.

The process often begins when a weighted linear combination of samples, called

compression measurement, is taken on a different basis than the one on which the signal

is weak. The comparison process allows you to discover small measures, which usually

contain critical information to reconstruct the signal. As the domain is scanned, it

becomes possible to convert the image back to the intended domain[72].This is a

possibility even if the compression measurements will be less than the number of pixels

or other types of information found in the signal.

For the theory of compressed detection to work effectively, it is sometimes necessary to

apply a parsimony constraint. This process is initiated when a system of

underdetermined or undetermined linear equations must be solved. The application

Page 68: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

51

makes it possible to minimize the number of non-zero components and to find the

solution.

The premise of CS is a mathematical projection that attempts to explain how sparsity

signals are recovered when signals are under-sampled in accordance with the theory of

Nyquist.Assuming that the complex space Cn contains a signal denoted as x and also

assuming that the signal is in the Ψ domain such that

𝑠𝑠 = Ψ𝑥𝑥 (3.1)

Ψ is the 𝑚𝑚 × 𝑚𝑚 matrix indicating the sparse transformation. The signal is a Fourier

coefficient and they are well above zero in value. A system of measurement quantifies

the signal z in p dimensional space only by taking m projections of signal x as

𝑧𝑧 = Φ𝑥𝑥 (3.2)

𝑧𝑧 ∈ 𝐶𝐶𝑛𝑛 ,𝐾𝐾 < 𝑝𝑝 < 𝑚𝑚 and Φis the 𝑝𝑝 × 𝑚𝑚 matrix. Φis often sub-sampled in the magnetic

resonance imaging technique. The above expression can be further transformed to

𝑧𝑧 = ΦΨ*s, x =Ψ*s (3.3)

The symbol * means the transpose operation and the signal x is not very condensed in

the Ψregion. Image obtained using the MRI technique may be poorly characterized in

the wavelength domain with the aid of the wavelength transformation matrix.

The Theory of compressive sensing offers an exceptional solution to the problem

through the formula below.

min𝑥𝑥

||Ψ x̂||l1 (3.4)

||Ψx̂||l1∶= ∑ |𝑥𝑥𝛾𝛾|𝑖𝑖 is the l1 norm of x with x1 the ith element of the image reconstruction of

the signal x is achievable if certain mathematical conditions are met.

Page 69: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

52

3.3 Restricted Isometry in Compression Detection

An essential satisfactory condition for the precise reconstruction of the signal x is the

property called restricted isometry[24, 28, 73]. For the matrix Φ, the Restrictive

Isometry can be expressed as

(1 − δ𝑘𝑘)||𝑥𝑥||𝑑𝑑22 ≤ ‖Φ 𝑥𝑥𝑑𝑑22 ‖ ≤ (1 + δ𝑘𝑘)||𝑥𝑥||𝑑𝑑22 (3.5)

δ𝑘𝑘∈ (0, 1) is a constant known as the Restricted Isometry Property and the term l2 is

described as ||𝑥𝑥||𝑑𝑑2 = (∑ |𝑥𝑥𝑖𝑖 i|2)1/2. The Restrictive Isometry Property is similar to the

equation below[24, 74, 75],

(1 − δ𝑘𝑘) ≤ σ𝑚𝑚𝑖𝑖𝑛𝑛2 [Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)] ≤ σ𝑚𝑚𝑚𝑚𝑥𝑥

2 [Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝑘𝑘)] ≤ (1− δ𝑘𝑘) (3.6)

The factors σ𝑚𝑚𝑖𝑖𝑛𝑛[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)] and σ𝑚𝑚𝑚𝑚𝑥𝑥[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝑘𝑘)] stands for the lowest and the highest

quantitative values that the factor Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾) can assume.Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾) in itself is the sub-

matrix obtained from the K regions ofΦ. The factorδ2k which is the Restricted Isometry

Property constant is the lowest factor that modify the equation above for each sub-

matrix of Φ and is crucially bound on the space between 1 and all the discrete values

that the factor Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝑘𝑘) can assume. When δ2k<1, it denotes that a signal x is sparse in

the K region and can be precisely reconstructed from the possible values ofΦ.

The factor δ k∈ (0, 1), offers the precise reconstruction of the signal x and the

measurements of δ k defines the stability of the reconstruction of the signal. The

interference of noise which can assume the value 𝜉𝜉, 𝑧𝑧 = Φ𝑥𝑥 + ξ and the reconstructed

signal 𝑥𝑥� fulfils the equation below[76].

�|𝑥𝑥 − 𝑥𝑥�|�𝑑𝑑22≤ 4ξ2

1−δ2𝑘𝑘 (3.7)

Hence, the lower the value of δ k, the lower the error encountered upon the

reconstruction of the signal and the higher the value of δ k, the higher the error

encountered upon the signal reconstruction. But then, the calculation of δk for a possible

Page 70: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

53

value of Φ is NP-hard and therefore inflexible. Because δkcritically abound on the space

between unity and the possible values of Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)s, the scope of the constant δkcan be

measured by the space between unity and σ𝑚𝑚𝑖𝑖𝑛𝑛[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)]s and σ𝑚𝑚𝑚𝑚𝑥𝑥[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)]s. The

lower space, the lower the value of the constant δk and then the higher the outputΦ.

Because of this equation( (1− δ𝑘𝑘) ≤ σ𝑚𝑚𝑖𝑖𝑛𝑛2 [Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)] ≤ σ𝑚𝑚𝑚𝑚𝑥𝑥

2 [Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝑘𝑘)] ≤ (1− δ𝑘𝑘) )

must be in place, the values of σ𝑚𝑚𝑖𝑖𝑛𝑛[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)] and σ𝑚𝑚𝑚𝑚𝑥𝑥[Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝐾𝐾)] over arbitrarily

sampled Φ𝑠𝑠𝑠𝑠𝑠𝑠(𝑘𝑘)’s are used in [25, 74, 75] to evaluate the performance of δkin a given

measurement matrix Φ. This technique is also implemented in this thesis.

3.4 Incoherence in Compressive Sensing

Anotheressential condition for the precise reconstruction of the signal x is the crucial

factor known as incoherence[30, 77]. For a couple of values of the matrix Φ and

sparsifying transform matrix Ψ, fulfilling Φ*Φ = nl and Ψ*Ψ = l, their incoherence is

outlined as

µ(Φ,Ψ) = 𝑚𝑚𝑎𝑎𝑥𝑥𝑘𝑘.𝑗𝑗| <Φk, ψj>| (3.8)

where ψj and Φk are the jth and kth columns of Φ and Ψ respectively and µ(Φ,Ψ)∈[1,

�𝑚𝑚]. The factor µ(Φ,Ψ) = 1 is known as maximal incoherence. When m ≥ 𝐶𝐶.µ2(Φ,Ψ)

. K .log(𝑚𝑚). andC is a constant then a sparsity signal x in the k-space can be precisely

reconstructed. Hence, µ(Φ,Ψ) decides the least number of measurement essential for

the exact reconstruction of the signal x. The lower the µ(Φ,Ψ), the lower the value of

m required for the precise reconstruction of the signal x. It is paramount to know that

both Incoherence and Restrictive Isometry Property are adequate conditions on the

measurement matrix. Hence, they are similar and both or either of them can be used to

investigate, design and evaluate the measurement matrix for the precise reconstruction

of the signal x.

Page 71: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

54

3.5 Compressed Sensing in Magnetic Resonance

Imaging

Accelerating the speed of data acquisition in Magnetic Resonance Imaging and

reconstruction by Compressed Sensing is anactive area of research that is increasingly

capturing the attention of researchers and data scientists recently with the sole

objectives being to increase the speed and precision of the image reconstruction

effectively. Compressed sensing is a method that allows for the reconstruction of the

intended signal using the data obtained below the Nyquist-Shannon sampling rate.

Magnetic Resonance Imaging is a perfect scheme for the application of Compressed

Sensing because it acquires images that are already in the Fourier domain instead of the

pixel domain [32, 35]. The CS in MRI is first implemented in [32], where VD random

sub-sampling of phase encoding is proposed as a data sampling technique.The

combination of Parallel MRI Technique with Compressive Sensing has had an

accelerating possibility in the encoding carried out by MRI[27, 78]. Also, the CSMRI

can be used for dynamic imaging while using the k-t space sparsity.

The theory of Compressed Sensing offers the way around certain problems by using

previous information of the compressibility of the signal. This concept promises ideal

reformation of the signal from sub-sampled signals when some conditions are in place.

Images obtained from the biomedical analysis are inherently sparsity in certain

transform region (Fourier, Wavelet, etc.) while angiograms are sparsity in the pixel

region. As reported in some events, the signal may be insufficiently sparsified with

slight loss of information. The mode of acquiring image data using the method of MRI

inherently gains encoded signals instead of the signal values. Therefore sub-sampled

sparse reconstruction techniques can be used in Magnetic Resonance Imaging for the

reconstruction of images from sparse value. It can be demonstrated that the signal in

MRI follows the Fourier formula

s(t) =∫ 𝑚𝑚(𝑟𝑟)𝑅𝑅 ei2πk(t)r dr where k(t) ∞∫ 𝐺𝐺(𝑠𝑠)𝑑𝑑0 𝑑𝑑𝑠𝑠 (3.9)

Page 72: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

55

m(r) represents the object while r stands for the co-ordinate in the spatial domain, k(t)

represents the data in the k region while s(t) represents the signal.

3.6 Application of Compressive Sensing

3.6.1 Sparse Error Correction

This method is proficient in identifying and amending errors in data that seems

inaccurate in the channel of transmission. This method of correcting data depends on the

pattern of recurrence of the data, redundancy checks, or closest-neighbour code search.

Let’s take into account a certain scenario where a signal xm with M entries areencoded

by taking length-N encryptions linearly independent (φ1, ... ..,φM) such that N> M and

adding them up utilizing the entries of the signal x as coefficients. The received signal is

encoded of length N code

𝑦𝑦 = ∑ φ𝑖𝑖𝑀𝑀𝑚𝑚=1 𝑥𝑥𝑖𝑖 = Φf,

whereΦ is a matrix with differentcodewords for columns. Let’s say that the channel of

transmission get to corrupt the values of y in anadditive way, in such a way that the

collected data is

𝑦𝑦 = Φ𝑥𝑥 + 𝑒𝑒,

where 𝑒𝑒 represents an error vector. The method implement for the recovery of the sparse

signal in thecontext of compressed sensing offers a number of scheme that is effectively

utilized in the estimation of error vector 𝑒𝑒 , hence, making it possible to rectify it and

acquire the signal x, where the error 𝑒𝑒is adequately sparse. In order to estimate the error,

we construct a matrix Θ that we use as the foundation for the orthogonal subspace of the

span of the matrix Φ, i.e., an (N − M) × N matrix Θ which places Θ Φ = 0. We can work

on the values upon the acquisition of such matrix when we multiply them with the

matrix to get the expression

Page 73: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

56

𝑦𝑦� = Θ Φx + Θ𝑒𝑒 = Θ𝑒𝑒.

Assuming that the matrix Θ is perfect for compressive sensing (i.e. is fulfils the

requirement of the RIP) and 𝑒𝑒 is largely sparse, then 𝑒𝑒 can precisely be approximated

with the method of compressive sensing. As soon as 𝑒𝑒 is calculated, the error-free

measurements can be deduced as

𝑦𝑦�= 𝑦𝑦 − �̂�𝑒,

and the recovery of the signal is represented as

𝑥𝑥� = Φϯy− Φϯ�̂�𝑒

For illustration purpose, when the encoded data φm possess identical and random

distribution in the sub-Gaussian entries. A k-sparse error can be rectified when M< 𝑁𝑁 −

𝐶𝐶𝐾𝐾 log𝑁𝑁/𝐾𝐾 for a constant C that satisfies the Restrictive Isometry Property.

3.6.2 Linear Regression and Model Selection

A substantial amount of the algorithms for sparse recovery of signals initially were

designed to solve the problem of model selection and sparse linear regression of

statistical problems. Contextually, data involving a set of input and output variables

were provided to us. We will assume that there exist a sum of N variables, and we

notice that a sum of M input and response pairs.

If we choose to denote the set of input variable observation to be a 𝑀𝑀 × 𝑁𝑁 matrix Φ and

the set of response variable observations as a 𝑀𝑀 × 1 vector y, then according to analysis

using linear regression, y may be an estimate of a linear function of variables.

Nevertheless, if the input variables are big in contrast to the number of observations,

i.e., 𝑁𝑁 ≫ 𝑀𝑀 then it becomes overwhelmingly difficult, and this is because we intend to

approximate N parameters using a relatively few N observations. Broadly speaking, this

may be hard to surmount, however practically speaking, it is frequent that only small

input variables are significant to forecast the response variable. This means that in this

scenario, the signal x which we intend to approximate is sparse and we can be able to

Page 74: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

57

use every method we have learned thus far for sparse recovery to approximate the value

of the signal x. Statistically, the sparsity of the signal does not only assist us in our

objective of getting a linear regression, but it also executes model selection by

recognizing the essential variables in forecasting the response.

3.6.3 Group Testing and Data Stream Algorithms

One more case where the method of CS and sparse recovery algorithms can be

potentially useful is in its application to data stream algorithm and group testing and

then the associated cases of data stream computation.

3.6.3.1 Group Test

Some of the most antiquated ideas that address how to recover sparse algorithms were

implemented with the knowledge of combinatorial group testing[79-81]. Assuming that

there exists an N total list of items and K irregular elements in which we attempt to find.

Foe example, we are attempting to identify the problematic products in an industry or

the small group of the infected tissue sample in the body of a patient. In two of the

above scenarios the vector x shows the elements which are defective, i.e. xi ≠ 0 for the

defective K elements and xi = 0. In the easiest real-world scenario, these tests are

denoted with the binary matrix Φ whose entries φij are equivalent to unity on the

condition that the jth list is used in the ith test. When the output of the test is straight in

comparison to the inputs, then it is safe to say that the vector x is very much identical to

the standard recovery problem in CS.

3.6.3.2 Computation on Data Streams

One other area where the knowledge of CS has been found to be very important is in the

data stream computation[82, 83]. As in case of a characteristic data streaming challenge,

assume that xi stands for the number of packets moving through a network router with

destinationi. Storing the vector x is not possible because the sum of the probable

destinations (denoted by 32-bit Internet Protocol Address) is about 𝑁𝑁 = 232. Hence, as

Page 75: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

58

a substitute for intending to store x directly, the researcher can rather store𝑦𝑦 = Φ𝑥𝑥. Φ is

an 𝑀𝑀 ×𝑁𝑁 matrix with 𝑁𝑁 ≫ 𝑀𝑀. Vector y is frequently known as a sketch. Explicitly in

the case of the network traffic, we do not have to detect x directly, instead, we measure

increments to xi. Hence we build y iteratively by appending the ith column to y every

time we detect an increment to xi,which is a possibility for us because 𝑦𝑦 = Φ𝑥𝑥 is linear.

The takeover of network traffic by traffic to a small number of destinations, the vector x

is compactable and hence the case of recovering the signal x from the sketch Φ𝑥𝑥 is

again fundamentally the equivalent as the sparse recovery problem in CS.

3.6.4 Coronary Heart Imaging

CS can speed up the rate of acquiring data, permitting the whole heart to be visualized

in just one hold of breadth. Per cardiac prompt, one k-space spiral obtained for every

slice. The restrictions imposed by time always allows CS sub-sampling reconstruction

in the two-level double k-space that eliminates the interference induced by under

sampling without degrading the quality of the image.

3.6.5 Brain Imaging

The concept of Compressive Sensing in performing brain imaging promises to decrease

the time needed for the analysis while increasing the quality of the image so that

appropriate brain activities and structures can be determined without any error.

3.6.6 Rapid 3-D Angiography

CS is particularly suitable in angiography, where it provides advanced temporal and

spatial imaging quality at the cost of under-sampled artifacts[35]. This is one of

thesuperior advantages of the application of compressive sensing in MRI.

Page 76: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

59

3.7 Some Limitations of CSMRI The following problems have not been fully solved in the present work: In conventional CS-MRI, Fourier matrix as a sensing matrix and Wavelet matrix as a

sparsifying transform matrix respectively are not mutually incoherent. Moreover,

Fourier encodingconcentrates signal energy on the centre of the k-space known as low

frequencies region. This restricts the subsampling to fully sample the low frequency

region and insufficiently sample the high frequency region at high acceleration

factors,resulting in degraded image quality. This limitation can cause a huge loss in

image resolution.

Reduction factor is also crucial for high quality reconstruction in compressed sensing

even with PMRI. When it is increased to a certain label, the artefacts of the image

increaseand the SNR decreases simultaneously.

To overcome the limitations of present works found through literature survey,this thesis

introduces three new procedures.

1. In order to improve the quality of image reconstruction of compressive sensing

for high reduction factor, this thesis has proposed a novel compressive sensing

bunched phase encoding (CSBPE) technique.

2. Due to magnetic field inhomogeneity and imprecise field gradients, the phase of

the zigzag trajectories is different from theoretically calculated ones, which

lowers image quality. To overcome this problem, this thesis has introduced a

new technique called cross-correlation to remove the phase deviation from the

zigzag data position. In addition, to smooth and remove the aliasing from the

image, a modulation map estimation method has been incorporated in compress

sensing based image reconstruction.

3. In order to further improve the image quality of CSBPE, this thesis has

introduced a non-Fourier encoding scheme called chirp modulated Fourier

Page 77: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

60

encoding. The simulation result shows that the chirp modulated CSBPE

outperform the Fourier method.

3.8 Conclusion

Compression sampling is a mathematical model that reconstructs signals without

fulfilling Nyquist-Shannon theorem. This technique will undoubtedly revolutionize in

the field of signal processing, which is one of the areas that benefited most from data

compression. Currently, the CS is applied in areas such as biology, geology,

telecommunications and imaging. Moreover, for the proper functioning of the CS, it is

necessary to choose an appropriate technique to transform the data into rare data; this

work has shown that DCT is the best technique for physiological data, as it creates an

ideal sparse space for the needs of CS.

The idea and advances outlined in this thesis could possibly drive the novel uses of

CSMRI, which are previously considered difficult. The CS-MRI technique is still in it is

early stages. A lot of important questions remain unresolved [28]. These include

creating novel inconsistent sparse responses for the sampling operator, optimizing

sampling trajectories, studying the quality of the reconstruction with regards to its

clinical importance and increasing the fastness of the reconstruction algorithms. The

signal processing community has a big advantage here. There exist captivating

hypothetical and practical issues, assuring tangible gains in increasing medical care.

Page 78: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

61

Chapter 4 Compressive Sensing MRI using Fourier Based Bunched Phase Encoding Compressive Sensing MRI using Fourier Based Bunched Phase Encoding

4.1 Introduction

As discussed in the previous chapter, the data acquisition speed of imaging is critical to

the performance of practical MRI applications. The process of collecting data in MRI is

primarily limited by physical constraints (sweep rate, amplitude and gradient) and

physiological (nerve stimulation). High-speed MRI scanning can save a significant

amount of money and time for the patient, and can reduce patient discomfort and

imaging artifact caused by the patient’s motion. As a result, lots of researchers are

looking for new procedures that can reduce the amount of data requirement without

compromising the quality of the image for any reason.

As we know k-space sampling is performed by phase and frequency encodings. The key

to accelerating MRI data acquisition is to reduce the number of phase encodings, which

leads to k-space under sampling. But, when the k-space is under sampled the

reconstructed image shows aliasing artifacts.Various kind of methods have been

developed to reduce scanning time of MRI. But every methods has some limitation to

obtain certain level of acceleration.

Therefore, in this thesis,we will represent a new acceleration technique for the structural

MRI. This new method is developed by combining compressive sensing (CS) with

Fourier based bunched phased encoding (BPE) technique. Hence we call it CSBPE

method.

Page 79: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

62

We intend to exploit the inherent parsimony in magnetic resonance imaging and to

implement a method for integrating different approaches to improving the image

quality.

BPE is a non-linear sampling technique which was proposed by H. Moriguchi in 2006

[63]. This technique asserts that a number of phase encoding line can be reduced, if k-

space data are sampled along a zigzag trajectory during each readout and samples are

acquired at a sampling frequency higher than that of the conventional rectilinear

acquisition. That means we can reduce the number of TR cycles in rectilinear MRI data

acquisition by setting the PE steps size larger than that of the conventional acquisition

technique, and also an increased number of readout samples are acquired along a zigzag

k-space trajectory at a higher sampling rate than in conventional acquisition. This

technique opens opportunities to reduce the total number of TR cycles and hence the

MR scan time.

Among other popular methods, CS-MRI is a new framework for data sampling and

signal recovery. MRI is one of the suitable applications for CS because it meets two

main conditions such as sparsity and incoherency of CS. Hence, the MR images can be

reconstructed using CS technique from data sampled at a rate well below the Nyquist

rate. Because BPE and CS-MRI reduce sampling based on different ancillary

information (non-linear zigzag sampling for BPE and image sparseness for CS), it is

desirable to combine BPE and CS for further reduction of data reduction.

In conventional MRI, compressive sensing is much more efficient due to the facts that

most MR images are sparse in the transforms domain and data in k-space allows a

certain level of incoherent sampling [26, 32, 35, 84]. In CSMRI, the incoherence

between measurement or sensing matrix and sparsifying transform matrix are not

optimal. Moreover, the energy of Fourier encoding usually concentrates in the centre of

the k-space known as the low-frequency region.

This creates a restriction on the subsampling pattern to acquire adequate sample data

from the low-frequency region and insufficiently sample data in the high-frequency

region at high acceleration factors, which results in degraded image quality. Such

limitation can cause a massive loss in image resolution at a higher reduction factor.

Page 80: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

63

Besides, some approaches can achieve acceleration without phase array coils.

Bunchedphase encoding is one of them. It works on a wave-like sampling pattern to

reduce sampling data, and reconstruct aliased free image using Papoulis’ generalised

sampling theory [63]. Bunched Phase encoding takes advantage of the theoretical

concept of [85] to reduce the number of phase encodings and Repeating Time (TR)

cycles in Cartesian MRI data acquisitions. In this method, the phase encoding (PE) step

size can be set larger than the conventional rectangular acquisition, hence reducing the

number of phase encodings needed. To acquire sufficient information from the reduced

phase encodings, an additional number of readout samples are acquired, along the

zigzag or oscillating trajectories in the k-space.

This technique can decrease the total number of encodings and TR cycles and hence the

total MRI scan time. However, SNR of BPE approach becomes considerably reduced as

the reduction factor is increased generally if reduction factor is higher than two, the

noise of the reconstructed image becomes very pronounced even through aliasing

artefact are not significant [63]. To take advantages of the coil sensitivity map, Bunch

phase encoding has also been combined with sensitivity encoding (SENSE) in the

readout direction to further accelerate data acquisition in MRI [64]. However, BPE-

SENSE methods also have some crucial problem when the reduction factor is increased.

Even the signal to noise ratio of the BPE approach becomes considerably reduced as the

reduction factor is increased. We observe that if R > 2, the noise of the reconstructed

image is gradually increased even through aliasing artifact are not exist.

In this chapter, we proposed a novel method named CSBPE, which has been partially

presented[86], to combine BPE and CS-MRI for the zigzag trajectory with guaranteed

incoherence. CSBPE sequentially carries out CS reconstruction using sparse MRI for

the aliased image and BPE for the final unfolded image. Simulation and experimental

results show that CSBPE can achieve better resolution a reduction factor higher than

those achieved by CSMRI and BPE individually.

Page 81: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

64

4.2 BPE and CS Techniques

The sequential step diagram in Figure 4.1 shows the CSBPE method, where image

sampling and reconstruction methods of compressive sensing and bunched phased

encoding technique are used together to form a novel approach to achieve accelerated

MR imaging. In the following, the compressive sensing and bunched phase encoding

technique will be briefly described separately and then the overall scheme of CSBPE

will be represented.

Figure 4.1: Sequential steps of CSBPE image reconstruction process

In bunched phase encoding technique, a subsampled two dimensional image signal Iis

generated by zigzag changes of phase directional gradient and oversampling of

frequency domain data in readout direction to acquire N/R1 rows, where R1 is the

reduction factor and its values always greater than 1. Figure 4.2(a) and (b) illustrate the

rectangular grid line data and the zigzag data, where some datadefined as baseline or

reference data and rest are shifted data, respectively.

BPE Pulse Sequence Design

BP Encoded MRI data acquisition

subsampled K-space data

Further VD subsampling

Aliasing coefficient correction

CS Reconstruction

Image reconstruction

from BP Encoded MRI data

SOS of all channel Output image

Page 82: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

65

Figure 4.2:Schematic diagram of k-space trajectories of bunched phase encoding.

(a) Rectilinear sampling data. (b) zigzag sampling datain the BPE scheme.

Bunched phase approach isbased on the Papoulis’ generalized sampling theory[87],

which states that under the condition that the original function f(t) is zero outside a finite

interval (i.e., Time-limited), the signal f(t) can be reconstructed without aliasing

artifactseven though the Nyquist criterion is violated in the Fourier domain. For

example, if n number of bunched samples are collected at 1/n-th the Nyquist rate in the

Fourier domain, then the originalsignal f (t)can reconstruct without aliasing artifacts

[63].

Bunched Phase encoding takes advantage of this theoretical concept to reduce the

number of Repeating Time (TR) cycles in Cartesian MRI data acquisitions. In this

method, the phase encoding (PE) step size can be set larger than in the conventional

acquisition, i.e., 1/FOVy = ∆ky, here FOVy is the field-of-view along the phase direction.

At the same time, an additional number of readout samples are acquired at a higher

sampling rate than in a conventional acquisition, and along with a zigzag or oscillating

trajectories in the k-space domain. This technique enables one to decrease the total

number of TR cycles and hence the total MR scan time. Thek-space trajectories of a

Page 83: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

66

standard rectilinear and the bunched phase encoding acquisition methods are

schematically shown in Figure 4.2.

In the BPE scheme, k-space data are acquired along a zigzag trajectory, and the distance

between two neighbouring PE lines is qΔky, where q is usually greater than 1. As shown

in Figure 4.2, more data are sampled at higher sampling frequencies using the BPE

technique than a standard sampling method during the same readout period. Because of

the gradient alternation, this is corresponding to acquisition of multiple phase encoding

lines in a single readout to the BPE method. In Figure 4.2, rΔky is defined as the width

of a zigzag k-space trajectory band. In the formerly proposed hybrid fast-scan methods

[88, 89], q - r was less than or equal to 1.Moreover, as the total number of TR cycles

was reduced as q was increased as well as r was usually required to increase. However,

in bunched phase encoding, the width of the phase encoding band, r, can be set to any

value provided that r is greater than 0. But, it is noted that the value of r is independent

of the reduction factor R. The typical sampling rate must satisfy the Nyquist criterion to

reconstruct the signals without aliasing artifacts. Therefore, if the reduction factor of the

bunched phase encoding method is R, the total number of samples acquired for each

TR, m, must be at least N*R.

However, SNR of BPE approach becomes considerably reduced as the reduction factor

is increased.Often when R>2 the noise of the reconstructed image is significantly

increased even through aliasing artefact does not exist[63]. To overcome this difficulty

we will try to further decrease the scan time while attaining noiseless aliasing free

image by combine BPE with other new approaches such as compressed sensing, and

SENSE, and non-Fourier encoding method such as noiselet, and chirp modulation

encoding.

Page 84: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

67

Figure 4.3: Generalized 1D bunched phase encoding technique in k-space

Figure 4.3 show the generalized 1D BPE technique in k-space where the data

subsampledat a factor 2 and shifted from its original position 𝛼𝛼Δ𝑘𝑘.The mathematical

derivation of the generalized BPE technique will be explained step by step in the

following. Nyquist sampling and Fourier transform of the generalised 1D signal can be

represented as

𝑀𝑀�𝑘𝑘𝑗𝑗� = �𝐼𝐼(𝑥𝑥𝑖𝑖)𝑛𝑛

𝑖𝑖=1

𝑒𝑒−𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖𝑘𝑘𝑗𝑗 (4.1)

𝐼𝐼(𝑥𝑥𝑖𝑖) = �𝑀𝑀�𝑘𝑘𝑗𝑗�𝑛𝑛

𝑖𝑖=1

𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖𝑘𝑘𝑗𝑗(4.2)

Where 𝐼𝐼(𝑥𝑥𝑖𝑖) and 𝑀𝑀�𝑘𝑘𝑗𝑗� are image and k-space data respectively, i = 1,2,3….n, j =

1,2,3….nand𝑘𝑘𝑗𝑗 = 𝑗𝑗∆𝑘𝑘. The equation of the aliasing signal due to downsampling is given

as

𝐼𝐼𝑑𝑑(𝑥𝑥𝑖𝑖) = �𝑀𝑀𝑑𝑑(𝑘𝑘𝑑𝑑)

𝑁𝑁𝑝𝑝

𝑑𝑑=1

𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖𝑘𝑘𝑑𝑑

Page 85: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

68

= � � 𝐼𝐼(𝑥𝑥𝑚𝑚)𝑒𝑒−𝑖𝑖2𝜋𝜋𝑥𝑥𝑚𝑚𝑘𝑘𝑑𝑑𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖𝑘𝑘𝑑𝑑𝑁𝑁

𝑚𝑚=1

𝑁𝑁𝑝𝑝

𝑑𝑑=1

(4.3)

where 𝑘𝑘𝑑𝑑 = 𝑑𝑑.𝑝𝑝∆𝑘𝑘. If we consider the bunched Nyquist sampling, the equation will be

𝐼𝐼𝑠𝑠1(𝑥𝑥𝑖𝑖) = �𝑀𝑀𝑠𝑠1�𝑘𝑘𝑗𝑗 + 𝛼𝛼∆𝑘𝑘�𝑁𝑁

𝑗𝑗=1

𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖�𝑘𝑘𝑗𝑗+𝛼𝛼∆𝑘𝑘�

= � � 𝐼𝐼(𝑥𝑥𝑚𝑚)𝑒𝑒−𝑖𝑖2𝜋𝜋𝑥𝑥𝑚𝑚�𝑘𝑘𝑗𝑗+𝛼𝛼∆𝑘𝑘�𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖�𝑘𝑘𝑗𝑗+𝛼𝛼∆𝑘𝑘�𝑁𝑁

𝑚𝑚=1

𝑁𝑁

𝑗𝑗=1

(4.4)

Now, the equation can be represented for the bunched phased down sampling

trajectories as

𝐼𝐼𝑠𝑠𝑑𝑑1(𝑥𝑥𝑖𝑖) = �𝑀𝑀𝑠𝑠𝑑𝑑1(𝑘𝑘𝑑𝑑 + 𝛼𝛼∆𝑘𝑘)

𝑁𝑁𝑝𝑝

𝑑𝑑=1

𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖(𝑘𝑘𝑑𝑑+𝛼𝛼∆𝑘𝑘)

= � � 𝐼𝐼(𝑥𝑥𝑚𝑚)𝑒𝑒−𝑖𝑖2𝜋𝜋𝑥𝑥𝑚𝑚(𝑘𝑘𝑑𝑑+𝛼𝛼∆𝑘𝑘)𝑒𝑒𝑖𝑖2𝜋𝜋𝑥𝑥𝑖𝑖(𝑘𝑘𝑑𝑑+𝛼𝛼∆𝑘𝑘)𝑁𝑁

𝑚𝑚=1

𝑁𝑁𝑝𝑝

𝑑𝑑=1

(4.5)

Equation (4.5) can be represented in matrix form

⎜⎜⎛ 𝑀𝑀𝛼𝛼0𝑀𝑀𝛼𝛼1⋮

𝑀𝑀𝛼𝛼𝑏𝑏−1⎠

⎟⎟⎞

=1𝑝𝑝

⎜⎜⎛

1 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼0

𝑝𝑝⋯ 𝑒𝑒

𝑖𝑖2𝜋𝜋𝛼𝛼0(𝑝𝑝−1)𝑝𝑝

1 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼1𝑝𝑝

… 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼1(𝑝𝑝−1)

𝑝𝑝

⋮1

𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼𝑏𝑏−1

𝑝𝑝…

𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼𝑏𝑏−1(𝑝𝑝−1)

𝑝𝑝 ⎠

⎟⎟⎞

×

⎜⎜⎜⎜⎛

𝐼𝐼(𝑥𝑥)

𝐼𝐼 �𝑥𝑥 − 𝐹𝐹𝑝𝑝�

⋮𝐼𝐼(𝑥𝑥 − 𝐹𝐹(𝑝𝑝−1)

𝑝𝑝 ⎠

⎟⎟⎟⎟⎞

(4.6)

Page 86: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

69

In Equation (4.6), I(x), I(x-F/p), . . . , I(x-F(p-1)/p) (all for 0 ≤ x ≤ F/p) can be found by

computing an inverse matrix (when p =m) or a pseudo inverse matrix (when p< m)

from Equation (4.6). Hence I(x) for the entire range –F/2 ≤ x≤ F/2 can be reconstructed

without aliasing artifacts as

⎜⎜⎜⎜⎛

𝐼𝐼(𝑥𝑥)

𝐼𝐼 �𝑥𝑥 − 𝐹𝐹𝑝𝑝�

⋮𝐼𝐼(𝑥𝑥 − 𝐹𝐹(𝑝𝑝−1)

𝑝𝑝 ⎠

⎟⎟⎟⎟⎞

= 𝑝𝑝

⎜⎜⎛

1 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼0

𝑝𝑝⋯ 𝑒𝑒

𝑖𝑖2𝜋𝜋𝛼𝛼0(𝑝𝑝−1)𝑝𝑝

1 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼1𝑝𝑝

… 𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼1(𝑝𝑝−1)

𝑝𝑝

⋮1

𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼𝑏𝑏−1

𝑝𝑝…

𝑒𝑒𝑖𝑖2𝜋𝜋𝛼𝛼𝑏𝑏−1(𝑝𝑝−1)

𝑝𝑝 ⎠

⎟⎟⎞

−1

×

⎜⎜⎛ 𝑀𝑀𝛼𝛼0𝑀𝑀𝛼𝛼1⋮

𝑀𝑀𝛼𝛼𝑏𝑏−1⎠

⎟⎟⎞

or in simpler form,

𝐼𝐼1 = 𝐶𝐶𝑇𝑇𝐷𝐷1 (4.7)

where𝐼𝐼1 is the reconstructed 1D image data matrix, 𝐷𝐷1 is the BP encoded 1D data matrix

and C is the aliasing coefficient matrix constructed based on the zigzag trajectories and

the reduction factor R1for the 1D case.

Similarly for the 2D case, if the image signal is I(x,y) then the Fourier transformed k-

space signal can be represente as

𝑀𝑀�𝑘𝑘𝑖𝑖 , 𝑙𝑙𝑗𝑗� = ��𝐼𝐼�𝑥𝑥𝑖𝑖, 𝑦𝑦𝑗𝑗�𝑒𝑒−𝑖𝑖2𝜋𝜋𝑁𝑁 〖(𝑥𝑥〗𝑖𝑖,𝑘𝑘𝑖𝑖+𝑥𝑥𝑗𝑗𝑑𝑑𝑗𝑗�

𝑁𝑁

𝑗𝑗=1

(4.8)𝑁𝑁

𝑖𝑖=1

where 𝑘𝑘𝑖𝑖 = 𝛾𝛾∆𝑘𝑘 and 𝑙𝑙𝑗𝑗 = 𝑗𝑗∆𝑘𝑘.

If we apply inverse Fourier transform onfully encoded 2D MRI data, the equation will

be

𝐼𝐼�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� =1𝑚𝑚2��𝑀𝑀�𝑘𝑘𝑖𝑖, 𝑙𝑙𝑗𝑗�𝑒𝑒

𝑖𝑖2𝜋𝜋𝑁𝑁 (𝑥𝑥𝑖𝑖,𝑘𝑘𝑖𝑖+𝑥𝑥𝑗𝑗𝑑𝑑𝑗𝑗) (4.9)

𝑁𝑁

𝑗𝑗=1

𝑁𝑁

𝑖𝑖=1

Page 87: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

70

Similarly, if we apply IFFT on subsampled 2D MRI data,the reconstructed image will

be aliasing and the equation will be

𝐼𝐼𝑑𝑑�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� =1𝑚𝑚2��𝑀𝑀𝑑𝑑�𝑘𝑘𝑖𝑖, 𝑙𝑙𝑗𝑗� 𝑒𝑒

𝑖𝑖2𝜋𝜋(𝑥𝑥𝑖𝑖,𝑘𝑘𝑖𝑖𝑁𝑁 𝑝𝑝�

+𝑥𝑥𝑗𝑗𝑙𝑙𝑗𝑗𝑁𝑁 )

(4.10)𝑁𝑁

𝑗𝑗=1

𝑁𝑁𝑝𝑝

𝑖𝑖=1

where p is the reduction factor.

Now,if we apply IFFT on 2D Bunched Nyquist sampled MRI data, the equation will be,

𝐼𝐼𝑠𝑠�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� =1𝑚𝑚2��𝑀𝑀𝑠𝑠�𝑘𝑘𝑖𝑖 + 𝛼𝛼∆𝑘𝑘, 𝑙𝑙𝑗𝑗 + 𝛽𝛽∆𝑘𝑘� 𝑒𝑒𝑖𝑖2𝜋𝜋(

𝑥𝑥𝑖𝑖(𝑘𝑘𝑖𝑖+𝛼𝛼∆𝑘𝑘)𝑁𝑁 +

𝑥𝑥𝑗𝑗(𝑙𝑙𝑗𝑗+𝛽𝛽∆𝑘𝑘)𝑁𝑁 ) (4.11)

𝑁𝑁

𝑗𝑗=1

𝑁𝑁

𝑖𝑖=1

Also, if we apply the down-sampled 2D Bunched phase encoded MRI data, the equation

will be,

𝐼𝐼𝑠𝑠𝑑𝑑�𝑥𝑥𝑖𝑖,𝑦𝑦𝑗𝑗� =1𝑚𝑚2��𝑀𝑀𝑠𝑠𝑑𝑑�𝑘𝑘𝑖𝑖 + 𝛼𝛼∆𝑘𝑘, 𝑙𝑙𝑗𝑗 + 𝛽𝛽∆𝑘𝑘� 𝑒𝑒

𝑖𝑖2𝜋𝜋(𝑥𝑥𝑖𝑖�𝑘𝑘𝑖𝑖+𝛼𝛼∆𝑘𝑘�

𝑁𝑁 𝑝𝑝�+𝑥𝑥𝑗𝑗�𝑙𝑙𝑗𝑗+𝛽𝛽∆𝑘𝑘�

𝑁𝑁 )𝑁𝑁

𝑗𝑗=1

𝑁𝑁𝑝𝑝

𝑖𝑖=1

(412)

Now, IFFT of the acquired data gives the aliased image data D with N/R1 rows. With

linear equation based reconstruction[63], the image I can be recovered by solvingthe

following equation for the image I,

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷, (4.13)

where C is the aliasing coefficient matrix constructed based on the zigzag trajectories

and the reduction factor R1.Linear equation based reconstruction technique can produce

aliasing free image from the single receive or multi-receivecoil data without using coil

sensitivity.

Page 88: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

71

Now for reminiscing the compressive sensing method, the methamatical model will be

explained from the begaining. Suppose an n-dimensionalcomplex signalIin the space

ℂ𝑛𝑛 can be sparsely represented in sparsifying transformation𝛹𝛹 as

𝑥𝑥 = 𝛹𝛹𝐼𝐼 (4.14)

where the matrix 𝛹𝛹is the orthogonal wavelet basis or the frame of column vectors.

That means a number of coefficients k in signal x are sparse. A mathematical

measurement scheme can measure mdimensional signal yby selectingonly mprojections

of the signal Ias

𝑦𝑦 = Φ𝐼𝐼 (4.15)

where Φ ∈ ℂ𝑚𝑚×𝑛𝑛 is the sensing matrix. In a specific type of signal such as MRI, 𝐼𝐼 ∈ ℂ𝑛𝑛

is the vectorised complex image data with the dimension 𝑚𝑚 = 𝑟𝑟𝑟𝑟for a𝑟𝑟 × 𝑟𝑟dimensional

image, and Φ ∈ ℂ𝑚𝑚×𝑛𝑛 is usually a restricted randomly subsampled DFT matrix obtained

from the encoding process, where𝑚𝑚 = 𝑝𝑝𝑟𝑟 and 𝑝𝑝 < 𝑟𝑟 is the number of phase encoding

linesused to acquire y.Therefore, for the image dimension𝑟𝑟 × 𝑟𝑟, the subsampling ratio

(𝑛𝑛𝑚𝑚

= 𝑟𝑟𝑝𝑝

) is determined by the number of phase encoding lines, p, which ultimately

determines the acquisition time [32]. Now, Eq (1) can be further represented as

𝑦𝑦 = ΦΨ∗𝑥𝑥, where 𝐼𝐼 = Ψ∗𝑥𝑥 (4.16)

where * denotes the conjugate transpose operator and the signal I is sparse in the 𝛹𝛹

domain. Image signal can be sparsely represented in the wavelet domain using the

wavelet transformmatrix. The signal equation becomes under-determined when the data

are under-sampled. Therefore, it is near impossible to recover the exact signal using

linear reconstruction approaches. The compressive sensing technique provides full

recovery of a sparse signal from subsampled measurements by providing a unique

solution to the ill-posed nonlinear recovery problem:

min𝑉𝑉�Ψ𝐼𝐼�𝑑𝑑1 𝑠𝑠. 𝑡𝑡. �𝑦𝑦 −Φ𝐼𝐼�

2𝑑𝑑2≤ 𝜀𝜀2 (4.17)

Page 89: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

72

Where �𝐼𝐼�𝑑𝑑1 ∶= ∑ |𝐼𝐼𝑖𝑖|𝑖𝑖 is the l1norm of I,Iiis the ith element of I, Ψ is the sparsifying

transform matrix, and𝜀𝜀 is the allowed data discrepancy. The constrained optimisation

problem of Eq. (4.17) is converted to an unconstrained problem with regularisation

penalties for computation purposes:

arg𝑚𝑚𝛾𝛾𝑚𝑚 ��𝑦𝑦 − Φ𝐼𝐼�2𝑑𝑑2

+ 𝜆𝜆1�Ψ𝐼𝐼�𝑑𝑑1 + 𝜆𝜆2𝑇𝑇𝑉𝑉(𝐼𝐼)� (4.18)

where TV is the total variation of the signal and 𝜆𝜆1, 𝜆𝜆2 are the sparsity and TV

regularisation penalties respectively [32]. However, accurate reconstruction of the

signalIis only achievable if the conditions of restricted isometry property and

incoherency are satisfied in compressive sensing.

Besides R1 reduction in BPE, we can also use an additional variable density (VD)

undersampling to randomly omit the BPE rows by a factor R2> 1, i.e. the rows of Dcan

be randomly subsampled by a factor R2 to acquire

𝐷𝐷𝑠𝑠 = 𝐸𝐸𝐷𝐷, (4.19)

using the undersampling matrix E. The Ds thus acquired can be used in the iterative CS

reconstruction given below to recover the aliased image𝐷𝐷� .

arg𝑚𝑚𝛾𝛾𝑚𝑚𝐷𝐷� ��𝐷𝐷𝑠𝑠 − 𝐷𝐷�𝐸𝐸�𝑑𝑑22

+ 𝜆𝜆1�ΨD��𝑑𝑑1 + 𝜆𝜆2�𝐷𝐷��𝑇𝑇𝑉𝑉� (4.20)

Where 𝜆𝜆1 and 𝜆𝜆1 are the weighting parameters that can be tuned to control the sparsity

and smoothness of the solution 𝐷𝐷� . The reconstructed aliased image data matrix D� is in

turn used in

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷� (4.21)

to reconstruct the image I.

Page 90: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

73

N×N image I, ∆Ky =1/FOVy, ∆Kx =1/FOVx, R1 > 1, r > 0

BPE reconstruction D:

Solve 𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷

Reduction Factor R2

(Variable down-sampling)

CS+ BPE reconstruction

using Ds:

arg𝑚𝑚𝛾𝛾𝑚𝑚𝐷𝐷� ��𝐷𝐷𝑠𝑠 − 𝐷𝐷�𝐸𝐸�𝑑𝑑22

+ 𝜆𝜆1�ΨD��𝑑𝑑1 + 𝜆𝜆2�𝐷𝐷��𝑇𝑇𝑉𝑉�

𝑆𝑆𝑆𝑆𝑙𝑙𝑆𝑆𝑒𝑒 𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷�

Figure 4.4: Schematic diagram of compressive sensing BPE

Page 91: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

74

The physically acquired data in this approach isDs,with overall reduction 𝑅𝑅 = 𝑅𝑅1𝑅𝑅2.

The image I can be precisely reconstructed by the above described two step CS plus

BPE reconstructions. Therefore, we call this approach CSBPE. Shown schematicallyin

Figure 4.4.

4.3 Multichannel CSBPFE

For a multichannelCSBPFE system, multiple measurements of the signalare acquired

simultaneously from all independent channels. Each channel in the multichannelCSMRI

has a sensitivity map associated with the location of this channel with respect to the

object scanned. The acquired signal from N-receiver MRI system can be modelled as

follows,

𝐷𝐷𝑠𝑠𝑖𝑖 = E𝑆𝑆𝑖𝑖𝐷𝐷, i ∈ [1, L] (4.22)

where, 𝐷𝐷 is the BPE encoded k-space data set, S𝑖𝑖 = diag[γ𝑖𝑖𝑗𝑗]𝑗𝑗=1,2,…,𝑛𝑛 is the complex-

valued of sensitivity map matrixof theith receiver channel with γ𝑖𝑖𝑗𝑗 being the sensitivity

of the ith channel at the jth pixel of the vectorised image, 𝐷𝐷𝑠𝑠𝑖𝑖 is the subsampled data

acquired from the coil of the ith channel, and L denotes the number of the receiver

channels, Therefore, Dsi can be represented in a matrix form,

𝔇𝔇𝑠𝑠 ≔

⎣⎢⎢⎢⎢⎢⎢⎢⎡𝐷𝐷𝐴𝐴1

𝐷𝐷𝑠𝑠2

.

.

.𝐷𝐷𝑠𝑠𝐿𝐿⎦

⎥⎥⎥⎥⎥⎥⎥⎤

=

⎣⎢⎢⎢⎢⎢⎢⎡E𝑆𝑆1

E𝑆𝑆2

:..

ES𝐿𝐿⎦⎥⎥⎥⎥⎥⎥⎤

𝐷𝐷 =:𝐹𝐹𝐷𝐷(4.23)

As obtained from the above equation, the multichannel measurement matrix, denoted as

F for D, comes from the measurement matrix ES𝐿𝐿 of L column with dimension Lm×n.

Page 92: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

75

The sub-matrices, ES𝐿𝐿 s, share a common measurement matrix E ∈ ℂ𝑚𝑚×𝑛𝑛 resulting

from the encoding process, so they measure the same I simultaneously with the same

under-sampling pattern and under-sampling ratio n/m. The under-sampling ratio is

defined by the number of PE lines used in data acquisition and it does not depend on the

number of channel L.

Sensitivity map SLis complex-valued and 𝑆𝑆𝑖𝑖 ≠ 𝑆𝑆𝑗𝑗for 𝛾𝛾 ≠ 𝑗𝑗 in general. Therefore, E𝑆𝑆𝑖𝑖 ≠

E𝑆𝑆𝑗𝑗for 𝛾𝛾 ≠ 𝑗𝑗 and they are independent base on the specific values of 𝑆𝑆𝑖𝑖 and 𝑆𝑆𝑗𝑗. Hence,

measurement matrix F can provide more individual measurement than single-channel

matrix 𝐸𝐸, and multichannel measurement can reduce the number of measurements, m,

which is essential for each channel for exact reconstruction of 𝐷𝐷� . Hence, the

multichannel CSBPE is considered for reconstructing the image I from the multichannel

measurement of MRI. The equation of the multichannel CSBPE MRI is given by

Arg𝑚𝑚𝛾𝛾𝑚𝑚 ��𝔇𝔇𝑠𝑠 − 𝐹𝐹𝐷𝐷��2𝑑𝑑2

+ 𝜆𝜆1�Ψ𝐷𝐷��𝑑𝑑1 + 𝜆𝜆2�𝑇𝑇𝑉𝑉�𝐷𝐷���𝑑𝑑1� (4.24)

where TV is the total variation of the signal and 𝜆𝜆1, 𝜆𝜆2 are the sparsity and TV

regularisation penalties respectively [32]. However, accurate reconstruction of the

signal 𝐷𝐷� is only achievable if the conditions of restricted isometry property and

incoherency are satisfied in compressive sensing.The compressive sensing reconstructed

aliased image data matrix 𝐷𝐷� are used in following equation to reconstruct the final

image I.

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷� (4.25)

Daubechies-4 (db-4) wavelet are used all the simulation and reconstruction of this work,

because the performance of db-4 is superior in sparsifying the MR results. But here,

there is some issues to obtain high resolution precise image. In the next step, we will

discuss these issues and their effective solutions.

Page 93: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

76

4.4 Aliasing Coefficient Correction

In BPE imaging, the k-space trajectory for each RO line (x-direction) is modified to

make a variation to form the zigzag trajectories from the rectlinear trajectories. A

precise knowledge of the zigzag coordinate is needed in order to reconstruct artifact free

images from BPE k-space data. Butthe physically realizedzigzag trajectories are

generally not the same as the theoritically calculated ones due to gradient imprecision

and field inhomogeneity, which results in poor image quality. A solution to this problem

is to measure the actualcoordinates of the trajectories and use them to construct the

coefficient matrix C. However such measurements generally requiresprescan calibration

that complicates the process and increases operation cost.

To overcome this difficulty, we canuse a technique known as cross-correlation to

estimate theactual coordinates of BPE trajectories from standard scan data.Cross

correlation is a technique used in signal processing to measure the degree of similarity

and the Euclidean distance between two signals [90].

The basic concept of obtaining a quantitative measure of similarity between two signals

is fundamentally the same as statistical correlation of two arbitrary variables. However,

it is not quantified explicitly by statistical measures like convariance or standard

deviations. Generally, the quantative comparison between any two image signals or

waveforms may be based upon the amount of the component of one signal contained in

the other signal. Consider two signals x(m) and y(m)of finite energy. It is possible to

make a correlation sequence that would indicate the similarity of the two signals at

differents time intants. If the two signals are diffetrent from each other, the correlation

sequence is known as cross correletion and is represented as

𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙) = � 𝑥𝑥(𝑚𝑚− 𝑙𝑙) 𝑦𝑦 (𝑚𝑚)∞

𝑚𝑚=−∞

𝑙𝑙 = 0, ±1, ±2, … … (4.25)

Page 94: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

77

where l is the shift or lag parameter.If the length of the signal is not infinite, the length

of the correlation is not infinite either. In that case, the finite cross-correlation equation

is given as

𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙) = � 𝑥𝑥(𝑚𝑚 + 𝑙𝑙) 𝑦𝑦 (𝑚𝑚)𝑀𝑀

𝑚𝑚=1

𝑙𝑙 ∈ [0,𝑀𝑀 − 1] (4.26)

𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙) can be used to measure the shift between two signals. If 𝑦𝑦 (𝑘𝑘) = 𝑥𝑥(𝑘𝑘 + 𝜏𝜏)

where 𝑘𝑘𝑘𝑘[1,𝑀𝑀]and 𝜏𝜏 are constant, the shift 𝜏𝜏 between x(k) and y(k) can be found by

finding 𝑙𝑙∗ = 𝜏𝜏 such that

𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙∗) = max𝑑𝑑∈[1,𝑀𝑀]

𝜑𝜑𝑥𝑥𝑥𝑥�𝑙𝑙 � (4.27)

In the frequency domain, we can represent cross-correlationof 1D signalsas

Φ𝑥𝑥𝑥𝑥 = 𝐹𝐹𝑇𝑇�𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙)� = ℱ(𝑥𝑥).ℱ(𝑦𝑦) (4.28)

where ℱ denotes the Fourier transform. When M is large, computation of 𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙)cab be

very heavily. Computing 𝜑𝜑𝑥𝑥𝑥𝑥(𝑙𝑙) using ℱ -transform can speed up computation,

especially in MRI which requires complex valued computations. Cross correlation can

also be defined for 2D signals. For 2D signals x(m,n) and y(m,n), their cross correlation

is defined as

𝐶𝐶(𝑘𝑘, 𝑙𝑙) = � �𝑥𝑥(𝑚𝑚,𝑚𝑚) 𝑦𝑦 (𝑚𝑚− 𝑘𝑘,𝑚𝑚 − 𝑙𝑙)𝑁𝑁

𝑛𝑛=1

𝑀𝑀

𝑚𝑚=1

(4.29)

The cross-correlation of 2D signals can be computed efficiently in the spatial domain by

𝐶𝐶(𝑘𝑘, 𝑙𝑙) = ℱ−1 { ℱ (𝑥𝑥)ℱ∗(𝑦𝑦)} (4.30)

whereℱ is the Fourier transform. The complex conjugate of ℱ accomplished reversal of

the feature via the Fourier transform property ℱ𝑦𝑦∗(−𝑚𝑚) = 𝐹𝐹∗𝑦𝑦(𝜔𝜔).

Page 95: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

78

k-space

…… Image-

space

Figure 4.5: Baseline data and shifted data in the Zigzag trajectory of Bunched

phase encoding. The image from all PE baseline data on left and rest of all

images from shifted PE line data and RO line data for all four sets of image. The

whole 256 × 1028 BPE data image is showing on below.

Page 96: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

79

Similarly in 1D case, the shift between x(m,n) and y(m,n) 𝑦𝑦(𝑚𝑚, 𝑚𝑚) = 𝑥𝑥(𝑚𝑚 − 𝜏𝜏𝑚𝑚 , 𝑚𝑚 −

𝜏𝜏𝑛𝑛) can be found by finding 𝑘𝑘∗ = 𝜏𝜏𝑚𝑚 and 𝑙𝑙∗ = 𝜏𝜏𝑛𝑛 such that

𝐶𝐶(𝑘𝑘∗, 𝑙𝑙∗) = max𝑘𝑘∈[1,𝑀𝑀]𝑑𝑑∈[1,𝑁𝑁]

𝐶𝐶 �𝑘𝑘, 𝑙𝑙 � (4.31)

We will use the cross correlation technique described above to determine the actual

zigzag trajectories of the acquired BPE k-space data.

In order to determine the individual data positional deviation of acquired BPE k-space

data from the reference calculated data position is decomposed into eight individualk-

space volumes of matrix size N x Ns by sorting the PE k-space lines into individual

matrices.Each k-space matrix represents the same aliased image with the entire k-space

shifted by the step sizedk(n). Figure 4.5 shows the acquired zigzag data orientation and

subsampled image.

Figure 4.6 shows the deviation along the PE direction as line plots along or near the

center ofk-space for each individual k-space matrix. The deviation in PE direction from

the cartesian RO k-spaceline can be computed from two matrices by cross-correlation.

The peak position of the cross correlationmatrix then corresponds to the pixel shift

between the two k-space matrices.

The deviation of the k-space data from the calculated zigzag trajectories can be

computed usingthe cross-correlation shown in Eq. (4.29).The peak position of the cross-

correlation matrix then corresponds to the pixel shift between the two k-space matrices.

That means we will find a maximum column value and a maximum row value. Then a

sub-pixel analysis of the cross-correlation matrices using a weighted centre of mass

approach can be applied to determine with a high sub-pixel resolution ofthe BPE k-

space shifts in every step. The deviaion of the matrices N =1: ncan be determined

relative to the baseline with N = 0. In the method section, we will elaborate the cross-

correlation technique to recover the precise zigzag coordinate of k-space.

Page 97: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

80

Figure 4.6: K-space lines on or near the k-space centre for eight

individual k-space volume of the matrix represents the same image

with the entire k-space shifted by the step size dk(n) and n=[0:7],

top: n=0 and bottom: n=7.

4.5 Smoothness Enhancing

There is another issue that we have observed: the smoothness of the reconstructed

image in compress sensing relies on the accuracy of the inverse Fourier transform of

shifted k-space data is known as modulation maps. Usually, data shift in k-space will

create a modulation in the image domain. Therefore, the proper estimation of

modulation maps is necessary for unaliasing and smoothing of the image. We can

mathematically represent modulation maps as,

Page 98: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

81

M(x𝑖𝑖, y𝑖𝑖) = ��𝐼𝐼(𝑘𝑘𝑚𝑚 ,𝑘𝑘𝑑𝑑)𝑒𝑒−𝑖𝑖𝜔𝜔𝑥𝑥𝑖𝑖(𝑘𝑘𝑚𝑚+𝛼𝛼∆𝑘𝑘𝑚𝑚)𝑒𝑒𝑖𝑖𝜔𝜔𝑥𝑥𝑖𝑖(𝑘𝑘𝑙𝑙+𝛼𝛼∆𝑘𝑘𝑙𝑙)𝑁𝑁

𝑑𝑑=1

𝑁𝑁

𝑚𝑚=1

(4.32)

where M(x𝑖𝑖 , y𝑖𝑖) is the image domain and 𝐼𝐼(𝑘𝑘𝑚𝑚 ,𝑘𝑘𝑑𝑑) is the frequency domain data set.

The incoherency in multichannel CS cannot be guaranteed in bunch phase encoding

because the encoding matrix is channel dependent and can vary from scan to scan. Thus,

an accurate estimation of phase modulation maps can solve this problem. We can

estimate thel-th coil modulation map, Ml,via

𝑀𝑀�𝑑𝑑 ≜ 𝑎𝑎𝑟𝑟𝑎𝑎𝑚𝑚 𝑚𝑚𝛾𝛾𝑚𝑚12‖𝑆𝑆𝑑𝑑 − 𝑑𝑑𝛾𝛾𝑎𝑎𝑎𝑎{𝑆𝑆𝑧𝑧𝑒𝑒𝑟𝑟𝑧𝑧}𝑀𝑀‖2 + 𝛽𝛽𝑅𝑅(𝑚𝑚)(4.33)

where𝛾𝛾 ∈ [1,2,3 … . . 𝐿𝐿l] and𝑅𝑅(𝑚𝑚) is a spatial roughness penalty function with weighting

factor β. The reference image,𝑆𝑆𝑧𝑧𝑒𝑒𝑟𝑟𝑧𝑧 can be obtained by taking the sum of squares or

geometric mean of individual coil images 𝑆𝑆𝑖𝑖’s. The modulation maps were estimated

from subsampled BPE image using the method of [78]. The phase of one coil’s image

can further be incorporated into this term to prevent inclusion of the underlying object’s

phase in the modulation map. The smoothness of the modulation map can be easily

improvedby varying β. Further, we have used a variable density randomly subsampled

technique in PE direction to reduce the data sample and reconstruct the image using CS

reconstruction technique, where we use individual channel reconstruction and add them

together to obtain the final image.In the Multichannel CS part, estimated modulation

maps are applied to improve the robustness of the image.

4.6 Methods

Both simulated BPE data and real BPE scan data were used in our studies. For

simulated BPE data, the in vivo data was used to simulate the BPE DAT. The in vivo

data was obtained from a 2D spin-echo brain scan of a healthy volunteer on 3T Skyra

(Siemens Healthcare, Erlangen, Germany) with the 32-channel head coil (FOV: 240

mm, Flip angle: 10o, image matrix: 256 x 256). The 256 x 256 k-space data were

interpolated to a 1024 x 1024 matrix, with the added data points to simulate the

Page 99: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

82

underlying continuous k-space within the original sampling grid. The interpolated data

matrix were zigzag-sampled, with R1 = 2, r = 1 (confer figure 4.2), and then variable

downsampled(VD) along rows by R2 = 2 to obtain the 64 x 1024 data matrix DS with

total reduction R = R1×R2 = 4. The DSthus obtained is used in the iterative CS

reconstruction in Eq.(4.25) to compute the image I. We also used variable density

random sampling CS with R = 2, 3,4 and FFT2 to reconstruct the images from the

original gradient trajectories k-space data.

Pulse sequence was designed to obtain real BPE data for the study of CSBPE. The pulse

sequence design of CSBPE method is implemented by minor modification of the

gradient waveforms of the standard sequence. In BPE, a zigzag gradient is incorporated

in the phase encoding direction during readout and the sampling rate is increased more

than double. Figure 4.7 shows a timing diagram of a sample BPE sequence.

In this study, it is considered that the reconstructed image size is N in Phase encoding

(PE) direction and N in Readout (RO) direction. That means a standard rectilinear

sampling method acquired N samples in each PE. If NBconsiders as a total number of

TR cycles in the BPE method, the reduction factor will be R = N/NB.

Figure 4.7: A sample pulse sequence design of bunched phase encoding; (Gse) Gradient in slice direction; (Gpe) Gradient in phase encoding direction; (Gro) Gradient in readout direction. A zigzag gradient is incorporated in the phase encoding direction during readout. The amplitude and period of the oscillatory

Page 100: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

83

gradient in this figure are not to scale.

Figure 4.8:(a)Inverse Fourier transform of the acquired all 32 channel data gives the aliased image; (b) Inverse Fourier transform of the acquired channel 4 data

gives the aliased image; (c) k-space of channel 4.

Figure 4.9: Phase encoded data sampling technique; (a) Random sampling; (b) Variable density Random sampling.

Page 101: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

84

In this data acquisition technique, k-space data are acquired along a zigzag trajectory in

the PE direction. Also, four time’s oversampled data are taken in a readout direction.

The distance between two adjacent PE lines is R1Δky, where R1 is typically higher than

1. As shown inFigure 4-4, more data are sampled at a higher sampling rate than the

standard acquisition method during the same readout period. Inverse Fourier transform

of the acquired data gives the aliased image shown in Figure 4.8, which can be further

processed by the linear equation based (LEB) reconstruction technique to produce the

unaliased image.

Also, an aliased free image can be produced from the multi-coil data without using a

coil sensitivity map. A variable density random sampling pattern hasbeen used in the

phase encoding direction according to a Gaussian distribution function in this research

which is shown in Figure 4.9 because the most important information of the image is

preserved in the centre of k-space. The nonlinear iterative method (4.24)is solved to

reconstruct the aliased image for acceleration factors of 2, 4, 6, and 8. In this case, the

encoding matrix E does not have any sensitivity mapinformation (i.e. E = Ф).

It has been obseved in the experiment, that some trajectory deviationhasoccuredin the

physical trajectories compared to the calculated ones due to field inhomogeneity and

imprecise gradient.Such deviationhas developedan artifact in the reconstructed image as

shown inFigure 4.13 and Figure 4.14.

To solve the trajectory deviation of real data, a new recovery technique has been applied

to reconstructan artifact-free image. Suppose, a schematic diagram of BPE k-space data

isas shown inFigure 4.10, where the red dashed lines are the calculated (simulated)

trajectory, ×’s are the measured data positions. The solid black horizontal line in the

figure is actually an underlying rectangular baseline grid, which contains the baseline

data points shown by the black x’s.

Page 102: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

85

Figure 4.10: BPE k-space data acquisition. Ccalculated (simulated) trajectorywhich

shown in red marks (----);Original (measured) trajectory where data positions are

marked as (×’s, ×’s, ×’s, ×’s), baseline (zero shift) data points are marked as ×’s,

measured shifted data points are marked as ×’s, ×’s, ×’s.

As explained earlier in Sect 4.4, if the image signal of M is band limited, the other data

points, shown by the blue x’s, green x’s and red x’s can be regarded as the baseline data,

i.e. the black x’s shifted in the x and y directions. After applying BPE in MRI data

acquisition, all the data are shifted towards PE ditrection to form a zigzag pattern except

the black x’s which is consider as the baseline data of zigzag trajectory.The baseline

data points can be used as a reference to calculate the x and y directional shift.

For example, there are four sets of k-space data: black, blue, green and red as shown in

Figure4.10 and the black x’s areconsideredas a reference or baseline data.Using the

BPE acquisition technique some data are shifted to the different position so the data

alignment will be zigzag and mathematically the data position should be on red dash(---

) line. But, due to some factors, the data are shifted to different positions from the

mathematically calculated position as shown in Figure 4.10.

The 2D cross-correlation technique described in section 4.4 has been used to estimate

the correct data position of themeasured shifted data with respect to the baseline (black

x’s) data. In order to determine the individual k-space shifts, the whole k-space is

decomposed into eight individual k-space sub matrix size 256 x 128 by sorting the PE

k-space lines into individual matrices such that for N=[0 :7].The matrix kn includes

k=128 PE k-space lines kn=[n+1,n+2,…,n+k].Each 256×128 submatrix of the 256×1028

Page 103: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

86

data matrix represents the same aliased image but shifted in k-space by the step size

dk(n).

Figure 4.10shows that the data shift along the PE direction as line plots along or near

the centre of k-space for each individual k-space matrix. The data shift in PE direction

from the cartesian RO k-space line can be computed from two matrices by cross-

correlation. Hence, the data deviation from its original calculated position can be

determined by applying the cross-correlation technique.

The peak position of the cross-correlationmatrix then corresponds to the pixel shift

between the two k-space matrices. A sub-pixel analysis of thecross-correlation matrices

using a weighted centre-of-mass approach can be used to determine with highsub-pixel

resolution the BPE k-space deviation in every step. The shifts of the matrices N=1:7

weredetermined relative to the first with n=0. The sequential steps of the cross-

correlation technique is described in Figure 4.11.

Figure 4.12 shows a deviation between the Synthetic and the actual measured k-space

data position. The figure shows that calculated data follow the estimated sin(x) graph

but the measured data points are not symmetrical with the estimated sin(x) line.

Themathematically calculated k-space shifts were Kc = [0.0, 0.13021, 0.5, 0.86979, 1.0,

0.86979, 0.5, 0.13021], almost representing asinusoidal function.

The calculation of the cross-correlations was carried out for each matrix, N = 1...7,

withrespect to the first matrix, n=0 , and for every channel separately. The final values

were then averagedover all channels and resulted in the following k-space shifts Km

=[0.0, 0.1191, 0.5634, 0.8235, 0.8514, 0.7382, 0.1978, 0.0341]. These also

approximately followed a sinusoidal function.These parameters are used to reducethe

artifacts in the image reconstruction using the linear equation, but the result for

suppression of the artifactswas not optimal.

Page 104: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

87

Figure 4.11: The sequential steps of Aliasing Coefficient Correction by cross correlation technique

Step 1• Load the BPE k-space data ( Data array size: 256×1024×Number of

Channel).

Step 2• Split k-space data in 8 zigzag sub k-spaces for each channel.• Puts those data into the individual k-space volumes.

Step 3

• Calulate the offset in k-space between w.r.t the first zero shift k-space loop over number of zigzag steps upto 8.

• Select the central part of k-space with the peak.

Step 4

• Calculate the cross correlation between zero shift or baseline k-space with rest of 7 shifted k-space images. The resulting matrix can be analysed for a shift between the k-space images by finding the location/ index of its peak.

Step 5 • Observed near the center or k-space to gain sub-pixel resolution precision.

Step 6• Calculate weighted center of mass• Calculate deviation from reference zero shift.

Step 7• Calculate average shift for all 32 channel.• Results

Page 105: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

88

Figure 4.12:Calculated (red) and measured (blue) BPE k-space positions.

The data were shifted to be symmetrical around zero PE shift.

Table 1: Calculated trajectory vs corrected trajectory of all 32 channel.

Calculated

Trajectory

Corrected

trajectory

(Ch 1)

Corrected

trajectory

(Ch 2)

………..

Corrected

trajectory

(Ch 32)

Corrected

trajectory

(Mean)

0.00000 0.00000 0.00000 ………. 0.00000 0.00000

0.13021 0.12665 0.11766 ……….. 0.11721 0.1191

0.50000 0.32806 0.72221 …………. 0.60141 0.5634

0.86979 0.83531 0.87044 …………. 0.7866 0.8235

1.00000 0.86801 0.90716 …………. 0.81606 0.8514

0.86979 0.75521 0.78237 …………. 0.71495 0.7382

0.50000 0.23095 0.17618 ………… 0.24334 0.1978

0.13021 0.053406 0.030958 …………. 0.037712 0.0341

Page 106: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

89

It was found empirically that an additional multiplicative factor had to be applied in

order to optimize the image qualities. This factor had to be chosen such that the peak-to-

peak amplitude of the measured parameters would result in a value of 1√2

as opposed to a

value of 1 as set for the synthetic data. In Figure 4.12, both data sets were positioned

symmetrically around the shift value S=0. Estimated fits of sine functions to the two

data sets are shown for a better comparison of the data.

Figure 4.13 and Figure 4.14show reconstructed phase maps and images respectively

using the two parameter sets calculated BPE shifts, Kc , and measured and adjusted BPE

shifts, Km. The images reconstructed with the measured and adjusted BPE shifts show

sharply reduced artifacts compared to the images reconstructed with the calculated BPE

shifts and appear to be close to artifact free in PE direction as can be seen in the single

channel phase images Figure 4.13, right. Remaining artifacts in RO direction are weak

and might be specific to other parameters of the BPE method.

Figure 4.13: BPE image reconstruction (channel 4), phase images, using the

synthetic BPE k-space PE shifts, Ks (left) and using the measured and corrected

BPE k-space PE shifts, Kc (right).

Page 107: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

90

Figure 4.14: BPE image reconstruction, all channels combined, R = 2, matrix 128 x 128, using the synthetic BPE k-space PE shifts, Ks (left), and using the measured and

corrected BPE k-space PE shifts, Kc (right).

Now, using the trajectory corrected BPE data and the coefficient matrix C, we can

reconstruct the image I.The 2D cross-correlation technique is a useful and reliable

method for estimating the relative shifts of 2D signals. It can be performed by

convolution and also by Fourier transform for fast computation.

There was another issue that has been observed in compress sensing part which is

responsible for some aliasing and roughness in the output image. In BPE, the phase-

encode lines are regularly under-sampled by a reduction factor of R1 (= 2). After

obtaining k-space data through BPE, we can estimate modulation maps using equation

(4.33) and that maps can be used in the CS part where the phase-encoded lines are

undersampled by the reduction factor R2 = 2, 3, 4 etc.But in compressed sensing part

under sampling should be variably dense because most of the image information

preserved in of k-space. Thus, the overall reduction factor will be R=R1×R2= 4, 6, 8.

Page 108: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

91

4.7 Simulations and Experimental Results

We have categorized our study into two stages:(i) A simulation study with a

multichannel data using Fourier based CSBPE reconstructiontechnique and then

compare with CSMRI reconstruction. Simulations were performed on a (256 × 256×8)

multichannelbrain image to study the performance of CSMRI using Fourier based

bunched phase encoding technique. (ii) Anexperimental study with a multichannel

encoded data usingFourier based CSBPE reconstructiontechnique and then compare

with multichannel CSMRI reconstruction image.The experimental study was

implemented on a (256 × 256×32) multichannel brain image. Before present the result

we will observe some results after using correction technique in different stage of

CSBPE MRI technique.

4.7.1 Multichannel CSBPE Simulation

The simulation was performed using a MATLAB R2017b software and a workstation of

Intel ® Xeron ® CPU E5-16600 @ 3.30 GHz, 64 bit Operating System, and 64.0 GB

memory. In addition, the Rice_Wavelet_Toolbox_2.4 is used to simulate the

compressed sensing reconstruction. A brain image (size: 256 × 256× 8) was used to

analyze the performance of Fourier encoding in CSMRI and CSBPE MRI for different

acceleration factors. The simulation were perform by using eight complex sensitivity

maps which are obtained from the head coil of a Siemens Skyra 3T scanner shown in

Figure 4.15. The multichannel data was simulated by calculating the phase shift

position in CSBP encoding and reconstruct the image using equation (4.23) and (4.24)

with a differentdownsampling factor.

Quantitative analysis of the relative error in the CSBPE reconstructed image and

CSMRI reconstructed image is defined as a metric:

𝑅𝑅𝑒𝑒𝑙𝑙𝑎𝑎𝑡𝑡𝛾𝛾𝑆𝑆𝑒𝑒 𝑒𝑒𝑟𝑟𝑟𝑟𝑆𝑆𝑟𝑟 =‖𝑥𝑥0−𝑥𝑥�‖𝑙𝑙2‖𝑥𝑥0‖𝑙𝑙2

(4.25)

Page 109: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

92

Figure 4.15: The coil Sensitivity maps used in CSBPE simulation. The sensitivity maps are estimated from data acquired on MR scanner

The graph shows in Figure 4.16, the mean relative error of CSBPE is less than the

CSMRI reconstruction method at difference reduction factor. There is another way to

calculate the image reconstruction error, which is calculated by artifact power (AP)

method based on the following equation:

𝐴𝐴𝑃𝑃 = �∑𝑁𝑁𝑧𝑧𝑧𝑧=1 ∑ ∑ �𝑥𝑥𝑥𝑥𝑥𝑥𝑧𝑧 −𝑥𝑥𝑥𝑥𝑥𝑥𝑧𝑧𝑟𝑟𝑟𝑟𝑟𝑟�

2𝑁𝑁𝑥𝑥𝑥𝑥=1

𝑁𝑁𝑥𝑥𝑥𝑥=1

𝑁𝑁𝑧𝑧 ×𝑁𝑁𝑥𝑥×𝑁𝑁𝑥𝑥 (4.26)

where 𝑥𝑥𝑥𝑥𝑥𝑥𝑧𝑧 is the CSBPE reconstructed image data and 𝑥𝑥𝑥𝑥𝑥𝑥𝑧𝑧𝑟𝑟𝑒𝑒𝑟𝑟 is the CSMRI

reconstructed data. Nx is the total number of frequency encodinglines, Ny is the total

number of phase encoding lines, and Nz is the totalnumber of slices. The artifact power

of CSBPE MRI at different reduction factor is quite better than general CSMRI, which

is present in table 2.

Page 110: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

93

Figure 4.16: The mean relative error versus the acceleration factor in CS-

MRI, CSBPE for simulation data.

Table 2: Comparison of the artifact power (AP) between proposed CSBPE-MRI methods and the reference CS-MRI in a simulated brain image at

different acceleration factor (R).

Acceleration factor (R)

Artifact Power (AP)

CSBPE-MRI CS-MRI

R2 0.0052 0.0116

R4 0.0113 0.0137

R6 0.0128 0.0172

R8 0.0226 0.0264

Figure 4.17 and Figure 4.18 show the simulation results of Fourier based compressed

sensing reconstruction using BPE data. Figure 4.18 is magnitude image and Figure 4.18

is phase image respectively. The BPE acquired data images are all aliased images which

has obtained before the CS reconstruction and the middle column show the aliased

images after CS reconstructionat reduction factor R = 4, 6, 8 and 10 respectively. Right

column shows the error images of Fourier based CS reconstruction for acceleration

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0 2 4 6 8 1 0 1 2 1 4

MEA

N R

ELAT

IVES

ERR

OR

FOR

1000

TR

IALS

ACCELERATION FACTOR

cs

CSBPE(Fourier)

Page 111: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

94

factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from

BPEdata can preserve image resolution.

Before CS

reconstruction

(Magnitude)

CS reconstructed images (Magnitude)

Error images (Magnitude)

(a)

R = 4

(b)

R = 6

(c)

R = 8

(d)

R = 10

Figure 4.19: Simulation results of Fourier based compressed sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the BPE acquired data image and CS reconstructed images at reduction factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from BPEdata can preserve image resolution.

Page 112: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

95

Before CS

reconstruction

(Phase)

After CS reconstructed images (Phase)

Error images (Phase)

(a)

R = 4

(b)

R = 6

(c)

R = 8

(d)

R = 10

Figure 4.20: Simulation results (phase image)of Fourier based compressed

sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the

BPE acquired data phase image and CS reconstructed phase images at reduction

factor R = 4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of

Fourier based CS reconstruction for acceleration factors of 4, 6, 8 and 10

respectively; The result show the CS reconstructed images from BPE data can

preserve image resolution.

Page 113: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

96

CSMRI CSBPE-MRI Reconstructed Error Reconstucted Error

Figure 4.21: Simulation results for comparing the conventional CS and Fourier

based CSBP encoding technique. (a) reference image; (b)-(d) reconstructed images

ofFourier based CS for acceleration factors of 4, 6, and 8 respectively; (e)-(g) error

images of Fourier based CS for acceleration factors of 4, 6, and 8 respectively; (h)-

(j) reconstructed images ofFourier based CSBP encoding for acceleration factors

of 4, 6, and 8 respectively; (k)-(m) error images of Fourier based CSBP encoding

for acceleration factors of 4, 6, and 8 respectively; The result of reconstructed

images of CSBPE outperforms the CS for preserving image resolution.

Page 114: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

97

Figure 4.19shows the reconstructedimages using Fourier encoded CS and CSBPE for

the acceleration factors 4, 6 and 8 respectively. The simulation results indicate that

Fourier based CSBP encoding technique is more capable of preserving resolution than

Fourier encoded CS at a higher reduction factor. These results suggest that CSBPE is

better in terms of preserving the resolution of the reconstructed images from under

sampled k-space data.

4.7.2 Multichannel CSBPE Experiment The experiments were performed on the phantom and the brain image datasets to

validate the proposed method. The data set was acquired on a Siemens Skyra 3T MRI

scanner using 32 channel head coils with a maximum gradient strength of 40 mT/m and

a maximum slew rate of 200 mT/m/sec. A256 points RF excitation pulse have used.

The flip angle was 10° calculated by the equal integral rule, with the SAR level checked

by Siemens’ RF pulse programming software IDEA to be well below the safety limit

and about 5% higher than that of Fourier encoding RF pulse. We also acquired the

Fourier encoded data using the spin echo (SE) sequence to compare the quality of the

reconstructed image from the data acquired by the bunched phase encoding sequence.

A slice selective sinc RF excitation pulse was used in the spin echo sequence with a flip

angle of 10°. The protocol parameters of the bunched phase encoding of the Fourier

encoding SE sequence were in vivo experiments FOV = 240 mm, TE/TR = 26/750 ms,

image matrix = 256 × 256. In general,bunched phase encoding is sensitive to field

inhomogeneity, but careful design of the sequence and correcting phase shifting which

has discussed in section 4.5 can result in high-quality images. To reconstruct the

bunched phaseencoded data, the inverse Fourier transform was taken along both phase

and frequency encoding axis.

Page 115: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

98

Figure 4.22: Comparison between with and without aliasing coefficient

correction of reconstructed MRI images from BPE acquired in vivo k-space

data; (a) reference image; (b) Reconstructed MRI image using calculated

(simulated) trajectories; (c) Reconstructed MRI image after aliased coefficient

correction using cross-correlation technique; (d) and (e) are error images. The

maximum error in the calculated trajectory is 0.2095 and in corrected trajectory

is 0.0423.

Page 116: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

99

Figure 4.23 : Comparison between CSBPE image reconstruction using

estimated modulation maps and non-estimated modulation maps; (a)

Original image (b) CSBPE image reconstruction without an estimated

modulation map at reduction factor R = 4, (c) CSBPE image

reconstruction using the estimated modulation map at reduction factor

R = 4. (d) and (e) are error images.

Page 117: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

100

Figure 4.20 shows the comparison between reconstructed image with and without

aliasing coefficient correction after obtaining BP encoded data. The reconstructed image

without correction shows ghost and noise in the image but after correction most of the

noise and ghost has been disappear. The image (d) and (e) shows the error image. The

maximum error in the calculated trajectory is 0.2095 and in corrected trajectory is

0.0423

Similarly,Figure 4.21(b)also shows some ghost and noise in the reconstructed image

when bunched phase encoding technique are applied before randomly subsampling for

CS reconstruction. However, we can remove noise in image and make a smooth image

using estimated modulation maps using equation (4.33). InFigure 4.21(c) shows that the

reconstructed image using estimation modulation has no ghost image and the image is

sharper than the conventional reconstruction technique.

4.7.3 Phantom Scan Data Experiment

Figure 4.22 shows experimental results of Fourier based CSBPEphantom data and the

reconstructed images for the different subsampling of experimental phantom data set

using CSBPE technique. The results show that the CSBPE reconstructed image using

high reduction factor is possible and this technique can produce artifact-free images.The

left column top imagerepresentsthe reference image and rest of the left column shows

reconstruction images using Fourier based CSBPE for down sampling R = 2, 4, 6, 8

respectively and the right column show the error images. The error for the reduction

factor 2, 4, 6 and 8 are 0, 0.021, 0.3637, and 0.3867 respectively. Phantom results are

confirmed that Fourier based CSBPE can preserved resolution of image better that

CSMRI reconstructed images

Page 118: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

101

Reference Image

Reconstructed Image Error image Reduction Factor

R = 2

R = 4

R = 6

R = 8

Figure 4.24: Experimental results of Fourier based CSBPEphantom data. The left column top imagerepresentsthe reference image and Fig. Rest of the left column shows reconstruction images using Fourier based CSBPE for down sampling R =

Page 119: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

102

2, 4, 6, 8 respectively and the right column show the error images. The error for the reduction factor 2, 4, 6 and 8 are 0, 0.021, 0.3637, and 0.3867 respectively.

Figure 4.25: Experimental results of Fourier based CSBP encoding. The left

column represents the reference image and Fig. (a) show image reconstruction

usingFourier based CSBPE for down sampling R = 4 and Fig. (b) – (c) show

reconstruction imagesof Fourier based CSBPencoding for acceleration factors of R

=6, 8 respectively and Fig.(d) -(f) (right Column): show the error images. The error

for the reduction factor 4, 6 and 8 are 0.021, 0.478 and 0.523 respectively.

Page 120: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

103

4.7.4 Brain in vivo Scan Data Experiment

Figure 4.23 shows the reconstructedimages from the subsampledbunched phase

encoded data sets. These results demonstrate that the bunched phaseencoding

reconstructions can produce artifact-free images. After retrospective under-sampling,

the unconstrained optimization program equation (4.23) is solved using the non-linear

conjugate gradient method to reconstruct the desired image for different acceleration

factors.

Figure 4.23 (a) -(c)shows the reconstructed images for the acceleration factors of 4, 6

and 8 on the CSBPE data while Figure 4.23 (d) – (f)shows the corresponding error

images. The relative errors at reduction factor 4, 6 and 8 are 0.021, 0.478 and 0.523,

respectively.

4.8 Computation Complexity of CSBPE as Compared

with CSMRI.

Roughly speaking, the computation of CSMRI is mainly that of the CS iterative

reconstruction algorithms (4.22)-(4.24), while CSBPE needs to use both (4.22)-(4.24)

and (4.25) in image reconstruction. The additional BPE unpacking computation in

(4.25) does increase computation complexity and hence computation time. But this is

the price we have to pay in order to accelerate imaging speed and enhance image

quality. MR image reconstruction is generally performed offline due the physical

constraints in MR signal generation and data acquisition process. The extra computation

time incurred by more complex computation is therefore not a major problem. In the

clinic applications where time critical image reconstruction is needed, the computation

can be accelerated by modern parallel computing hardware, e.g. GPU (Graphics

Processing Unit) array, which is now widely available on state-of-the-art MRI scanners.

Page 121: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

104

4.9 Conclusion

This chapter haspresented a new methodto accelerate data acquisition for MR

imaging.This method uses a novel compressed sensing approach in Fourier based

zigzag sampling of k-space and a new way to remove the artifactand noise from

physical scanning data at an acceleration factor up to 8. Fourier encoded CSMRI has

previously been used only for magnitude images. In the proposed method, both phase

and magnitude information is utilised in the reconstruction process.

Here, we have used the cross-correlation technique to estimate theexact coordinates of

non-rectangular periodic trajectories from standard scan data and reconstruct a high-

quality image from ZIGZAG in vivo data without prescan. Also, we have used the

estimatedmodulation map to smoothing the image and remove aliasing in compressed

sensing reconstruction stage.

The simulation and experimental results demonstrate that the new acceleration and

artifact removing techniquesof CSBPE-MRI outperform the conventional Fourier

encoded CSMRI. The encoding and reconstruction technique of the proposed CSBPE

scheme preserves the spatial resolution far better than the conventional Fourier encoded

CSMRI scheme. The proposed method is applicable to general non-rectangular periodic

trajectories. This CSBPE method has a significant potential for MRIacceleration.

Page 122: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

105

Chapter 5 Compressive Sensing MRI using Non Fourier Based Bunched Phase Encoding Compressive Sensing MRI using Non-Fourier based Bunched Phase Encoding

5.1 Introduction

There are three continuous steps used for data acquisition in MRI. RF pulse excitation is the

first step, which is typically a constant excitation of the field of view (FOV), and encoding

typically Fourier encoding is the second stage which is achieved by the imaging gradients. At

the last stage, data are acquired by analogue to digital converter. In fact, some specific case, a

selective RF pulse can also be used to perform excitation and encoding simultaneously [91,

92]. The spatially selective RF pulse was used in different wavelet-based encoding, and SVD

encoding was applied to increase the speed of MRI scans[93, 94].

Non-Fourier encoding technique can be combined with the parallel imaging method as

second steps, which relies on the sensitivity information of the receiver coil [95]. Besides, the

performance of the compressive sensing depends on the incoherence [30] between the

measurement domain and the sparsifying transform matrices some non-Fourier encoding

methods such as Noiselet encoding [96], random encoding [33, 97, 98], and chirp encoding

[99, 100]have better incoherence than Fourier encoding. Among them, the chirp modulation

technique has good strength to preserve incoherency compare to others technique.

Basically, chirp modulation can spread the k-space signal energy throughout the region k-

space, whereas Fourier encoding only concentrates the signal energy in the low-frequency

region. Chirp modulation can modify the sensing matrix and also can be used as a

controllable spread spectrum technique to achieve energy spread. Recently, there are some

non-Fourier encoding schemes such as Gaussian random encoding, Toeplitz random

encoding have been proposed for compressive sensing MRI have been investigated for

Page 123: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

106

incoherent sampling. But those methods are challenging to implement and requires long RF

pulse and unrealistic computation power and storage memory for reconstruction [33, 97,

101]. In our work, we have applied the chirp modulation technique due to the better

incoherency performance for CSMRI. In the CSBPE method, the use of chirp modulated RF

pulses to spread the energy along the phase encoding direction by exciting the scanned object

along phase encoding direction with different profile at each excitation.Therefore, application

of non-Fourier encoding to CSBPE is expected to outperform conventional Fourier encoding.

5.2 Chirp Modulationbased CSBP Encoding

In paper [99]the authors suggested a new spread spectrum procedure using Chirp modulation.

Spread spectrum method uses second-order shim coils to reform the radio frequency pulse. In

order to distribute the signal energy in the k-space and allow for incoherent sampling, a

spread spectrum technique is also presented to modify the encoding matrix. Chirp radio

frequency (RF) modulation can be used as a controllable spread spectrum method to achieve

an energy spread [96, 99, 100, 102].To reduce the reconstruction error for unsystematically

sampled k-space. The Fourier encoded MRI signal can be described as,

𝑌𝑌 = F𝑝𝑝𝑒𝑒I 𝐹𝐹𝑟𝑟𝑧𝑧 (5.1)

where Yis the acquired k-space matrix, I is the desired 2D image matrix; Fpe is the encoding

matrix in phase encoding direction and Fro is the encoding matrix in readout direction.

Inconventional MRI dataacquisition,Fpe and Fro are the Fourier transform matrices in the PE

and RO direction, respectively. Here Fpedata matrix is replaced by chirp modulated encoding

matrix Fc, which is defined as:

𝐹𝐹𝑐𝑐 = φ𝐹𝐹𝑝𝑝𝑒𝑒 (5.2)

where𝜑𝜑 is used to modulate the Fourier transform with its diagonal entry given as:

𝜑𝜑(𝑟𝑟, 𝑟𝑟) = 𝑑𝑑𝛾𝛾𝑎𝑎𝑎𝑎 [ 𝑒𝑒−𝑖𝑖�Δ𝑐𝑐𝑟𝑟2+Δ𝑐𝑐�] (5.3)

Page 124: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

107

where ∆c is the strength of the Chirp modulation, and r ∈ [0,1, … . N], ris the location of the

acquired point along the direction of encoding [103].The Fourier transform of each row of the

chirp encoding matrix Fc, constitutes a unique RF pulse that is subsequently applied along

the primary PE direction during data acquisition. After that, a non-linear conjugate gradient

iteration has used to reconstruct images from under sampled k-space data for different

acceleration factors.

Chirpmodulation spreads the energy of the k-space along the phase encoding direction. This

thesis proposesa random subsample the k-space along the phase encoding direction with a

uniform probability using chirp modulation to increase the sampling incoherency and hence,

increase the image quality of thereconstructed image of the CSBPE technique.

Figure 5.1: Sequential steps of Chirp modulated CSBPE image reconstruction process

BPE Pulse Sequence Design

BP Encoded MRI data acquisition

Chirp modulation of BPE data

subsampled K-space data

Further VD subsampling

Aliasing coefficient correction

CS Reconstruction

Image reconstruction

from BP Encoded MRI

data

SOS of all channel

Output image

Page 125: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

108

In this research, we have used the chirp modulation method to modify the bunched phase

encoded k-space energy. Therefore, the chirp modulation allows spreading the energy of the

bunched phaseencoded k-space along the phase encoding direction.Figure 5.1 shows the

sequential steps of chirp modulated CSBPE image reconstruction process.

According to equation (4.7) in chapter 4, the chirp modulated BPE equation will be

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷𝑐𝑐 (5.4)

where 𝐷𝐷𝑐𝑐 ∈ 𝕔𝕔𝑀𝑀 is bunched phase and chirp modulated Fourier base (CMFB) encoded k-

space dataset𝐷𝐷𝑐𝑐 = 𝐹𝐹𝑐𝑐𝐷𝐷 ,𝐶𝐶𝑇𝑇 ∈ 𝕔𝕔𝑁𝑁×𝑁𝑁 is the appropriate chirp modulated co-efficient matrix

constructed based on the sample co-ordinates. Besides R1 reduction in chirp modulated BPE,

we can also use an additional random undersampling to randomly omit the chirp modulated

Fourier based BPE rows by a factor R2> 1, i.e. the rows of 𝐷𝐷𝑐𝑐 can be randomly subsampled

by a factor R2 to acquire

𝐷𝐷𝑠𝑠 = 𝐸𝐸𝐷𝐷𝑐𝑐 (5.5)

where E is the undersampling matrix. The Ds thus acquired can be used in the iterative CS

reconstruction given below to recover the aliased image 𝐷𝐷�𝑐𝑐 . Atypical chirp modulated

CSBPE-MRI reconstruction attempts to solve

arg𝑚𝑚𝛾𝛾𝑚𝑚𝐷𝐷�𝑐𝑐 ��𝐷𝐷𝑠𝑠 − 𝐸𝐸𝐷𝐷�𝑐𝑐�𝑑𝑑22

+ 𝜆𝜆1�ΨD�𝑐𝑐�𝑑𝑑1 + 𝜆𝜆2�𝐷𝐷�𝑐𝑐�𝑇𝑇𝑉𝑉� (5.6)

where 𝜆𝜆1 and 𝜆𝜆1 are the weighting parameters that can be tuned to control the sparsity and

smoothness of the solution 𝐷𝐷�𝑐𝑐 . The reconstructed aliased image data matrix 𝐷𝐷�𝑐𝑐 is in turn

used in

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷�𝑐𝑐 (5.7)

to reconstruct the image I.

Page 126: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

109

(a)

(b)

(c)

Figure 5.2: (a) Fourier modulated k-space(Mesh view), (b)Chirp modulated k-

space(Mesh view), (c) Chirp modulated k-space

Page 127: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

110

5.3 Chirp Modulation ofMultichannel CSBPE

For a multi-receiver MRI system, multiple measurements of the signal can be acquired

simultaneously from different independent channels. Each channel in the multi-receiver MRI

has a sensitivity map associated with the location of this channel with respect to the object

scanned. The acquired signal from N-receiver MRI system can be modelled as follows,

𝐷𝐷𝑐𝑐𝑠𝑠𝑖𝑖 = E𝑆𝑆𝑖𝑖𝐷𝐷𝑐𝑐 , i ∈ [1, L] (5.8)

where,𝐷𝐷𝑐𝑐 is the chirp modulated BPE encoded k-space data set,S𝑖𝑖 = diag[γ𝑖𝑖𝑗𝑗]𝑗𝑗=1,2,…,𝑛𝑛 is the

complex-valued of sensitivity map matrixof theith receiver channel with γ𝑖𝑖𝑗𝑗 being the

sensitivity of the ith channel at the jth pixel of the vectorised image, 𝐷𝐷𝑐𝑐𝑠𝑠𝑖𝑖 is the chirp

modulated subsampled data acquired from the coil of the ith channel, and L denoted as the

number of the receiver channels, Therefore, Dsi can be represented as a matrix form,

𝔇𝔇𝑐𝑐𝑠𝑠 ≔

⎣⎢⎢⎢⎢⎢⎢⎢⎡𝐷𝐷𝑐𝑐𝑠𝑠1

𝐷𝐷𝑐𝑐𝑠𝑠2

.

.

.𝐷𝐷𝑐𝑐𝑠𝑠𝐿𝐿⎦

⎥⎥⎥⎥⎥⎥⎥⎤

=

⎣⎢⎢⎢⎢⎢⎢⎡E𝑆𝑆1

E𝑆𝑆2

:..

ES𝐿𝐿⎦⎥⎥⎥⎥⎥⎥⎤

𝐷𝐷𝑐𝑐 =:𝐹𝐹𝐷𝐷𝑐𝑐(5.9)

As obtained from the above equation, the multichannel measurement matrix denoted as F for

𝐷𝐷𝑐𝑐, which is come from the measurement matrix ES𝐿𝐿 of L column and dimension is Lm×n.

The sub-matrices, ES𝐿𝐿s, share a common measurement matrix E ∈ ℂ𝑚𝑚×𝑛𝑛 resulting from the

encoding process, so they measure the same 𝐷𝐷𝑐𝑐 simultaneously with the same under-

sampling pattern and under-sampling pattern ratio n/m. The under-sampling ratio is defined

by the number of PE lines used in data acquisition and it does not depend on the number of

channel L.

Page 128: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

111

Sensitivity map SLis complex-valued and 𝑆𝑆𝑖𝑖 ≠ 𝑆𝑆𝑗𝑗for 𝛾𝛾 ≠ 𝑗𝑗 in general. Therefore, E𝑆𝑆𝑖𝑖 ≠ E𝑆𝑆𝑗𝑗for

𝛾𝛾 ≠ 𝑗𝑗 and they are independent base on the specific values of 𝑆𝑆𝑖𝑖and𝑆𝑆𝑗𝑗. Hence, measurement

matrix F can provide more individual measurement than single-channel matrix 𝐸𝐸, and

multichannel measurement can reduce the number of measurements, m, which is essential for

each channel for exact reconstruction of 𝐷𝐷�𝑐𝑐. Hence, the multichannel CSBPE is considered

for reconstructing the aliased image 𝐷𝐷�𝑐𝑐 from the multichannel measurement of MRI. The

equation of the chirp modulated multichannel CSBPE MRI can denote as

arg𝑚𝑚𝛾𝛾𝑚𝑚 ��𝔇𝔇𝑐𝑐𝑠𝑠 − 𝐹𝐹𝐷𝐷�𝑐𝑐�2𝑑𝑑2

+ 𝜆𝜆1�Ψ𝐷𝐷�𝑐𝑐�𝑑𝑑1 + 𝜆𝜆2�𝑇𝑇𝑉𝑉(𝐷𝐷�𝑐𝑐)�𝑑𝑑1� (5.10)

where TV is the total variation of the signal and 𝜆𝜆1, 𝜆𝜆2 are the sparsity and TV regularisation

penalties respectively [32]. However, accurate reconstruction of the signal 𝐷𝐷�𝑐𝑐 is only

achievable if the conditions of restricted isometry property and incoherency are satisfied in

compressive sensing.Hence , the final image I can be reconstructed using the following

equation,

𝐼𝐼 = 𝐶𝐶𝑇𝑇𝐷𝐷� 𝑐𝑐 (5.11)

Daubechies-4 (db-4) wavelet are used all the simulation and reconstruction of this work. The

performance of db-4 is superior in sparsifying the MR results.

5.4 Chirp Modulated CSBPE Sampling Pattern The energy of Fourier encoded signal concentrate in the centre of the k-space and it is not

universal. So Fourier encoding is weakly incoherent with some sparsifying transform such as

Daubechies-4 (db-4) wavelet. In CSBPE scheme the variable density probability distribution

function is used to design a pattern for sampling the k-space that the low spatial frequency

always has to be fully sampled. However, sampling pattern using variable density probability

distribution function unable to adequate sampling of the high spatial frequency at high

acceleration factor. But, the proposed chirp modulated encoding method is designed to

optimally spread the energy of the signal along the desired under-sampling direction and

hence allowing the uniformly random sampling.Figure5.3 shows that the Fourier encoded

variable density sampling pattern and the chirp modulated random sampling pattern which

energy is spread along the phase direction.

Page 129: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

112

Figure 5.3: (a) shows that the Fourier encoded variable density sampling

pattern and (b) shows the chirp modulated random sampling pattern which

energy is spread along the phase direction.

5.5 Simulation and Experimental Results Simulations were performed to observethe proposal of chirp encoded CSBPE encoding

scheme in the multi-channel CS-MEI framework. Figure 5.2shows the acquired k-space data

for chirp modulated k-space. It is proved that the energy is spread out in PE direction for the

chirp encoded data, and it increases with chirping factor. The characteristic of energy

spreading of chirp modulation is a favourable condition for CS data acquisition and

reconstruction. Simulations of CSBPE using chirp modulated Fourier base was implemented

on a (256 × 256) brain image to study the performance of CS-MRI using chirp encoded based

BPE. The simulation study was divided into two parts: a simulation study with a multi-

channel chirp modulated BPE data set,and an experimental study with multiple channels

chirp modulated BPE data set, where the sensitivity profiles wereestimated from the

reference image and calculated phase modulation maps for BPE reconstruction.

A chirp modulation of the image is taken in the phase encoding direction to obtain chirp

encoding. Anabsolutely random sampling pattern need to use to sample the chirp encoded

data in the phase encoding direction. A non-linear solution of Eq. (5.11) need to solve to

reconstruct the final image for acceleration factors of 2, 4, 5.3, 6, and 8. Here, the encoding

matrix E does not have any sensitivity information (i.e. E = Ф).

Page 130: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

113

Simulations were performed on a 256×256×8complex brain image. The brainimage data for

Fourier encoding was simulated by variable density random under-sampling and taking the

2D Fourier transform of the images. While for simulating chirp modulation encoded k-space

data, a chirp transform was taken in phase directionfollowed by a Fourier transform in

readout direction. The pseudo-random under samplingscheme that is known to be optimal for

the Fourier encoding scheme was performed to under sample the Fourier encoded data. While

complete random under sampling was performed for chirp encoded data. Equation (5.11) was

solved using non-linear conjugate gradient iterations to reconstruct images from under

sampled k-space data for different acceleration factors.

The quantitative performance of both the encoding schemes can determine by used the

relative error defined as a metric:

𝑅𝑅𝑒𝑒𝑙𝑙𝑎𝑎𝑡𝑡𝛾𝛾𝑆𝑆𝑒𝑒 𝐸𝐸𝑟𝑟𝑟𝑟𝑆𝑆𝑟𝑟 = ‖𝑥𝑥0−𝑥𝑥�‖𝑙𝑙2‖𝑥𝑥0‖𝑙𝑙2

Figure 5.4The mean relative error versus the acceleration factor in CS-MRI, CSBPE

Fourier encoding and CSBPE chirp encoding for simulation data.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0 2 4 6 8 1 0 1 2 1 4

MEA

N R

ELAT

IVES

ERR

OR

FOR

1000

TRI

ALS

ACCELERATION FACTOR

cs

CSBPE(Fourier)

CSBPE(chirp)

Page 131: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

114

5.5.1 Simulation Results

The mean relative errorversus the acceleration factor is plotted in Figure 5.4and highlights

that CSBPE using chirp encodingoutperforms CSBPE using Fourier encoding for

acceleration factors up to R = 6 after that the image resolution going down in chirp encoding

technique than Fourier encoding.The mean relative error of output images is showing that the

error will increase as the reduction factor R increases. In addition, chirp modulation encoding

outperforms Fourier encoding.But both scheme is far better than convention CS scheme for

all acceleration factors.

Figure 5.5and Figure 5.6 show the simulation results of chirp modulated Fourier based

compressed sensing reconstruction using BPE data. Figure 5.5is magnitude image and Figure

5.6 is phase image respectively. The BPE acquired data images are all aliased images which

has obtained before the CS reconstruction and the middle column show the aliased images

after CS reconstructionat reduction factor R = 4, 6, 8 and 10 respectively. Right column

shows the error images of chirp modulated Fourier based CS reconstruction for acceleration

factors of 4, 6, 8 and 10 respectively; The result show the CS reconstructed images from

BPEdata can preserve image resolution.

Figure 5.7 shows the reconstructed images using chirp modulated Fourier based CSBP

encoding for the acceleration factors of 2, 4, 5.3, 6 and 8. In addition, the simulation results

illustrate that chirp modulated Fourier based CSBP encoded images is able to preserve

resolution better than Fourier based CSBP encoded imagesas is explicit from the enlarged

region of the reconstructed images. This result suggests that chirp encoding is better in term

of preserving the resolution of the reconstructed images from under sampled k-space data.

These simulation results show that chirp modulation encoding can preserve the sensitivity

information while performing CS. Also, table 3 shows the less error of chirp modulated

CSBPE compare to the Fourier based CSBPE and therefore, chirp modulation is better

encoding scheme for CSBPE-MRI. We have investigated the performance of Fourier

encoding, and the chirp modulated Fourier encoding schemes using single and multi-channel

data.

Page 132: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

115

Figure 5.8shows the comparison of image quality between Fourier encodedCSBPE and chirp

modulated CSBPE method at different reduction factor and it shows that the quality of chirp

modulated Fourier based CSBP encoded images are better than Fourier based CSBP encoded

images.Also Figure 5.9 shows the comparison of image quality between chirp encoded

CSBPE and CS-MRI method at different reduction factor.The result of reconstructed images

of chirp modulated CSBPE outperforms the CSMRI for preserving image resolution.

Table 3: Compare the simulation results of Fourier based CSBPE and chirp modulated CSBPE for single channel data.

Reduction factor (R) Error using Fourier based

CSBPE (Mean relative error)

Error using Chirp modulated

CSBPE (Mean relative error)

R =2 0 0

R =4 0.0515 0.0461

R = 6 0.0738 0.0719

R = 8 0.0925 0.0992

R = 10 0.1244 0.1197

5.5.2 Experimental Results

Figure 5.5 shows the comparing real phantom results between the Fourier based CSBPE and

chirp modulation based CSBP encoding schemes.Figure 5.6(b)-(d) showChirp modulated

Fourier based CSBPE reconstructed images at different acceleration factors.Figure 5.7(e)-(g)

show error images with Chirp modulated Fourier based CSBPE for acceleration factors of 4,

6, and 8 respectively.Figure 5.8(h)-(j) show images reconstructed with Fourier encoding

based CSBPE using for acceleration factors of 4, 6, and 8 respectively.Figure 5.9(k)-(m):

show error images with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8

respectively. But, for the real phantom data case as shown in Figure 5.10, The Chirp

modulated CSBPE results are not significantly high quality compare to Fourier based CSBPE

methods.

Page 133: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

116

Chirp based BPE

images before CS

reconstruction

Chirp based BPE

images after CS

reconstructed image Error image

(a)

R = 4

(b)

R = 6

(c)

R = 8

(d)

R = 10

Figure 5.10:Simulation results for performance evaluation of chirp based compressed

sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the chirp

based BPE acquired data image and CS reconstructed images at reduction factor R =

4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS

reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show

the CS reconstructed images from chirp based BPE data can preserve image

resolution.

Page 134: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

117

Chirp based BPE images before CS

reconstruction(Phase)

Chirp based BPE images CS

reconstructed image (Phase)

Error image

(a)

R = 4

(b)

R = 6

(c)

R = 8

(d)

R = 10

Figure 5.11: Simulation results for performance evaluation of chirp based compressed

sensing reconstruction using BPE data; (a)-(d) (left two column) comparing the chirp

based BPE acquired data image and CS reconstructed images at reduction factor R =

4, 6, 8 and 10 respectively; (a)-(d) (right column) error images of Fourier based CS

reconstruction for acceleration factors of 4, 6, 8 and 10 respectively; The result show

the CS reconstructed images from chirp based BPE data can preserve image

resolution.

Page 135: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

118

Figure 5.12: Simulation results for Chirp modulated Fourier encoding based CSBPE.

Left column represents the reference image 256 × 256 (up/down: phase encodes,

left/right: frequency encode), Fig. (a) - (d) (middle column): shows reconstructed

images byChirp modulated Fourier basedCSBPE with random down sampling

patterns for acceleration factors of R= 4, 5.7, 6, 8 respectively and Fig.(a)-(d) (right

Column): shows the error images.

(a)

(b)

(c)

(d)

R =

R =

R =

R =

Page 136: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

119

Chirp encoded CSBPE Fourier encoded CSBPE Reconstructed image Error Reconstructed image Error

Figure 5.13: Comparing results between the Fourier based CSBPE and chirp

modulation based CSBPE encoding schemes (up/down: phase encodes, left/right:

frequency encode). (b)-(d): show images reconstructed withChirp modulated based

CSBPE for acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images

withChirp modulated based CSBPE for acceleration factors of 4, 6, and 8

respectively; (h)-(j): show images reconstructed withFourier encoding based CSBPE

using for acceleration factors of 4, 6, and 8 respectively; (k)-(m): show error images

withFourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively;

b

c

d

e

f

g

R = 4

R = 6

R = 8

Page 137: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

120

Chirp encoded CSBPE CS-MRI Reconstructed image Error Reconstructed image Error

Figure 5.14: Comparing results between the CS-MRI and chirp modulation based

CSBPE encoding schemes (up/down: phase encodes, left/right: frequency encode).(b)-

(d): show images reconstructed with Fourier encoding based CSBPE using for

acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images with Fourier

encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively; (h)-(j):

show images reconstructed with Chirp modulated based CSBPE for acceleration

factors of 4, 6, and 8 respectively; (k)-(m): show error images with Chirp modulated

based CSBPE for acceleration factors of 4, 6, and 8 respectively;

h

i

J

k

l

m

R = 4

R = 6

R = 8

Page 138: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

121

Chirp encoded CSBPE Fourier encoded CSBPE

Reconstructed image Error Reconstructed image Error

Figure 5.15: Compare the results between the Fourier based CSBPE and chirp

modulation based CSBPE encoding schemes (up/down: phase encodes, left/right:

frequency encode). (b)-(d): show images reconstructed with Chirp modulated based

CSBPE for acceleration factors of 4, 6, and 8 respectively; (e)-(g): show error images

with Chirp modulated based CSBPE for acceleration factors of 4, 6, and 8

respectively; (h)-(j): show images reconstructed with Fourier encoding based CSBPE

using for acceleration factors of 4, 6, and 8 respectively; (k)-(m): show error images

with Fourier encoded based CSBPE for acceleration factors of 4, 6, and 8 respectively;

R = 4

R = 6

R = 8 R = 8

R = 6

R = 4

Page 139: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

122

5.6 Conclusion

The study presented in this chapter was motivated by the fact that non-Fourier encoding

schemes have demonstrated better performance in CS reconstruction. There are a number of

previous works, including random spectrum, spread spectrum, Toeplitz encoding and

Noiselet encoding, that have validated the advantages of non-Fourier encoding for CS

imaging. Most of the studies presented before have used non-Fourier encoding for parallel

MRI or CS MRI, which have improved performance. Here we have introduced chirp

modulated non-Fourier in CSBPE to spread the signal energy in BPE space, aiming to further

improve the image quality of CSBPE method (Figure 5.3).

A key ingredient of CS MRI is to under-sample the Fourier encoded k-space that has most

signal energy at the central low frequencies. To acquire more energy, the under-sampling is

dense at the central low frequencies and sparse at the peripheral high frequencies. The k-

space data thus acquired results in the loss of high frequency energy and hence the loss of

image details. In contrast, chirp modulated non-Fourier encoding can spread the energy

evenly in the k-space, and hence allows for uniform random under-sampling to acquire the

signal energy evenly over the entire frequency range.

The k-space data thus acquired gives more image details and hence better image quality. In

principle, spreading signal energy in BPE space by chirp modulation should have the same

effect as it does in Fourier encoded k-space. However, the experiment results of this chapter

have shown that its impact on the improvement of image quality is minor. This is because the

signals in BPE space is distributed in non-rectangular grid, e.g. zig-zag grid, with its energy

already spread, to some extent, by the irregular grid.

Page 140: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

123

Chapter 6 Conclusion and Future work

Conclusion and Future work In this chapter, we will review the main contributions of this thesis and discuss some possible

future direction based on our research.

6.1 Conclusion MRI is an innovative medical imaging technology widely used in contemporary clinical

diagnosis and biological research. It has the capability to non-invasively produce high-

resolution images of the internal soft organ of human or animal in any direction. However, it

suffers from its lengthy data acquisition time caused by physical constraints. As discussed in

Chapters 1-3, despite the great research effort in recent years to shorten the data acquisition

time, it is still far away from achieving this goal. This thesis has presented a number of new

solutions to address this issue.

In Chapter 4, we have developed a new class of CSBPE methods to further reduce the data

acquisition time of MRI without compromising the image quality. These methods combine

the regular under-sampling scheme of BPE with the random under-sampling scheme of CS to

acquire significantly less k-space data than that of conventional BPE and CS MRI schemes.

The k-space data thus acquired are used in the iterative CS image reconstruction and then

BPE image reconstruction to obtain high quality image. For simulated zig-zag k-space data,

these new methods have produced high quality image at the high reduction (acceleration)

factor up to eight, which outperforms the conventional BPE and CS MRI schemes.

To overcome the difficulty of image ghost caused by the imprecise BPE trajectories in real

MRI scan, we have used cross-correlation technique to develop a new aliasing coefficient

correction method. This new method only requires the k-space data from normal scan to

effectively and precisely calculate the physically realised true trajectories. Hence, it avoids

the pre-scan calibration commonly required by the conventional aliasing coefficient

correction methods, and can significantly reduce the preparation time and operation cost of

MRI. The experiments on the real zig-zag k-space data have shown that this new correction

Page 141: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

124

method can effectively remove the image ghost of CSBPE caused by the imprecise BPE

trajectories in real MRI scans, yielding high quality images.

In chapter 5, we have investigated further improvement of CSBPE scheme by using chirp

modulated non-Fourier BPE. This investigation is based on the fact that chirp modulation can

spread the signal energy in the Fourier encoded k-space where the data is acquired/measured,

and hence improve the sampling incoherence and image quality of CS reconstruction. The

experiments on simulated data have shown that such approach has only resulted in some

minor improvement of image quality. This is because BPE itself can spread some signal

energy in the Fourier encoded k-space and hence improve the incoherence of signal sampling.

The effect of chirp modulation in BPE is therefore limited. Nevertheless, the chirp

modulation in CSBPE still yields some improvement in image quality.

6.2 Future Work Listed below are some possible directions for future research.

• The CSBP encoding sequence can be designed in 3D volume imaging. The simulation

results of CSBPE demonstrate that images reconstructed from 3D sub-sampled

Fourier and non-Fourier encoded data are of better quality than those from 2D sub-

sampled CSBPE data. The first step is to design the pulse sequence of Fourier

bunched phase encoding or chirp modulated non-Fourier bunched phase encoding for

volume imaging. After that, phantom and human body scanning should be conducted

in order to test the performance of the pulse sequence. The reconstructed image from

volume imaging data could be used to determine whether the expected image can be

obtained by directly applying the inverse Fourier transform and inverse chirp

transform on the volume imaging data.

• Zig-zag GRAPPA can be introduced into CSBPE to reconstruct the image and

compared with the CSBPE method presented in this thesis. We may randomly under-

sample the k-space encoded by the zig-zag BPE, and use CS reconstruction to

generate the aliased image. We can then use zig-zag GRAPPA to reconstruct the

original image. The image thus reconstructed and be then compared with that of

Page 142: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

125

CSBPE. This might lead to a zig-zag GRAPPA based CSBPE method with better

image quality.

• Some other non-Fourier encoding such as Noiselet can also be used to study the

performance of CSBPE method.

Page 143: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

126

REFERENCES

[1] [Online]. Available: https://www.nobelprize.org/prizes/physics/

[2] B. MacWilliams, "Russian claims first in magnetic imaging," Nature, vol. 426, pp. 375-375, 2003/11/01 2003.

[3] P. C. Lauterbur, "Image formation by induced local interactions: examples employing nuclear magnetic resonance," 1973.

[4] H. Pearson, "Magnetic pioneers net Nobel for putting medicine in the picture," ed: Nature Publishing Group, 2003.

[5] G. A. Wright, "Magnetic resonance imaging," IEEE Signal Processing Magazine, vol. 14, pp. 56-66, 1997.

[6] A. E. Chang, Y. L. Matory, A. J. Dwyer, S. C. Hill, M. E. Girton, S. M. Steinberg, et al., "Magnetic resonance imaging versus computed tomography in the evaluation of soft tissue tumors of the extremities," Ann Surg, vol. 205, pp. 340-8, Apr 1987.

[7] G. M. Bydder, R. E. Steiner, I. R. Young, A. S. Hall, D. J. Thomas, J. Marshall, et al., "Clinical NMR imaging of the brain: 140 cases," American Journal of Roentgenology, vol. 139, pp. 215-236, 1982/08/01 1982.

[8] B. Fischl and A. M. Dale, "Measuring the thickness of the human cerebral cortex from magnetic resonance images," Proceedings of the National Academy of Sciences of the United States of America, vol. 97, pp. 11050-11055, Sep 26 2000.

[9] S.-G. Kim and J. J. H. Ackerman, "Quantification of regional blood flow by monitoring of exogenous tracer via nuclear magnetic resonance spectroscopy," Magnetic Resonance in Medicine, vol. 14, pp. 266-282, 1990.

[10] S. Ogawa, T. M. Lee, A. R. Kay, and D. W. Tank, "Brain magnetic-resonance-imaging with contrast dependent on blood oxygenation," Proceedings of the National Academy of Sciences of the United States of America, vol. 87, pp. 9868-9872, 1990.

[11] C. L. G. Ham, J. M. L. Engels, G. T. van de Wiel, and A. Machielsen, "Peripheral nerve stimulation during MRI: Effects of high gradient amplitudes and switching rates," Journal of Magnetic Resonance Imaging, vol. 7, pp. 933-937, 1997.

[12] J. E. Bishop, G. E. Santyr, F. Kelcz, and D. B. Plewes, "Limitations of the keyhole technique for quantitative dynamic contrast-enhanced breast MRI," Jmri-Journal of Magnetic Resonance Imaging, vol. 7, pp. 716-723, Jul-Aug 1997.

[13] D. C. Noll, D. G. Nishimura, and A. Macovski, "Homodyne Detection in Magnetic-Resonance-Imaging," Ieee Transactions on Medical Imaging, vol. 10, pp. 154-163, Jun 1991.

[14] P. Mansfield, "Multi-Planar Image Formation Using NMR Spin Echoes," Journal of Physics C-Solid State Physics, vol. 10, pp. L55-L58, 1977 1977.

Page 144: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

127

[15] J. Hennig, A. Nauerth, and H. Friedburg, "RARE Imaging - A Fast Imaging Method for Clinical MR," Magnetic Resonance in Medicine, vol. 3, pp. 823-833, Dec 1986.

[16] D. A. Feinberg and K. Oshio, "GRASE (Gradient-Echo and Spin-Echo) MR Imaging - A New Fast Clinical Imaging Technique," Radiology, vol. 181, pp. 597-602, Nov 1991.

[17] P. B. Roemer, W. A. Edelstein, C. E. Hayes, S. P. Souza, and O. M. Mueller, "The NMR phased-array," Magnetic Resonance in Medicine, vol. 16, pp. 192-225, Nov 1990.

[18] D. K. Sodickson and W. J. Manning, "Simultaneous acquisition of spatial harmonics (SMASH): Fast imaging with radiofrequency coil arrays," Magnetic Resonance in Medicine, vol. 38, pp. 591-603, Oct 1997.

[19] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, "SENSE: sensitivity encoding for fast MRI," Magnetic resonance in medicine, vol. 42, pp. 952-962, 1999.

[20] M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang, et al., "Generalized autocalibrating partially parallel acquisitions (GRAPPA)," Magnetic Resonance in Medicine, vol. 47, pp. 1202-1210, 2002.

[21] R. W. Brown, Y.-C. N. Cheng, E. M. Haacke, M. R. Thompson, and R. Venkatesan, Magnetic resonance imaging: physical principles and sequence design: John Wiley & Sons, 2014.

[22] Z.-P. Liang and P. C. Lauterbur, Principles of magnetic resonance imaging: a signal processing perspective: SPIE Optical Engineering Press, 2000.

[23] A. Macovski, "Noise in MRI," Magnetic Resonance in Medicine, vol. 36, pp. 494-497, 1996.

[24] E. J. Candes, J. Romberg, and T. Tao, "Robust uncertainty principles: exact signal reconstruction from highly incomplete frequency information," IEEE Transactions on Information Theory, vol. 52, pp. 489-509, 2006.

[25] E. J. Candes and T. Tao, "Near-Optimal Signal Recovery From Random Projections: Universal Encoding Strategies?," IEEE Transactions on Information Theory, vol. 52, pp. 5406-5425, 2006.

[26] D. L. Donoho, "Compressed sensing," IEEE Transactions on Information Theory, vol. 52, pp. 1289-1306, 2006.

[27] D. Liang, B. Liu, J. Wang, and L. Ying, "Accelerating SENSE using compressed sensing," Magnetic Resonance in Medicine, vol. 62, pp. 1574-1584, 2009.

[28] E. J. Candès, "The restricted isometry property and its implications for compressed sensing," Comptes Rendus Mathematique, vol. 346, pp. 589-592, 2008 2008.

[29] E. Candes, M. Rudelson, T. Tao, and R. Vershynin, "Error correction via linear programming," in 46th Annual IEEE Symposium on Foundations of Computer Science (FOCS'05), 2005, pp. 668-681.

[30] C. Emmanuel and R. Justin, "Sparsity and incoherence in compressive sampling," Inverse Problems, vol. 23, p. 969, 2007.

[31] M. Elad, "Optimized projections for compressed sensing," IEEE Transactions on Signal Processing, vol. 55, pp. 5695-5702, 2007.

Page 145: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

128

[32] M. Lustig, D. Donoho, and J. M. Pauly, "Sparse MRI: The application of compressed sensing for rapid MR imaging," Magnetic Resonance in Medicine, vol. 58, pp. 1182-1195, 2007.

[33] J. P. Haldar, D. Hernando, and Z.-P. Liang, "Compressed-sensing MRI with random encoding," IEEE Transactions on Medical Imaging, vol. 30, pp. 893-903, 2011.

[34] M. Sandilya and S. R. Nirmala, "Compressed sensing trends in magnetic resonance imaging," Engineering Science and Technology, an International Journal, vol. 20, pp. 1342-1352, 2017/08/01/ 2017.

[35] M. Lustig, D. L. Donoho, J. M. Santos, and J. M. Pauly, "Compressed Sensing MRI," IEEE Signal Processing Magazine, vol. 25, pp. 72-82, 2008.

[36] L. Feng, T. Benkert, K. T. Block, D. K. Sodickson, R. Otazo, and H. Chandarana, "Compressed sensing for body MRI," Journal of Magnetic Resonance Imaging, vol. 45, pp. 966-987, 2017.

[37] M. Murphy, M. Alley, J. Demmel, K. Keutzer, S. Vasanawala, and M. Lustig, "Fast l(1)-SPIRiT Compressed Sensing Parallel Imaging MRI: Scalable Parallel Implementation and Clinically Feasible Runtime," IEEE Transactions on Medical Imaging, vol. 31, pp. 1250-1262, Jun 2012.

[38] F. Bloch, "Nuclear Induction," Physical Review, vol. 70, pp. 460-474, 10/01/ 1946.

[39] E. M. Purcell, H. C. Torrey, and R. V. Pound, "Resonance Absorption by Nuclear Magnetic Moments in a Solid," Physical Review, vol. 69, pp. 37-38, 01/01/ 1946.

[40] G. A. WRIGHT, "Magnetic Resonance Imaging," IEEE, p. 10, 1997.

[41] A. J. Bohris, U. Goerke, P. J. McDonald, M. Mulheron, B. Newling, and B. Le Page, "A broad line NMR and MRI study of water and water transport in portland cement pastes," Magnetic Resonance Imaging, vol. 16, pp. 455-461, 6// 1998.

[42] C. Ahn, J. Kim, and Z. Cho, "High-speed spiral-scan echo planar NMR imaging-I," IEEE transactions on medical imaging, vol. 5, pp. 2-7, 1986.

[43] C. H. Meyer, B. S. Hu, D. G. Nishimura, and A. Macovski, "Fast spiral coronary artery imaging," Magnetic resonance in medicine, vol. 28, pp. 202-213, 1992.

[44] J. I. Jackson, C. H. Meyer, D. G. Nishimura, and A. Macovski, "Selection of a convolution function for Fourier inversion using gridding (computerised tomography application)," IEEE transactions on medical imaging, vol. 10, pp. 473-478, 1991.

[45] N. Seiberlich, F. A. Breuer, M. Blaimer, K. Barkauskas, P. M. Jakob, and M. A. Griswold, "Non-Cartesian data reconstruction using GRAPPA operator gridding (GROG)," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 58, pp. 1257-1265, 2007.

[46] C. A. McKenzie, E. N. Yeh, M. A. Ohliger, M. D. Price, and D. K. Sodickson, "Self-calibrating parallel imaging with automatic coil sensitivity extraction," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 47, pp. 529-538, 2002.

Page 146: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

129

[47] M. Weiger, K. P. Pruessmann, and P. Boesiger, "Cardiac real-time imaging using SENSE," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 43, pp. 177-184, 2000.

[48] Y. Kurihara, Y. K. Yakushiji, I. Tani, Y. Nakajima, and M. Van Cauteren, "Coil sensitivity encoding in MR imaging: advantages and disadvantages in clinical practice," American Journal of Roentgenology, vol. 178, pp. 1087-1091, 2002.

[49] J. S. Van den Brink, Y. Watanabe, C. K. Kuhl, T. Chung, R. Muthupillai, M. Van Cauteren, et al., "Implications of SENSE MR in routine clinical practice," European journal of radiology, vol. 46, pp. 3-27, 2003.

[50] M. Weiger, K. P. Pruessmann, and P. Boesiger, "2D SENSE for faster 3D MRI," Magnetic Resonance Materials in Physics, Biology and Medicine, vol. 14, pp. 10-19, 2002.

[51] H. Köstler, J. J. Sandstede, C. Lipke, W. Landschütz, M. Beer, and D. Hahn, "Auto-SENSE perfusion imaging of the whole human heart," Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 18, pp. 702-708, 2003.

[52] F. H. Lin, K. K. Kwong, J. W. Belliveau, and L. L. Wald, "Parallel imaging reconstruction using automatic regularization," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 51, pp. 559-567, 2004.

[53] F. A. Breuer, P. Kellman, M. A. Griswold, and P. M. Jakob, "Dynamic autocalibrated parallel imaging using temporal GRAPPA (TGRAPPA)," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 53, pp. 981-985, 2005.

[54] F. Huang, J. Akao, S. Vijayakumar, G. R. Duensing, and M. Limkeman, "k-t GRAPPA: A k-space implementation for dynamic MRI with high reduction factor," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 54, pp. 1172-1184, 2005.

[55] Z. Wang, J. Wang, and J. A. Detre, "Improved data reconstruction method for GRAPPA," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 54, pp. 738-742, 2005.

[56] J. Park, Q. Zhang, V. Jellus, O. Simonetti, and D. Li, "Artifact and noise suppression in GRAPPA imaging using improved k-space coil calibration and variable density sampling," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 53, pp. 186-193, 2005.

[57] M. Blaimer, F. A. Breuer, M. Mueller, N. Seiberlich, D. Ebel, R. M. Heidemann, et al., "2D-GRAPPA-operator for faster 3D parallel MRI," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 56, pp. 1359-1364, 2006.

[58] D. Huo and D. L. Wilson, "Robust GRAPPA reconstruction and its evaluation with the perceptual difference model," Journal of Magnetic Resonance Imaging, vol. 27, pp. 1412-1420, 2008.

[59] T. Zhao and X. Hu, "Iterative GRAPPA (iGRAPPA) for improved parallel imaging reconstruction," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 59, pp. 903-907, 2008.

Page 147: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

130

[60] U. Gamper, P. Boesiger, and S. Kozerke, "Compressed sensing in dynamic MRI," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 59, pp. 365-373, 2008.

[61] J. Yang, Y. Zhang, and W. Yin, "A fast alternating direction method for TVL1-L2 signal reconstruction from partial Fourier data," IEEE Journal of Selected Topics in Signal Processing, vol. 4, pp. 288-297, 2010.

[62] D. S. Smith, J. C. Gore, T. E. Yankeelov, and E. B. Welch, "Real-time compressive sensing MRI reconstruction using GPU computing and split Bregman methods," International journal of biomedical imaging, vol. 2012, 2012.

[63] H. Moriguchi and J. L. Duerk, "Bunched phase encoding (BPE): A new fast data acquisition method in MRI," Magnetic Resonance in Medicine, vol. 55, pp. 633-648, 2006.

[64] H. Moriguchi, J. Sunshine, and J. Duerk, "Further scan time reduction of bunched phase encoding using sensitivity encoding," in Proceedings of the 13th Annual Meeting of ISMRM, Miami Beach, Florida, USA, 2005, p. 287.

[65] N. Seiberlich, F. A. Breuer, P. Ehses, H. Moriguchi, M. Blaimer, P. M. Jakob, et al., "Using the GRAPPA operator and the generalized sampling theorem to reconstruct undersampled non-Cartesian data," Magnetic resonance in medicine, vol. 61, pp. 705-715, 2009.

[66] A. Trakic, B. K. Li, E. Weber, H. Wang, S. Wilson, and S. Crozier, "A rapidly rotating RF coil for MRI," Concepts in Magnetic Resonance Part B: Magnetic Resonance Engineering: An Educational Journal, vol. 35, pp. 59-66, 2009.

[67] A. Trakic, H. Wang, E. Weber, B. Li, M. Poole, F. Liu, et al., "Image reconstructions with the rotating RF coil," Journal of Magnetic Resonance, vol. 201, pp. 186-198, 2009.

[68] M. Li, Z. Zuo, J. Jin, R. Xue, A. Trakic, E. Weber, et al., "Highly accelerated acquisition and homogeneous image reconstruction with rotating RF coil array at 7 T—A phantom based study," Journal of Magnetic Resonance, vol. 240, pp. 102-112, 2014.

[69] B. Bilgic, B. A. Gagoski, S. F. Cauley, A. P. Fan, J. R. Polimeni, P. E. Grant, et al., "Wave-CAIPI for highly accelerated 3D imaging," Magnetic resonance in medicine, vol. 73, pp. 2152-2162, 2015.

[70] S. Chen, B. Qiu, F. Zhao, C. Li, and H. Du, "Fast Compressed Sensing MRI Based on Complex Double-Density Dual-Tree Discrete Wavelet Transform," International Journal of Biomedical Imaging, vol. 2017, p. 13, 2017.

[71] S. Geethanath, "Novel Applications Of Compressed Sensing To Magnetic Resonance Imaging And Spectroscopy," 2012.

[72] E. J. Candes and J. K. Romberg, "Signal recovery from random projections," 2005, pp. 76-86.

[73] E. J. Candes and T. Tao, "Decoding by linear programming," IEEE Transactions on Information Theory, vol. 51, pp. 4203-4215, 2005.

[74] K. Ni, S. Datta, P. Mahanti, S. Roudenko, and D. Cochran, "Efficient deterministic compressed sensing for images with chirps and Reed–Muller codes," SIAM Journal on Imaging Sciences, vol. 4, pp. 931-953, 2011.

Page 148: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

131

[75] L. Applebaum, S. D. Howard, S. Searle, and R. Calderbank, "Chirp sensing codes: Deterministic compressed sensing measurements for fast recovery," Applied and Computational Harmonic Analysis, vol. 26, pp. 283-290, 2009.

[76] M. Elad, Sparse and redundant representations: from theory to applications in signal and image processing: Springer Science & Business Media, 2010.

[77] T. Tuma and P. Hurley, "On the incoherence of noiselet and Haar bases," in SAMPTA'09, 2009, p. General session.

[78] R. Otazo, D. Kim, L. Axel, and D. K. Sodickson, "Combination of compressed sensing and parallel imaging for highly accelerated first-pass cardiac perfusion MRI," Magnetic Resonance in Medicine, vol. 64, pp. 767-776, 2010.

[79] Y. Erlich, N. Shental, A. Amir, and O. Zuk, "Compressed sensing approach for high throughput carrier screen," in 2009 47th Annual Allerton Conference on Communication, Control, and Computing (Allerton), 2009, pp. 539-544.

[80] R. M. Kainkaryam, A. Bruex, A. C. Gilbert, J. Schiefelbein, and P. J. Woolf, "poolMC: Smart pooling of mRNA samples in microarray experiments," BMC bioinformatics, vol. 11, p. 299, 2010.

[81] N. Shental, A. Amir, and O. Zuk, "Identification of rare alleles and their carriers using compressed se (que) nsing," Nucleic acids research, vol. 38, pp. e179-e179, 2010.

[82] S. Muthukrishnan, "Data streams: Algorithms and applications," Foundations and Trends® in Theoretical Computer Science, vol. 1, pp. 117-236, 2005.

[83] G. Cormode and M. Hadjieleftheriou, "Finding the frequent items in streams of data," Communications of the ACM, vol. 52, pp. 97-105, 2009.

[84] M. Lustig and J. M. Pauly, "SPIRiT: Iterative self-consistent parallel imaging reconstruction from arbitrary k-space," Magnetic Resonance in Medicine, vol. 64, pp. 457-471, 2010.

[85] T. B. Pi, "Generalized Sampling Expansion," IEEE Transactions on Circuits and Systems, vol. 24, 1977.

[86] K. R. I. Jingxin Zhang, Kai Zhu, "Compressed Sensing MRI Using Bunched Phase Encoding," ISMRM 25th annual Meeting and Exhibition, April 2017 2017.

[87] A. Papoulis, "Generalized sampling expansion," IEEE transactions on circuits and systems, vol. 24, pp. 652-654, 1977.

[88] C. M. J. V. Uijen, J. H. D. Boef, and F. J. J. Verschuren, "Fast fourier imaging," Magnetic Resonance in Medicine, vol. 2, pp. 203-217, 1985.

[89] E. M. Haacke, F. H. Bearden, J. R. Clayton, and N. R. Linga, "Reduction of MR imaging time by the hybrid fast-scan technique," Radiology, vol. 158, pp. 521-529, 1986.

[90] J. N. Sarvaiya, S. Patnaik, and S. Bombaywala, "Image Registration by Template Matching Using Normalized Cross-Correlation," in 2009 International Conference on Advances in Computing, Control, and Telecommunication Technologies, 2009, pp. 819-822.

Page 149: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

132

[91] L. P. Panych, G. P. Zientara, and F. A. Jolesz, "MR image encoding by spatially selective rf excitation: An analysis using linear response models," International journal of imaging systems and technology, vol. 10, pp. 143-150, 1999.

[92] L. P. Panych, "Theoretical comparison of Fourier and wavelet encoding in magnetic resonance imaging," IEEE transactions on medical imaging, vol. 15, pp. 141-153, 1996.

[93] L. P. Panych, G. P. Zientara, P. Saiviroonporn, S. S. Yoo, and F. A. Jolesz, "Digital wavelet-encoded MRI: a new wavelet-encoding methodology," Journal of Magnetic Resonance Imaging, vol. 8, pp. 1135-1144, 1998.

[94] L. P. Panych, C. Oesterle, G. P. Zientara, and J. Hennig, "Implementation of a fast gradient-echo SVD encoding technique for dynamic imaging," Magnetic resonance in medicine, vol. 35, pp. 554-562, 1996.

[95] D. Mitsouras, W. S. Hoge, F. J. Rybicki, W. E. Kyriakos, A. Edelman, and G. P. Zientara, "Non-Fourier-encoded parallel MRI using multiple receiver coils," Magnetic Resonance in Medicine: An Official Journal of the International Society for Magnetic Resonance in Medicine, vol. 52, pp. 321-328, 2004.

[96] K. Pawar, G. Egan, and J. Zhang, "Multichannel Compressive Sensing MRI Using Noiselet Encoding," PLoS ONE, vol. 10, pp. 1-27, 2015.

[97] H. Wang, D. Liang, K. F. King, and L. Ying, "Three-dimensional hybrid-encoded MRI using compressed sensing," in 2012 9th IEEE International Symposium on Biomedical Imaging (ISBI), 2012, pp. 398-401.

[98] D. Liang, G. Xu, H. Wang, K. F. King, D. Xu, and L. Ying, "Toeplitz random encoding MR imaging using compressed sensing," in 2009 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2009, pp. 270-273.

[99] G. Puy, J. P. Marques, R. Gruetter, J. P. Thiran, D. V. D. Ville, P. Vandergheynst, et al., "Spread Spectrum Magnetic Resonance Imaging," IEEE Transactions on Medical Imaging, vol. 31, pp. 586-598, 2012.

[100] S. A. A. Hasani, G. F. Egan, and J. Zhang, "Spread Spectrum Using Chirp Modulated RF Pulses for Incoherent Sampling Compressive Sensing MRI," International Journal of Signal Processing Systems, vol. 4, pp. 1-5, 2016.

[101] F. Sebert, Y. Zou, and L. Ying, "Compressed sensing MRI with random B1 field," Proc ISMRM,(Toronto, Canada, 2008), p. 3151, 2008.

[102] Y. Wiaux, G. Puy, R. Gruetter, J. P. Thiran, D. V. D. Ville, and P. Vandergheynst, "Spread spectrum for compressed sensing techniques in magnetic resonance imaging," in 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2010, pp. 756-759.

[103] Y. Chen, J. Li, X. Qu, L. Chen, C. Cai, S. Cai, et al., "Partial Fourier transform reconstruction for single-shot MRI with linear frequency-swept excitation," Magnetic Resonance in Medicine, vol. 69, pp. 1326-1336, 2013.

Page 150: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

133

List of Publications

1. Jingxin Zhang and Kazi Rafiqul Islam, "Compressed Sensing MRI Using

Bunched Phase Encoding", Presented in ISMRM 25th Annual Meeting and

Exhibition, Honolulu, USA, 30 May-05 June, 2017.

2. Kazi Rafiqul Islam and Jingxin Zhang, "Simple and effective trajectory

estimation for image reconstruction of accelerated k-space acquisition on

non-rectangular periodic trajectories", Presented in ISMRM 27th Annual

Meeting and Exhibition, Vancouver, Canada, 11May-17May, 2019.

Page 151: Compressed Sensing Magnetic Resonance Imaging UsingFourier … · 2020. 8. 13. · Compressed Sensing Magnetic Resonance Imaging UsingFourier and Non -Fourier Based Bunched Phase

134

THE END