RECONSTRUCTION METHODS

FOR FAST MAGNETIC RESONANCE IMAGING

A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ELECTRICAL
ENGINEERING
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

Philip James Beatty

December 2006
c Copyright by Philip James Beatty 2007
All Rights Reserved

ii
I certify that I have read this dissertation and that, in my opinion, it
is fully adequate in scope and quality as a dissertation for the degree
of Doctor of Philosophy.

(Dwight Nishimura) Principal Adviser

I certify that I have read this dissertation and that, in my opinion, it

is fully adequate in scope and quality as a dissertation for the degree
of Doctor of Philosophy.

(John Pauly)

I certify that I have read this dissertation and that, in my opinion, it

is fully adequate in scope and quality as a dissertation for the degree
of Doctor of Philosophy.

(Brad Osgood)

Approved for the University Committee on Graduate Studies.

iii
iv
Abstract

Magnetic Resonance Imaging (MRI) is a very successful method for imaging the
body, due in large part to the excellent soft tissue contrast that can be obtained.
A major challenge for MRI is reducing the long scan times that can be required
to obtain an image. Fast Magnetic Resonance Imaging uses sophisticated encoding
techniques, such as non-Cartesian k-space trajectories and parallel imaging, to reduce
the scan times required for MRI and requires advanced MRI scanner hardware and
reconstruction methods.
This work focuses on reconstruction methods for fast MRI. In this dissertation,
improvements that can be made to the gridding method for reconstructing MR images
encoded using non-Cartesian k-space trajectories are described. In addition, a new
method called Anti-aliasing Partially Parallel Encoded Acquisition Reconstruction
(APPEAR) is introduced and developed for reconstructing magnetic resonance images
encoded using non-Cartesian k-space trajectories and parallel imaging.
The improvements to the gridding method described in this work include using
a minimal oversampling ratio, improved design for a sampled convolution kernel, re-
duced field-of-view reconstruction and using block grid storage. Used together, these
improvements can result in a three-fold reduction in computation memory require-
ments and can reduce the reconstruction time by a factor of approximately thirty
times for three-dimensional (3-D) image reconstruction, compared to the use of a
Kaiser-Bessel convolution kernel on a 2X oversampled grid using conventional line-
by-line and slice-by-slice grid storage.
The APPEAR method is a parallel imaging reconstruction method that can be
used with arbitrary k-space trajectories, is non-iterative and does not need to estimate

v
coil sensitivity functions. In this work, the mathematical framework for parallel
imaging reconstruction is extended and this extended framework is used to develop
and justify the APPEAR method. The concept of correlation values is introduced and
used to improve the efficiency of the APPEAR method. Phantom and in-vivo results
are shown for 1-D non-Cartesian k-space trajectories and variable-density spiral k-
space trajectories.

vi
Acknowledgements

This dissertation is the final work of my graduate school career, which has spanned
a five year period of great change and growth in my life. These years have been very
good to me, in large part because of the help and support of many people.
Supporting oneself financially while pursuing graduate studies can be challenging
and so I am very grateful for the generous financial support I have received, making
it possible for me to come to Stanford and focus on my studies. The financial aid
for my first year at Stanford was through the Department of Electrical Engineering
and was provided by the Rose M. Chappelear Memorial Fund. I was subsequently
named a William R. Hewlett Fellow and the financial aid for my second, third and
fourth years was provided by the William R. Hewlett Stanford Graduate Fellowship.
In my fifth year, I was supported through a research associateship under my advisor,
Professor Dwight Nishimura.
Professors Dwight Nishimura, John Pauly and Brad Osgood, who form my dis-
sertation reading committee and sat on my oral defense committee, have spent a lot
of time reading various drafts of this dissertation and discussing this work with me
and have helped to make this dissertation much stronger than when the first draft
was ejected from the printer. I would also like to thank Professor Donald Cox, who
was the chair of my oral defense committee, for his helpful feedback.
This dissertation and the research behind it would not have been possible without
the continued support and encouragement of Professor Dwight Nishimura. I met
Dwight in the Spring of 2001, when I visited Stanford to explore the possibility of
graduate studies at Stanford. That meeting had a large influence on my decision
to attend Stanford. During my time at Stanford, Dwight has been my principal

vii
advisor in title as well as in action. Dwight is a clear thinker and this manifests
itself in his accessible teaching style, his germane questions and his ability to read
a long manuscript and quickly identify the weak points. I have learned a great deal
from him, both in and out of the classroom and I would like to thank him for his
encouragement, support and commitment to me, my education and this work.

It was through the Medical Image Reconstruction class taught by Professor John
Pauly, my associate advisor, that I started on the first project that is a part of this
dissertation: using a minimally oversampled grid in gridding reconstruction. John
has been a great help in shaping the ideas contained in this dissertation. John is
enthusiastic, friendly and most importantly, patient. There have been many times
when it has taken me weeks of thinking to figure out how insightful John’s comments
had been in a previous discussion. Both Dwight and John have worked hard for many
years to build a strong research group with the right people and equipment for MR
research and without the infrastructure of the Magnetic Resonance Systems Research
Laboratory (MRSRL), my work could not have started.

There are many people in the MRSRL that have helped me with problems both
big and small. For helpful discussions full of great advice that have helped me to
find my way, I would like to thank Krishna Nayak, Brian Hargreaves, Greig Scott,
and Steve Conolly. In addition to great advice, Brian has directly and significantly
helped my work, providing me with pulse sequence code and Matlab scripts and has
come in early in the morning to scan with me on more than one occasion.

For their friendship and all of their research related and non-research related help I
would like to thank my office mates and colleagues: Jin-Hyung Lee, Sharon Ungersma,
Paul Gurney, Jongho Lee, Logi Viarsson, Michael Lustig, Peter Larson, Julie DiCarlo,
Juan Santos, Bob Schaffer, Jong Park, Nathaniel Matter and Ross Venook. I would
like to especially thank Paul Gurney, with whom I shared an adjoining cubicle for
the majority of my time at Stanford and hotel rooms during the ISMRM conferences.
Paul has been the first person that I turn to to discuss a new idea or problem and he
has always been very generous with his time and brilliant with his ideas.

In addition to the aforementioned professors, mentors and colleagues at Stanford,

viii
I would like to acknowledge and thank Professor R. Mark Henkelman of the Univer-
sity of Toronto for his advice over the years and for engaging me in the problems
of MR image reconstruction during my internship with the Mouse Imaging Centre
(MICe) in Toronto in the summer of 2002. As well, I would like to thank Jean Brit-
tain for inviting me to work on parallel imaging reconstruction at General Electric’s
Applied Science Laboratory in Menlo Park during the summer of 2005. Working with
Anja Brau, a lot of progress was made during this summer in understanding parallel
imaging reconstruction methods and this has greatly influenced my work.
Lastly, I would like to acknowledge my family for their support and encourage-
ment. I would especially like to thank my loving wife, Tara Tappuni, who has made
more sacrifices than anyone, including myself, to give me the opportunity to come to
Stanford and accomplish this work.

ix
x
Contents

Abstract v

Acknowledgements vii

1 Introduction 1
1.1 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Setting the Stage 5

2.1 Creating Transverse Magnetization . . . . . . . . . . . . . . . . . . . 6
2.2 The Emitted Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 The Received Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3 Gridding 23
3.1 The Gridding Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.2 Aliasing Amplitude: A Metric for Gridding Accuracy . . . . . . . . . 31
3.3 Choosing a Kaiser-Bessel Kernel . . . . . . . . . . . . . . . . . . . . . 33
3.4 Presampling the Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5 Choosing the Optimal Sampled Kernel . . . . . . . . . . . . . . . . . 45
3.6 The Block Storage Format . . . . . . . . . . . . . . . . . . . . . . . . 47
3.7 FOV and Resolution Flexibility . . . . . . . . . . . . . . . . . . . . . 49
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4 APPEAR 55
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

xi
 4.2 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.1 SENSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.2 APPEAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.2.3 Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5 Correlation Values 97
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.2 Analysis of Computation . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 What are Correlation Values? . . . . . . . . . . . . . . . . . . . . . . 102
5.4 Fast Computation of Correlation Values . . . . . . . . . . . . . . . . 106
5.5 Graphical Interpretation and Regularization . . . . . . . . . . . . . . 108
5.6 Gridding and Local Projection Interpolation . . . . . . . . . . . . . . 118
5.7 Approximating Correlation Values . . . . . . . . . . . . . . . . . . . . 124
5.8 Computing Approximate Correlation Values . . . . . . . . . . . . . . 129
5.9 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
5.10 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
5.11 Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . . 135

6 Summary and Future Recommendations 139

6.1 Future Recommendations . . . . . . . . . . . . . . . . . . . . . . . . 140

A Aliasing Amplitude 143

A.1 Nearest-Neighbor Interpolation . . . . . . . . . . . . . . . . . . . . . 145
A.2 Linear Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . 145

B Apodization Correction 147

C Calculating the Optimal Sampled Kernel 149

xii
D Noise in APPEAR Reconstructions 151

Bibliography 153

xiii
xiv
List of Tables

3.1 General Variables and Functions . . . . . . . . . . . . . . . . . . . . . 27

3.2 Gridding Kernel Variables and Functions . . . . . . . . . . . . . . . . 28
3.3 Comparison of β Values . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.1 General Variables and Functions . . . . . . . . . . . . . . . . . . . . . 58

4.2 Data Synthesis Variables . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.3 Local Projection Calibration Variables . . . . . . . . . . . . . . . . . 66
4.4 Central Region Interpolation Variables . . . . . . . . . . . . . . . . . 78

xv
xvi
List of Figures

2.1 Axial knee with two different types of contrast . . . . . . . . . . . . . 6

2.2 Creating transverse magnetization . . . . . . . . . . . . . . . . . . . . 7
2.3 Gradient encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Tracing a path in k-space . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Example k-space trajectories . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 DFT point-spread function . . . . . . . . . . . . . . . . . . . . . . . . 17
2.7 Spiral point-spread function . . . . . . . . . . . . . . . . . . . . . . . 19
2.8 Sensitivity encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

3.1 Circular convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2 Computation improvement . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Aliasing amplitude versus image reconstruction error . . . . . . . . . 34
3.4 Aliasing amplitude for a selection of kernels . . . . . . . . . . . . . . 35
3.5 Reconstruction of a sagittal knee image . . . . . . . . . . . . . . . . . 36
3.6 Choosing a kernel appropriate for the oversampling ratio . . . . . . . 37
3.7 Convolution kernel design . . . . . . . . . . . . . . . . . . . . . . . . 38
3.8 Determining the kernel sampling density . . . . . . . . . . . . . . . . 41
3.9 Nearest-neighbor and triangle kernel reconstruction errors . . . . . . 43
3.10 Error comparison using sampled kernels . . . . . . . . . . . . . . . . . 44
3.11 Error comparison: Kaiser-bessel kernel versus optimal sampled kernel 46
3.12 Line-by-line storage versus block storage . . . . . . . . . . . . . . . . 48
3.13 Spiral-PR and Cones Trajectories . . . . . . . . . . . . . . . . . . . . 49
3.14 Memory page access . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.15 FFT with the block storage format . . . . . . . . . . . . . . . . . . . 51

xvii
3.16 FFT compute time versus size . . . . . . . . . . . . . . . . . . . . . . 52
3.17 Reduced FOV gridding . . . . . . . . . . . . . . . . . . . . . . . . . . 53

4.1 Acquisition, local and pattern sets . . . . . . . . . . . . . . . . . . . . 65

4.2 Central and fit regions . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.3 A pattern set not always in the acquisition set . . . . . . . . . . . . . 74
4.4 Block diagram of APPEAR . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Oversampling and apodization . . . . . . . . . . . . . . . . . . . . . . 79
4.6 Oversampling and apodization: image space view . . . . . . . . . . . 82
4.7 1-D trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.8 Numerical phantom results . . . . . . . . . . . . . . . . . . . . . . . . 86
4.9 Error in estimated encoding function images . . . . . . . . . . . . . . 88
4.10 Ball phantom results . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.11 in vivo brain results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.12 Error in estimated encoding function with rotation . . . . . . . . . . 93

5.1 Graphical interpretation of encoding . . . . . . . . . . . . . . . . . . 110

5.2 Graphical representation of a local subspace . . . . . . . . . . . . . . 111
5.3 Choice of linear combination weights . . . . . . . . . . . . . . . . . . 112
5.4 Graphical interpretation of regularization . . . . . . . . . . . . . . . . 114
5.5 Effect of regularization on gain . . . . . . . . . . . . . . . . . . . . . 115
5.6 Uneven 1-D sampling pattern . . . . . . . . . . . . . . . . . . . . . . 121
5.7 Local projection interpolation with uneven sampling . . . . . . . . . . 123
5.8 Shepp-Logan phantom results with uneven sampling . . . . . . . . . . 124
5.9 Graphical comparison of correlation value approximations . . . . . . . 128
5.10 Block diagram of efficient APPEAR implementation . . . . . . . . . . 129
5.11 Correlation values and a calibration region of limited extent . . . . . 130
5.12 Diagonal interpolation of correlation values . . . . . . . . . . . . . . . 133
5.13 Shepp-Logan spiral results . . . . . . . . . . . . . . . . . . . . . . . . 135
5.14 Head coil results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
5.15 Cardiac coil results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

xviii
A.1 Sampled and interpolated kernel . . . . . . . . . . . . . . . . . . . . . 144
A.2 Accuracy of ε1 approximations . . . . . . . . . . . . . . . . . . . . . . 146

xix
xx
Chapter 1

Introduction

Magnetic Resonance Imaging (MRI) is a very successful method for imaging the body,
due in large part to the excellent soft tissue contrast that can be obtained. However,
a major drawback of MRI is the long scan times that can be required to obtain an
image.
In order to reduce the time needed to acquire sufficient information to reconstruct
an image, fast MRI uses sophisticated encoding techniques, such as non-Cartesian k-
space trajectories and encoding using receiver-coil sensitivities. Accurate and efficient
image reconstruction from signals acquired using these sophisticated encoding tech-
niques is a challenging problem and the focus of this work. A brief outline describing
the focus of each chapter is provided below.

1.1 Thesis Outline

Chapter two provides a brief introduction to MRI, showing how the MR signal is
produced, acquired and modeled. The graphical and mathematical perspective used
in this chapter sets the stage for the later chapters, which develop and describe
methods for constructing images from the acquired MR signal.
The first method that I develop in detail is the gridding method, the topic of
chapter three. The gridding method has been used in medical imaging reconstruction

1
2 CHAPTER 1. INTRODUCTION

for over twenty years [1], as an efficient way to reconstruct images encoded with non-
Cartesian k-space trajectories. After describing the basic gridding method, I describe
a number of improvements that can be made to the method that result in signifi-
cant reductions in computation memory and time without degrading the quality of
the reconstructed image. These improvements include: using a minimal oversam-
pling ratio, improved design for a sampled convolution kernel, reduced field-of-view
reconstruction and using block grid storage. Used together, these improvements can
reduce gridding reconstruction time by a factor of approximately thirty times for
three-dimensional (3-D) image reconstructions.
Chapters four and five develop and explore a new method for parallel imaging
reconstruction—reconstruction of images that have been encoded by a combination
of gradient encoding and receiver coil sensitivities. This new method, called Anti-
aliasing Partially Parallel Encoded Acquisition Reconstruction (APPEAR), can be
used with arbitrary k-space trajectories, is non-iterative and does not need to estimate
coil sensitivity functions.
Chapter four gives a brief overview of previous work on parallel imaging and then
extends the theory of multiple receiver coil reconstruction. The APPEAR method
is developed and described using the extended theory introduced in this chapter.
Phantom and in vivo results for a 2D trajectory with variable density sampling in
the phase-encode direction show that the APPEAR method is able to remove the
aliasing artifacts caused by reduced gradient encoding.
While chapter four develops and describes the APPEAR method and shows that
the method is able to achieve good image quality, chapter five deals with the compu-
tation required by the APPEAR method. Insight into the computation required by
APPEAR is given by a new concept: correlation values. In chapter five, the concept
of correlation values is defined and developed. While the proposed implementation
of the APPEAR method given in chapter four is pedagogically useful, chapter five
shows that correlation values lead to an implementation of the APPEAR method that
improves the computational efficiency ten-fold. As well as being significant from a
computation point of view, the concept of correlation values can give helpful insight
into the APPEAR method in particular and MR image reconstruction in general.
1.1. THESIS OUTLINE 3

Significant portions of chapter five are devoted to developing this insight.

Chapter six concludes this work, summarizing what I have accomplished in the
area of reconstruction methods for fast magnetic resonance imaging and giving my
thoughts on future directions in this area.
4 CHAPTER 1. INTRODUCTION
Chapter 2

Setting the Stage

When an MRI scanner images a part of the body, it obtains an image of the transverse
magnetization at each spatial location. Unlike the X-ray attenuation coefficient, which
is a simple material property that can be imaged using computed tomography (CT),
the transverse magnetization imaged by MRI is not a simple material property.
Magnetic fields are used by an MRI scanner to produce, or excite, transverse
magnetization in the tissue being imaged. The amount of transverse magnetiza-
tion excited in the tissue depends on material properties of the tissue as well as the
magnetic fields that are applied to the tissue. Moreover, by controlling the applied
magnetic fields, it is possible to vary how the transverse magnetization relates to the
material properties of the tissue, allowing MRI scanners to achieve different types of
contrast between tissue types as shown in Fig. 2.1.
An MRI scanner works by first creating transverse magnetization that provides
a useful contrast in the body part to be imaged, and then collecting information to
form an image of the transverse magnetization that has been created. This work
is concerned with the latter stage: understanding the information that is collected
about the transverse magnetization and how this information can be used to form
an image. In this chapter, a model that describes the collected information is de-
veloped, providing a foundation for the following chapters, which focus on methods
for efficiently transforming the collected information into images. The model used in
this chapter is based on a classical description of MR. Although not as rigorous as a

5
6 CHAPTER 2. SETTING THE STAGE

(a) (b)

Figure 2.1: Axial knee images shown with two different types of contrast that can be
obtained using an MRI scanner. (a) T2-weighted image with suppression of signal
from fat tissue. (b) T1-weighted image.

quantum mechanical description, the classical description leads to an intuitive model

that is sufficient for the purposes of this work.

2.1 Creating Transverse Magnetization

To create transverse magnetization, a typical MRI scanner uses three complimentary
systems: 1) a magnet that creates a strong homogeneous magnetic field B0 , 2) a radio-
frequency pulse generator and transmit coil that can produce a time-varying magnetic
field at the resonant frequency of the magnetization and 3) a gradient system with x,
y and z-axis gradient coils that can produce a magnetic field that varies in time and
linearly in space.
The B0 field is perpetually existent within an MRI scanner, its direction defining
the longitudinal direction, typically plotted along the z-axis. When an object is
placed in the scanner, the B0 field causes the magnetic moments of certain nuclei to
tend to align along the longitudinal direction, producing a net magnetic moment, or
2.1. CREATING TRANSVERSE MAGNETIZATION 7

z z

1H 1
H
1

1H
H

H
1
1H
1

1H
1 H
H

1
H
y y

x x

(a) (b)
z z
T1
T2

1
H 1
H
1H

1
1H H

H
1
1

1H
1
H H
1
H

y y

x x

(c) (d)

Figure 2.2: (a) Strong external magnetic field aligns magnetic moments, producing
a net magnetization in the direction of the field. (b) Radio-frequency pulse tips the
magnetization out of equilibrium. (c) Tipped out of equilibrium, the magnetization
precesses in the transverse (xy) plane, emitting a radio-frequency signal at a frequency
proportional to the magnetic field. (d) In addition to precessing, the magnetization
also returns to equilibrium.

magnetization, in the longitudinal direction as shown in Fig. 2.2(a).

In the body, the most abundant, by far, of the affected nuclei are hydrogen nu-
clei, found in both water and lipids. Each of the nuclei has a resonant frequency
proportional to the surrounding magnetic field. The constant of proportionality γ
is called the gyromagnetic ratio and for hydrogen nuclei, γ/2π = 4257 Hz / Gauss.
Thus, in a 1.5T magnetic field, a common value for B0 , hydrogen nuclei will have
a resonant frequency of about 64 MHz. The resonant frequency is also called the
Larmor frequency.
The radio-frequency (RF) pulse generator and transmit coil are used to tip the
magnetization out of equilibrium, creating transverse magnetization as shown in Fig.
8 CHAPTER 2. SETTING THE STAGE

2.2(b). While the magnitude of the RF magnetic field is many orders of magnitude
smaller than the B0 field, it succeeds in tipping the magnetization by operating at the
resonant frequency of the magnetization. Because only magnetization whose resonant
frequency is the same as the frequency of the RF pulse is tipped, the gradient system
can be used to limit the locations where transverse magnetization is created. For
example, applying a gradient field along the x-axis will cause the resonant frequency
to vary linearly in the x direction. Applying an RF pulse while this gradient field
is applied will create transverse magnetization in a yz plane at a single x location.
Together, the RF system and gradient system allow for great flexibility in creating
transverse magnetization.

Mathematically, the transverse magnetization is modeled as a complex-valued

function m(r), where r is an (x, y, z) three-tuple that gives the spatial location. By
convention, the real part of m(r) corresponds to the x-component of the transverse
magnetization and the imaginary part to the y-component of the transverse magne-
tization. As shown in Fig. 2.2(c), when the magnetic moments are tipped out of
equilibrium, they precess in the transverse (xy) plane at the Larmor frequency, f (r).
As previously noted, the Larmor frequency at a spatial location r is proportional to
the magnetic field at that location:

γ
f (r) = B(r), (2.1)
2π

where B(r) is the magnetic field and γ is the gyromagnetic ratio. As the magnetization
precesses, it emits a radio-frequency signal at the Larmor frequency with a magnitude
proportional to the magnitude of the transverse magnetization. It is this emitted
signal that is detected, providing information about the transverse magnetization.

In addition to precessing, the magnetization also returns to equilibrium, decaying

in the transverse plane with a time constant of T2 and re-growing in the longitudinal
direction with a time constant of T1. As the magnetization returns to equilibrium,
the net magnetization does not remain constant, as the coherency of the magnetic
moments of the nuclei can change during this time. The time constants T1 and T2
are material properties that differ from tissue to tissue and are the most common
2.2. THE EMITTED SIGNAL 9

material properties used to provide contrast in MR imaging.

In a typical MR scan, the magnetization will be repeatedly excited, or tipped into
the transverse plane, with some signal detected after each excitation. By adjusting
the time between repetitions and the time at which the signal is acquired, T1-weighted
and T2-weighted contrasts can be achieved.

2.2 The Emitted Signal

As discussed in the previous section, an MRI scanner can be programmed to cause

each spatial location in an object to emit a radio-frequency signal with a magnitude
proportional to the transverse magnetization at that particular spatial location. The
difficulty in forming an image arises from the fact that all excited spatial locations
emit a signal at the same time, making it challenging to differentiate the signal at
one spatial location from another spatial location. In this section, I describe how
the gradient system can be used to warp the phase of the emitted signal by location.
This procedure, called gradient encoding, can be used to encode spatial location
information.
The emitted signal, m(r, t) is the signal emitted from location r at time t, after
γ
being demodulated around the carrier frequency f0 = B.
2π 0
In the absence of the
gradient system, m(r, t) = m(r).
Gradient encoding uses the magnetic field produced by the gradient system to
change the Larmor frequency during the scan. The magnetic field produced by the
gradient system varies linearly in space and is expressed in terms of the gradient field
strength G(t), which has units of Gauss/cm. The magnetic field produced by the
gradient system is G(t) · r, changing the Larmor frequency by

γ
∆f (r, t) = G(t) · r. (2.2)
2π

By integrating the frequency over time, it is possible to obtain the change in the
10 CHAPTER 2. SETTING THE STAGE

phase of the signal caused by the gradient system:

Z t
∆φ(r, t) = 2π −∆f (r, τ ) dτ (2.3)
0
Z t
γ
= 2π G(τ ) · r dτ
− (2.4)
0 2π
= −2πk(t) · r, (2.5)

where Z t
γ
k(t) = G(τ )dτ . (2.6)
0 2π
The negative sign in Eq. 2.3 is the result of the precession direction of protons. The
change in phase expressed in Eq. 2.5 can be included in the equation for the emitted
signal:

m(r, t) = m(r)ei∆φ(r,t) (2.7)

= m(r)e−i2πk(t)·r . (2.8)

The next section, which looks at how the emitted signals are received by a receiver
coil, completes the encoding picture and shows how gradient encoding can be used to
encode spatial location information.

2.3 The Received Signal

At a given time, a radio-frequency receiver coil measures a single value: the voltage
induced in the coil as a result of the magnetic flux through the coil. The transverse
magnetization at each location in the excited object will make some contribution
to the magnetic flux through the coil and hence some contribution to the induced
voltage. The measured voltage is the sum of the contributions from all locations.
The contribution of the transverse magnetization at a spatial location to the measured
voltage depends on the coil geometry and loading as well as the spatial location of the
magnetization. Using j to index the receiver coils, the receiver-coil sensitivity function
for coil j, sj (r), is a complex-valued function that characterizes the contribution to
2.3. THE RECEIVED SIGNAL 11

the coil measurement from the emitted signal m(r, t). Summing the contributions
from all locations of the magnetization expressed in Eq. 2.8, the following equation
for the received MR signal dj (t) is derived:
Z
dj (t) = sj (r)m(r, t) dV (2.9)
V
Z
= sj (r)m(r)e−i2πk(t)·r dV . (2.10)
V

Although this simple signal equation does not take into account many things, includ-
ing the decay of the transverse magnetization and factors such as B0 inhomogeneity
and changes in magnetic susceptibility that can affect the rate of precession, it pro-
vides in many cases an adequate model for the received signal on an MR system and
is the model that will be used for the remainder of this work.

In many imaging situations only one receiver coil is used: the so called ‘body coil’
which is, ideally, uniformly sensitive to all spatial locations (sj (r) = 1). In this case,
Eq. 2.10 simplifies to Z
d(t) = m(r)e−i2πk(t)·r dV . (2.11)
V

In Eq. 2.11 an acquired datum d(t) is the result of multiplying the magnetization by
a complex exponential and summing the result. At each point in time the magneti-
zation is multiplied by a different complex exponential as shown on the left column
of Fig. 2.3. Recognizing e−i2πk(t)·r as a Fourier kernel, an acquired datum can also be
interpreted as a sample in spatial-frequency or k-space, as shown on the right column
of Fig. 2.3. Equation 2.6 gives the relationship between gradient field strength G(t)
and k-space location. After transverse magnetization is excited, waveforms are played
out on the gradient system, Fig. 2.4(a), that result in a desired k-space trajectory as
shown in Fig. 2.4(b).

Expressing the received signal in Eq. 2.11 in terms of k-space location, Eq. 2.11
can be rewritten as the Fourier transform:
Z
d(k) = m(r)e−i2πk·r dV . (2.12)
V
12 CHAPTER 2. SETTING THE STAGE

ky

kx

x
(a)
ky

kx

x
(b)
ky

kx

x
(c)

Figure 2.3: When only gradient encoding is considered, an acquired datum can be
modeled as the result of multiplying the magnetization by a complex exponential
image and summing the result. Alternatively, each datum can be viewed as acquiring
a sample of k-space. Shown in (a)-(c) are these two perspectives for three different
encoding locations.
2.3. THE RECEIVED SIGNAL 13

kx

Gx t

kx

Gy t

(a) (b)

Figure 2.4: Playing gradient waveforms (a) trace a trajectory through k-space (b).
The relationship between the gradient field strength and k-space trajectory is given
by Eq. 2.6.
14 CHAPTER 2. SETTING THE STAGE

Likewise, the magnetization can be expressed in terms of the emitted signal function
d(k) using the inverse Fourier transform:
Z
m(r) = d(k)ei2πk·r dkx dky dkz . (2.13)
k -space

Equation 2.13 provides useful insight into MR image reconstruction, even if it is an

unusable equation since an MR scanner can only collect a finite number of samples
of d(k) along a k-space trajectory and the entire d(k) function is never known.

Due to the flexibility of modern gradient systems, many different k-space trajec-
tories can be achieved, four of which are shown in Fig. 2.5. One advantage of the
DFT trajectory shown in Fig. 2.5(a) is that image reconstruction can be performed
with a two-dimensional Fast Fourier Transform (2-D FFT), since the data falls on a
Cartesian grid. The 2-D FFT computes evenly-spaced samples of ms (x, y):

Ns
X
ms (x, y) = d (kx,j , ky,j ) ei2π(kx,j x+ky,j y) , (2.14)
j=1

where(kx,j , ky,j ) select the Ns evenly-spaced k-space locations that have been acquired
along kx and ky . Equation 2.14 can be related to a 2-D version of m(r) = m(x, y) as
2.3. THE RECEIVED SIGNAL 15

ky ky

kx kx

(a) (b)

ky ky

kx kx

(c) (d)

Figure 2.5: Four possible k-space trajectories: (a) DFT, (b) variable-density in the
phase-encode direction, (c) radial, (d) variable-density spiral.
16 CHAPTER 2. SETTING THE STAGE

follows:
Ns
X
ms (x, y) = d (kx,j , ky,j ) ei2π(kx,j x+ky,j y)
j=1
Ns Z
X
= d (kx , ky ) δ (kx − kx,j , ky − ky,j ) ei2π(kx x+ky y) dkx dky
j=1 k -space
Z Ns
X
= d (kx , ky ) δ (kx − kx,j , ky − ky,j ) ei2π(kx x+ky y) dkx dky
k -space j=1
( Ns
)
X
= IFT d (kx , ky ) δ (kx − kx,j , ky − ky,j ) (x, y)
j=1
(N )
X s

= IFT {d (kx , ky )} (x, y) ∗ IFT δ (kx − kx,j , ky − ky,j ) (x, y)

j=1
= m(x, y) ∗ PSF(x, y), (2.15)

where (N )
X s

PSF(x, y) = IFT δ (kx − kx,j , ky − ky,j ) (x, y). (2.16)

j=1

The point-spread-function PSF(x, y) describes how the k-space sampling of the

trajectory affects the reconstructed image. As shown in Fig. 2.6, the point-spread-
function for the DFT trajectory consists of equally-spaced points. The sharpness
of the points determines the resolution of the reconstructed image and the distance
between points determines the size of object that can be imaged without introducing
aliasing artifacts, shown in Fig. 2.6(b).

An advantage of the variable-density spiral trajectory shown in Fig. 2.5(d) is that

k-space can be sufficiently sampled with fewer magnetization excitation repetitions,
resulting in faster data acquisition. Such an advantage is very important in fast MRI.
A disadvantage of such a non-Cartesian k-space trajectory is that the reconstruction
becomes more complicated. Just as the FFT is an efficient method for calculating
the sum in Eq. 2.14 for evenly spaced values of x and y, gridding, covered in detail
2.3. THE RECEIVED SIGNAL 17

m(x,y) PSF(x,y) ms(x,y) m(x,y) PSF(x,y) ms(x,y)

= * =
*

(a) (b)

Figure 2.6: Graphical illustration of Eq. 2.15 with (a) sufficient gradient encoding
and (b) insufficient gradient encoding, leading to aliasing artifacts.

in the next chapter, is an efficient method for calculating

Ns
X
ms (x, y) = d (kx,j , ky,j ) w (kx,j , ky,j ) ei2π(kx,j x+ky,j y) (2.17)
j=1

for evenly spaced values of x and y when (kx,j , ky,j ) do not fall on a Cartesian grid.
The density-compensation weight values w(kx,j , ky,j ) allow a more ideal point-spread-
function to be achieved. An ideal point-spread-function has a sharp point at x = y = 0
and no or very little content within a radius of the desired field-of-view around the
central point. Starting with Eq. 2.17 and imitating the derivation of Eq. 2.15, an
image reconstructed using gridding can be expressed in the same form as Eq. 2.15
where
(N )
X s

PSF(x, y) = IFT w (kx,j , ky,j ) δ (kx − kx,j , ky − ky,j ) (x, y). (2.18)
j=1

Figure 2.7 illustrates convolution with a spiral trajectory point-spread-function. In (a)

18 CHAPTER 2. SETTING THE STAGE

sufficient gradient encoding is performed such that convolving with the point-spread-
function does not lead to aliasing artifacts. When insufficient gradient encoding is
performed, as shown in (b), the aliasing artifacts produced by convolving with the
spiral point-spread-function have a very different structure compared to the aliasing
artifacts associated with the DFT trajectory, shown in Fig. 2.6(b).

Expressing ms (x, y) as the magnetization function m(x, y) convolved with a point-

spread-function PSF(x, y) as in Eq. 2.15 gives a useful visualization of the operation
performed by gridding. However it is more useful for the development of the gridding
method to view ms (x, y) as the inverse Fourier transform of the sampled, density-
compensated k-space data function Ms (kx , ky ):

ms (x, y) = IFT {Ms (kx , ky )} (x, y), (2.19)

where
Ns
X
Ms (kx , ky ) = d (kx , ky ) w (kx,j , ky,j ) δ (kx − kx,j , ky − ky,j ) . (2.20)
j=1

The equivalence of Eq. 2.19 to Eq. 2.17 can be shown by substituting Eq. 2.20 into
Eq. 2.19 and simplifying. Calculating the inverse Fourier transform of Ms (kx , ky )
using a direct implementation of Eq. 2.17 is computationally prohibitive; the next
chapter discusses how the gridding method efficiently computes this transform.

As discussed above, one way to reduce the encoding time necessary to fully en-
code an image is by using non-Cartesian trajectories. The encoding time can also
be reduced by using multiple receiver coils. When more than one coil is present, the
receiver-coil sensitivity functions can be used to encode spatial location information,
allowing an image to be reconstructed without being fully gradient encoded. Such
encoding schemes are often called ‘parallel imaging’, since at any one time Nc sepa-
rately encoded data values are acquired, where Nc is the number of receiver coils. To
see more clearly how the magnetization is encoded in such a scheme, Eq. 2.10 can be
2.3. THE RECEIVED SIGNAL 19

m(x,y) PSF(x,y) ms(x,y)

(a)

* =

(b)

* =

Figure 2.7: Graphical illustration of convolving with a spiral k-space trajectory point-
spread-function with (a) sufficient gradient encoding and (b) insufficient gradient
encoding, leading to aliasing artifacts.
20 CHAPTER 2. SETTING THE STAGE

rewritten in terms of an encoding function ej (k, r):

Z
dj (t) = sj (r)m(r)e−i2πk(t)·r dr (2.21)
ZV
= m(r)ej (k, r) dr (2.22)
V

where
ej (k, r) = sj (r)e−i2πk(t)·r (2.23)

In this case, each acquired datum can be viewed as multiplying the magnetization
by an encoding function and summing the result. The encoding function ej (k, r)
is a product of both the receiver-coil sensitivity function and the gradient encoding
complex exponential, as shown in Fig. 2.8. The methods described in chapters 4 and
5 take advantage of both gradient encoding and receiver-coil sensitivity encoding,
allowing for faster acquisitions.
2.3. THE RECEIVED SIGNAL 21

(a)

+
-V
Coil 1
(b)

+
-V
(c) Coil 2

Figure 2.8: (a) As in Fig. 2.3, a datum acquired using the body coil can be modeled
as multiplying the magnetization by a complex exponential image and summing the
result. (b), (c) For coils with non-uniform coil sensitivity, an acquired datum can be
modeled as multiplying the magnetization by an encoding function (Eq. 2.23) and
summing the result.
22 CHAPTER 2. SETTING THE STAGE
Chapter 3

Gridding Reconstruction

In magnetic resonance imaging (MRI), the data are obtained in the spatial frequency
domain, often called k-space. The image reconstruction problem can be posed as a
Fourier inversion problem which can be solved efficiently using the fast Fourier trans-
form (FFT) when the data lie on a Cartesian grid. When the data samples do not fall
on a Cartesian grid, the MRI community has largely turned to the gridding algorithm
due to its efficiency and accuracy [1, 2]. Gridding is simple and robust and has pa-
rameters, the grid oversampling ratio and the kernel width, that can be used to trade
accuracy for computational memory and time reductions. In this chapter, I describe
a number of techniques that can be used to reduce the computation memory and
time needed by the gridding method while maintaining high reconstruction accuracy.
These techniques include using a minimal oversampling ratio, presampling the kernel
and taking advantage of block storage. Although I focus primarily on MRI recon-
struction, gridding can also be used in cone-beam CT reconstruction and ultrasound
tomography [3, 4].

The gridding algorithm can be divided into two stages: sampling density com-
pensation and inverse Fourier transform; in this chapter I consider the later stage
only, which is equivalent to the nonuniform FFT (NUFFT) problem type one [5].
Good treatments of density compensation are given by Pipe and Menon and Rasche
et al. [6, 7]. In the gridding algorithm, the inverse Fourier transform is estimated by

23
24 CHAPTER 3. GRIDDING

convolving the density compensated data with a finite kernel, sampling onto a Carte-
sian grid, performing a fast Fourier transform and multiplying by an apodization
correction function.
Reconstruction methods similar to gridding also exist, which do not separate the
density compensation from the inverse Fourier transform. This can be accomplished
by using shift-variant interpolators or iterative methods [8, 9, 10]. Shift-variant in-
terpolators require a significant precomputation stage and enough memory to hold a
unique set of weights for each data sample. When enough memory is available and
the same data sample locations are repeatedly reconstructed, as is the case in real-
time 2D imaging, such methods are appropriate. The work in this chapter has been
motivated by a desire to efficiently reconstruct three-dimensional (3-D) non-Cartesian
data samples. In this case, the memory required to store precomputed weights for
shift-variant interpolators can limit the size of the image one is able to reconstruct.
Also limiting is the use of a grid oversampling ratio of two, traditionally used in
gridding reconstruction [11]. A grid oversampling ratio of two leads to an eight-times
increase in memory required for a 3-D image. By using a minimal oversampling ratio
(1.125-1.375), the required memory drops to 1.4 - 2.6 times the memory required
for the final image. Using a minimal oversampling ratio also leads to a significant
reduction in computation time due to the reduced size of the FFT that is performed.
However, maintaining high accuracy on a minimally oversampled grid requires a well-
suited convolution kernel; in sections 3.3 and 3.5 I develop two kernel design methods.
On the one hand, memory constraints make undesirable the use of precomputed
weights for each data sample; on the other hand, evaluating the Kaiser-Bessel function
several times for each data point can account for most of the computation time.
Greengard and Lee have shown that using a Gaussian kernel can be quite efficient;
in one dimension, the Gaussian kernel requires two exponential evaluations per data
sample point plus two multiplications for each point on the grid to which this data
sample is deposited [12]. In this chapter, I show that a similar efficiency can be
achieved without being limited to a particular kernel shape by using a presampled
kernel. Interpolation between these presampled values is used to approximate the
continuous gridding kernel. For linear interpolation, interpolating between two stored
 25

values involves one multiplication and one addition, similar to the two multiplications
required for the Gaussian kernel. In some cases the speed improvement is superior
to precomputed weights since the presampled kernel can easily fit in the computer
cache. I examine both nearest-neighbor and linear interpolation of the presampled
kernel and develop expressions for how finely the kernel must be sampled so as not
to degrade the reconstruction.
Ideally the gridding method is implemented such that the reconstruction computa-
tion time is as short as possible while being confident that no reconstruction artifacts
are noticeable in the final image. I define the aliasing amplitude, a metric for gridding
accuracy, which allows one to analyze how the grid oversampling ratio, kernel width
and kernel sampling affect the reconstruction accuracy. By choosing gridding param-
eter combinations that have a low aliasing amplitude, one can have great confidence
in the fidelity of the reconstruction. For the aliasing amplitude requirements of MR
image reconstruction, parameter combinations with minimal oversampling ratios are
much more computationally expedient than combinations that use an oversampling
ratio of two. For 3-D images, the memory constraint of the reconstruction computer
might also have to be taken into account in the selection of the oversampling ratio.
When a minimal oversampling ratio and presampled kernel are used to perform
a 3-D reconstruction, the computation time required for both gridding and FFT is
limited, not by the ability of the processor to accomplish the computation, but by
the time it takes to get the appropriate data from the memory to the processors for
the computation. By organizing the grid in a block storage format, data likely to be
accessed concurrently is put in the same page of memory. The block storage format
reduces processor-to-memory traffic and can lead to a factor of four reduction in com-
putation time. By combining the use of a minimal oversampling ratio, a presampled
kernel and block storage, the computation time for gridding can be reduced by a
factor of approximately thirty, compared to a less sophisticated approach that does
not take advantage of any of these techniques.
In addition to being simple and robust, gridding also allows for a certain amount
of flexibility since the resolution and field-of-view of the final image can be chosen
independently of the resolution of the acquired data and the spatial extent of the
26 CHAPTER 3. GRIDDING

object being imaged. In this chapter, I show some practical situations that take ad-
vantage of this flexibility of the gridding method, including the ability to reconstruct
an image with a field-of-view smaller than the extent of the object being imaged.

3.1 The Gridding Algorithm

For the rest of this chapter, I shall assume that the data has already been density
compensated and I will use ‘gridding’ to refer to the inverse Fourier transform step
of the reconstruction. The basic gridding algorithm is straightforward, consisting of
the following steps:

1. Convolve the k-space samples with a gridding kernel.

2. Sample the result onto evenly spaced sample locations (the grid).

3. Perform an FFT (I use fftw) [13].

4. Multiply by an apodization correction function.

Throughout this chapter I treat the one dimensional case for pedagogical simplicity;
it is easy to extend the result to two and three dimensions when separable gridding
kernels are used. I define the variables and functions used throughout this chapter
in Tables 3.1 and 3.2. In addition, I define the image to have unit resolution, so the
image pixel locations are at −N/2, −N/2 + 1, . . . , N/2 − 1, for even N , and I sample
k-space at G evenly spaced points in the range [−0.5, 0.5], since I measure k-space
in cycles per unit distance. With these conventions, the gridding algorithm can be
expressed mathematically as

1
b s [i] = IFT {[Ms (kx ) ∗ C(kx )] III (Gkx )} (i)
m . (3.1)
c(i)

Note that the apodization correction function, 1/c(i), is the reciprocal of the inverse
Fourier transform of the kernel, C(kx ), and is determined completely by the choice of
kernel. One would like gridding to perform the mathematical operations given in Eq.
3.1, however in practice the FFT algorithm is used, not a continuous inverse Fourier
3.1. THE GRIDDING ALGORITHM 27

Table 3.1: General Variables and Functions

D Number of k-space data points.
N Width of the image, pixels.
G Number of grid sample locations in k-space.
(samples are 1/G inverse pixels apart)
α = G/N Grid oversampling ratio.
kx k-space location, kx ∈ R.
x Image location, x ∈ R.
Pixel location, i ∈ − N2 , − N2 + 1, . . . , N2 − 1 .

i
Ms (kx ) Sampled, density corrected k-space data.
Ms (kx ) = D
P
i=1 Md (kx,i )δ(kx − kx,i )
Md (kx,i ) Density corrected k-space datum value at kx,i .
ms (x) Inverse Fourier transform of Ms (kx ).
ms (x) = D
P
i=1 Md (kx,i ) exp(j2πkx,i x)
m
b s [i] Gridding estimate of ms (x = i).
ε[i] Aliasing amplitude.
E[i] Image reconstruction error.
E[i] = |ms (i) − m
b s [i]|
P∞
III (x) Shah function, III (x) = p=−∞ δ(x − p).
FT {f (x)} (kx ) Fourier transform.
R∞
FT {f (x)} (kx ) = −∞ f (x) exp(−j2πkx x)dx
IFT{f (kx )}(x) Inverse Fourier transform.
R∞
IFT{f (kx )}(x) = −∞ f (kx ) exp(j2πkx x)dkx
N.B. Parenthesis, (), are used for functions that take real arguments.
Square brackets, [], are used for integer arguments.
28 CHAPTER 3. GRIDDING

Table 3.2: Gridding Kernel Variables and Functions

C(kx ) Continuous valued kernel function.
(e.g., Kaiser-Bessel, truncated sinc, Gaussian)
c(x) Inverse Fourier transform of C(kx ).
W Width of kernel in grid units.
(one grid unit is 1/G, the distance between samples)
S Kernel sampling density.
(number of kernel samples per grid unit)
Cs (kx ) Sampled kernel.
Cs (kx ) = C(kx ) III (SGkx )
cs (x) Inverse Fourier transform of Cs (kx ).
H(kx ) Interpolation function.
H(kx ) = u (SGkx ) for nearest neighbor interpolation.
H(kx ) = ∧ (SGkx ) for linear interpolation.
h(x) Inverse Fourier transform of H(kx ).
C(k
b x ) Continuous valued kernel after sampling and interpolation.
b x ) = Cs (kx ) ∗ H(kx )
C(k
c(x)
b Inverse Fourier transform of C(k
b x ).
c(x) = cs (x)h(x)
b
3.1. THE GRIDDING ALGORITHM 29

C(kx)

kx
-0.5 0.5

Figure 3.1: The convolution with the gridding kernel is a circular convolution. Thus,
when the gridding kernel goes beyond the extent of the grid, it wraps around to the
other side of the grid.

transform. Because of this, one only computes m

b s [i] for integer values of i—that is,
one computes image values at discrete pixel locations. As well, there is the following
caveat: if, due to the width of the kernel, [Ms (kx ) ∗ C(kx )] III (Gkx ) extends outside of
the range [−0.5, 0.5], the values outside of this range must be added to the opposite
edge of the grid as shown in Fig. 3.1. This is because sampling in the image domain
is equivalent to a repetition in k-space.

The computation time is the sum of two terms. The first term is due to the
convolution and sampling step of the algorithm and is proportional to DW 2 for 2D
images and DW 3 for 3D images, where D is the number of data points and W is the
width of the kernel on the oversampled grid. This term can be reduced by shrinking
W and by reducing the constant of proportionality. The second term is computing
the FFT. For grid oversampling ratio α and image width N , it is proportional to
α2 N 2 log αN for 2D images and α3 N 3 log αN for 3D images and can therefore be
reduced by choosing a smaller oversampling ratio. As the bulk of the computer
memory requirement is in storing the grid, which is of size α2 N 2 for 2D images and
α3 N 3 for 3D images, a smaller α also reduces the memory requirements.
30 CHAPTER 3. GRIDDING

150 MB 40 sec
35 sec
120 MB
30 sec
FFT 25 sec
90 MB
20 sec
60 MB 15 sec
10 sec
30 MB *
FFT 5 sec
0 MB * 0 sec
α=1.375 α=2 α=1.375 α=2
(a) (b)

Figure 3.2: Reconstruction of a 128 × 128 × 128 image from the 3-D cones trajectory
with 2,304,000 points [14]. With a minimal oversampling ratio of 1.375, I used a
sampled Kaiser-Bessel kernel, W = 5, with linear interpolation. With the typical
oversampling ratio of two, I used a Kaiser-Bessel kernel, W = 4. Both parameter
choices have similar aliasing amplitudes, resulting in similar reconstruction errors
(Fig. 3.11). (a) Memory requirements, (b) reconstruction time on PC with one 1.67
GHz AMD processor. ‘*’ refers to the convolution and sampling step of the gridding
algorithm. This demonstrates the large gains that can be achieved, especially in 3-
D image reconstruction, by using a minimal oversampling ratio and a presampled
gridding kernel, but without the advantage of the block storage format.
3.2. ALIASING AMPLITUDE: A METRIC FOR GRIDDING ACCURACY 31

When one moves to a minimal oversampling ratio to decrease the computation

time of the FFT, if one wishes to maintain the same accuracy, one must increase
the width of the kernel, thereby increasing the computation time of the convolu-
tion. When the kernel is not presampled, the speed gain due to a smaller FFT can be
negated by the increased computation of the convolution. This is why using a presam-
pled kernel is important when using a minimal oversampling ratio: it speeds up the
convolution, keeping it from dominating the reconstruction time. By using a minimal
oversampling ratio and a presampled kernel, both the convolution and FFT stages of
reconstruction are considerably reduced, without loss of accuracy. Figure 3.2 shows
that we obtain an eight times reduction in computation time for a 128×128×128 3-D
reconstruction. Some further computation reductions can be obtained at the expense
of accuracy by decreasing α and W , although there is a limit, since one cannot go
below an oversampling ratio of one and interpolation must take some positive amount
of time.

3.2 Aliasing Amplitude: A Metric for Gridding

Accuracy

As noted in the previous chapter, the goal of the gridding algorithm is to find the
inverse Fourier transform of Ms (kx ). Put simply, the goal is to have m
b s [i] = ms (x)
when x = i and the extent to which this is not true is the extent to which gridding is
inaccurate. For a given data acquisition, one can calculate the image reconstruction
error, E[i] = |ms (i) − m
b s [i]|, where ms (i) is calculated directly using the sum in
Table 3.1 and m
b s [i] is found by performing the gridding reconstruction. It should
be noted that while the direct summation method computes an exact inverse Fourier
Transform, it takes an excessive amount of computation.
While the image reconstruction error gives the inaccuracy of the gridding algo-
rithm for a certain object and trajectory, it offers little insight into the causes of this
error and little guidance in designing a more suitable gridding kernel. To gain some
32 CHAPTER 3. GRIDDING

insight into the causes of this error, I rewrite Eq. 3.1 completely in the image domain:
   
1 i 1
m
b s [i] = [ms (i)c(i)] ∗ III . (3.2)
G G c(i)

The sampling onto the grid in k-space causes a repetition of ms (x)c(x) in image
space every G pixels. These shifted versions of ms (x)c(x) alias back into the image,
accounting for the error in the gridding estimate. Using Eq. 3.2, one can rewrite the
image reconstruction error as

E[i] = |ms (i) − m

b s [i]|
1 X
= ms (i + Gp)c(i + Gp) .
c(i) p6=0

Since the error depends on ms (x), it depends on the k-space sampling pattern and
the object being imaged. As one tries to design gridding solutions to have low errors
over a range of sampling patterns and objects, it is most helpful to have a metric
which is independent of ms (x), but gives an estimate of the reconstruction errors one
can expect. For this purpose, Jackson et al. used the relative amount of aliasing
energy, which is independent of the image and sampling pattern used and provides
a good relative ordering of accuracy [11]. However, the aliasing energy metric used
by Jackson has a couple of shortcomings. Since Jackson averaged the aliasing energy
over the entire image, one does not get an idea of how the error will vary from
position to position in the image. As well, it is desirable to have a metric which will
predict, at least in terms of the order of magnitude, what one can expect for the
image reconstruction error when using a given kernel. By taking the square root of
the aliasing energy at each pixel location, I obtain a metric that is both a function of
position and can be used to estimate E[i]. I call this metric the aliasing amplitude,
ε[i]: s
1 X
ε[i] = 2 [c(i + Gp)]2 . (3.3)
[c(i)] p6=0

One way to view this metric is to model ms (x) as a Gaussian distributed random
3.3. CHOOSING A KAISER-BESSEL KERNEL 33

complex number, independent along x, with unit variance. In this case, the expected
value of m
b s [i] is ms (i) and the standard deviation of location i in the image is identical
to the aliasing amplitude at this pixel location.
The aliasing amplitude gives an order-of-magnitude estimate for the image recon-
struction error. That is, although the error in a reconstructed image is dependent on
the object and k-space sampling pattern, one can reliably predict the order of magni-
tude and profile of the error from the aliasing amplitude, as shown in Fig. 3.3. This
allows one to design kernels using the aliasing amplitude as a metric for accuracy.
Figure 3.4 provides a reference of the maximum aliasing amplitude of an image for
various oversampling ratios and kernel widths. With the typical signal-to-noise ratio
(SNR) of MR images, an aliasing amplitude of 0.1 will likely suppress visible artifacts,
allowing the use of a kernel of width three on a grid with an oversampling ratio of
1.125. However, in cases of high dynamic range, or when the reconstruction field-of-
view is smaller than the extent of the object, more accuracy may be needed. To cover
such cases, an oversampling ratio of 1.25 with a width four kernel will reduce the
maximum aliasing amplitude to under 0.01. A maximum aliasing amplitude under
0.001 can be achieved with an oversampling ratio of 1.375 and a width five kernel,
giving higher accuracy than is likely necessary in MRI with a much lower computation
time than can be achieved using an oversampling ratio of two.
In Fig. 3.5, I compare the accuracy of the parameter combinations with aliasing
amplitudes of 0.1 and 0.01 in reconstructing an in vivo sagittal knee image. Although
using a parameter combination with an aliasing amplitude of 0.1 performs quite well,
it is possible in some cases to see slight degradations in reconstruction quality. This
degradation is not apparent when the aliasing amplitude is 0.01 or lower.

3.3 Choosing a Kaiser-Bessel Kernel

One of the difficulties in designing a gridding solution is that a different gridding
kernel must be chosen for each choice of oversampling ratio, α, and kernel width, W .
For example, using a kernel designed for α = 2 on a grid with α = 1.375 can lead
to a large loss in accuracy, as shown in Fig. 3.6. The error images in (b) and (d)
34 CHAPTER 3. GRIDDING

Figure 3.3: The aliasing amplitude, ε[i], is a good predictor for the image reconstruc-
tion error, E[i]. Sampling the two numerical phantoms (a) and (b) with both radial
and spiral k-space trajectories, produces four data sets. In order to look at the 1D
error pattern, I used a very low error method in the y-direction (oversampling ratio
α = 2, with a Kaiser-Bessel kernel of width W = 8). In the x-direction I used the
oversampling ratio and kernel shown in (c)-(f), where SOCP refers to the optimal
sampled kernels described in section VI. For each of the kernels in the x-direction,
I reconstructed four 256 × 256 images (two phantoms × two trajectories) and sub-
tracted them from the ideal reconstructions using the exact inverse Fourier transform.
Taking the absolute value I obtained the image reconstruction error, E[i]. This gave
1024 error values (4 images, 256 lines per image) at each x-location for each kernel.
The Max Error is the maximum of these 1024 values at each x location. This is very
close to the aliasing amplitude plot, showing that the aliasing amplitude is a good
indicator of the order of magnitude of the maximum reconstruction error one is likely
to see.
3.3. CHOOSING A KAISER-BESSEL KERNEL 35

Figure 3.4: Maximum aliasing amplitude for a selection of kernels. The shape param-
eter for the Kaiser-Bessel kernels is found using Eq. 3.5. SOCP refers to the optimal
sampled kernels found using second-order cone programming (SOCP), a technique
that is discussed in detail in section 3.5.

show that the large errors in (a) and (c) can be avoided by designing the kernel for
the oversampling ratio used. In the case that the errors in (a) and (c) are acceptable,
one would do better to design for this target accuracy and decrease the computation
time.

Jackson et al. examined many different functions that could be used for gridding
kernels and found the Kaiser-Bessel window preferable since it is easily computable
and offers near optimal performance [11]. Jackson also calculated the optimal shape
parameter for a range of W for oversampling ratios of one and two, as shown in Table
3.3. The shape parameter, β, determines the shape of the Kaiser-Bessel window.
Here, I derive a simple equation which gives near optimal β for any α ∈ [1, 2]. A
similar equation for β has been developed by Wajer et al., but this equation does not
work well for minimal oversampling ratios [15]. It should be noted that our metric for
optimality of the kernel is slightly different than that used by Jackson, accounting for
the discrepancy in β values in Table 3.3. Jackson searched for the kernel that gave
the lowest average aliasing energy over the image whereas I search for the kernel that
36 CHAPTER 3. GRIDDING

Figure 3.5: Sagittal knee images obtained with a 2D spiral sequence. (a),(c) were
reconstructed using α = 1.125 and W = 3, giving a maximum aliasing amplitude of
about 0.1. (b),(d) were reconstructed using α = 1.25 and W = 4, giving a maximum
aliasing amplitude of about 0.01. When the full field-of-view is reconstructed in (a)
and (b), no difference in the images is discernible. A more challenging test is to
reconstruct a region of interest which is smaller than the extent of the object (c),(d).
(e),(f) are error images of (c),(d) windowed up 10 times. In this case, the α = 1.125
method gives a slight loss of visible accuracy, whereas the α = 1.25 method is still
exceedingly accurate.
3.3. CHOOSING A KAISER-BESSEL KERNEL 37

Figure 3.6: The k-space data for the numerical phantom in Fig. 2(a) was acquired
using a radial trajectory in (a) and (b) and a spiral trajectory in (c) and (d). These
two data sets were reconstructed on a grid with grid oversampling ratio α = 1.375
using two different Kaiser-Bessel kernels, both of width W = 5. In (a) and (c),
the shape parameter, β, is 11.4410, the shape parameter appropriate when the grid
oversampling ratio is two (which it is not in this case). In (b) and (d), the shape
parameter is 9.5929, the shape parameter appropriate for the grid oversampling ratio
α = 1.375 (Eq. 3.5). Shown is the image reconstruction error windowed up 1000 times
relative to Fig. 2(a). By not choosing the kernel appropriate for the oversampling
ratio used, (a) and (c) have large reconstruction errors around the edges of the image.

gives the lowest maximum aliasing amplitude at any point in the image. I refer to this
as optimal in the rest of this chapter. I choose this metric because I am interested
in ensuring that the gridding reconstruction does not introduce any visible artifacts
into the image. When the average aliasing energy is minimized, much of the aliasing
energy is deposited around the edges of the image and the aliasing amplitude at these
points is much higher than the average.
A good convolution kernel acts as a multi-bandpass filter in the image domain with
passband x ∈ [−N/2, N/2] and stop-bands x ∈ [Gp − N/2, Gp + N/2], for all integer
values of p 6= 0. Choosing a larger oversampling ratio allows for wider transition
bands. The equations for the Kaiser-Bessel window and its inverse Fourier transform
are given by

G  p  W
C(kx ) = I0 β 1 − (2Gkx /W )2 for |kx | ≤
W q 2G
sin (πW x/G)2 − β 2
c(x) = q (3.4)
(πW x/G)2 − β 2

where I0 (kx ) is the zero-order modified Bessel function of the first kind.
38 CHAPTER 3. GRIDDING

Figure 3.7: (a) Positioning the first zero crossing at the edge of the first stop-band
results in a good kernel design, since it allows the main lobe to be as wide as possible
without any of the main lobe contributing to the aliasing amplitude. A wide main lobe
lessens apodization in the passband. Since apodization correction increases artifacts
aliasing in from the stop-bands as well as the passband signal, less apodization in the
passband is preferable. (b) The kernel design in (a) can be slightly improved upon
by allowing the main lobe to enter slightly into the first stop-band. In this case the
objective is to keep the aliasing amplitude due to the main lobe in the first stop-band
equal to the aliasing amplitude caused by the first side-lobe.
3.3. CHOOSING A KAISER-BESSEL KERNEL 39

Table 3.3: Comparison of β Values

β
α=1 α=2
Wa Jackson Eq. 3.5 Jackson Eq. 3.5
2 2.3934 1.4050
3 4.2054 3.7830 6.6875 6.4861
4 5.7567 5.6199 9.1375 8.9962
5 7.4302 7.3341 11.5250 11.4410
6 13.9086 13.8551
7 16.2734 16.2522
8 18.5547 18.6389
a
I use the kernel width parameter, W , which is equal to twice
the window width parameter used by Jackson et al. for α = 2.

Intuitively, the best Kaiser-Bessel window is the one for which the central lobe
of the inverse Fourier transform of the window, c(x), extends to the edge of the first
stop-band, as shown in Fig. 3.7. A narrower central lobe leads to increased error
through apodization of the passband whereas a wider central lobe allows the large
central lobe into the
q first stop-band. From Eq. 3.4, the first
q zero of the
Kaiser-Bessel
2
window is when (πW x/G)2 − β 2 = π. Setting β = π W 2 α − 12 /α2 − 1 will
position this zero crossing at x = G − N/2, as shown in Fig. 3.7(a). It turns out that
it is possible to do slightly better if one allows some of the main lobe to enter into
the first stop-band, as illustrated by Fig. 3.7(b). In essence the aliasing amplitude
at the edge of the image is allowed to rise until it reaches the aliasing amplitude
caused by the first side-lobe. The point at which this happens is difficult to solve
analytically, however I have found that a slight modification of the equation I obtain
by positioning the first zero crossing at x = G − N/2 works very well in practice. The
modified equation is: s  2
W2 1
β=π α− − 0.8 (3.5)
α2 2
Equation 3.5 has the advantage that it gives a β value which is very close to the
optimal β value and is simple to calculate. Equation 3.5 can be used to generate a
kernel that will ensure that the gridding solution is highly accurate, considering the
40 CHAPTER 3. GRIDDING

α and W selected. Once the α and W have been selected, the only way one has
of reducing the computation time is to reduce the constant of proportionality of the
DW 2 term (DW 3 term for 3D). That is, the computation time can only be reduced
by decreasing the time it takes to deposit a data sample onto a grid point. The next
section shows how this can be done by presampling the kernel and using interpolation.

3.4 Presampling the Kernel

Even though the Kaiser-Bessel window is relatively inexpensive to calculate, evalu-
ating it several times for each data point can significantly slow down the algorithm,
often accounting for the majority of the computation time. A simple solution is to
finely sample the kernel before starting the reconstruction process. The computer can
then look up these sampled values and use nearest-neighbor or linear interpolation
when a kernel value is needed. This amounts to replacing C(kx ) with the sampled
and interpolated kernel C(k
b x ) (Table 3.2). While it is clear that this scheme offers
a significant computation time savings, it is not clear how the aliasing amplitude
is affected. I demonstrate that if not done carefully, this sampling can become the
dominant source of inaccuracy in the gridding algorithm. I also show how to imple-
ment this scheme so that all of the computation time savings are realized without a
noticeable loss in accuracy.
The aliasing amplitude of a sampled kernel is both a function of the sampling plus
interpolation and the underlying kernel being sampled. With a low kernel sampling
density, the dominant source of the aliasing amplitude is the interpolation function
and kernel sampling density. As the kernel sampling density increases, the aliasing
amplitude becomes more defined by the underlying kernel being sampled. This is easy
to see if one considers the two extremes: 1) sampling with one point and using linear
interpolation, the resulting kernel is a triangle kernel regardless of the kernel function
being sampled; 2) sampling with infinitely many points, the sampled kernel becomes
the underlying kernel function itself and the aliasing amplitude has no dependence
on the interpolation function. Appendix A provides a mathematical analysis of the
aliasing amplitude of a sampled and interpolated kernel. Here I describe the results
3.4. PRESAMPLING THE KERNEL 41

Figure 3.8: (a) Dependence of the two components, ε1 and ε2 , of the aliasing ampli-
tude on the kernel sampling density, S, for a sampled Kaiser-Bessel kernel with linear
interpolation. The kernel width is W = 5 and β was obtained using Eq. 3.5. Shown
here is pixel position i = 64, with N = 128, G = 176 (α = 1.375). Since ε2 has
a weak dependence on the kernel sampling density, once ε1 drops below ε2 , little is
gained by increasing the kernel sampling density. (b) Decrease in maximum aliasing
amplitude as kernel sampling density increases (shown for a sampled Kaiser-Bessel
kernel with width W = 5 on a grid with oversampling ratio α = 1.375). When the
kernel is under-sampled, ε ≈ ε1 and the aliasing amplitude corresponds to Eqs. 3.7
and 3.8. Once the kernel is sufficiently sampled, ε ≈ ε2 , which has little dependence
on the interpolation method or kernel sampling density. The advantage of linear in-
terpolation is that it reaches the domain where ε ≈ ε2 with a much smaller kernel
sampling density than using nearest-neighbor interpolation.

of that analysis and give a simple expression for ensuring that a kernel is sampled
finely enough such that its aliasing amplitude is that of the underlying kernel function
being sampled.
As shown in appendix A, Eq. A.3, one useful property of the aliasing amplitude
of a sampled kernel is that it can be written as the quadrature sum of two terms,
q
ε[i] = (ε1 [i])2 + (ε2 [i])2 , (3.6)

where the first term, ε1 , is only a function of the interpolation function and sampling
density. The second term is a function of the kernel being sampled, but is mainly
independent of the interpolation function and sampling density, as shown in Fig.
42 CHAPTER 3. GRIDDING

3.8(a). Roughly, one can take ε1 to be the aliasing amplitude due to sampling of the
kernel and ε2 to be the aliasing amplitude inherent in the choice of kernel. The effect
of increasing the sampling density is to decrease the value of ε1 . Once ε1 is an order
of magnitude less than ε2 , ε[i] ∼
= ε2 [i] and there is little benefit to further increasing
the kernel sampling density, also illustrated in Fig. 3.8(a).
The fundamental difference between nearest-neighbor and linear interpolation is
how quickly ε1 decreases as the kernel sampling density is increased. For both in-
terpolation schemes, ε1 is greatest at the edge of the image, when i = −N/2. For
the two interpolation schemes this maximum value of ε1 over the image is derived in
appendix A to be

0.91
max(nearest-neighbor ε1 ) = (3.7)
αS
0.37
max(linear ε1 ) = . (3.8)
(αS)2

Since ε1 is inversely proportional to the sampling density S for nearest-neighbor

interpolation, a large kernel sampling density is needed to decrease ε1 below the
value of ε2 . For linear interpolation, ε1 is inversely proportional to S 2 , and requires a
far smaller kernel sampling density for a comparable decrease in ε1 , as shown in Fig.
3.8(b).
This point can be illustrated with a simple design example; the objective is to
design a sampled kernel to achieve a maximum aliasing amplitude of about 10−3 on
a grid with oversampling ratio α = 1.25. From Fig. 3.4, this can be achieved with
a Kaiser-Bessel kernel with W = 6. To ensure that sampling the kernel does not
adversely affect the aliasing amplitude, one can specify that max(ε1 ) should be 10−4 .
For nearest-neighbor interpolation, this would require a kernel sampling density of
7280. With W = 6, this would entail sampling the kernel with 43,680 points. Using
linear interpolation, the sampling density would need to be at least 49, requiring only
294 points for a kernel width of six.
Using a nearest-neighbor interpolation kernel or a triangle kernel as the gridding
kernel are special cases of presampled kernels, when S = 1. In both of these cases,
ε = ε1 and Eq. 7 and 8 can be used to predict the kernel accuracy, as shown in Fig.
3.4. PRESAMPLING THE KERNEL 43

Figure 3.9: Two common convolution kernels, the nearest-neighbor kernel and the
triangle kernel, are simple cases of presampled and interpolated kernels. In both of
these cases, the kernel sampling density, S = 1. From Eq. A.5 in appendix A, when
S = 1, ε2 = 0 and so ε = ε1 . (a) Reconstruction with nearest-neighbor interpolation
kernel on a grid with oversampling ratio, α = 2. (b) Error in (a) windowed up 5 times.
(c) Reconstruction with a triangle kernel, width W = 2, on a grid with oversampling
ratio, α = 2. (d) Error in (c) windowed up 25 times. These errors agree well with
the aliasing amplitude. From Eqs. 3.7 and 3.8, the maximum error for the nearest-
neighbor kernel should be about 5 times that of the triangle kernel. Comparing (b)
and (d) shows that this is the case.
44 CHAPTER 3. GRIDDING

Figure 3.10: Numerical phantom in Fig. 2(a) acquired with a radial trajectory. Shown
is the image reconstruction error resulting from different gridding parameters. (a)
α = 2, Kaiser-Bessel kernel, W = 6. (b) Same parameters as (a), but now the kernel
is sampled with 600 points and linear interpolation is used. This gives no increase in
the reconstruction error. (c) Same kernel as (a), but here the kernel is sampled with
3000 points and nearest-neighbor interpolation is used. Even though far more points
are sampled, using nearest-neighbor interpolation increases the image reconstruction
error. (d) α = 1.375, Kaiser-Bessel kernel, W = 5, sampled with 300 points and
linearly interpolated. Compared to the method in (c),(d) uses a smaller oversampling
ratio, a narrower kernel, and fewer kernel samples making it much quicker to compute.
However, it still has the same level of reconstruction error. All image reconstruction
error images have been windowed up 10,000 times relative to Fig. 2(a).

3.9.
I recommend using linear interpolation since the number of samples required to
reduce the aliasing amplitude due to nearest-neighbor sampling can be very large. As
well, it should be noted that in terms of computation time and memory it is preferable
to lose accuracy through choosing a smaller oversampling ratio or kernel width than
through overly coarse kernel sampling. Figure 3.10 highlights this concept. With
an oversampling ratio of two and a kernel width of six, sampling with 3000 points
and using nearest-neighbor interpolation (c) leads to a significant increase in error
over the unsampled kernel (a) and oversampling with 600 points and using linear
interpolation (b). If this level of error is acceptable for the given application, a
computation time and memory improvement can be seen by reconstructing using an
oversampling ratio of 1.375, a kernel width of five, sampling with 300 points and using
linear interpolation, (d), since this has the same level of error.
For the Kaiser-Bessel kernel, Eq. 3.4 can be used for apodization correction.
When the kernel is sampled and interpolated, the apodization correction function
should change accordingly. In appendix B, I give an accurate method for determining
3.5. CHOOSING THE OPTIMAL SAMPLED KERNEL 45

the apodization correction from the kernel sample values.

3.5 Choosing the Optimal Sampled Kernel

Thus far the sampled kernel has been designed by sampling a continuous valued
kernel, like the Kaiser-Bessel window. By sampling finely enough and through proper
interpolation, the sampled kernel can achieve the same accuracy as its continuous
valued template. However, this is not the only way to design a sampled kernel.
Starting with a certain kernel width and kernel sampling density one can design
the sample values directly to achieve the minimum value for the maximum aliasing
amplitude for a given oversampling ratio.
Once the interpolation method has been chosen, the kernel sampling density and
oversampling ratio ε1 is fixed and modifying the sample values can only reduce ε2 .
Because of this, the problem can be written as minCs (kx ) max(ε2 ), where Cs (kx ) is the
sampled kernel. Equation A.5 in appendix A gives the following expression for ε2 in
terms of the inverse Fourier transform of Cs (kx ):
v #2
u S−1 "
uX b h(i + Gp)c s (i + Gp)
ε2 [i] = t
p=1
h(i)c s (i)

where b
h(x) is a simple function of h(x) as shown in appendix A. From this equation,
one can write the following second-order cone program (SOCP) [16]:

find Cs (kx )
v
u S−1 h i2
uX
subject to t h(i + Gp)cs (i + Gp) ≤ εmax
b
2 h(i)cs (i)
p=1

and cs (i) > 0

 
N N N
for i ∈ − , − + 1, . . . , − 1 .
2 2 2

This SOCP finds a kernel, Cs (kx ), if it exists, such that the maximum value of ε2 [i] ≤
46 CHAPTER 3. GRIDDING

Figure 3.11: Comparison of the image reconstruction error for four different kernels.
(a)-(d) used a radial acquisition and (e)-(h) used a spiral acquisition; the object
sampled is shown in Fig. 2(a). In the first column, a width four Kaiser-Bessel kernel
was used on a grid with an oversampling ratio of two. In the second column a width
five Kaiser-Bessel kernel was used on an α = 1.375 grid, showing a similar error level.
In the third column, the kernel used in (b),(f) was sampled with 300 points and
linear interpolation was used, not changing the error noticeably. In the last column,
the kernel was also of width five, sampled with 300 points and linearly interpolated,
but the optimal sampled kernel was used (section 3.5). The optimal sampled kernel
distributes the errors more evenly over the image and results in a lower maximum
error. All reconstruction error images are windowed up 5000 times relative to Fig.
2(a).

εmax
2 . See appendix C for a MATLAB implementation of this SOCP. By decreasing
εmax
2 until the smallest value is found where a solution exists, the optimal sampled
kernel is found.
The disadvantage of this method is that a series of SOCP problems must be solved
for every α, W , and S that is to be used. Each solution can take a couple of minutes
to compute. However, once the solution is computed, the sample values can be stored
in a file for future use. The advantage of this method is that it does provide a slightly
lower maximum aliasing amplitude than the Kaiser-Bessel window. As well, it tends
to distribute the aliasing amplitude more evenly over the entire image as shown in
Fig. 2 and 3.11.
For the examples given in Fig. 3.11, the maximum reconstruction error of the
optimal sampled kernel (d),(h) is 75% of the maximum reconstruction error of the
3.6. THE BLOCK STORAGE FORMAT 47

comparable unsampled Kaiser-Bessel kernel (b),(f). This leads to the surprising no-
tion that sampling the kernel can actually lead to lower errors in practice than if one
had not sampled the kernel. This is not an inherent advantage of sampled kernels,
since a sampled and interpolated kernel is a special type of continuous kernel. Rather,
sampling the kernel allows the use of the design method developed above, which gives
a kernel design not readily available when a continuous approach is taken.

3.6 The Block Storage Format

In practice, the amount of time it takes to complete an image reconstruction compu-
tation on a modern computer is dependent on both the amount of computation that
must be performed and the memory bandwidth and latency. Using a reduced over-
sampling ratio and a sampled kernel reduces the computation time by reducing the
amount of computation that must be performed. These techniques are so successful
that even for moderately sized 3-D images of 128 × 128 × 128, the computation time
is not limited by the ability of the central processing unit (CPU) of the computer
to perform the necessary calculations, but by the memory bandwidth and latency:
the time it takes to get data from the computer memory to the CPU for processing.
This section develops the block storage format, which is a reorganization of the grid
data in the computer memory. The block storage format reduces the traffic between
the CPU and memory, reducing the computation time approximately four-fold when
reconstructing a 256 × 256 × 256 image.
In standard practice, the grid data is stored in a line-by-line storage format as
shown in Fig. 3.12. In this format, the data for each line of the 3-D grid is stored
contiguously in memory. Figure 3.12 also shows the block storage format, where data
for a small 3-D block of the grid is stored contiguously in memory. In the current
implementation 8 × 8 × 8 voxel blocks are used, corresponding to one page of memory.
In the line-by-line format, data is very spread out in memory when traversing the
data perpendicular to the lines. By storing the data in the block storage format,
data likely to be accessed concurrently is put in the same page of memory. Numerical
experiments were performed using two 3-D trajectories: the spiral-PR trajectory
48 CHAPTER 3. GRIDDING

Line-by-line Storage Block Storage

Figure 3.12: Block storage stores the same information as the common line-by-line
format, but the information is organized differently in the computer memory. In the
current implementation, 8 × 8 × 8 blocks of grid locations are stored concurrently in
memory.

and the cones trajectory, both illustrated in Fig. 3.13. Figure 3.14 shows how the
number of pages accessed during the gridding process is reduced by using the block
storage format, speeding up the entire reconstruction. Gridding was performed with
a kernel width of four and oversampling ratio of 1.25. In addition, with the line-by-
line storage format, the computation time is very sensitive to the orientation of the
grid. For example, when the spirals are rotated about the contiguous-line direction
to form the spiral-PR trajectory, far fewer memory pages need to be accessed than
when the spirals are rotated around an axis normal to the contiguous-line direction.
Block storage is immune to this effect.
In order to perform the FFT while the data is stored in the block format, one can
simply copy a row of blocks to a buffer, perform a 1-D FFT, using fftw, and copy the
result back, as shown in Fig. 3.15. Repeating this for all three dimensions gives a 3-D
FFT. Performing the FFT in this way is 4 times faster than performing a 3-D FFT
on data stored in a line-by-line format, for a 320 × 320 × 320 grid. Figure 3.16 shows
the FFT computation time for the grid stored using both line-by-line format and the
block storage format. Less dependent on the CPU-memory traffic constraints, the
computation time for the FFT with block storage format varies consistently with the
3.7. FOV AND RESOLUTION FLEXIBILITY 49

Orientation Direction
(a) (b)

Figure 3.13: (a) Cones trajectory. (b) Spiral-PR trajectory.

amount of computation, roughly N 3 log N .

3.7 FOV and Resolution Flexibility

In the gridding algorithm, data is convolved with a gridding kernel and the result is
sampled onto a Cartesian grid. Since the sampling onto the grid causes a repetition
in image space, the gridding kernel must be chosen to suppress these repetitions from
aliasing back into the image. Usually, the FOV is set to the FOV of the object and the
kernel is designed to suppress artifacts from the point-spread-function associated with
the k-space trajectory. As shown in Fig. 3.17, if an FOV smaller than the object
is chosen, it is possible to use a gridding kernel for this smaller FOV to suppress
other parts of the object (in addition to artifacts from the point-spread-function)
from aliasing into the smaller FOV. Choosing the appropriate part of the image to
reconstruct is accomplished by adding a linear phase in k-space, corresponding to a
shift in image space.
With the gridding method of reconstruction, the FOV and resolution of the final
50 CHAPTER 3. GRIDDING

|| to memory lines 3000 to memory lines

3000

# Pages Accessed
# Pages Accessed

2000
2000

to memory lines 1000

1000
|| to memory lines
0 0 0
0 2000 4000 6000 8000 1000 2000 3000
(a) Interleave (b) Interleave
3000
3000
# Pages Accessed
# Pages Accessed

2000
2000

1000
1000 ,|| memory lines ,|| memory lines

0 0 0
0 2000 4000 6000 8000 1000 2000 3000
(c) Interleave (d) Interleave

Figure 3.14: Far fewer pages of memory need to be accessed while gridding an in-
terleave when block storage is used instead of line-by-line storage. Plots of the num-
ber of memory pages accessed for each interleave are shown for (a) cones trajectory
with line-by-line storage, (b) spiral-PR trajectory with line-by-line storage, (c) cones
trajectory with block storage, (d) spiral-PR trajectory with block storage. For the
line-by-line format, the number of pages accessed depends on whether the orientation
direction of the trajectory (Fig. 3.13) is parallel or perpendicular to the orientation
of the memory lines. When block storage is used, the number of pages accessed is
very immune to the orientation direction of the trajectory.
3.7. FOV AND RESOLUTION FLEXIBILITY 51

FFT Buffer
Figure 3.15: When the grid is stored in block format, the FFT is performed in each
direction (x, y, z) separately. Blocks in the direction along which the FFT is to be
performed are copied to an FFT Buffer, converting these blocks to line-by-line format.
The FFT is performed on these lines using FFTW and then the lines are copied back
to block storage.

image are not tied to the FOV or resolution of the k-space trajectory used to acquire
the data. This makes gridding a very flexible reconstruction method. In applications
such as cardiac imaging, where the region of interest (the heart) is smaller than the
FOV of the k-space trajectory (the chest), reconstruction time can be reduced by
using a reduced FOV. Since the FOV can be easily shifted by applying a linear phase
shift, gridding makes it easy to implement reconstruction methods such as PILS [17]
that rely on localized coil sensitivities. To balance the competing objectives of high
resolution, good signal-to-noise ratio and fast imaging, often the image space resolu-
tion of the acquired data is chosen to be different along the x, y, and z dimensions.
When this acquired data is reconstructed using gridding, the resolution of the grid
can be chosen to be equal in all dimensions. This provides a simple way to achieve the
correct aspect ratio regardless of how the data is acquired. By separating the choice
of FOV and resolution for the final image from the design choices of the k-space
trajectory, gridding simplifies the imaging process.
52 CHAPTER 3. GRIDDING

2
W
10

B kN 3logN
1 640
10 W
FFT (seconds)

B
0
10 320
W

B
−1 160
10 W FFTW
B Blocked
B
W
80
−2
10
2 3
10 NxNxN 10

Figure 3.16: Time required to compute a 3-D FFT plotted against FFT size. The
computation grows as N 3 log N , where N is the size of the FFT along one dimension.
When N = 80, the time to compute the FFT is similar between line-by-line FFTW
and when block storage format is used. As the size of the FFT increases, the com-
putation time grows as N 3 log N when the block storage format is used. When the
line-by-line format is used, the computation time grows at a faster rate as memory
management becomes an issue.
3.7. FOV AND RESOLUTION FLEXIBILITY 53

+ +
...
...
+
...
+
...
FOV FOV
1.25 x FOV 1.25 x FOV 1.25 x FOV1.25 x FOV
(a) (b)
Gridding Kernel Appodization
Object * PSF

Figure 3.17: Sampling onto a grid causes the image convolved with the point-spread-
function (PSF) to be repeated and summed every 1.25×FOV, for a grid oversampling
ratio of 1.25. The gridding kernel is designed to suppress these repetitions from
entering into the FOV. Typically, the FOV is set to capture the entire object (a), but
can be set to capture a part of the object (b), allowing the use of a smaller grid.
54 CHAPTER 3. GRIDDING

3.8 Conclusion
In this chapter, a number of technical refinements to the basic gridding algorithm are
brought together. These include using a minimal oversampling ratio, presampling the
kernel, two methods for designing high performance gridding kernels, the use of the
block storage format and flexibility in specifying the field-of-view for the reconstructed
image. All of these developments have been placed in a consistent framework and I
have developed the aliasing amplitude as a useful metric for evaluating the accuracy
of the gridding algorithm.
These technical improvements complement each other, giving an accurate and
efficient method for reconstructing MR images when the data is acquired using non-
Cartesian trajectories. The computation time and memory savings are especially
significant in 3D imaging, where a thirty-fold reduction in computation time and
three-fold reduction in memory requirements can be achieved.
Chapter 4

Anti-aliasing Partially Parallel

Encoded Acquisition
Reconstruction

4.1 Introduction
In magnetic resonance imaging (MRI), it is well known that multiple receiver coils can
be used to reduce the gradient encoding required and, consequently, the time needed
to acquire an image. This is because multiple receiver coils enable the image to be
encoded in parallel—at each sample time each receiver collects a differently encoded
datum. The acquired data are the result of encoding with both the gradients and the
spatial sensitivity of the receiver coils (coil sensitivities) and no longer correspond to
samples in a common k-space.
Over the past few years, many reconstruction methods that take advantage of sen-
sitivity encoding have been developed and improved. The original SMASH method
[18] is limited to Cartesian k-space trajectories and places requirements on the coil
sensitivity functions that are difficult to achieve in practice. SENSitivity Encoding
(SENSE) theory [19, 20] relaxes the requirement that the coil sensitivities have spe-
cific profiles and provides an iterative method for reconstructing arbitrary k-space
trajectories. The SPACE-RIP [21] method allows non-iterative reconstruction for

55
56 CHAPTER 4. APPEAR

k-space trajectories which do not fall on a Cartesian grid in one dimension, while
the recently introduced PARS method [22] can reconstruct one, two and three di-
mensional arbitrary k-space trajectories without iteration. All of the above methods
require the coil sensitivity functions to be found to a high degree of accuracy. Find-
ing the coil sensitivities can be accomplished by performing a full field-of-view (FOV)
initial calibration scan before acquiring the image data [19] or by designing the k-
space trajectory such that a low resolution full FOV scan can be extracted from the
acquired data [23]. In practice it is difficult to obtain the coil sensitivity functions
without errors and, even when small errors exist in the coil sensitivity functions used
in the reconstruction process, these errors can lead to visible image artifacts.
PILS [17] is a non-iterative method that does not require the coil sensitivities to be
known with great accuracy and works with arbitrary k-space trajectories. However,
the PILS method is only viable when the coil sensitivities are sufficiently localized in
space.
GRAPPA, both single-column [24] and multi-column [25] versions, as well as pre-
decessors AUTO-SMASH [26] and VD-AUTO-SMASH [27], uses an autocalibration
technique that does not require the coil sensitivities to be known and avoids prob-
lems in coil sensitivity estimation that affect the previous methods. For Cartesian
trajectories, autocalibration is able to remove the aliasing artifacts caused by reduced
gradient encoding. While the autocalibration technique has been extended to specific
non-Cartesian trajectories [28, 29, 30, 31, 32, 33], in doing so, it loses some of its
ability to successfully remove all of the aliasing artifacts.
In this chapter, I develop a new calibration technique which I call the local pro-
jection calibration technique. I show that this new technique is able to remove the
aliasing artifacts from arbitrary k-space trajectories, without needing to estimate the
coil sensitivity functions.
In their development of SENSE theory [19], Pruessmann et al. provide a general
formulation for encoding with coil sensitivities. In this chapter, I extend the linear
algebra framework of SENSE to develop the local projection calibration technique.
Using this linear algebra framework, I show that the local projection calibration tech-
nique is fundamentally different from techniques that use low resolution images to
4.2. THEORY 57

construct coil sensitivity estimations. Moreover, the local projection calibration tech-
nique avoids the main difficulties in estimating coil sensitivities from low resolution
data: Gibbs ringing distortion and an inability to deal with sensitivity maps with
high frequency content.
Finally, I describe the APPEAR (Anti-aliasing Partially Parallel Encoded Acqui-
sition Reconstruction) method for reconstructing parallel encoded acquisitions as a
practical method that uses the local projection calibration technique to remove the
aliasing artifacts caused by reduced gradient encoding. The non-iterative APPEAR
method is unique in being able to take advantage of the local projection calibration
technique in reconstructing arbitrary k-space trajectories. I show phantom and in
vivo reconstruction results using APPEAR for trajectories which are non-Cartesian
in the one phase-encode dimension.

4.2 Theory
In this section, the pertinent SENSE theory is reviewed and then extended to pro-
vide an understanding of the local projection calibration technique. The APPEAR
method is introduced, demonstrating how local projection calibration can be used to
reconstruct images from arbitrary k-space sampling.

4.2.1 SENSE
In MRI, data acquisition can be viewed as analyzing the spatially varying transverse
magnetization, m(r), by projecting it onto a collection of encoding functions which
take the form ej (k, r) = sj (r)g(k, r) where sj (r) is the sensitivity of receiver coil j and
g(k, r) = exp(−i2πkT r) is the gradient encoding function for location k, as described
in Table 4.1. When only one receiver coil is used, each encoding function can be
uniquely identified by a specific location in k-space. When using multiple coils, the
notion of location can be extended by defining an encoding location to be the pair of
values, (j, k), consisting of an integer coil index j and continuous-valued three-tuple
k-space location k, that uniquely identifies an encoding function. The projection of
58 CHAPTER 4. APPEAR

Table 4.1: General Variables and Functions

r Spatial location. 3 × 1
j Coil index.
k k-space location. 3 × 1
m(r) Magnetization being analyzed.
sj (r) Sensitivity of coil j.
g(k, r) Gradient encoding function
for k-space location k.
T
g(k, r) = exp −i2πk r
ej (k, r) Encoding function.
ej (k, r) = sj (r)g(k, r)
(j, k) Encoding location, coil j at k-space location k.
dj (k) Data acquired at encoding location (j, k).
A Acquisition set.
Contains all encoding locations of data (from all coils)
acquired in the scan.
NA Number of elements in A
(also number of data values acquired in scan).
Nv Number of voxels in reconstructed image.
Nc Number of receiver coils.
dA Vector containing acquired data values. NA × 1
m Vector containing voxel values of image being analyzed. Nv × 1
ej (k) Encoding vector. Nv × 1
sj Sensitivity vector. Nv × 1
g(k) Gradient encoding vector. Nv × 1
EA Encoding matrix for acquired data. NA × Nv
Each row consists of eTj (k) for a (j, k) ∈ A.
RA SENSE reconstruction matrix for acquired data. Nv × NA
m
b Vector containing reconstructed voxel values. Nv × 1
DSFT {a} (k) Discrete space Fourier transform.
DSFT {a} (k) = gT (k)a
a·b Hadamard or entry-wise vector product.
4.2. THEORY 59

m(r) onto an encoding function is accomplished by the integral

Z
dj (k) = ej (k, r)m(r) dr, (4.1)
V

where dj (k) is the resulting datum acquired at encoding location (j, k).

Reconstruction can be viewed as an attempt to synthesize the magnetization image

from the results of this analysis. Typically, a finite number of data points are acquired
and a finite number of voxel values are reconstructed. The acquisition set, A, contains
the encoding locations, from all coils, acquired in the scan,

A = {(j, k)|dj (k) was acquired in the scan} , (4.2)

and the acquired data can be assembled in the vector dA . Since the encoding is linear,
a linear reconstruction is appropriate, which can be written as

m
b = RA dA , (4.3)

where the vector m

b contains the reconstruction estimate of the voxel values and RA
is the reconstruction matrix, each row of which synthesizes an image voxel value from
the acquired analysis data.

Pruessmann et al. showed that when the magnetization, m(r), can be approx-
imated with a Dirac delta function at the center of each voxel, the reconstruction
matrix, RA , that minimizes the noise in the reconstructed image can be expressed as

†
RA = EA , (4.4)

† H −1

−1 H −1
where EA = EA ΨA EA EA ΨA is a pseudo-inverse of the encoding matrix EA . I
will discuss EA in more detail shortly. ΨA , the sample noise matrix, is included to
calculate the pseudo inverse that minimizes the noise in the reconstructed image. In
general, this chapter assumes the Dirac delta voxel approximation, allowing functions
of r to be written as vectors with a total number of elements of Nv , the total number
60 CHAPTER 4. APPEAR

Table 4.2: Data Synthesis Variables

dbj (k) Synthesized datum at location (j, k).
Linear combination of data values.
wj,j 0 ,A (k, k0 ) Linear combination weight applied to dj 0 (k0 ) when synthesizing
dbj (k) from data values acquired at locations in A.
wj,A (k) Vector containing all weights to synthesize dbj (k) from data
values acquired at locations in A. NA × 1
ej,A (k)
b Estimated encoding vector for location (j, k). Nv × 1
Linear combination of encoding vectors at locations in A.
j,A (k) Error in estimated encoding vector b ej,A (k).

of voxels. With this discretization of the magnetization, Eq. 4.1 can be written as

dj (k) = eTj (k)m, (4.5)

where m is the Nv × 1 magnetization vector, containing the magnetizations of the

voxels, and ej (k) is the Nv × 1 encoding vector for encoding location (j, k). Note
that the encoding vector can still be broken up into the vector representation of the
sensitivity and gradient encoding terms, i.e., ej (k) = sj · g(k), where sj is multiplied
by g(k) using the Hadamard, or entry-wise, product.
Each row of the encoding matrix, EA , consists of an encoding vector in the acqui-
sition set such that
dA = EA m. (4.6)

SENSE theory assumes that the coil sensitivities are known, allowing the encoding
matrix EA and any encoding vector ej (k) to be calculated.
When the magnetization is fully encoded, the set of encoding vectors comprising
EA span the space of the magnetization vector m. In this case, Eq. 4.4 can be used
to find the reconstruction matrix and the image can be reconstructed using Eq. 4.3.
Since the final objective of any reconstruction method is to reconstruct the image,
this is usually how SENSE theory is employed.
As shown later, the APPEAR method involves synthesis of unacquired data. Due
to this property of APPEAR, it is helpful at this point to note that SENSE theory can
4.2. THEORY 61

be employed to synthesize, from the acquired data, any datum dbj (k) for an unacquired
location k within the extent of acquired k-space. As SENSE theory assumes that
the coil sensitivities are known (and hence the encoding functions are known), the
magnetization vector in Eq. 4.5 can be replaced with the SENSE estimate of the
magnetization from Eq. 4.3:

dbj (k) = eTj (k)m

b (4.7)
= eTj (k)RA dA (4.8)
X
= wj,j 0 ,A (k, k0 )dj 0 (k0 ), (4.9)
(j 0 ,k0 )∈A

where wj,j 0 ,A (k, k0 ) is a complex scalar value formed by multiplying eTj (k) by the
column of RA corresponding to the encoding location (j 0 , k0 ). From Eq. 4.9 it is clear
that the estimated datum, dbj (k), is simply a linear combination of the acquired data.
T
The linear combination weights can be denoted by the vector wj,A (k) = eTj (k)RA ,
allowing Eq. 4.9 to be rewritten as

T
dbj (k) = wj,A (k)dA . (4.10)

While Eq. 4.10 is very useful for showing how an unacquired datum can be synthe-
sized from the acquired data, it does not provide much intuition as to the quality of
the synthesis. A better intuition can be obtained by expressing the synthesized datum
in terms of an estimated encoding vector. This can be accomplished by substituting
Eq. 4.6 into Eq. 4.10:

T
dbj (k) = wj,A eTj,A (k)m.
(k)EA m =b (4.11)

ej,A (k) is a linear combination of the encoding vectors

The estimated encoding vector b
of the acquired data
eTj,A (k) = wj,A
b T
(k)EA . (4.12)

Since I use the idea of an estimated encoding vector extensively as I develop

the theory for local projection calibration, it is useful to discuss it in further detail.
62 CHAPTER 4. APPEAR

While both the magnetization and the encoding vector are written as vectors to take
advantage of matrix notation, both are best viewed as images. Acquiring a datum
as in Eq. 4.5 can then be viewed as multiplying the magnetization image by an
encoding vector image and summing all of the resultant voxel values. By expressing
a synthesized datum in terms of an estimated encoding vector, as in Eq. 4.11, the
synthesized datum can also be viewed as the result of summing the multiplication of
two images (the magnetization image and the estimated encoding vector image). If
the estimated encoding vector for an encoding location is not the same as the true
encoding vector for that location, the synthesized datum can be in error. This can
be made more concrete by defining the error in the estimated encoding vector as the
vector
Tj,A (k) = b
eTj,A (k) − eTj (k). (4.13)

The error in the estimated encoding vector can also be viewed as an image and the
error in the synthesized datum can then be calculated by multiplying the magnetiza-
tion image by this error image and summing the voxel values. By Parseval’s theorem,
the total energy of the artifacts in image space is equal to the total energy of all of the
data synthesis errors. Although this does not give an indication of the level of struc-
ture in the artifacts, it points out that reducing the errors of the estimated encoding
vectors directly reduces the energy of the artifacts in the reconstructed image. To
ensure accurate synthesis, the error in an estimated encoding vector image should be
zero, or sufficiently close to zero, at any voxel locations that contain magnetization.
Note, however, that an estimated encoding vector and its associated error vector are
dependent solely on the linear combination weights and the encoding vectors of the
acquired data; they do not depend on the magnetization itself. Thus, the error in
the estimated encoding vector, which can be viewed as an image, gives us the abil-
ity to see the quality of a set of linear combination weights independently from the
magnetization being imaged.

When the magnetization is fully encoded, and the coil sensitivity functions known,
the linear combination weights calculated by SENSE give a perfect estimate of the
encoding vector. Since typically the magnetization is over-encoded, with more data
4.2. THEORY 63

acquired than voxels reconstructed, the SENSE weights are not the only weights that
would give a perfect estimate of the encoding vector. However, when the acquired data
contains noise, these SENSE weights do give the linear combination that minimizes
the noise in the synthesized datum.
As can be seen from Eq. 4.10, when SENSE is used to synthesize data at unac-
quired k-space locations, it uses a linear combination, similar to GRAPPA, PARS and
APPEAR. Since all of these methods can be expressed as procedures for synthesizing
unacquired data using a linear combination, the differences in image quality between
these methods can be attributed to different choices for linear combination weight
values. When the coil sensitivities are known, the weights calculated by SENSE al-
low the encoding vectors to be perfectly estimated, resulting in removal of aliasing
artifacts. However, when errors exist in the coil sensitivity estimates, these errors
are propagated to the weights calculated by SENSE, leading to imperfect estimations
of the encoding vectors and possible residual aliasing artifacts. In the next section
I show how APPEAR uses local projection calibration to obtain linear combination
weights without estimating the coil sensitivities, and that these weights give estimated
encoding vectors with very little error.

4.2.2 APPEAR

APPEAR generates the linear combination weights needed to synthesize unacquired

data from data acquired on all of the coils by using local projection calibration together
with central region interpolation, both of which are described in this section. Using
these linear combination weights, unacquired data can be synthesized—a process I
refer to as local projection interpolation. The current implementation of APPEAR
synthesizes data across a full-FOV grid for each coil; separate coil images can be
obtained by computing the Fourier transform of each grid. These coil images can
then be combined to form a composite image; currently this combination is done
using sum-of-squares.
In the first part of this section, I develop the local projection calibration technique
and show that the technique is able to find high quality linear combination weights.
64 CHAPTER 4. APPEAR

While the local projection calibration technique does not assume that the coil sensi-
tivity functions are known, it does make the assumption that within a central region
of k-space, the data values are known or can be found for every encoding location
(j, k). Since invariably only a finite amount of data is acquired, one must be able to
synthesize appropriate values for unacquired locations within the central region.
In the second part of this section, I develop the idea of central region interpolation,
whereby a data value at any encoding location within a sufficiently sampled central
region of k-space can be synthesized from only the data acquired on the same coil.
The APPEAR method uses central region interpolation to synthesize data at encoding
locations within the central region as needed to perform local projection calibration.
In turn, local projection calibration generates linear combination weights used to
synthesize unacquired data throughout k-space using data acquired from all coils.

Local Projection Calibration

Despite Eq. 4.12, a small subset of the acquired encoding vectors (a few rows of EA )
is sufficient for estimating an unacquired encoding vector, when the chosen subset of
acquired encoding vectors are local to the k-space location of the encoding vector to
be estimated. By limiting the number of acquired encoding vectors used in the esti-
mate, local projection calibration can find high-accuracy weights using the sufficiently
sampled central region of k-space. When linear combination weights are found in this
way, the estimated encoding vector is a projection of the unacquired encoding vector
onto the subspace spanned by the local acquired encoding vectors. Self-calibrating
PARS [22] also synthesizes an unacquired datum from acquired data in a limited local
neighborhood in k-space, and uses a sufficiently sampled central region of k-space for
calibration. However, the self-calibration technique [23] used by PARS, which uses
the sufficiently sampled central region to estimate the coil sensitivities, is different
from the local projection calibration technique (which does not estimate the coil sen-
sitivities) and generates a different set of linear combination weights. In this section,
I develop the local projection calibration technique, leaving a comparison between
self-calibrating PARS and local projection calibration for the discussion section.
I start by introducing some notation that allows us to work with local acquired
4.2. THEORY 65

Acquisition Set, A Local Set at k, L (k) Pattern Set P (k,k’)

Pattern of L (k) shifted to k’

Coil 4

Coil 4
Coil 4
ky ky ky

k’

Coil 3

Coil 3
Coil 3
kx kx kx

Coil 2

Coil 2
Coil 2

Coil 1

Coil 1
Coil 1

(a) (b) (c)

Location is a member of the set

Figure 4.1: (a) The acquisition set, A, contains all of the acquired data locations for
all coils. (b) The local set, L(k), contains the acquired data locations for all coils
in a radius κ around k. (c) The pattern set, P(k, k0 ), contains the locations in the
pattern of L(k), but centered around k0 . The locations in P(k, k0 ) are not necessarily
in the acquisition set and therefore can be different from L(k0 ).

data. I then specify how the weights for the local neighborhood are calculated. Fi-
nally, I show that when the weights are calculated in this way, the errors in the
estimated encoding vectors weighted by the magnetization are minimized.
Instead of using the entire acquisition set, A, I define the local set, L(k), as a set
of acquired locations local to k-space location k:

L(k) = {(j, k + ∆k)|(j, k + ∆k) ∈ A, k∆kk < κ} . (4.14)

Essentially, L(k), as described in Table 4.3 and illustrated in Fig. 4.1, contains
all of the acquired locations from all coils within a radius κ of k. By replacing the
66 CHAPTER 4. APPEAR

Table 4.3: Local Projection Calibration Variables

L(k) Local set.
Contains all the acquired locations, from all coils,
within radius κ of k.
NL(k) Number of elements in L(k).
EL(k) Encoding matrix for data at locations in L(k). NL(k) × Nv
dL(k) Vector containing acquired data values local to k. NL(k) × 1
wj,L(k) (k) Vector containing all weights to synthesize dbj (k) from data
values at locations in L(k). NL(k) × 1
ej,L(k) (k)
b Estimated encoding vector for location (j, k). Nv × 1
Linear combination of encoding vectors at locations in L(k).
j,L(k) (k) Error in estimated encoding vector b ej,L(k) (k). Nv × 1
P(k, k0 ) Pattern set.
L(k) shifted by k0 − k in k-space.
dP(k,k0 ) Vector containing acquired data values at locations in P(k, k0 ).
NL(k) × 1
When locations in P(k, k0 ) are not part of the acquired data,
e P (k, k0 ) is used as described in the Central Region Interpolation
d
section.
EP(k,k0 ) Encoding matrix for data at locations in P(k, k0 ). NL(k) × Nv
ej,P(k,k0 ) (k0 )
b Estimated encoding vector for location (j, k0 ). Nv × 1
Linear combination of encoding vectors at locations in P(k, k0 ).
j,P(k,k0 ) (k0 ) Error in estimated encoding vector b ej,P(k,k0 ) (k0 ). Nv × 1
4.2. THEORY 67

acquisition set, A, with the local set, L(k), in Eq. 4.10, the synthesis of a datum
from locally acquired data can be written as

T
dbj (k) = wj,L(k) (k)dL(k) (4.15)

and the associated estimated encoding vector can be expressed by replacing the ac-
quisition set, A, with the local set, L(k), in Eq. 4.12:

eTj,L(k) (k) = wj,L(k)

b T
(k)EL(k) . (4.16)

Each row of EL(k) contains an encoding vector at one of the locations in L(k). Note
that EL(k) has far fewer rows than EA and the encoding vectors comprising EL(k) are
eTj,L(k) (k)
not expected to span the space of the magnetization vector m. As such, b
will not be a perfect estimate of eTj (k). Rather, it is desirable to find the linear
T
combination weights, wj,L(k) (k), that minimize the magnitude of the error in the
estimated encoding function:

min Tj,L(k) (k) = min T

wj,L(k) (k)EL(k) − eTj (k) . (4.17)
wj,L(k) (k) wj,L(k) (k)

Both local projection calibration and local projection interpolation are named
eTj,L(k) (k) is always in
using ‘local projection’ because the estimated encoding vector b
the space spanned by the encoding vectors in the local set and can be visualized as
a projection of the true encoding vector onto this ‘local subspace’. Local projection
calibration is a method for generating the linear combination weights that would make
eTj,L(k) (k), as calculated in Eq. 4.16, a projection of eTj (k) onto the local subspace.
b
While Eq. 4.17 would generate such a set of linear combination weights, for it to
be possible to use Eq. 4.17 directly, the encoding vectors would need to be known,
implying that the coil sensitivities would also need to be known. Local projection
calibration does not use Eq. 4.17 directly. Instead, it provides a way to find near
optimal local weights without needing to know the encoding vectors.
Local projection calibration takes advantage of the fact that linear combination
weights which are found in one region of k-space can be applied in a completely
68 CHAPTER 4. APPEAR

different region. The reason such an approach works is because the element-by-
element magnitude of the error in the estimated encoding vector, j,L(k) (k) , which I
refer to as the magnitude image of the error in the estimated encoding vector, does not
change when the encoding vector location and its neighborhood is shifted in k-space
from position k to position k0 . This can be shown by noting that shifting an encoding
vector in k-space is equivalent to multiplying the vector by a gradient encoding term,
allowing us to write ej (k0 ) = ej (k) · g(k0 − k). Since j,L(k) (k) is a linear combination
of encoding vectors, the shifted error vector can be written as j,L(k) (k) · g(k0 − k).
Finally, since g ∗ (k, r)g(k, r) = exp(i2πkT r) exp(−i2πkT r) = 1

j,L(k) (k) · g(k0 − k) = j,L(k) (k) . (4.18)

Alternatively, one can recognize that a shift in k-space is equivalent to applying

a linear phase shift to the image-space vector j,L(k) (k), which does not affect the
magnitude of any element.

Because shifting in k-space does not affect the magnitude of an image-space vector,
the same weights that minimize the magnitude of the shifted error vector will minimize
the magnitude of the original error vector. Thus, what matters in determining the
weights is not where the encoding vectors are in k-space, but the pattern they form
relative to the k-space location of the encoding vector being synthesized. To take
advantage of the ability to freely shift in k-space, the pattern set is defined as

P(k, k0 ) = {(j, k0 + ∆k)|(j, k + ∆k) ∈ L(k)} , (4.19)

which takes the pattern of acquired locations around location k and centers them
around location k0 . Illustrations of A, L(k) and P(k, k0 ) are given in Fig. 4.1. Using
the pattern set, a modification of Eq. 4.16 can be written, in which all of the encoding
vectors have been shifted in k-space from position k to position k0 :

eTj,P(k,k0 ) (k0 ) = wj,L(k)

b T
(k)EP(k,k0 ) . (4.20)

Equation 4.20 expresses the estimate of the encoding function at location (j, k0 ) from
4.2. THEORY 69

encoding functions in the local neighborhood of k0 that are of the same pattern as
the acquired encoding functions are around k. The error in the estimated encoding
vector in Eq. 4.20 can be written as

j,P(k,k0 ) (k0 ) = b
ej,P(k,k0 ) (k0 ) − ej (k0 ) (4.21)
ej,L(k) (k) − ej (k) · g(k0 − k)
 
= b (4.22)
= j,L(k) (k) · g(k0 − k). (4.23)

Thus, using the pattern set, Eq. 4.18 can be written as

j,P(k,k0 ) (k0 ) = j,L(k) (k) . (4.24)

Note that both Eq. 4.16 and Eq. 4.20 use the same set of weights, wj,L(k) (k), to
calculate the estimated encoding vector. When the same weights are used, Eq. 4.24
shows that the magnitude images of the error vectors for both encoding vectors are
the same.
Without knowing the coil sensitivities, no encoding vectors are known, so one can-
not find the linear combination weights using Eq. 4.17. However, there is sufficiently
sampled data within the low spatial frequency central region of k-space and this data
can be used to generate the linear combination weights.
All of the data used to find the linear combination weights comes from the so-
called central region of k-space. That is, local projection calibration assumes that
there is a region, called the central region, where dj (k) is known, or can be found,
for all encoding locations (j, k) as long as k is within the central region. In addition
to the central region, we define the fit region as a region within the central region as
shown in Fig 4.2. The fit region is important because for any location k0 within the
fit region, the pattern set P(k, k0 ) will always be in the central region.
Local projection calibration determines the local weights by minimizing
h i h i
T
wj,L(k) (k) dP(k,k01 ) . . . dP(k,k0N ) − dj (k01 ) . . . dj (k0Nk0 ) , (4.25)
k0

where k01 . . . k0Nk0 denote Nk0 k-space locations that fall on a full-FOV Cartesian grid
70 CHAPTER 4. APPEAR

ky

Fit Region kx

Central Region

Figure 4.2: The central region is the region within which all possible dj (k) are known.
The fit region is the region inside of the central region, with a margin of κ, the radius
of the pattern set.
4.2. THEORY 71

in the fit region and dP(k,k0 ) contains data at the locations in P(k, k0 ). I will develop
more insight into the choice of k01 . . . k0Nk0 shortly, as we analyze Eq. 4.25. In addition,
it is important to remember that in this section it is assumed that within the central
region of k-space, the data values are known or can be found for every encoding
location (j, k). When central region interpolation is discussed, it will become clear
how the data values in dP(k,k0i ) and dj (k0i ) are determined when they are not explicitly
acquired. The weights can be computed directly as
h ih i†
T
wj,L(k) (k) = dj (k01 ) . . . dj (k0Nk0 ) dP(k,k01 ) . . . dP(k,k0N ) , (4.26)
k0

where † denotes a pseudo-inverse.

While Eq. 4.26 shows how the local projection calibration technique finds linear
combination weights using data from the central region, it does not give much infor-
mation on the properties of these weights. I now take advantage of the notation I have
developed to demonstrate that the weights found by minimizing Eq. 4.25 will mini-
mize the error in the estimated encoding vector weighted by the magnetization. That
the minimization is performed on the magnetization-weighted error in the estimated
encoding vector and not on the unweighted error in the estimated encoding vector is
to some degree unavoidable. However, this weighting is not undesirable, as the esti-
mated encoding vector will tend to be more accurate where there are large amounts
of magnetization. Errors in the estimated encoding vectors at locations where there
are large amounts of magnetization can result in large errors in the synthesized data
values, so it is especially important that the estimated encoding vectors be accurate
at these locations.

To demonstrate that such a minimization is, in fact, accomplished by local projec-

tion calibration, I start by looking at the data estimation error for a k-space location
k0 , within the fit region and expanding each data value in terms of its encoding vector
72 CHAPTER 4. APPEAR

and the magnetization, dj (k) = eTj (k)m and dP(k,k0 ) = EP(k,k0 ) m.

T
wj,L(k) (k)dP(k,k0 ) − dj (k0 ) = wj,L(k)
T
(k)EP(k,k0 ) m − eTj (k0 )m (4.27)
eTj,P(k,k0 ) (k0 )m − eTj (k0 )m
= b (4.28)
= Tj,P(k,k0 ) (k0 )m (4.29)
= gT (k0 ) j,P(k,0) (0) · m .
 
(4.30)

The expression gT (k0 ) j,P(k,0) (0) · m can be interpreted in a very natural way.
 

First of all, note that j,P(k,0) (0) is a special case; since ej (0) = sj , the sensitivity
of coil j, b
ej,P(k,0) (0) is an estimate of the sensitivity of coil j and j,P(k,0) (0) is the
error in that estimate. j,P(k,0) (0) · m is then the sensitivity estimation error weighted
by the magnetization. Finally, multiplication by gT (k0 ) computes the discrete space
Fourier transform (DSFT) of the weighted sensitivity estimation error, evaluated at
k0 .
In Eq. 4.25, the magnitude is taken of a vector, each element of which contains
T
the data estimation error, wj,L(k) (k) dP(k,k0 ) - dj (k0 ), for one of the k0 locations in
k01 . . . k0Nk0 . Expressing the data estimation error for each k-space location, k01 . . . k0Nk0 ,
as a sample of the discrete space Fourier transform of the sensitivity estimation error
weighted by the magnetization, as in Eq. 4.30, Eq. 4.25 can be written as the magni-
tude of the discrete space Fourier transform of the magnetization-weighted sensitivity
estimation error, sampled at the low spatial frequency locations k01 . . . k0Nk0 :

N k0
X
DSFT j,P(k,0) (0) · m (k) δ(k0i − k) .

(4.31)
i=1

Since Eq. 4.25 is mathematically equivalent to Eq. 4.31, the weights that min-
imize Eq. 4.25 will also be the weights that minimize Eq. 4.31. Thus, the weights
obtained using local projection calibration minimize the low frequency components of
the magnetization-weighted sensitivity estimation error. Since the majority of the en-
ergy of the magnetization and sensitivity functions is contained in their low frequency
components, local projection calibration will tend to find weights that successfully
4.2. THEORY 73

minimize the overall magnetization-weighted sensitivity estimation error.

By setting k0 = 0 in Eq. 4.24, it becomes apparent that the magnitude image of
the sensitivity estimation error is equivalent to the magnitude image of the error in
the estimated encoding vector used to synthesize the datum at location k:

j,P(k,0) (0) = j,L(k) (k) (4.32)

Thus, the magnetization-weighted sensitivity estimation error is equal to the

magnetization-weighted error in the estimated encoding vector,

j,P(k,0) (0) · m = j,L(k) (k) · m (4.33)

and minimizing the magnetization-weighted sensitivity estimation error is equivalent

to minimizing the magnetization-weighted error in the estimated encoding vector.
Equation 4.31 also provides some insight into the choice of k01 . . . k0Nk0 . By choosing
k01 . . . k0Nk0 to cover a full-FOV grid in the fit region, one ensures that all of the low
frequency components of the magnetization-weighted sensitivity estimation error are
included in the minimization. Choosing the extent of the fit region in k-space is still
an inexact art, depending to some degree on the energy distribution in k-space of
the magnetization and coil sensitivities. However, by choosing Nk0 to be much larger
than the total number of elements in the local set, L(k), the error in the estimated
encoding vector will be determined more by the ability of the local encoding vectors,
EL(k) , to synthesize the estimated encoding vector than by the effect the extent of
the fit region has on the calculation of the weight values.
In the above analysis of local projection calibration, it was shown that high quality
local weights can be found without knowing the coil sensitivities, assuming all data
values required by Eq. 4.26 are known, or can be found. While interpolation might be
necessary when required data values have not been acquired directly, when all of the
acquired data falls on a full-FOV uniform Cartesian grid, acquiring data at all grid
locations in the central region is sufficient for local projection calibration. As shown in
Fig. 4.3(a), this is because all of the data values required by Eq. 4.26 are contained in
the acquisition set A. The autocalibration technique used by Cartesian GRAPPA is
74 CHAPTER 4. APPEAR

Uniform Sampling non-Uniform Sampling

ky ky

k’ k’
kx kx

k k

(a) (b)

Figure 4.3: (a) When data is acquired on a uniform Cartesian grid, it is possible to
choose the k0 such that the pattern set P(k, k0 ) is contained in the acquisition set A.
In this case, no interpolation within the central region is required to perform local
projection calibration. (b) When a non-uniform sampling trajectory is used, P(k, k0 )
is not contained in the acquisition set A and data at the locations in P(k, k0 ) must
be synthesized.

a practical implementation of local projection calibration that takes advantage of this

case. However, to use local projection calibration for arbitrary k-space trajectories,
one must find a practical way for dealing with the reality that data is not acquired
for every location in the central region. In the next section, I show how APPEAR
uses central region interpolation to deal with this problem.
4.2. THEORY 75

Central Region Interpolation

When data is acquired using a k-space sampling pattern that does not fall on a
uniform Cartesian grid, local projection calibration cannot be used directly since it
requires data values from locations in the central region that have not been acquired,
as shown in Fig. 4.3(b). To overcome this problem, the APPEAR method accurately
synthesizes data values in the central region from the acquired data. By being able
to synthesize any value in the central region, APPEAR is able to take full advantage
of the local projection calibration technique.
Note that APPEAR uses two types of data synthesis. Local projection interpola-
tion is used to synthesize unacquired data from data acquired on multiple coils, and
the linear combination weights for this are found using local projection calibration.
Central region interpolation uses central region data from a single coil to synthesize
central region data on that same coil. When data from multiple coils is used for
synthesis, the challenge is to perform accurate calibration and generate good linear
combination weights without knowing the coil sensitivities (which prevents one from
knowing the encoding functions). In the above section on local projection calibration,
I describe how APPEAR deals with this challenge. When an unacquired datum is
synthesized only from data acquired on the same coil, as is the case for central region
interpolation, only the gradient encoding is relevant and the gradient encoding vectors
are known. Thus, calibration is not a problem in this case. Rather, the challenge is to
ensure that the synthesis is accurate. Here, I show how central region interpolation
achieves this goal.
Before engaging the details of central region interpolation, I summarize the overall
APPEAR procedure. As illustrated in Fig. 4.4, for each point on the k-space grid
to be synthesized (from acquired data on all coils), APPEAR determines the local
pattern of acquired data and then synthesizes data (from acquired data on a single
coil) in that pattern at locations across a full-FOV grid in the fit region. In the
central region, a synthesized datum is denoted by dej (k), differentiating it from the
synthesized datum value dbj (k). Whereas dbj (k) is synthesized from data from all coils
in regions where the gradient encoding does not sufficiently encode for the FOV, dej (k)
is synthesized only from coil j data in the central region, where the gradient encoding
76 CHAPTER 4. APPEAR

Synthesizing dˆ j(k)
Find Local Set at k, L(k)

Synthesize data d˜ j(k’)

and d˜P(k,k’) at Pattern
Central Region
Sets P(k,k’) for a
Interpolation
full-FOV grid of k’ Local Projection
locations in the fit region Calibration

Calculate synthesis
weights (Eq. 4.34)

Synthesize datum from

Local Projection local acquired data and
Interpolation synthesis weights
(Eq. 4.15)

Figure 4.4: The APPEAR procedure for synthesizing an unacquired datum at location
k using local acquired data from multiple coils.
4.2. THEORY 77

does sufficiently encode for the FOV. The data values synthesized at the locations in
P(k, k0 ) can be assembled into a vector, denoted by d
e P(k,k0 ) . The weights can then
be calculated as
h ih i†
T
wj,L(k) (k) = dej (k01 ) . . . dej (k0N ) d e P(k,k0 ) . . . d
1
e P(k,k0 ) ,
N
(4.34)

which is identical to Eq. 4.26, except that the data values are synthesized. Using
Eq. 4.34, APPEAR determines a separate set of weights for each grid point to be
synthesized, repeating the procedure illustrated in Fig. 4.4 to synthesize data at all
of the grid points outside of the central region.

The success of the APPEAR method rests heavily on the ability to synthesize data
values within the central region of a coil’s k-space using only the data acquired on
that receiver coil. Techniques, such as gridding [1, 2] and BURS [8], for synthesizing
data in new k-space locations from a set of full-FOV data are commonly used for
reconstructing non-Cartesian k-space sampling patterns acquired using only one coil.
However, these methods are designed to minimize errors in a particular FOV in image
space, whereas the synthesis method used by APPEAR needs to minimize errors in the
low spatial frequency region of k-space. This is a subtle but important difference. For
example, when gridding is performed on an oversampled grid, errors can be pushed
outside of the FOV of the object, separating the errors from the object. However,
these errors cannot be separated from the data in such a clean fashion in k-space. I
now develop the procedure that APPEAR uses to get high quality synthesized data
suitable for use with the local projection calibration technique.

The single-coil data synthesis problem can be formulated in a similar way to the
multi-coil formulation. However, in the single-coil case, the unknown coil sensitiv-
ity is grouped with the magnetization by letting ms denote the sensitivity-weighted
magnetization, described in Table 4.4. Since only one coil is being considered at a
time, the coil index j has been dropped. The analysis of the sensitivity-weighted
magnetization is done by gradient encoding, such that d(k) = gT (k)ms . Similar to
the multi-coil case, one can synthesize an encoding function from acquired encoding
functions in the local k-space neighborhood, but in this case the encoding is solely
78 CHAPTER 4. APPEAR

Table 4.4: Central Region Interpolation Variables

dj (k) Synthesized data value at location (j, k) from coil j acquired data.
e
ms Sensitivity weighted image. Nv × 1
C(k) Kernel set.
Contains all the acquired k-space locations within radius κ of k.
NC(k) Number of elements in C(k).
GC(k) Gradient encoding matrix for data at locations in C(k). NC(k) × Nv
wC(k) (k) Vector containing all weights to synthesize g(k) from gradient
encoding vectors at locations in C(k). NC(k) × 1
bC(k) (k) Estimated gradient encoding vector for location k. Nv × 1
g

accomplished by the gradients. The single coil version of Eq. 4.16 can be written as

T T
g
bC(k) (k) = wj,C(k) (k)GC(k) , (4.35)

where C(k) is the k-space kernel set

C(k) = {k + ∆k|dj (k + ∆k) was acquired, k∆kk < κ} (4.36)

and GC(k) is the gradient encoding matrix, each row of which contains a gradient
encoding vector whose location is in the kernel set. The estimated gradient encoding
vector is denoted by g
bC(k) (k) and the linear combination weights, wj,C(k) (k), can
be found as the weights that minimize the magnitude of the error in the estimated
gradient encoding vector:

T
min g
bC(k) (k) − gT (k) = min T
wj,C(k) (k)GC(k) − gT (k) . (4.37)
wj,C(k) (k) wj,C(k) (k)

Unlike the multi-coil case where the encoding vectors are not known, in the single coil
case the gradient encoding vectors are known and the weights can be found directly
by minimizing Eq. 4.37.
While such a procedure finds the optimal weights for estimating g(k) from its sur-
rounding neighborhood, this estimate is still not of high enough accuracy when using
a small number of points. To illustrate this, I use a one dimensional example, show-
ing the accuracy that can be expected from this approach. Five different attempts
4.2. THEORY 79

k-space image space

kx ˆ
g(0)
1.5
kx
kx
1
* kx
kx
C(k) k=0 0.5
(a) *
0
real component -128 0 128
imaginary component rx
(b)
ˆ
apodized g(0)
1.5

*
0.5

0
-128 0 128
rx
(c)

Figure 4.5: Oversampling and apodization can greatly increase the accuracy of syn-
thesized data values in the central region. In this 1D illustration, gradient encoding
vectors at four acquired locations are used to synthesize the gradient encoding vector
at kx = 0. (a) The synthesis is performed for different shifted versions of the acquired
vector. (b) Without apodization or oversampling, the accuracy of the synthesis is
poor. (c) The acquired locations oversample the FOV with an oversampling factor
of 1.5 and an apodized version of the gradient encoding vector at kx = 0 is synthe-
sized. In this case, all of the synthesized gradient encoding vectors overlap with the
apodization vector target, showing that this synthesis can be done with high accuracy.
80 CHAPTER 4. APPEAR

are made to synthesize g(0), which equals unity, from four local gradient encoding
vectors. In each of the five attempts, the four local gradient encoding vectors are
shifted slightly in relation to g(0), as shown in Fig. 4.5(a). For each attempt, I find
the optimal weights by minimizing Eq. 4.37 and use these weights in Eq. 4.35 to find
the estimated gradient encoding vector. These five estimated encoding vectors are
plotted in Fig. 4.5(b). In the first attempt, where one of the local gradient encoding
vectors is at location k = 0, g
b(k) is unity throughout. However, as the local gradient
encoding vectors get shifted, the synthesis degrades. In the last attempt, where g(0)
is midway between two local gradient encoding vectors, the edges of the estimated
encoding vector are close to zero.
The accuracy of the data synthesis can be greatly increased using oversampling
and apodization. While both of these concepts are used in gridding, they must be
applied in a slightly different way in this case. Instead of depositing the data on
an oversampling grid, in this case oversampling is accomplished by acquiring data
in the central region more closely together (spaced for a larger FOV than the FOV
of the object). Since higher oversampling factors reduce the overall SNR efficiency
of the scan [34], it is important to use the smallest oversampling factor that still
allows for highly accurate data synthesis. Currently an oversampling ratio of 1.5 is
used, which has been shown to give high accuracy in gridding [35]. APPEAR takes
advantage of apodization by finding the weights that synthesize an apodized version
of the gradient encoding vector instead of the gradient encoding vector itself. By
synthesizing all of the data in the central region using similarly apodized versions of
the gradient encoding vectors, the resultant data, dej (k), is equivalent to data obtained
by encoding an apodized version of the sensitivity-weighted magnetization with the
non-apodized gradient encoding vectors.
Figure 4.5(c) shows that using oversampling and apodization lead to a dramatic
improvement in the fidelity of the synthesized data. In Fig. 4.5(c), all five apodized
estimated gradient encoding vectors and the target apodization function overlap,
appearing as one curve.
In a case such as the example in Fig. 4.5, where the samples in the central
region are uniformly spaced, finding the weights that fit to an apodized version of the
4.2. THEORY 81

gradient encoding vector can be accomplished by convolving the acquired data with
an appropriate gridding kernel and sampling at the unacquired location. Thus, the
result shown in Fig. 4.5(c) can be realized by using a Kaiser-Bessel window with shape
parameter β = 7.9, chosen according to Eq. 5 in [35] for an oversampling ratio of 1.5
and a kernel width of 4. In this case, the apodization will be equivalent to the Fourier
transform of the gridding kernel. Figure 4.6 shows an image space picture illustrating
how oversampling and apodization allow for highly accurate synthesis from uniformly
spaced samples. As I discuss in the next chapter, when the samples in the central
region are not uniformly spaced, the gridding approach cannot be expected to find
the optimal weights. However it is possible that for some k-space trajectories such as
radial and spiral, gridding with density compensation might give sufficient accuracy.
While a gridding approach is computationally attractive, the weights can alternatively
be found directly by solving for the weights that minimize the error in the estimated
apodized gradient encoding function:

T
min wj,C(k) (k)GC(k) − gaT (k) (4.38)
wj,C(k) (k)

where ga (k) = a·g(k) for apodization vector a. This approach is attractive because it
will find the optimal weights for any given kernel set, C(k), as long as it is sufficiently
oversampled for the FOV.

When fitting to an apodized gradient encoding function, the synthesized data

in the central region is equivalent to data that would be obtained with the same
coil sensitivity but with a magnetization that is an apodized version of the actual
magnetization. Thus, instead of the local projection calibration technique minimizing
the encoding estimate error weighted by the magnetization, in the APPEAR case
it minimizes the encoding estimate error weighted by an apodized version of the
magnetization. Because of this, one must be careful not to allow too severe of an
apodization, which can reduce the quality of the estimated encoding vector toward
the edges of the FOV. In all experiments to date, the apodization function used
(Fourier transform of a Kaiser-Bessel window) performs very well.

By slightly oversampling in the central region and synthesizing data that would
82 CHAPTER 4. APPEAR

m(x)*psf(x)
apodization function

Figure 4.6: An image space illustration of how oversampling and apodization allow
high quality synthesis of the apodized magnetization when the samples in the cen-
tral region are uniformly spaced. Oversampling the central region pushes the image
repetitions further away, allowing for a transition band. The apodization function
(the Fourier transform of the gridding kernel) then suppresses the image repetitions.
Unlike gridding, in k-space interpolation, m(r) ∗ psf(r) must be zero (or sufficiently
close to zero) in the transition band.
4.3. METHODS 83

be generated from an apodized version of the magnetization, instead of insisting

on synthesis of data that would be acquired by the scanner, one is able to provide
the local projection calibration technique with data values at any location in the
central region, as required by the local projection calibration technique. In doing so,
APPEAR gives a practical implementation of local projection calibration for arbitrary
k-space trajectories.

4.2.3 Noise
Noise in the acquired data samples affects APPEAR in two ways. Firstly, noise in
the central region can degrade the quality of the synthesis weights thereby increasing
the errors in the estimated encoding vectors. Since errors in the estimated encoding
vectors can manifest themselves as structured image artifacts, it is important that
the data acquired in the central region have a sufficiently high signal-to-noise ratio
to ensure that the estimated encoding vectors are accurate. Secondly, noise in the
data samples leads to noise in the reconstructed image. Similar to other multi-coil
reconstruction methods, the noise level varies from voxel to voxel. Since APPEAR
is non-iterative, the noise level at each voxel can be calculated in a straightforward
manner from the receiver noise matrix and the synthesis weights, as shown in the
appendix D.

4.3 Methods
Numerical simulations were performed in Matlab (MathWorks, Natick, MA) using a
Shepp-Logan phantom and the same coil architecture used for the numerical simu-
lations in [36]. The k-space trajectory used, shown in Fig. 4.7(a), consisted of 128
phase-encode lines, the central 20 lines spaced with an oversampling ratio of 1.5. The
remaining 108 lines were uniformly spaced, giving a constant acceleration of 2.25 in
the high spatial frequency region. As a full-FOV dataset for equivalent k-space cover-
age would require 256 phase-encode lines, the net acceleration was 2. Reconstructions
were performed using PARS and APPEAR with a local set radius, κ = 4/ FOV. For
84 CHAPTER 4. APPEAR

APPEAR, interpolation in the central region was performed using a kernel set ra-
dius of 4/(3FOV). The apodization function used was the Fourier transform of a
Kaiser-Bessel window with shape parameter, β = 7.9, illustrated in Fig. 4.5(c).
Scanner experiments were performed on a 1.5T GE Excite scanner using two
k-space trajectories: the trajectory used in the numerical simulations and a non-
uniformly spaced trajectory. The non-uniformly spaced trajectory was identical to
the uniformly spaced trajectory except that the 108 high spatial frequency phase-
encodes were non-uniformly spaced, giving an acceleration ranging from 1 to 3.5, as
seen in Fig. 4.7(c). A standard GRASS sequence was used for all scans.
The first experiment consisted of scanning an axial slice of a ball phantom using an
8-channel high-resolution knee coil (MRI Devices Corp., Waukesha, Wisconsin). In
this experiment, the frequency encoding was done in the A/P direction. The second
experiment consisted of scanning an axial slice of the brain of a healthy volunteer using
an 8-channel high-resolution head coil (MRI Devices Corp., Waukesha Wisconsin). In
this experiment, the frequency encoding was done in the R/L direction. A 5 mm slice
was excited for both experiments. The ball was scanned with a TE/TR of 10/1000
and an 11 cm FOV. The brain was scanned with a TE/TR of 10/50 and a 22 cm
FOV. The data was reconstructed using PARS, APPEAR and by separately gridding
the data from each coil followed by a sum-of-squares combination. The gridding was
performed as described in [35], with a Kaiser-Bessel kernel, an oversampling factor of
1.25 and a kernel width of 4.

4.4 Results
The results for the numerical simulations are shown in Fig. 4.8. While the AP-
PEAR method is able to remove the aliasing artifact, the PARS reconstruction still
has visible aliasing artifact. Since the coil sensitivities are known in the case of the
numerical phantom experiment, it is possible to calculate the error in the estimated
encoding vector, j,L(k) (k), directly for a set of linear combination weights. The error
vector, j,L(k) (k), can be viewed as an image, similar to the magnetization vector m.
While the error vector images do not directly represent errors in the reconstructed
4.4. RESULTS 85

Uniform Acceleration Variable Acceleration

0.5 0.5

ky 0 ky 0

-0.5 -0.5
0 phase-encode 128 0 phase-encode 128
(a) (c)
2 2
sampling density

sampling density

1.5 1.5

1 1

0.5 0.5

0 0
-0.5 0 0.5 -0.5 0 0.5
ky ky
(b) (d)

Figure 4.7: (a) Phase-encode locations for trajectory with uniform acceleration out-
side of the central region. (b) Sampling density of the trajectory in (a). (c) Phase-
encode locations for trajectory with variable acceleration outside of the central region.
(d) Sampling density of the trajectory in (c).
86 CHAPTER 4. APPEAR

(a) (b)

Figure 4.8: (a) Reconstruction of Shepp-Logan numerical phantom using the PARS
method. (b) Reconstruction of the same data-set as in (a) using the APPEAR
method.

images, they do indicate the accuracy with which unknown data values are synthe-
sized. The error in a synthesized datum value can be found by multiplying the error
vector image, corresponding to the encoding location of the synthesized datum, by
the magnetization and summing the result. Thus, nonzero values in j,L(k) (k) where
there is no magnetization are unimportant.
Figure 4.9 compares the error in the estimated encoding vector at one encoding
location, obtained using the APPEAR weights, to the error obtained using the PARS
weights and to the error obtained using the weights found directly from Eq. 4.17. In
the region where there is magnetization, the error in the estimated encoding vector
obtained using the PARS weights, where the coil sensitivities are estimated from
low spatial frequency data, is significantly larger than the error obtained using the
APPEAR weights. As well, since APPEAR only minimizes the error in locations
where there is magnetization, it is able to do a better job of estimating the encoding
function at locations where there is magnetization than a strict application of Eq.
4.17.
4.5. DISCUSSION 87

The results of the phantom experiment are shown in Fig. 4.10. When the data
from each coil is reconstructed separately by gridding, the aliasing artifacts due to
sampling at a reduced FOV are evident. The variable acceleration trajectory reduces
the structure of the aliasing artifact and spreads it throughout the image. Thus, while
the PARS result for the uniform trajectory shows some residual artifact, the artifact
is not noticeable for the PARS result with the variable acceleration trajectory. The
cost of using a variable density trajectory is a reduction in the SNR efficiency [34].
The local projection calibration technique used by APPEAR is able to remove the
aliasing artifact for both trajectories.
Figure 4.11 shows the in vivo results. Aliasing artifact is clearly visible in the
gridding reconstructions and there is some aliasing of the skull into the middle of the
image in the PARS reconstruction when the uniform acceleration trajectory is used.
Once again, APPEAR is able to get rid of aliasing artifact for both trajectories.

4.5 Discussion
As a method for reconstructing images from partially parallel encoded acquisitions,
the APPEAR method performs well. In the experiments presented, APPEAR had
significantly less artifact than the PARS method. The main difference between PARS
and APPEAR is that APPEAR uses local projection calibration whereas PARS uses
the self-calibration technique [23], in which data in the central region is used to esti-
mate the in vivo coil sensitivities. My results indicate that using the local projection
calibration technique leads to less errors than using the self-calibration technique.
To understand these results, it is instructive to look more closely at the self-
calibration technique used by PARS. PARS finds the local weights that satisfy

T
min wj,L(k) (k)EPARS,L(k) − eTPARS,j (k) , (4.39)
wj,L(k) (k)

which is similar to Eq. 4.17, except that EL(k) and eTj (k) are replaced with EPARS,L(k)
and eTPARS,j (k). For the PARS encoding functions, the coil sensitivities are replaced
with low resolution images reconstructed by inverse Fourier transform from the central
88 CHAPTER 4. APPEAR

(a) (b) (c)

Figure 4.9: For the numerical simulation, the coil sensitivities are known. Here
magnitude image of the error in the estimated encoding function, j,L(k) (k) , for
a coil on the right-hand side of the image and the unacquired k-space location k =
[0, 21.3/ FOV]T , is shown for three different methods of finding the linear combination
weights. (a) Knowing the coil sensitivity functions, weights are found by minimizing
Eq. 4.17. (b) Weights are found using PARS, where the coil sensitivities are estimated
from the central region data. (c) Weights are calculated using the APPEAR method.
By minimizing the magnetization-weighted error in the estimated encoding function,
APPEAR is very accurate in the area where signal is being acquired. All images
have been windowed up 100X with respect to the magnitude of the encoding function
being synthesized.
4.5. DISCUSSION 89

Gridding Uniform Acceleration Variable Acceleration

(a) (d)
PARS

(b) (e)
APPEAR

(c) (f )

Figure 4.10: An axial slice of a ball phantom was scanned using an 8-channel high-
resolution knee coil. Data was obtained using both trajectories shown in Fig. 4.7. In
this experiment, the frequency encoding was done in the A/P direction. (a),(b),(c)
were all reconstructed from the same data, as were (d),(e),(f).
90 CHAPTER 4. APPEAR

Uniform Acceleration Variable Acceleration

Gridding

(a) (d)
PARS

(b) (e)
APPEAR

(c) (f )

Figure 4.11: An axial brain slice was scanned using an 8-channel high-resolution
head coil. Data was obtained using both trajectories shown in Fig. 4.7. In this
experiment, the frequency encoding was done in the R/L direction. (a),(b),(c) were
all reconstructed from the same data, as were (d),(e),(f). The artifact in (b) can be
found by looking at the same location as the artifact in (a).
4.5. DISCUSSION 91

region data. These images can be written as [m(r)sj (r)] ∗ psf(r), where psf(r) is the
point-spread-function associated with the sampling pattern in the central region. The
point-spread-function can be modified by applying different weightings to the k-space
sampling locations. As described in [23], a Kaiser window weighting can be used to
reduce the Gibbs ringing in the resultant image. Regardless of the weights chosen for
the k-space sampling locations, convolution with psf(r) makes it difficult to extract
sj (r). Self-calibration is built on the approximation that sj (r) is very low frequency
and
[m(r)sj (r)] ∗ psf(r) ≈ [m(r) ∗ psf(r)] sj (r), (4.40)

which separates sj (r) from the other terms. Note that Eq. 4.40 is simply Eq. 1 in [23],
written in notation consistent with this paper. Once sj (r) is separated in this way,
coil-by-coil images can be reconstructed [37]. The error in the approximation given
in Eq. 4.40 tends to be largest near locations where m(r)sj (r) changes abruptly, such
as at the edge of the object. These errors in turn lead to errors in the reconstruction
matrix and estimated encoding functions. While local projection calibration limits
itself to synthesizing estimated encoding functions from local neighborhoods, it is
able to minimize Eq. 4.31 without approximation. Because of this, local projection
calibration is not susceptible to errors due to the approximation in Eq. 4.40.

The autocalibration procedure used in Cartesian GRAPPA is a case of local projec-

tion calibration, where the pattern sets needed to find the synthesis weights align per-
fectly with the acquired data in the central region. However, the extension of the au-
tocalibration technique to radial, spiral and VIPR trajectories [28, 29, 30, 31, 32, 33],
is not a case of local projection calibration. These extensions of autocalibration focus
on finding weights that optimize the fit between raw unprocessed scanner-acquired
data, or data which has been interpolated in the time dimension, but do not insist on
the integrity of the pattern set. This is very different from local projection calibration,
which insists on the integrity of the pattern set, but accepts the use of synthesized
(and apodized) data generated in the central calibration region from single-coil data
in order to generate the linear combination weights used to synthesize unacquired
data from acquired data on multiple coils.
92 CHAPTER 4. APPEAR

When autocalibration is used for non-Cartesian trajectories, one is presented with

the complication of finding acquired data in the necessary patterns for the calibration
procedure. For radial [28] and VIPR [33] trajectories, the autocalibration procedure
requires an initial full-FOV, full-resolution calibration scan. This is in contrast to
both PARS and APPEAR, which are able to calibrate on the inherently oversampled
central region of these trajectories. In addition, as noted by Samsonov et al. [38],
the calibration data required by autocalibration for these trajectories often requires
extensive time averaging, since calibration is performed in outer portions of k-space,
where the signal-to-noise ratio is worse than the signal-to-noise ratio in the central
region.
For the dual-density spiral trajectory [31, 32], autocalibration is able to calibrate
on the central region. However, unlike PARS and APPEAR, which use the entire
central region in finding the linear combination weights for a particular pattern set,
the autocalibration procedure is only able to use a fraction of the central region, as
shown in Fig. 4 in [32].
In addition, in order to find a sufficiently large number of data fits for calibration,
these methods disregard small differences between pattern sets. For example, two
pattern sets that differ by a small rotation or dilation might be treated as the same
pattern set. In Fig. 4.12, it is shown that disregarding small differences between
pattern sets degrades the quality of the estimated encoding vectors. To generate this
figure, a common set of weights were found to minimize the magnetization-weighted
errors in the encoding vectors for a collection of pattern sets which only differ by a
slight rotation (±5◦ ). Comparing these results to Fig. 9, we see that this results in an
error level similar to PARS. By using local projection calibration, APPEAR is able
to achieve a smaller error level.
APPEAR provides a great deal of freedom in choosing the k-space trajectory
used to acquire the data. While the two trajectories used in this paper have the
same number of phase-encodes and hence same net acceleration, they have different
properties. The variable acceleration trajectory has worse SNR compared to the
uniform acceleration trajectory, but is more robust to errors in the estimated encoding
vectors and less likely to produce images with visible aliasing artifacts.
4.5. DISCUSSION 93

(a) (b) (c)

Figure 4.12: Disregarding small rotational differences in pattern sets. The pattern
set around location k = [0, 21.3/ FOV]T (Fig. 4.9) is rotated continuously through
10◦ (−5◦ to 5◦ ), creating a collection of pattern sets which only differ by a slight
rotation. Data is collected from the central region for each of these pattern sets.
Then, treating the collection of pattern sets as a common pattern set, the data is
concatenated and weights are found (Eq. 4.26). The magnitude image of the error in
the estimated encoding function is calculated and windowed, as in Fig. 4.9. Shown,
the error in the estimated encoding function for the pattern set rotated (a) −5◦ ,
(b) 0◦ , (c) 5◦ . By disregarding small rotational differences in pattern sets, as is done
with autocalibration for non-Cartesian trajectories, these weights lead to significantly
larger errors in the estimated encoding vectors. By ensuring the integrity of the
pattern sets, APPEAR achieves much smaller errors in the estimated encoding vector
as shown in Fig. 4.9(c).
94 CHAPTER 4. APPEAR

The APPEAR method is unique in requiring an oversampled region in the center

of k-space. For some trajectories, such as radial, VIPR and 3D cones [39], where
oversampling is inherent, this is not a disadvantage. However, for many trajectories,
this will require attention in the trajectory design. Specifically, the central region
should be designed to critically sample a FOV larger than the extent of the object
being imaged. In my experience, a FOV of 1.5 times the extent of the object is
sufficient, but may not be optimal.

In this chapter, I have focused on describing the APPEAR method and providing
some preliminary results. As the method matures, I am exploring the tradeoffs in
the choice of parameters, such as the extent of the local neighborhood and central
region. In addition, since APPEAR, like PARS, finds a unique set of weights in order
to synthesize each unacquired datum, it is a computationally intensive method. The
following chapter discusses the computation required by APPEAR in more detail.

It should also be noted that the utility of the APPEAR method is to reduce the
encoding time necessary to acquire an image and as such, should only be employed
when the encoding time needs to be reduced. Many scans are not limited by encod-
ing time, but by the time required to obtain sufficient SNR [40]. In this case, the
APPEAR method is of limited utility. Indeed, the first use of multiple receiving coils
in MRI was to increase SNR, not to reduce encoding time [41].

Still, there are many cases when it is advantageous to reduce the encoding time.
One case is cardiac and body imaging where motion during the acquisition can lead
to artifacts. Another case is contrast enhanced imaging where high signal levels are
present, but only for a limited time frame. For such cases, the results presented
here indicate that APPEAR will provide high quality reconstructions. APPEAR is
simple to use, since no coil sensitivities are needed and APPEAR is non-iterative.
This simplicity does not limit the flexibility of APPEAR, which is able to reconstruct
arbitrary k-space trajectories.
4.6. CONCLUSIONS 95

4.6 Conclusions
The theory behind local projection calibration has been derived. This theory shows
that local projection calibration will attain near optimal linear combination weights
for data synthesis, even though the coil sensitivities are not known. This theory also
shows that there is a fundamental difference between local projection calibration and
techniques which estimate coil sensitivities from low spatial frequency data. The re-
sults of this chapter indicate that local projection calibration produces reconstructions
with less aliasing artifact than reconstructions which use estimated coil sensitivities
from low spatial frequency data.
In this chapter, the APPEAR method is introduced as a way to take advantage
of local projection calibration with arbitrary k-space trajectories. APPEAR is a
straightforward method that does not require iteration. Preliminary experiments
confirm that APPEAR is able to produce high quality reconstructions.
96 CHAPTER 4. APPEAR
Chapter 5

Correlation Values

5.1 Introduction
The previous chapter focused on the image quality benefits of the APPEAR method,
giving only passing reference to the computation requirements. In this chapter the
computation required by a straightforward implementation of APPEAR is analyzed
and the concept of correlation values is developed and used to improve the com-
putational efficiency of APPEAR by a factor of approximately 10X. Although the
concept of correlation values was developed while looking to improve the computa-
tional efficiency of APPEAR, correlation values also lead to a novel way of viewing
the reconstruction problem that can lead to useful insights into image reconstruction
methods. In this chapter the concept of correlation values is developed and used
to provide insight into synthesis errors, reconstruction regularization, limitations in
the gridding method and how local projection calibration relates to calibration by
estimating coil sensitivity functions. A 2-D version of APPEAR is implemented and
phantom and in-vivo results are obtained for variable-density spiral trajectories.
Computationally, APPEAR is similar to the PARS method in that both methods
generate a set of linear combination weights, used to synthesize an unacquired datum,
by solving a linear system. Straightforward implementations of both APPEAR and
PARS can have a heavy computational burden for non-Cartesian trajectories because
many sets of linear combination weights must be generated. For a 2-D non-Cartesian

97
98 CHAPTER 5. CORRELATION VALUES

trajectory such as a spiral trajectory, where a unique set of linear combination weights
is found for each unacquired k-space location to be synthesized, this can involve
solving tens of thousands of linear systems and over an hour to reconstruct one 2-D
image [22].

One proposal for reducing the computation time of PARS for the radial trajec-
tory is given by Samsonov et al. [38]. In this proposal, linear combination weights
are found for a fraction of the unacquired k-space locations and these linear combi-
nation weights are then interpolated to generate the linear combination weights for
the remaining unacquired k-space locations. Such an approach has similarities to
non-Cartesian GRAPPA in that both methods reduce the computational burden of
reconstruction by reducing the number of pattern sets for which solving a system
of linear equations is used to find the linear combination weights. Since the only
difference between PARS and APPEAR is how the linear combination weights are
generated for a given pattern set, such an approach can be used with APPEAR. In
such a scheme, APPEAR could be used to calculate linear combination weights for a
fraction of the pattern sets and the weights for the remaining pattern sets could be
found using interpolation.

In this chapter, I focus on a complementary avenue to reducing the computation

of APPEAR: reducing the computation necessary to generate the weights for a given
pattern set by reducing the size of the linear system that must be solved. Just as
the scheme proposed by Samsonov for speeding up PARS can be used to speed up
APPEAR, so too can the scheme proposed in this chapter be used to speed up PARS
and GRAPPA. In addition, the scheme proposed in this chapter can further reduce
the computation required by the Samsonov scheme, by reducing the computation
required to calculate the fraction of linear combination weights that are not derived
from interpolation.

The following section provides an analysis of the computation required by a

straightforward implementation of APPEAR. Next, correlation values are introduced
and it is shown that a linear system used to find linear combination weights can
be distilled to a system composed entirely of correlation values. For scans that are
5.2. ANALYSIS OF COMPUTATION 99

entirely gradient encoded, it is shown that simple analytic expressions for the corre-
lation values can be found, leading to fast computation. This finding is significant for
APPEAR, which must interpolate data in the central region. Since gridding is com-
monly viewed as a k-space interpolation technique, a section is devoted to comparing
gridding with interpolation based on correlation values (local projection interpola-
tion). While correlation values are both pedagogically useful and useful when only
gradient encoding is considered, they are not immediately useful in dealing with sensi-
tivity encoding, where the encoding functions are not known. However, many parallel
imaging reconstruction methods can be expressed in terms of approximate correlation
values. Methods that are amenable to being expressed in terms of approximate cor-
relation values include those that estimate the coil sensitivity functions as well as the
APPEAR method, which uses local projection calibration. The difference between
local projection calibration and coil sensitivity estimation is studied by examining
the difference in how these methods approximate correlation values. Following this,
techniques for efficiently computing approximate correlation values are described. Fi-
nally, an implementation of APPEAR that takes advantage of correlation values is
described and results are shown for variable-density spiral imaging.

5.2 Analysis of Computation

Like PARS and GRAPPA, the APPEAR method can be split into two steps:

1. Calibration: find the linear combination weights.

2. Synthesis: synthesize unacquired data from surrounding k-space data and the
relevant linear combination weights.

The fact that direct methods such as APPEAR can be split into calibration and
synthesis steps makes them attractive, especially for real-time applications. With
such methods, the calibration can be performed once and used to synthesize multiple
images. Since the synthesis step is computationally similar to gridding, it can be
performed in real-time on modern computers for 2-D images and a moderate number
100 CHAPTER 5. CORRELATION VALUES

of receiver coils (for example, eight receiver coils). Iterative methods such as conjugate
gradient (CG) SENSE cannot be split up in this way, requiring iteration for each
frame. Nevertheless, for non-Cartesian trajectories CG-SENSE is often much faster
than APPEAR since the calibration step used in APPEAR is very computationally
intensive. As most of the computation is contained in the calibration step, I focus on
this step for the remainder of this section.
The calibration step consists of finding linear combination weights for each unique
pattern set. Since the problem of finding linear combination weights for one pattern
set is independent of finding the linear combination weights for a different pattern
set, the solution for each pattern set can be computed separately, making it easy
to parallelize the calibration computation. This is an advantage over iterative CG-
SENSE where the iterations must proceed serially. While APPEAR has the attractive
features of separating the calibration from the synthesis and being easy to parallelize,
it has the drawback that the amount of computation required can be too large for some
applications even when multiple processors are used to parallelize the computation.
The amount of computation grows with the number of unique pattern sets for
which linear combination weights need to be found. For a Cartesian k-space trajectory
with an acceleration of two outside the central region, calibration only needs to be
performed for a single pattern set. When the acceleration outside the central region
is increased to three, calibration must be performed on two pattern sets. Calibration
of Cartesian k-space trajectories is not prohibitive since linear combination weights
need to be found for only a handful of pattern sets.
For non-Cartesian trajectories, APPEAR takes the same approach as PARS, find-
ing a unique set of linear combination weights for each data location to be synthesized.
This significantly increases the computation necessary for the calibration step: for a
spiral trajectory with an FOV/resolution ratio of 256, this entails computing lin-
ear combination weights for about 50,000 different pattern sets. To understand how
much computation this requires, it is necessary to look at the computation needed to
calculate the linear combination weights for one pattern set.
By far, the majority of the computation in calculating the linear combination
weights for a pattern set using Eq. 4.34 is contained in two steps: synthesizing
5.2. ANALYSIS OF COMPUTATION 101

pattern set data for a full-FOV grid of k0 locations in the fit region and multiplying
 
H
d
e 0)
h i  P(k,k..
1 
e P(k,k0 ) . . . d
d . (5.1)
e P(k,k0  
1 N )  
k0  
eH 0
d P(k,kN k0)

which
h is the computationally
i significant step in calculating the pseudo-inverse of
dP(k,k01 ) . . . dP(k,k0N ) since generally Nk0 , the number of k0 locations that fall on
e e
k0
a full-FOV Cartesian grid in the fit region, is much larger than NL(k) , the length
of d
e P(k,k0 ) . Synthesizing the pattern set data in the central region is proportional to
2
Nk0 NL(k) and the matrix multiplication is proportional to Nk0 NL(k) . For both of these
steps, the computation grows linearly with Nk0 .

While synthesizing the pattern set data is a step unique to APPEAR, the matrix
multiplication step is also performed in GRAPPA. In order to reduce the computation
time for the calibration step, it is not uncommon for GRAPPA implementations to
use a smaller Nk0 , the number of fits in the fit region. The downside of such an
approach is that the quality of the calibration can suffer from calibrating with a
reduced fit region. In practice, this is not a significant downside, since trajectories
used with Cartesian GRAPPA typically have a fit region that goes to the edge of
sampled k-space in the readout direction. GRAPPA implementations that do not use
the portions of the fit region that are near the edges of sampled k-space in the readout
direction still calibrate effectively since the outer portions of k-space contain only a
small portion of the energy of the image. Nevertheless, one of the contributions of
correlation values is to make the computation less dependent on k0 .

It is helpful at this point to compare the computation required by APPEAR to

that required by PARS. Recall that PARS solves Eq. 4.39 to generate the linear
combination weights:

T †
wj,L(k) (k) = eTPARS,j (k)EPARS,L(k) . (5.2)
102 CHAPTER 5. CORRELATION VALUES

Since EPARS,L(k) is a very fat matrix, with the number of columns equal to Nv , the
number of voxels in the image and the number of rows equal to NL(k) , the num-
ber of elements in the local set, the most computationally significant part of this
H
calculation is multiplying EPARS,L(k) EPARS,L(k) in calculating the pseudo-inverse of
EPARS,L(k) . Thus, the cost for PARS to compute a set of linear combination weights
2
is proportional to Nv NL(k) . Since Nv is typically much larger than Nk0 , a straight-
forward implementation of PARS is typically more computationally intensive than a
straightforward implementation of APPEAR.
The computation required by APPEAR and PARS is similar, both being propor-
2
tional to NL(k) ; they differ in that APPEAR is also proportional to Nk0 whereas PARS
is proportional to Nv . Using the correlation value, the dependence of the computation
on Nk0 and Nv can be largely removed, leaving only the common (and typically much
smaller) dependence on NL(k) .

5.3 What are Correlation Values?

A correlation value is a complex number expressing the mutual relation between two
encoding vectors. Correlation values are useful because the optimal linear combina-
tion weights for a given pattern set can be found quickly once the correlation values
that relate the members of the pattern set are known. The correlation value relating
encoding location (j1 , k1 ) to encoding location (j2 , k2 ) is given by an inner or dot
product between the corresponding weighted encoding functions:
Z
c(j1 , k1 ; j2 , k2 ) = eej1∗ (k1 , r) eej2 (k2 , r) dr, (5.3)

where
eej (k, r) = f (r)ej (k, r) (5.4)

is an encoding function where each spatial location has been weighted by f (r). The
weighting function f (r) allows more importance to be placed on the correlation be-
tween encoding functions at certain spatial locations (for example within the field-of-
view or where more magnetization is expected). A discussion of how f (r) is chosen
5.3. WHAT ARE CORRELATION VALUES? 103

for a specific application is dealt with in the following sections. For discrete encoding
vectors, the correlation value can be written as

eH
c(j1 , k1 ; j2 , k2 ) = e ej2 (k2 ) .
j1 (k1 ) e (5.5)

The problem formulation for finding optimal linear combination weights for a local
set can be written completely in terms of correlation values. Before showing this, it
is helpful to develop a simplified notation. A local set L(k), defined in Eq. 4.14
and illustrated in Fig. 4.1(b) contains NL(k) encoding vectors, corresponding to the
acquired data locations for all coils in a radius κ around k. To simplify the notation
when dealing with encoding locations within a local set, each encoding location can
be specified by an index i in the local set. In this scheme, encoding location i would
be synonymous with encoding location (ji , ki ), the ith encoding location in L(k).
The index i is in the range 1 . . . NL(k) and can be used to simplify the notation for
encoding vectors and correlation values as:

eji (ki ) = ei ,
eji (ki ) = e
e ei ,
c (ji , ki ; j, k) = c (i; j, k) and
c (ji1 , ki1 ; ji2 , ki2 ) = c (i1 ; i2 ) .

Using this simplified notation, I now proceed to express the optimal linear com-
bination weights in terms of correlation values. I start with a version of Eq. 4.17,
modified such that the error in the estimated encoding function is weighted by f , the
vector version of f (r):

min Tj,L(k) (k) · f T = min T

wj,L(k) (k)E eTj (k) ,
eL(k) − e (5.6)
wj,L(k) (k) wj,L(k) (k)
104 CHAPTER 5. CORRELATION VALUES

where  
eT1
e
 .. 
E
eL(k) = 
 . .
 (5.7)
eTNL(k)
e

Including the f weighting makes Eq. 5.6 more flexible than Eq. 4.17. For example, if
it is known a priori that the magnetization is encompassed in a circular FOV, f can
be used to minimize the error only within this FOV, resulting in less error within the
FOV, where the error can affect the image quality. Finding the linear combination
T
weights wj,L(k) (k) that minimize Eq. 5.6 is a least-squares problem that can be solved
as:

T
wj,L(k) (k) = e e†
eTj (k)E (5.8)
L(k)
 −1
eTj (k)E
= e H
eL(k) eL(k) E
E H
eL(k) . (5.9)

Taking the transpose of Eq. 5.9 and expanding E

eL(k) as in Eq. 5.7, wj,L(k) (k) can be
5.3. WHAT ARE CORRELATION VALUES? 105

written as
  H   −1 T
eT1
e eT1
e
.. ..
 h i 
 T   
∗
wj,L(k) (k) = ee (k) .   .  e e∗NL(k)
 e1 . . . e
  (5.10)
 j      
eTNL(k)
e eTNL(k)
e
  −1  
eH
e 1 eH
e 1
.. h ..
 i  
= 

 .  e
 e 1 . . . e
e NL(k) 
 
 .  ej (k)
e (5.11)
eH
eNL(k) eH
eNL(k)
 −1  
eH
e1 e e1 ... eH
e eNL(k)
1 e eH
e ej (k)
1 e
 .. .. ..   .. 
= 
 . . . 


 . 
 (5.12)
eH
eNL(k) e eH
e1 . . . e eNL(k)
NL(k) e eH
e ej (k)
NL(k) e
 −1  
c(1; 1) ... c(1; NL(k) ) c(1; j, k)
 .. .. ..   .. 
= 
 . . .  
  . 
 (5.13)
c(NL(k) ; 1) . . . c(NL(k) ; NL(k) ) c(NL(k) ; j, k)
= C−1
aa cas , (5.14)

where
 
c(1; 1) ... c(1; NL(k) )
 .. .. .. 
Caa = 
 . . . ,
 (5.15)
c(NL(k) ; 1) . . . c(NL(k) ; NL(k) )
 
c(1; j, k)
 .. 
cas = 
 . .
 (5.16)
c(NL(k) ; j, k)

The correlation matrix Caa is composed of correlation values that relate every
member of the local set to every other member. The subscript indicates that the
matrix relates analysis locations to other analysis locations. The members of the
local set are considered analysis locations, as acquiring data at these locations can be
viewed as analyzing the magnetization, each acquired datum measuring a component
106 CHAPTER 5. CORRELATION VALUES

of the overall magnetization. The correlation matrix Caa is also Hermitian (CH
aa =
Caa ) and positive definite (all of the eigenvalues of Caa are positive).

The correlation vector cas is composed of correlation values that relate every mem-
ber of the local set (analysis locations) to the unacquired location (j, k). The subscript
indicates that the vector relates analysis locations to a synthesis location. The un-
acquired location is also referred to as the synthesis location since the datum at this
location is synthesized from the data at the analysis locations.

Once the correlation values that populate Caa and cas have been generated, the
computation time for computing the linear combination weights wj,L(k) (k) is propor-
3
tional to NL(k) , exchanging the Nk0 dependence in a straightforward implementation
of APPEAR for a dependence on the far smaller NL(k) .

5.4 Fast Computation of Correlation Values

In this section, I show how correlation values can be computed quickly when the
encoding functions are known and some or all of the encoding is gradient (Fourier)
encoding.

When only gradient encoding is employed, the encoding functions have the form:

e(k, r) = g(k, r) (5.17)

= e−i2πk·r . (5.18)

In this case, the weighting function f (r) should be chosen to take into account the
correlation only within the FOV of interest. This can be accomplished with a rect
function. I start with a simple 1-D example, where
r
1  x 
f (x) = u . (5.19)
FOV FOV
5.4. FAST COMPUTATION OF CORRELATION VALUES 107

Substituting Eqs. 5.18 and 5.19 into Eq. 5.3 gives

Z r
1  x 
c (kx,1 ; kx,2 ) = u ei2πkx,1 x
FOV FOV
r
1  x 
u e−i2πkx,2 x dx (5.20)
Z FOV FOV 
1 x
= u e−i2π(kx,2 −kx,1 )x dx (5.21)
FOV FOV
  x 
1
= FT u (kx,2 − kx,1 ) (5.22)
FOV FOV
= sinc(FOV(kx,2 − kx,1 )). (5.23)

For 2-D imaging, when a circular FOV is chosen,

c(k1 ; k2 ) = jinc(FOV kk2 − k1 k) (5.24)

and when a rectangular FOV is chosen

c(k1 ; k2 ) = sinc(FOVx (kx,2 − kx,1 )) sinc(FOVy (ky,2 − ky,1 )). (5.25)

For spiral scans, which have a circular FOV, central region interpolation used in
APPEAR takes advantage of Eq. 5.24 to reduce the computation for central region
interpolation.

When receiver coil sensitivity encoding is also employed, the encoding functions
can be written as

ej (k, r) = sj (r)g(k, r) (5.26)

= sj (r)e−i2πk·r . (5.27)
108 CHAPTER 5. CORRELATION VALUES

Substitution into Eq. 5.3 yields

Z
c(j1 , k1 ; j2 , k2 ) = f ∗ (r)s∗j1 (r)ei2πk1 ·r f (r)sj2 (r)e−i2πk2 ·r dr (5.28)
Z
= |f (r)|2 s∗j1 (r)sj2 (r)e−i2π(k2 −k1 )·r dr (5.29)

= FT |f (r)|2 s∗j1 (r)sj2 (r) (k2 − k1 ).


(5.30)

Thus, when the coil sensitivity of each coil is known, correlation values can be quickly
computed using the Fourier transform.
Choosing f (r) = m(r) corresponds to the weighting used by in vivo coil sen-
sitivities, where images acquired by the coils are used in place of coil sensitivities.
Correlation values for the PARS method, which uses low resolution in vivo coils
sensitivities, can be efficiently calculated using Eq. 5.30.
Thus far, correlation values have been introduced and it was shown that the
linear combination weights used in local projection interpolation could be generated
by solving a problem composed of correlation values. This section demonstrated
techniques for quickly computing correlation values in useful cases when the encoding
functions are known. The next section explores correlation values from a different
angle, giving them a graphical interpretation that can make it more intuitive to work
with correlation values.

5.5 Graphical Interpretation and Regularization

Expressing image space as discrete voxels allows us to express the encoding functions
and magnetization in terms of vectors, making it possible to express the reconstruction
computation in terms of vector and matrix operations. A further advantage of vectors
is that they also have a simple geometric interpretation as a line segment in a multi-
dimensional space. In this section, I use the graphical interpretation of vectors to
give a visual picture for correlation values as well as local projection interpolation
and calibration. The use of regularization with APPEAR is introduced and justified
using a graphical approach.
5.5. GRAPHICAL INTERPRETATION AND REGULARIZATION 109

Both the magnetization vector m and all encoding vectors ej (k) are vectors in
the same Nv -dimensional “voxel” space. Voxel space is a multi-dimensional space,
where each axis corresponds to a single voxel value. In fact, since voxel values are
complex-valued in MRI, each voxel is defined by two axis and the dimensionality of
m and ej (k) is really 2 × Nv . The doubling of the number of dimensions due to
complex values is ignored in this section, simplifying the explanations without any
loss in generality. Still, the dimensionality of voxel space Nv is usually very large:
for a 2-D image, Nv might be 256 × 256 = 65, 536 and for a 3-D image, Nv might
be 256 × 256 × 256 = 16, 777, 216. Because of the difficulty in viewing such a large
vector space, simple examples in two and three dimensions are chosen to develop
a graphical interpretation of the mathematical operations being performed. Though
these examples are trivial, they can help develop a useful intuition into reconstruction
methods.

In a vector space view, a datum value can be seen as measuring the projection of
the magnetization vector along the direction of the encoding vector, when the encod-
ing vector is of unit magnitude. When the encoding vector is not of unit magnitude,
the data value is scaled by the magnitude of the encoding vector. As shown in Fig.
5.1, a geometric view of encoding the magnetization is of projecting the magnetiza-
tion along numerous encoding vectors, each pointing in a different direction. When an
unacquired datum dbj (k) is synthesized by computing a linear combination of acquired
data, the encoding vector b
ej,L(k) (k) for this unacquired datum is a linear combina-
tion of the encoding vectors corresponding to the acquired data (the encoding vectors
constituting EL(k) ). Because of this, b
ej,L(k) (k) is always in the subspace spanned by
EL(k) . However, the actual encoding vector of the target synthesis location ej (k) is
not necessarily in the subspace spanned by EL(k) and when ej (k) is not in the sub-
space spanned by the encoding vectors of the local set, there is necessarily some error
in the estimated encoding vector, j,L(k) (k). Local projection calibration draws its
name from the fact that b
ej,L(k) (k) is a projection of ej (k) onto the subspace spanned
by the local set, as shown in Fig. 5.2. The error in b
ej,L(k) (k) can be larger than
necessary if a non-ideal set of linear combination weights is used, as shown in Fig.
5.3. The objective of the calibration step is to generate a set of linear combination
110 CHAPTER 5. CORRELATION VALUES

Two Voxel Image

voxel
0 1
voxel 1 (a) voxel 1
m m
DC DC
High Frequency High Frequency
n
t io
c n
je e t io ea1
ea2 p ro a1 ea2 ec
oj
pr

voxel 0 voxel 0

(b) (c)

Figure 5.1: Graphical interpretation of gradient encoding a simple two voxel image.
(a) The image consists of only two voxels, voxel 0 and voxel 1. This image can be
fully encoded using two Fourier components: a low frequency sum of the voxels and
a high frequency difference of the voxels. Encoding can be viewed graphically as a
projection in voxel space of the magnetization vector m along the direction of the
low frequency encoding vector ea1 (b) and high frequency encoding vector ea2 (c).
5.5. GRAPHICAL INTERPRETATION AND REGULARIZATION 111

pixel 1

ea2
et
projection of et
ea1
pixel 0 pixel 1

Figure 5.2: Graphical representation when Nv = 3, and there are only two local
set encoding vectors, ea1 and ea2 . The target encoding vector, et is not contained
in the subspace spanned by ea1 and ea2 , although the difference between et and
the projection of et onto the subspace spanned by the local set encoding vectors is
small. The optimal linear combination of ea1 and ea2 estimates et by synthesizing the
projection of et .

weights that minimize j,L(k) (k).

While the fact that b

ej,L(k) (k) is restricted to the subspace spanned by the local set
encoding vectors has the disadvantage that some estimation error is almost certainly
unavoidable, it greatly simplifies the dimensionality of the problem. Since the best
one can achieve is to fit to the projection of ej (k) onto the subspace spanned by the
local set encoding vectors, the only information one needs to know about ej (k) is its
projection onto that subspace. In fact the entire problem becomes a problem in the
subspace spanned by the local set encoding vectors. PARS maintains the problem
in the large Nv -dimensional vector space; APPEAR expresses the problem in the
112 CHAPTER 5. CORRELATION VALUES

pixel 1

ea2
et error
es
ea1 projection of et
pixel 0 pixel 1

(a)

pixel 1

ea2
et es
ea1 projection of et
pixel 0 error pixel 1

(b)

Figure 5.3: The choice of linear combination weights can have a large effect on the
error in the estimated encoding vector. The synthesized encoding vector es is a linear
combination of ea1 and ea2 . In (a), the linear combination weights are chosen such
that es is close to the projection of et , leading to a small error (dotted line). In (b),
the linear combination weights are chosen such that es is not close to the projection
of et , leading to a much larger error vector (dotted line).
5.5. GRAPHICAL INTERPRETATION AND REGULARIZATION 113

smaller Nk0 subspace. However, expressing the problem in terms of correlation values
corresponds to working directly in the local set subspace.
Vectors in voxel space are defined by their values along the Nv orthogonal axes. In
the local set subspace, no coordinate system is defined. Instead, each vector, including
the local set encoding vectors, is defined by its correlation to the local set encoding
vectors. Because all of the vectors are defined in a relative way, the absolute position
of the local set subspace within the entirety of voxel space is lost. Geometrically, the
subspace corresponding to the pattern set P(k, k0 ) is a rotated version of the local set
L(k) subspace. When these two spaces are written in terms of correlation values, the
absolute position information is lost and the two subspaces become indistinguishable.
The correlation matrix Caa describes the local set subspace. The eigenvectors of
Caa provide an orthonormal basis for the subspace and the eigenvalues indicate the
quality of measurement in each dimension. Analyzing Caa by eigenvalues provides
insight into the appropriateness of regularization. For small eigenvalues, large linear
combination weights are required to synthesize a component along the correspond-
ing eigenvector direction. As noted in the previous chapter, large linear combination
weight values amplify noise in the data and can lead to images with a poor signal-to-
noise ratio (SNR). To maintain high SNR, one approach is to disregard the compo-
nents along the eigenvector directions with small eigenvalues. The problem with this
approach is that it requires an explicit calculation of the eigenvectors and values of
Caa , which is time consuming. Tychonov regularization achieves much the same effect
in a computationally efficient manner. Other regularization approaches are described
by Qu et al. [42].
A simple 2-D example helps to give a graphical picture of the effect of regular-
ization. Suppose two data samples are acquired at the same k-space location and
one datum value is required at this same location for reconstruction. In such a sit-
uation, the two data samples could be averaged to reduce the noise in the sample
√
and increase the signal-to-noise ratio by a factor of 2. Now suppose that the two
data samples are acquired very close to each other in k-space and for reconstruction
a data value is required very close to the two acquired data samples. Figure 5.4(a)
shows such a situation, where the three points are each separated by 0.01/ FOV and
114 CHAPTER 5. CORRELATION VALUES

0.05
k1 k2 k3
e(k1)
k
e(k2)
0.01 0.01
FOV FOV er(k1)
0
e(k3)

0 1
(a) (b)

Figure 5.4: (a) Data is acquired at k2 and k3 (black circles). A datum needs to be
synthesized at k1 (black x). All three points are very close in k-space. (b) Local set
subspace view. The local subspace is the space spanned by e(k2 ) and e(k3 ). The
projection of e(k1 ) onto the local set subspace is b e(k1 ). Synthesizing be(k1 ) from
e(k2 ) and e(k3 ) requires large linear combination weights. The regularized solution
synthesizes ber (k1 ) using much smaller linear combination weights. Note that the plot
is scaled in the ordinate direction.
5.5. GRAPHICAL INTERPRETATION AND REGULARIZATION 115

20 105

16
103
12

Gain
Gain

101
8
10-1
4

0 10-3
0 1 2 10-5 10-3 10-1 101
 
(a) (b)

Figure 5.5: Both (a) and (b) plot the gain applied to cas in an eigenvector direction
with eigenvalue λ, when computing the linear combination weights. (a) Linear axes
scaling. (b) Logarithmic axes scaling. The solid line is the gain without regularization
(1/λ). The dashed line is the gain with Tychonov regularization, where λave = 1 and
δ is found according to Eq. 5.41.

a datum value at k1 is to be synthesized from acquired data values at k2 and k3 .

Since all three points are so close to being at the same location, intuition leads to
a reasonable solution: average the two data values acquired at k2 and k3 . Applying
gridding to re-sample onto k1 would also approximate an averaging solution, taking
into account density compensation. However, without regularization, local projection
interpolation produces a different solution which is much more sensitive to noise in
the acquired data values. To show this, we use the expression for correlation values
with 1-D gradient encoding in Eq. 5.23 and form Caa and cas :

" #
sinc(0) sinc(0.01)
Caa = (5.31)
sinc(0.01) sinc(0)
" #
sinc(0.01)
cas = . (5.32)
sinc(0.02)
116 CHAPTER 5. CORRELATION VALUES

The weights can then be calculated as

w = c−1 cas (5.33)

"as #
1.9995
= . (5.34)
−0.9999

Rather than increasing the SNR, the SNR of the synthesized datum is reduced by
a factor of kwk = 2.24. With two members in the local set, the local set subspace
spans two dimensions, shown in Fig. 5.4(b). The projection of e(k1 ) onto the local
set subspace is shown in Fig. 5.4(b) as b
e(k1 ). Although the local set subspace spans
two dimensions, one of those dimensions has a very small eigenvalue. Regularization
suppresses the projection along the direction of the small eigenvalue, producing an
estimated encoding vector b
er (k1 ) as shown in Fig. 5.4(b). In doing so, the local
set subspace is effectively reduced to one dimension, in the direction of the large
eigenvalue and the SNR benefit of two measurements is realized. When Caa is better
conditioned, regularization has a minimal effect on the solution.

Tychonov regularization computes the weights as

−1
w = CH
aa Caa + δI CH
aa cas (5.35)

where I is the NL(k) ×NL(k) identity matrix and δ is a parameter of the regularization.
Continuing with our 2-D example, we start by analyzing Caa in Eq. 5.32 in terms of
eigenvectors and eigenvalues:
" #" # " #
1 1 1 λ1 0 1 1 1
Caa =√ √ (5.36)
2 1 −1 0 λ2 2 1 −1

where the two eigenvalues are λ1 = 1+sinc(0.01) = 1.99984 and λ2 = 1−sinc(0.01) =

0.00016. The inverse of Caa is found by taking the inverse of the eigenvalues:
" #" # " #
1
1 1 1 0 1 1 1
C−1
aa =√ λ1
1
√ . (5.37)
2 1 −1 0 λ2
2 1 −1
5.5. GRAPHICAL INTERPRETATION AND REGULARIZATION 117

Since λ2 = 0.00016 is very small, 1/λ2 = 6080 is very large and can lead to large
weight values. Using Tychonov regularization, C−
aa 1 is replaced with

" # λ1
 " #
1 1 1 0 1 1
 √1

−1 λ21 +δ
CH
aa Caa + δI CH
aa =√ 
λ2
. (5.38)
2 1 −1 0 λ22 +δ
2 1 −1

√
When an eigenvalue λ  δ,
λ 1
≈ . (5.39)
λ2 + δ λ
√
When an eigenvalue λ  δ

λ λ
≈ ≈ 0. (5.40)
λ2 + δ δ
Thus, δ should be chosen to be significantly larger than the square of small eigenvalues
and significantly smaller than the square of large eigenvalues. While the aforemen-
tioned principle for choosing δ is true, it is very vague and does not give a useful
expression for δ. In practice, I have found that

δ = 0.01λ2ave (5.41)

works well, where λave is the average eigenvalue of Caa and can be quickly calculated
as
trace(Caa )
λave = . (5.42)
NL(k)
Using Eq. 5.41 in the 2-D example, δ = 0.01 and
" #
0.499
w= , (5.43)
0.499
√
which is very close to the intuitive averaging solution, achieving a 2 SNR improve-
ment. Figure 5.4(b) shows the estimated encoding vector found by regularization,
er (k1 ). For our 2-D example, Fig. 5.5 compares the gain of the non-regularized and
b
regularized solutions as a function of the eigenvalue, showing that the regularized
118 CHAPTER 5. CORRELATION VALUES

solution suppresses the gain from small eigenvalues.

While the correlation matrix Caa describes the local set subspace, the correla-
tion vector cas tells one what the projection of the synthesis encoding vector ej (k)
is onto the local set subspace. When projecting ej (k) onto the local set subspace,
information about ej (k) is lost. For example, assuming Nv = 256 × 256 = 65, 536
and NL(k) = 100, there are 65, 536 − 100 = 65, 436 dimensions in voxel space that
are orthogonal to the local set subspace. The value of all components of ej (k) along
these 65, 436 dimensions are lost when ej (k) is projected onto the local set subspace.
Calibration methods that estimate the receiver-coil sensitivity functions attempt to
estimate all Nv dimensions of ej (k). When a local kernel interpolation approach is
taken, most of the information contained in the receiver-coil sensitivity function is
never used since it is orthogonal to the local set subspace. In contrast to the sen-
sitivity function estimation approach, local projection calibration used by APPEAR
focuses on accurately estimating the projection of ej (k) onto the local set subspace
(or identically the projection of the receiver-coil sensitivity function onto the pattern
set P(k, 0) subspace). Local projection calibration is not concerned with estimating
any component of ej (k) orthogonal to the local set subspace. While APPEAR is able
to accurately estimate the projection of the receiver-coil sensitivity function onto the
pattern set P(k, 0) subspace, it is difficult to obtain an estimate of the receiver-coil
sensitivity functions from this estimate, both because of the loss of information by
projection and because the projection result is only known in the pattern set sub-
space. While local projection calibration can estimate correlation values to put the
problem into the local set subspace, recall that the absolute relation of this subspace
to the entire voxel space is not known.

5.6 Gridding and Local Projection Interpolation

Both gridding and local projection interpolation can be used to synthesize k-space
data falling on a Cartesian grid from non-Cartesian acquired k-space data. Moreover,
both methods synthesize data from acquired data within a local k-space neighbor-
hood. However, the two methods have some important differences, which I outline
5.6. GRIDDING AND LOCAL PROJECTION INTERPOLATION 119

in this section. The purpose of this section is not to show that one method is supe-
rior to the other, but rather to show the strengths and weaknesses of both methods,
allowing the reader to make a more informed choice between methods for a specific
application.

One property of gridding is that it is data driven, convolving the acquired data
with a gridding kernel and sampling the result onto a Cartesian grid. In contrast,
local projection interpolation is grid driven, finding the acquired data local to a grid
point and using it to synthesize a datum at the grid point. The difference between
being data or grid driven is more a matter of implementation since, once all of the
linear combination weights for local projection interpolation are generated, they can
be reorganized in the computer memory to allow for a data driven implementation.
However, the “gridding kernel” that is generated from local projection interpola-
tion linear combination weights is not necessarily spatially invariant like the gridding
kernels discussed in chapter 3. While a data driven approach is attractive for imple-
mentation, allowing the data to be processed as it is acquired from the scanner, it
is more illuminating to compare gridding and local projection interpolation using a
matrix formulation.

When the image is encoded solely with gradient encoding, the acquired data can
be expressed with a modified version of Eq. 4.6:

dA = GA m

where GA is the gradient encoding matrix for the acquisition set A and dA is the
acquired data for the single coil. Unlike the multiple coil case, where the entries of
the encoding matrix EA are not known, GA is completely known. From a matrix
perspective then, the reconstructed image m
b can be computed as

b = G†A dA .
m

The reason such a computation is not often attempted is that the matrix G†A is very
large, making such an approach very computationally intensive. Gridding computes
120 CHAPTER 5. CORRELATION VALUES

m
b as
b = GH
m A WdA (5.44)

where W is a diagonal matrix that contains the density compensation weights. Note
that GH
A , the Hermitian conjugate of GA , is the inverse Fourier transform that grid-
ding computes quickly using convolution interpolation and the Fast Fourier Trans-
form. Gridding works well when

†
GH
A W ≈ GA (5.45)

and Sedarat and Nishimura [43] show that Eq. 5.45 can be used to find optimal
†
density compensation weights. How well GH
A W approximates GA depends on the
sampling trajectory, S(k). For a DFT trajectory, which samples a Cartesian grid,
†
GH
A = GA . For radial and spiral trajectories, suitable density compensation can
†
make GH
A W a good approximation for GA .

The gridding approximation for G†A can break down when uneven sampling is
used. Consider the sampling function
    
1 2 1 1 2 1
S(k) = III (Nx kx ) III Ny k y − + III Ny k y + . (5.46)
2 3 3 2 3 3

The magnetization is assumed to be in the Nx × Ny FOV. This sampling function

samples evenly at the Nyquist rate in the kx -direction and unevenly in the ky -direction.
In the ky -direction, the sampling alternates between the Nyquist rate and twice the
Nyquist rate, as shown in Fig. 5.6(a). Since the Fourier transform is separable in kx
and ky , we can consider only the ky -dimension without loss of generality:
   
1 2 1 1 2 1
S(ky ) = III Ny k y − + III Ny k y + . (5.47)
2 3 3 2 3 3

For this sampling function, it is optimal for all of the density compensation weights to
be equal, as can be seen by the symmetry of the sampling function. With this density
compensation, the point-spread function psf(y) is the inverse Fourier transform of
5.6. GRIDDING AND LOCAL PROJECTION INTERPOLATION 121

S(ky) psf(y)
1 0.5
Ny Ny

2 2
... ... ... N N ...
3 y 3 y
ky y
(a) (b)
m(x) * psf(x)

... ...
y

(c)

Figure 5.6: (a) The sampling function S(ky ) (Eq. 5.47) samples k-space unevenly.
Note that the samples in S(ky ) are never separated by more than 1/ Ny , never sam-
pling less than the Nyquist rate. (b) The point-spread function psf(y) which is the
Fourier transform of the sampling function. (c) The point-spread function convolved
with magnetization, showing the aliasing artifact produced in the image when grid-
ding is used for reconstruction.
122 CHAPTER 5. CORRELATION VALUES

S(ky ):
   
3 3 −iπy/Ny 3 3
psf(y) = III y e + III y eiπy/Ny (5.48)
4Ny 2Ny 4Ny 2Ny
   
3 3 π
= III y cos y (5.49)
2Ny 2Ny Ny
∞  n 
X n
= cos 2π δ y − 2Ny , (5.50)
n=−∞
3 3

which is plotted in Fig. 5.6(b). Recall that the final image for the gridding method
is
m(y)
b = m(y) ∗ psf(y). (5.51)

There are three delta functions in psf(y) that contribute to image content within the
FOV. Since we are only concerned with the image reconstruction within the FOV, we
can write
        
2π 2Ny 2π 2Ny
m(y)
b = m(y) ∗ δ(y) + cos − δ y+ + cos δ y−
3 3 3 3
   
1 2Ny 1 2Ny
= m(y) − m y + − m y− . (5.52)
2 3 2 3

Thus gridding corrupts the reconstruction with two scaled copies of the magnetiza-
tion aliasing into the reconstructed image, shown in Fig. 5.6(c). This poor quality
reconstruction is a limitation of gridding and is a particular case of the approximation
in Eq. 5.45 being a poor approximation.
I turn now to the local projection interpolation approach to solving this problem,
clearly showing how this approach can attain qualitatively different results from grid-
ding. For this example, I take the approach that we wish to synthesize unacquired
data at the locations:  
2
III Ny k y (5.53)
3
to create an evenly spaced data set. Figure 5.7 shows one of these locations being
synthesized from a local neighborhood of eight points. The linear combination weights
have been calculated using Eq. 5.35 with the correlation values found using Eq. 5.23.
5.6. GRIDDING AND LOCAL PROJECTION INTERPOLATION 123

ky
-0.057 ×

0.108
×

-0.356
× 1

0.797
×
∑
0.797
×

-0.356
× 0
y
0.108
×

-0.057
×
y

Figure 5.7: Local projection interpolation of g(k, r)(0, r) = e−i2π0·r = 1 from eight
local neighbors in S(ky ) (Eq. 5.47).

Figure 5.7 illustrates how each of the gradient encoding functions for the local
neighborhood are multiplied by a linear combination weight and the results summed
to approximate the gradient encoding function for the unacquired location, in this
case unity over the field of view. Figure 5.8 shows that the results with this approach
are markedly better than gridding this case. Whereas gridding has the expected
aliasing artifacts, local projection calibration can be used to reconstruct the image
without the aliasing artifacts seen in the gridding reconstruction.
For trajectories where Eq. 5.44 is a good approximation, gridding is a very ef-
ficient reconstruction method that does not require any matrix inverse calculations.
124 CHAPTER 5. CORRELATION VALUES

(a) (b) (c)

Figure 5.8: (a) Shepp-Logan numerical phantom. (b) Gridding reconstruction for
sampling trajectory S(ky ) in Eq. 5.47. (c) Local projection interpolation reconstruc-
tion for sampling trajectory in Eq. 5.47. Local projection interpolation reconstruction
is able to suppress the aliasing artifacts visible in the gridding reconstruction.

However, when the approximation in Eq. 5.44 is poor, local projection interpolation
can be used to reconstruct the image with less aliasing artifact, albeit at the expense
of a more computationally intensive method.
Up until this point, I have dealt with correlation values in situations where the en-
coding functions are known, allowing the correlation values to be explicitly calculated.
As highlighted in the previous chapter, the challenge with reconstructing images en-
coded with receiver-coil sensitivity functions arises from the fact that the encoding
functions are not fully known. In the next section, I look at how the correlation values
can be approximated in this situation.

5.7 Approximating Correlation Values

In this section, local projection calibration is put in the form of approximating cor-
relation values. By putting local projection calibration in the form of approximating
correlation values, a graphical picture can be developed which is useful in understand-
ing local projection calibration and critical to the following section which derives a
5.7. APPROXIMATING CORRELATION VALUES 125

fast method for computing the local projection calibration correlation value approxi-
mations.
Equation 4.34 shows how APPEAR generates linear combination weights from
synthesized data in the central calibration region using local projection calibration.
Just as Eq. 5.13 expresses Eq. 5.8 in terms of correlation values, Eq. 4.34 can be
expressed as
 −1  
c(1; 1)
b ... c(1; NL(k) )
b c(1; j, 0)
b
 .. ... ..   .. 
wj,L(k) (k) = 
 . . 


 . ,
 (5.54)
c(NL(k) ; 1) . . . b
b c(NL(k) ; NL(k) ) c(NL(k) ; j, 0)
b

where b
c(l1 ; l2 ) is the correlation value approximation made by local projection cali-
bration, relating encoding location l1 to encoding location l2 , where both l1 and l2 are
in the pattern set P(k, 0). I have chosen to base the encoding locations on P(k, 0)
instead of L(k) as this simplifies the analysis while being mathematically identical.
Each of the correlation value approximations in Eq. 5.54 can be expressed as

k0N
Xk0

c(j1 , k1 ; j2 , k2 ) =
b d∗j1 (k0 + k1 )dj2 (k0 + k2 ) (5.55)
k0 =k01
Z Z ∗
0 −i2π(k0 +k1 )·r1
= Sfit (k ) mj1 (r1 )e dr1
Z 
−i2π(k0 +k2 )·r2
mj2 (r2 )e dr2 dk0 (5.56)
Z Z
= m∗j1 (r1 )mj2 (r2 )
Z
0 0
Sfit (k0 )ei2π(k +k1 )·r1 e−i2π(k +k2 )·r2 dk0 dr1 dr2 (5.57)
Z Z
= m∗j1 (r1 )mj2 (r2 )ei2πk1 ·r1 e−i2πk2 ·r2
Z
0
Sfit (k0 )e−i2πk ·(r2 −r1 ) dk0 dr1 dr2 (5.58)
Z Z
= m∗j1 (r1 )mj2 (r2 )δfit (r2 − r1 )

ei2πk1 ·r1 e−i2πk2 ·r2 dr1 dr2 , (5.59)

126 CHAPTER 5. CORRELATION VALUES

where Sfit (k) is the sampling function for the fit grid locations (k01 . . . k0Nk0 ) and δfit (r)
is its Fourier transform:
δfit (r) = FT {Sfit (k)} (r). (5.60)

Note that I have not included the effect of apodization due to data synthesis in the
central calibration region. I have left this out for simplicity. When included, the
apodization function a(r) apodizes the correlation weighting function f (r), changing
it from m(r) to m(r)a(r). Equation 5.55 expresses a straightforward calculation of
the correlation value approximation. Equation 5.59 puts the correlation value approx-
imation in a form that allows for comparison with the definition of the correlation
value. To aid in this comparison, Eq. 5.3 can be expanded as
Z
c(j1 , k1 ; j2 , k2 ) = eej1∗ (k1 , r) eej2 (k2 , r) dr (5.61)
Z Z
= m∗j1 (r1 )mj2 (r2 )δ(r2 − r1 )

ei2πk1 ·r1 e−i2πk2 ·r2 dr1 dr2 . (5.62)

Local projection calibration approximates the δ(r2 − r1 ) function in Eq. 5.62

with δfit (r2 − r1 ), which is the Fourier transform of the sampling function for the fit
examples in the fit region. Note that when the fit region is expanded to cover all of
k-space, δfit (r2 − r1 ) → δ(r2 − r1 ).

In the previous chapter, I showed that the approximation made by local pro-
jection calibration is different from the approximation made by PARS, which uses
low-resolution in vivo coil sensitivities. The correlation value approximation used by
PARS can be written as
Z
c (j1 , k1 ; j2 , k2 ) =
b [mj1 (r) ∗ psf cal (r)]∗ ei2πk1 ·r

[m (r) ∗ psf cal (r)] e−i2πk2 ·r dr (5.63)

Z j2
= [mj1 (r1 ) ∗ psf cal (r1 )]∗ [mj2 (r2 ) ∗ psf cal (r2 )]

ei2πk1 ·r1 e−i2πk2 ·r2 δ(r2 − r1 )dr1 dr2 . (5.64)

5.7. APPROXIMATING CORRELATION VALUES 127

Unlike local projection calibration which approximates the delta function, when cali-
bration is performed using low resolution in vivo coil sensitivities, the magnetization
weighted coil sensitivities are approximated with low resolution versions of the mag-
netization weighted coil sensitivities.

For both the correlation value definition and the PARS approximation, the corre-
lation value relating encoding locations (j1 , k1 ) and (j2 , k2 ) is a function of j1 and j2
and the separation in k-space k2 − k1 . That is,

c (j1 , k1 + ∆k; j2 , k2 + ∆k) = c (j1 , k1 ; j2 , k2 ) . (5.65)

For the local projection calibration correlation value approximation, this shift invari-
ance is not generally true, and points to a difference in approach. Heberlein and Hu
show that there is some similarity between Cartesian GRAPPA, a case of local pro-
jection interpolation, and the “kriging” method used in geostatistics [44]. However,
since the cross-covariance terms used in kriging are shift invariant whereas the cor-
relation value approximations used with Cartesian GRAPPA are not shift invariant,
one can deduce that the two methods are not identical. Moreover, the correlation
matrix Caa generated by local projection calibration is guaranteed to be Hermitian
positive definite, whereas the kriging matrix is guaranteed Hermitian but not positive
definite.

Equations 5.59, 5.62 and 5.64 all express a correlation value as a 2-D Fourier
transform of similar but slightly varied functions. Figure 5.9 illustrates the similarity
and variation of these functions graphically. The shift-invariance expressed in Eq.
5.65 corresponds to having non-zero values only along the diagonal where r1 = r2 in
Fig. 5.9. While the local projection calibration approximation has non-zero values
off of the diagonal, the values diminish quickly as the distance from the diagonal is
increased. This property is used in the next section, which develops a method to
efficiently calculate the correlation value approximations.
128 CHAPTER 5. CORRELATION VALUES

coil j2 coil j1
m(r) mj1(r1)
)
-r 1
(r 2


+ +
- V - V r2 mj2(r2)

r1, r2
r1
(a) (b)

)
-r 1
mj1(r1) mj1(r1) (r 2
)  fit
-r 1
r( 2


r2 mj2(r2) r2 mj2(r2)

r1 r1
(c) (d)

Figure 5.9: This figure contains a graphical comparison between the ideal correlation
value (Eq. 5.62) and the correlation value approximation from PARS (Eq. 5.64)
and APPEAR (Eq. 5.59). (a) Plot of 1-D magnetization m(r) and coil sensitivity
for coil j1 (solid line) and coil j2 (dotted line). (b) The 2-D Fourier transform of
the correlation values (Eq. 5.62) has three parts: mj1 (r1 ), plotted along r1 , mj2 (r2 ),
plotted along r2 and δ(r2 − r1 ), which sets all values to zero which are not along the
diagonal r2 = r1 . (c) The PARS approximation replaces mj1 (r1 ) and mj2 (r2 )with low
resolution versions of the magnetization-weighted sensitivities. (d) The approxima-
tion used by APPEAR replaces the delta function with δfit (r2 − r1 ), which is a sinc
function when the fit region is a rect. A wider fit region results in a more narrow sinc
function, which more effectively suppresses off-diagonal content.
5.8. COMPUTING APPROXIMATE CORRELATION VALUES 129

Multi-Coil Array Data Data Synthesis

coil
synthesis location analysis location

ky Interpolate Correlation
Values, Solve for
Local Pattern
Reconstruction Weights
kx
Calibration
coil j1 coil j2
ky
k1
k2

kx
Calibration Gridded Calibration Correlation values
Data dj(k) Data d˜j(k) c(j1,k1; j2,k2)

Figure 5.10: The APPEAR procedure modified from Fig. 4.4 to take advantage of
more efficient computation using correlation values. For the 2-D spiral trajectories
used, this modification reduces the computation time by a factor of ten.

5.8 Computing Approximate Correlation Values

The approach that I take to implement APPEAR with correlation values is to first
calculate correlation values relating each coil pair, across a k1 , k2 grid. Then, the
APPEAR method is modified to calculate the linear combination weights from cor-
relation values, where appropriate correlation values are interpolated from the k1 , k2
grids of correlation values. The modified APPEAR implementation is shown in Fig.
5.10. In this section, I develop a method to quickly calculate the k1 , k2 correlation
value grids and show that any required correlation value can be interpolated from
these grids.
To show how I calculate the correlation values for a k1 , k2 grid, I start by rewriting
Eq. 5.56 as
130 CHAPTER 5. CORRELATION VALUES

Scal(k’)
(a)
k‘

Sfit(k’)

(b) k‘

Mj(k’)

(c) k‘

Mj(k’+)

(d) k‘

Mj(k’-)

(e) k‘

Figure 5.11: 1-D illustration demonstrating that Mj∗1 (k0 + k1 ) can be replaced with
0 +k

Mj∗1 ,cal (k0 + k1 ) = u k cal 1
Mj∗1 (k0 + k1 ) in the integrand in Eq. 5.66. The value
‘cal’ specifies the extent of the calibration region in k-space. (a) Sampling function
for calibration region, Scal (k0 ), spans extent of the calibration region (light gray). (b)
The fit region is of smaller extent (dark gray). (c) k-space data Mj (k0 ). Note that
Mj (k0 ) extends past the calibration region. The portion in the calibration region is
shaded in black. (d),(e) Mj (k0 + κ) and Mj (k0 − κ) are equivalent to Mj (k0 ) shifted
left and right by κ (the radius of the local set). Note that kk1 k ≤ κ and that only the
domain of Mj (k0 ) within the calibration region enters into the fit region when Mj (k0 )
is shifted. Because of these properties, Sfit (k0 )Mj (k0 +k1 ) = Sfit (k0 )Mj,cal (k0 +k1 ) and
Mj∗1 (k0 + k1 ) can be replaced by Mj∗1 ,cal (k0 + k1 ) in the integral in Eq. 5.66 without
changing the value of integrand.
5.8. COMPUTING APPROXIMATE CORRELATION VALUES 131

Z
c(j1 , k1 ; j2 , k2 ) =
b Sfit (k0 )Mj∗1 (k0 + k1 )Mj2 (k0 + k2 )dk0 , (5.66)

where Z
Mj (k) = mj (r)e−i2πk·r dr. (5.67)

As shown in Fig. 5.11, Mj∗1 (k0 + k1 ) can be replaced with Mj∗1 ,cal (k0 + k1 ) which is
zero outside of the central calibration region. In addition, we change variables, letting
k = k0 + k2 and define

Mj2 ,fit(k2 ) (k) , Sfit (k − k2 )Mj2 (k), (5.68)

giving
Z
c(j1 , k1 ; j2 , k2 ) =
b Mj∗1 ,cal (k − k2 + k1 )Mj2 ,fit(k2 ) (k)dk (5.69)
Z Z ∗
−i2π(k−k2 +k1 )·r1
= mj1 ,cal (r1 )e dr1
Z
mj2 ,fit(k2 ) (r2 )e−i2πk·r2 dr2 dk (5.70)
Z Z
= m∗j1 ,cal (r1 )mj2 ,fit(k2 ) (r2 )
Z
ei2π(k−k2 +k1 )·r1 e−i2πk·r2 dk dr1 dr2 (5.71)
Z Z
= m∗j1 ,cal (r1 )mj2 ,fit(k2 ) (r2 )e−i2π(k2 −k1 )·r1
Z
ei2πk·(r2 −r1 ) dk dr1 dr2 (5.72)
Z Z
= m∗j1 ,cal (r1 )mj2 ,fit(k2 ) (r2 )e−i2π(k2 −k1 )·r1
δ(r2 − r1 )dr1 dr2 (5.73)
Z
= m∗j1 ,cal (r)mj2 ,fit(k2 ) (r)e−i2π(k2 −k1 )·r dr (5.74)

= FT m∗j1 ,cal (r)mj2 ,fit(k2 ) (r) (k2 − k1 ).


(5.75)
132 CHAPTER 5. CORRELATION VALUES

Equation 5.75 expresses the correlation value approximation as the Fourier trans-
form of m∗j1 ,cal (r)mj2 ,fit(k2 ) (r). Both mj1 ,cal (r) and mj2 ,fit(k2 ) (r) can be calculated by
a Fourier transformation of central calibration region data. While k2 is fixed by
mj2 ,fit(k2 ) (r), the Fourier transform of m∗j1 ,cal (r)mj2 ,fit(k2 ) (r) contains the correlation
values for all values of k1 . Since a discrete Fourier transform is used in practice,
correlation values for all values of k1 are not generated; the current implementation
uses a grid with an oversampling ratio of 3, allowing fine sampling of k1 , and the
ability to interpolate using convolution interpolation without apodization.
By taking the Fourier transform of m∗j1 ,cal (r)mj2 ,fit(k2 ) (r) one is able to compute all
of the needed correlation values for a given j1 , j2 and k2 . By repeating this procedure,
varying j1 , j2 and k2 , all of the correlation value grids can be generated. In the interest
of reducing computation time, it is preferable to repeat this procedure as few times
as possible, sampling k2 coarsely while retaining high accuracy. It is possible to take
advantage of the diagonal nature of the Fourier transform of the correlation values
on a k1 , k2 grid, as shown in Fig. 5.12. By interpolating diagonally on k1 , k2 grid,
instead of vertically in the k2 direction, a diagonal stop-band and diagonal pass-bands
are created. Even with an oversampling ratio of one or less in the k2 direction, it is
possible to achieve high accuracy interpolation.

5.9 Methods
APPEAR for 2-D non-Cartesian trajectories was implemented in C++. Computation
of linear combination weights was implemented using the straightforward calculation
described in Fig. 4.4 and using correlation values to reduce the computation (Fig.
5.10).
Numerical simulations were performed using a Shepp-Logan phantom and the
same coil architecture used for numerical simulations in [36]. A variable density
spiral trajectory was used to encode the Shepp-Logan phantom. This 16-interleave
spiral trajectory was designed to have an FOV of 1.5X the FOV of the object, to
a resolution of 6.25 pixels, providing the central calibration region. Each interleave
then continued in k-space to a resolution of 1 pixel with 0.5X the FOV of the object.
5.9. METHODS 133

r2 k2

r1
k1

(a)
r2 k2

r1
k1

(b)
r2 k2

r1
k1

(c)

Figure 5.12: (a) (right) Correlation values sampled on a k1 , k2 grid for a particular
j1 , j2 coil pair. (left) 2-D Fourier transform of these correlation value samples. Even
sampling on the k1 , k2 grid leads to repetition in the Fourier transformed space. Since
the k1 , k2 grid is sampled finely in the k1 direction, repetitions in the r1 direction are
spaced far apart and the unwanted repetitions can be suppressed using convolution
interpolation (gridding). In the k2 direction, the sampling is more coarse leading to
more closely spaced repetitions in the r2 direction. (b) Convolution interpolation in
the k2 direction requires the stop bands (gray boxes) to border on the pass band,
making it difficult to achieve a high accuracy interpolation. (c) By using diagonal
convolution interpolation, it is possible to take advantage of the diagonal nature of
the Fourier transform of the correlation values. In the diagonal direction, the pass
band and stop bands can be much more narrow, allowing for a generous transition
band.
134 CHAPTER 5. CORRELATION VALUES

The optimal trajectory, subject to slew rate and maximum gradient limitations, that
met the design criteria was used, designed using the method outlined in [39]. For
this trajectory, the k-space area of the central calibration region is under 2.6% of the
total k-space covered.
Scanner experiments were performed on a 1.5T scanner (Signa HDx, GE Health-
care, Waukesha, WI) using a stack-of-spirals trajectory, with a 16-interleave variable-
density spiral trajectory having the same constraints as in the numerical experiment.
Reconstruction was performed by taking the Fourier transform in the stack direc-
tion and then treating each spiral as a separate data set, reconstructing each plane
individually using gridding and APPEAR.
Axial scans of a ball phantom and brain were taken using an 8-channel high-
resolution head coil. The FOV of these scans was 20 cm × 20 cm × 12.8 cm and the
resolution was 1 mm × 1 mm × 2 mm. A 50 ms TR was used, giving a total scan
time of 52 seconds. The readout time for each interleave was 5.4 ms.
Axial scans of a resolution phantom were taken using an 8-channel cardiac coil
(MRI Devices Corp., Waukesha Wisconsin). The FOV of these scans was 12 cm ×
12 cm × 9.6 cm and the resolution was 0.6 mm × 0.6 mm × 1.5 mm. A 50 ms TR
was used, giving a total scan time of 52 seconds. The readout time for each interleaf
was 6.2 ms.

5.10 Results
Figure 5.13 shows the results of the numerical simulation. The aliasing artifacts from
reduced gradient encoding, shown in the gridding reconstruction in (a) have been
removed in the APPEAR reconstruction shown in (b).
Results for the 8-channel head coil are shown in Fig. 5.14. The APPEAR method
(b), (d), is able to remove the aliasing artifacts, seen in the gridding reconstructions
(a), (c). Figure 5.15 shows the results with the 8-channel cardiac coil. Once again,
APPEAR is able to remove the aliasing artifacts shown in the gridding reconstruction.
There was no difference between image quality between images reconstructed us-
ing the straightforward APPEAR procedure (Fig. 4.4) and the modified procedure
5.11. DISCUSSION AND CONCLUSIONS 135

(a) (b)
Figure 5.13: (a) Gridding reconstruction of the Shepp-Logan numerical phantom,
encoded with variable-density spiral trajectory. (b) APPEAR reconstruction of the
same dataset as in (a).

(Fig. 5.10) which takes advantage of correlation values. However, the straightforward
procedure took approximately 30 minutes on a 3 GHz Pentium 4 processor to recon-
struct one spiral image, whereas the modified procedure was able to reconstruct the
same spiral image in approximately 3 minutes.

5.11 Discussion and Conclusions

In this chapter, the concept of correlation values was defined and developed. Cor-
relation values are useful aids for understanding local projection calibration and in-
terpolation. In addition, correlation values provide a common framework with which
to view many different parallel imaging methods. For example, APPEAR and PARS
cannot be compared in terms of their sensitivity function estimation, as APPEAR
136 CHAPTER 5. CORRELATION VALUES

(a) (b)

(c) (d)

Figure 5.14: Single axial slice from stack-of-spirals trajectory with 8-channel head
coil. (a) Ball phantom, gridding reconstruction (b) ball phantom, APPEAR recon-
struction. (c) Brain, gridding reconstruction. (d) Brain, APPEAR reconstruction.
5.11. DISCUSSION AND CONCLUSIONS 137

(a) (b)
Figure 5.15: (a) Single axial slice of a resolution phantom from stack-of-spirals tra-
jectory with 8-channel cardiac coil. (a) Gridding reconstruction. (b) APPEAR re-
construction.
138 CHAPTER 5. CORRELATION VALUES

makes no such estimation. However, both methods can be formulated as approxi-

mating correlation values and can be compared in terms of how the correlation value
approximations differ.
In addition to providing a useful framework for understanding methods such as
GRAPPA, PARS and APPEAR, correlation values lead to a simple way to reduce
the computation burden of APPEAR. By using correlation values as outlined in this
chapter, the computation time was reduced ten-fold compared to a straightforward
implementation as outlined in chapter 4. Certainly there is more work to be done to
further reduce the computation time for an APPEAR reconstruction, but the use of
correlation values is a significant stride toward reducing computation concerns as an
impediment toward adoption of APPEAR.
Chapter 6

Summary and Future

Recommendations

Modern MR scanners currently include gradient systems that can encode the magneti-
zation information along arbitrary k-space trajectories and multiple RF receiver-coils
that can take advantage of receiver-coil sensitivity encoding. These components of
MR scanners enable fast acquisition of the data necessary to form an image.
Development and improvement of methods that can efficiently reconstruct high
quality images from MR data is an important area of MRI research. With the more
widespread use of non-Cartesian k-space trajectories and parallel imaging, recon-
struction methods for MRI have, out of necessity become more sophisticated. In this
dissertation, I have presented a number of advances in MRI reconstruction:

• A number of technical refinements to the basic gridding algorithm were devel-

oped. These refinements include using a minimal oversampling ratio, presam-
pling the kernel, two methods for designing high performance gridding kernels,
the use of the block storage format and the ability to take advantage of flexibly
choosing the field-of-view for the reconstructed image. For 3-D non-Cartesian
trajectories, these technical advances can be used together to achieve a thirty-
fold reduction in computation time and three-fold reduction in computer mem-
ory requirements.

139
140 CHAPTER 6. SUMMARY AND FUTURE RECOMMENDATIONS

• A new method, Anti-aliasing Partially Parallel Encoded Acquisition Recon-

struction (APPEAR), was developed for reconstructing images encoded with
both non-Cartesian k-space trajectories and multiple receiver coils. The AP-
PEAR method is non-iterative and does not require the receiver coil sensitivities
to be known. In the development of APPEAR, the theory behind local pro-
jection calibration was derived and interpolation in the calibration region was
introduced as a way to take advantage of local projection calibration for arbi-
trary k-space trajectories.

• The concept of correlation values was developed and shown to provide a useful
framework for understanding parallel imaging reconstruction methods. In ad-
dition, by using correlation values, the APPEAR method can be implemented
more efficiently, leading to a ten-fold reduction in computation time.

6.1 Future Recommendations

While this dissertation includes a number of advances in reconstruction methods for
fast MRI, combining non-Cartesian trajectories with parallel imaging remains an ac-
tive area of research. In particular, balancing reconstruction efficiency with image
quality requirements remains a challenge. While in the long term more computing
power might alleviate some of the efficiency requirements, in the short term recon-
struction efficiency is bound to be a limiting factor as it becomes more commonplace
to use a large number of coils in parallel and acquire large 3-D data sets. The following
list gives some specific areas of future work related to the APPEAR method:

• A number of parameters used in the APPEAR method have not been adequately
or systematically studied. While it is believed that the parameters chosen in this
work were reasonable, further study could lead to a more refined implementation
of the method. These parameters include: the radius, κ, used to define the local
set, the size of the central calibration region, the amount of apodization used
to increase the accuracy of interpolation in the central calibration region and
the oversampling factor in the central calibration region.
6.1. FUTURE RECOMMENDATIONS 141

• In the current implementation of APPEAR, data is synthesized directly onto

a Cartesian grid. Since APPEAR can synthesize data at any k-space location
within the extent of acquired k-space, other possibilities exists. For example,
unacquired arms of a spiral trajectory could be synthesized to form a full field-
of-view spiral trajectory for each coil; gridding could then be used to reconstruct
each coil’s image. More work needs to be done to determine where best to place
the synthesis locations.

• Related to the above point, it has been shown in [38] that for the radial trajec-
tory, filling in unacquired radial lines can lead to the option for interpolating
the linear coefficient weight values which can drastically reduce the computa-
tion burden. While such an approach can be used with linear combination
weights generated using APPEAR, the effect of interpolating weight values on
the error in the estimated encoding functions has not been systematically stud-
ied. In addition, it is not clear how well such an approach can be used with
variable-density spiral trajectories or the 3-D cones trajectory.

• Partial k-space encoding is another technique to reduce the encoding time nec-
essary to acquire enough data for an image. Future work is needed to determine
how partial k-space encoding can be combined with the APPEAR method.

• Thus far, APPEAR has been implemented for 2-D non-Cartesian trajectories,
but has not yet been implemented for 3-D non-Cartesian trajectories. 3-D scans
stand the most to gain from reduced encoding time and such an implementation
is important for future work.

• While much of the theory behind APPEAR has been developed and APPEAR
has been implemented for 2-D non-Cartesian trajectories, only limited in vivo
results have been obtained. The next step in evaluating APPEAR for clinical
use is to use it with clinically relevant scans and evaluate its performance.
142 CHAPTER 6. SUMMARY AND FUTURE RECOMMENDATIONS
Appendix A

Aliasing Amplitude of a Sampled

Kernel

When the kernel is sampled and interpolated in k-space, its inverse Fourier transform
in image space can be written as a periodic term, cs (x), of period SG modulated by
an envelope function, where the envelope function is the inverse Fourier transform of
the interpolation function. This is shown in Fig. A.1. The periodicity of cs (x) allows
us to write the infinite sum in Eq. 3.3 as a finite sum and thus quickly calculate the
aliasing amplitude for a sampled kernel.

Starting from Eq. 3.3, I derive the aliasing amplitude for a sampled interpolated
kernel, c(i) = cs (i)h(i), as follows:
s
1 X
ε[i] = c(i + Gp)2
c(i)2 p6=0
v
uX  2
u c(i + Gp)
=t − 1.
p
c(i)

I now group the terms of c(i) separated by SG; the periodicity of cs (x) allows us to

143
144 APPENDIX A. ALIASING AMPLITUDE

Figure A.1: The sampled and interpolated kernel, b c(x), can be viewed as the periodic
function cs (x) = c(x) ∗ III (SGx), with period SG, modulated by the envelope function
h(x), which is a sinc function for nearest-neighbor interpolation and a sinc-squared
function for linear interpolation.

simplify c(i + SG) = cs (i + SG)h(i + SG) = cs (i)h(i + SG).

v #2
u S−1 "
uX b h(i + Gp)c s (i + Gp)
ε[i] = t −1 (A.1)
p=0
h(i)c s (i)
sX
where b
h(i) = [h(i + SGq)]2 .
q

Computing b
h(x) is straightforward, realizing that
s   
2 x−i
h(i) =
b FT [h(x)] III (0). (A.2)
SG

Noting that the p = 0 term in Eq. A.1 is independent of cs (x), one can separate ε(i)
into the quadrature sum of two terms:
q
ε[i] = (ε1 [i])2 + (ε2 [i])2 , (A.3)
A.1. NEAREST-NEIGHBOR INTERPOLATION 145

where
v" #2
u
u bh(i)
ε1 [i] = t −1 (A.4)
h(i)
v #2
u S−1 "
uX b h(i + Gp)c s (i + Gp)
ε2 [i] = t . (A.5)
p=1
h(i)c s (i)

A.1 Nearest-Neighbor Interpolation

For nearest-neighbor interpolation, h(x) = sinc (x/SG). Substituting this into Eq.
A.2 and A.4, b
h(i) = 1 and
s
1
ε1 [i] = 2 − 1.
sinc (i/SG)

Since i  SG, one can approximate 1/ sinc2 (i/SG) − 1 as π 2 i2 /3S 2 G2 , using the first
nonzero term of the Taylor series about i = 0. Thus

π|i|
nearest-neighbor ε1 [i] ≈ √ . (A.6)
3SG

When i = −N/2, ε1 ≈ 0.91/αS as in Eq. 3.7.

A.2 Linear Interpolation

= sinc2 (x/SG). Once again, I substitute this into Eq.

For linear interpolation, h(x) q
h(i) = 2 + 1 cos 2π i and


A.2 and A.4 to obtain b 3 3 SG

s
2
+ 13 cos 2π SG
i


3
ε1 [i] = − 1.
sinc4 (i/SG)
146 APPENDIX A. ALIASING AMPLITUDE

Figure A.2: Although Eqs. A.6 and A.7 are approximations for ε1 [i], they are very
accurate approximations. In this example, N = 256, α = 1.25 and S = 60. (a) shows
the agreement for nearest-neighbor interpolation. Eq. A.6 is within 44 parts-per-
million of ε1 [i] in this case. (b) shows the agreement for linear interpolation. Eq. A.7
is within 209 parts-per-million of ε1 [i] in this case.

Since i  SG, one can approximate ε1 (i) using the first nonzero term of the Taylor
series about i = 0, giving
2
π2

i
linear ε1 [i] ≈ √ . (A.7)
3 5 SG

When i = −N/2, ε1 ≈ 0.37/(αS)2 as in Eq. 3.8.

Equations A.6 and A.7 provide simple expressions for ε1 [i], the part of the aliasing
amplitude due to nearest-neighbor and linear interpolation respectively. Figure A.2
demonstrates that these approximations are also highly accurate. Equation A.5 gives
a finite sum solution for ε2 [i], the part of the aliasing amplitude due to the kernel
sample values. These simple and accurate expressions for the aliasing amplitude
allow us to quickly evaluate the expected performance of a given gridding kernel and
to generate comparisons of kernels as in Fig. 3.4.
Appendix B

Apodization Correction Function

of a Sampled Kernel

With a continuous kernel, obtaining the exact apodization correction function requires
one to have an analytic solution for the inverse Fourier transform of the kernel. With
the sampled kernel one can use the FFT to obtain the apodization correction function
as follows:

1. Zero-pad the sampled kernel to length SG.

2. Compute the IFFT.

3. Select the N values from x = −N/2 to N/2 − 1.

4. Multiply by the Inverse Fourier transform of the interpolation function, h(x).

5. Take the reciprocal of these N values.

147
148 APPENDIX B. APODIZATION CORRECTION
Appendix C

Calculating the Optimal Sampled

Kernel in MATLAB

The SOCP in section 3.5 finds a kernel, Cs (kx ), such that ε2 [i] < εmax
2 . Although
this SOCP is easy to understand, it is not well suited for implementation. As such,
I implement the following SOCP which, though slightly more complicated, gives the
same result and is better suited for implementation:

minimize t
v
u S−1 h i2
uX
subject to t bh(i + Gp)cs (i + Gp) ≤ εmax
2 h(i)cs (i) + t
p=1

Cs (0) = 1
Cs (−kx ) = Cs (kx )
 
N
for i ∈ 0, 1, . . . , .
2

In this SOCP, when t is less than zero, a kernel exists such that ε2 [i] < εmax
2 .
Many optimization packages exist which will solve second-order cone programs.
Lobo et al. have written a package specifically for SOCP problems [45]. In addition,
the SeDuMi package is a freely available package that will solve many optimization
problems, including SOCP problems [46] [47]. Commercial packages, such as MOSEK

149
150 APPENDIX C. CALCULATING THE OPTIMAL SAMPLED KERNEL

(www.mosek.com), are also available. Here, I provide a MATLAB program which

implements the SOCP shown above, using the yalmip MATLAB toolbox and SeDuMi
[48]. N , G, W and S are as defined in Tables 4.1 and 3.2.

% f u n c t i o n Cs = f i n d K e r n e l (N, G, W, S , e2max )
% Find sampled k e r n e l , Cs ( kx ) , such t h a t
% \ e p s i l o n 2 [ i ] < e2max .
% I f no such k e r n e l can be found ,
% Cs i s s e t t o 0 .
f u n c t i o n Cs = f i n d K e r n e l (N, G, W, S , e2max )
F = set ( ’ ’ );
t = sdpvar ( 1 , 1 ) ;
Cs = sdpvar (W∗S /2 −1 ,1);
kx = [ 1 :W∗S/2 −1] / S / G;
p = [ 1 : S−1] ’;
f o r i = 0 : (N/ 2 ) ,
h h a t = s q r t ( 2/3 + 1/3 ∗ . . .
c o s ( 2 ∗ p i /S/G ∗ ( i+G∗p ) ) ) ;
h = s i n c ( i /S/G) . ˆ 2 ;
% Use t h e d i s c r e t e c o s i n e t r a n s f o r m
% t o g e t c s ( x ) from Cs ( kx )
c A l i a s = 2 ∗ c o s ( 2 ∗ p i ∗ ( i+G∗p ) ∗ kx ) ∗ . . .
Cs + 1 ;
cApodize = 2 ∗ c o s ( 2 ∗ p i ∗ i ∗ kx ) ∗ Cs + 1 ;
F = F + s e t ( cone ( h h a t . ∗ c A l i a s , . . .
e2max ∗ h ∗ cApodize + t ) ) ;
end
s o l v e s d p (F , t ) ;
i f ( d o u b l e ( t ) <= 0 )
Cs = [ f l i p u d ( d o u b l e ( Cs ) ) ; 1 ; d o u b l e ( Cs ) ] ;
else
Cs = 0 ;
end
Appendix D

Noise in APPEAR Reconstructions

The noise from each receiver coil is modeled as a random process with Gaussian
statistics and zero mean. The variance and cross-correlation of the noise between
the receiver coils is given by the Nc × Nc receiver noise matrix Ψ, as defined by
Pruessmann et al. [19]. Denoting the noise random process for coil j as nj (t), the
data acquired at location (j, k) can be written as dj (k) = eTj (k)m + nj (tk ) and the
expected value of n∗j (tk )nj 0 (tk0 ) is given by
(
Ψjj 0 if tk = tk0
E n∗j (tk )nj 0 (tk0 ) =
 
. (D.1)
0 otherwise
The noise variance for the voxel location r in the reconstructed image for coil j can
be expressed as
 
2
X
σj2 (r) = E  exp(i2πkT r)wj,L(k)
T
nL(k)  , (D.2)
k∈G

where G is the set of k-space locations forming a full-FOV grid and nL(k) is a vector
containing the noise values of the data acquired in the local neighborhood of k-space
location k. After some manipulation, σj2 (r) can be written as
X
σj2 (r) = exp(i2π∆kT r)φj (∆k), (D.3)
∆k∈G

151
152 APPENDIX D. NOISE IN APPEAR RECONSTRUCTIONS

where
X X X
∗ 0 0
φj (∆k) = Ψj 0 j 00 wj,j 0 (k, k )wj,j 00 (k − ∆k, k ). (D.4)
j 0 ,j 00 k∈G k0 ∈I(k,∆k)

The k-space set I(k, ∆k) = {k0 |(j, k) ∈ L(k) ∩ L(k − ∆k), k − ∆k ∈ G} is the in-
tersection set between two local neighborhoods, one centered at k and the other
centered at k − ∆k, given that both centers lie on the grid. Equation D.3 can be
viewed as the inverse discrete Fourier transform of φj (∆k). Since the set I(k, ∆k)
is null when k∆kk > 2κ, twice the radius of the local neighborhood used for data
synthesis, φj (∆k) = 0 for k∆kk > 2κ. Thus, σj2 (r) is a low frequency function of r.
When the noise from the coils is not correlated, the average noise variance in a voxel
is given by
X X 2
φj (0) = Ψj 0 j 0 |wj,j 0 (k, k0 )| . (D.5)
k∈G (j 0 ,k0 )∈L(k)

Thus, large weights will increase the noise in the image. While regularization can
be used to keep the weights from getting too large, this can result in weights that do
not sufficiently suppress the aliasing artifacts.
Bibliography

[1] J. O’Sullivan, “A fast sinc function gridding algorithm for Fourier inversion in
computer tomography,” Medical Imaging, IEEE Transactions on, vol. MI-4, pp.
200–207, 1985.

[2] H. Schomberg and J. Timmer, “The gridding method for image reconstruction
by Fourier transformation,” Medical Imaging, IEEE Transactions on, vol. 14,
no. 3, pp. 596–607, 1995.

[3] S. Schaller, T. Flohr, and P. Steffen, “An efficient Fourier method for 3-D radon
inversion in exact cone-beam CT reconstruction,” Medical Imaging, IEEE Trans-
actions on, vol. 17, no. 2, pp. 244–250, 1998.

[4] M. Bronstein, A. Bronstein, M. Zibulevsky, and H. Azhari, “Reconstruction in

diffraction ultrasound tomography using nonuniform FFT,” Medical Imaging,
IEEE Transactions on, vol. 21, no. 11, pp. 1395–1401, 2002.

[5] A. Dutt and V. Rokhlin, “Fast Fourier transforms for nonequispaced data,”
SIAM J. Sci Comput., vol. 14, no. 6, pp. 1369–1393, 1993.

[6] J. G. Pipe and P. Menon, “Sampling density compensation in MRI: rationale and
an iterative numerical solution,” Magn Reson Med, vol. 41, no. 1, pp. 179–86,
1999.

[7] V. Rasche, R. Proksa, R. Sinkus, P. Bornert, and H. Eggers, “Resampling of

data between arbitrary grids using convolution interpolation,” Medical Imaging,
IEEE Transactions on, vol. 18, no. 5, pp. 385–92, 1999.

153
154 BIBLIOGRAPHY

[8] D. Rosenfeld, “New approach to gridding using regularization and estimation

theory,” Magn. Reson. Med., vol. 48, no. 1, pp. 193–202, 2002.

[9] J. Fessler and B. Sutton, “Nonuniform fast Fourier transforms using min-max
interpolation,” Signal Processing, IEEE Transactions on [see also Acoustics,
Speech, and Signal Processing, IEEE Transactions on], vol. 51, no. 2, pp. 560–
574, 2003.

[10] H. Moriguchi and J. L. Duerk, “Iterative next-neighbor regridding (INNG): im-

proved reconstruction from nonuniformly sampled k-space data using rescaled
matrices,” Magn Reson Med, vol. 51, no. 2, pp. 343–52, 2004.

[11] J. Jackson, C. Meyer, D. Nishimura, and A. Macovski, “Selection of a convo-

lution function for Fourier inversion using gridding,” Medical Imaging, IEEE
Transactions on, vol. 10, no. 3, pp. 473–478, 1991.

[12] L. Greengard and J.-Y. Lee, “Accelerating the nonuniform fast fourier trans-
form,” SIAM Rev., vol. 46, pp. 443–454, 2004.

[13] M. Frigo and S. G. Johnson, The FFTW web page, 2004. [Online]. Available:
http://www.fftw.org/

[14] P. Irarrazabal and D. G. Nishimura, “Fast three dimensional magnetic resonance

imaging,” Magn Reson Med, vol. 33, no. 5, pp. 656–62, 1995.

[15] F. Wajer, E. Woudenberg, R. de Beer, M. Fuderer, A. Mehlkopf, and D. van

Ormondt, “Simple equation for optimal gridding parameters,” in Proceedings of
the 7th Annual Meeting of ISMRM, Philadelphia, 1999, p. 663.

[16] M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, “Applications of second-

order cone programming,” Linear Algebra and its Applications, vol. 284, no. 1-3,
pp. 193–228, 1998.

[17] M. A. Griswold, P. M. Jakob, M. Nittka, J. W. Goldfarb, and A. Haase, “Partially

parallel imaging with localized sensitivities (PILS),” Magn. Reson. Med., vol. 44,
no. 4, pp. 602–9, 2000.
BIBLIOGRAPHY 155

[18] D. K. Sodickson and W. J. Manning, “Simultaneous acquisition of spatial har-

monics (SMASH): fast imaging with radiofrequency coil arrays,” Magn. Reson.
Med., vol. 38, no. 4, pp. 591–603, 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,
no. 5, pp. 952–62, 1999.

[20] K. P. Pruessmann, M. Weiger, P. Bornert, and P. Boesiger, “Advances in sensi-

tivity encoding with arbitrary k-space trajectories,” Magn. Reson. Med., vol. 46,
no. 4, pp. 638–51, 2001.

[21] W. E. Kyriakos, L. P. Panych, D. F. Kacher, C. F. Westin, S. M. Bao, R. V.

Mulkern, and F. A. Jolesz, “Sensitivity profiles from an array of coils for encoding
and reconstruction in parallel (SPACE RIP),” Magn. Reson. Med., vol. 44, no. 2,
pp. 301–8, 2000.

[22] E. N. Yeh, C. A. McKenzie, M. A. Ohliger, and D. K. Sodickson, “Parallel

magnetic resonance imaging with adaptive radius in k-space (PARS): constrained
image reconstruction using k-space locality in radiofrequency coil encoded data,”
Magn. Reson. Med., vol. 53, no. 6, pp. 1383–92, 2005.

[23] C. A. McKenzie, E. N. Yeh, M. A. Ohliger, M. D. Price, and D. K. Sodick-

son, “Self-calibrating parallel imaging with automatic coil sensitivity extraction,”
Magn. Reson. Med., vol. 47, no. 3, pp. 529–38, 2002.

[24] M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang,

B. Kiefer, and A. Haase, “Generalized autocalibrating partially parallel acquisi-
tions (GRAPPA),” Magn Reson Med, vol. 47, no. 6, pp. 1202–10, 2002.

[25] Z. Wang, J. Wang, and J. A. Detre, “Improved data reconstruction method for
GRAPPA,” Magn. Reson. Med., vol. 54, no. 3, pp. 738–742, 2005.

[26] P. M. Jakob, M. A. Griswold, R. R. Edelman, and D. K. Sodickson, “AUTO-

SMASH: a self-calibrating technique for SMASH imaging. SiMultaneous Acqui-
sition of Spatial Harmonics,” Magma, vol. 7, no. 1, pp. 42–54, 1998.
156 BIBLIOGRAPHY

[27] R. M. Heidemann, M. A. Griswold, A. Haase, and P. M. Jakob, “VD-AUTO-

SMASH imaging,” Magn. Reson. Med., vol. 45, no. 6, pp. 1066–74, 2001.

[28] M. A. Griswold, R. M. Heidemann, and P. M. Jakob, “Direct parallel imaging

reconstruction of radially sampled data using GRAPPA with relative shifts,” in
Proc. Eleventh ISMRM, 2003, p. 2349.

[29] K. A. Heberlein, Y. Kadah, and X. Hu, “Segmented spiral parallel imaging using
GRAPPA,” in Proc. Twelfth ISMRM, 2004, p. 328.

[30] R. M. Heidemann, M. A. Griswold, G. Krger, S. Kannengiesser, B. Kiefer, and

P. M. Jakob, “Fast parallel image reconstructions for spiral trajectories,” in
Second International Workshop on Parallel MRI, 2004.

[31] K. A. Heberlein and X. Hu, “Auto-calibrated parallel imaging using dual-density

spirals,” in Second International Workshop on Parallel MRI, 2004.

[32] K. Heberlein and X. Hu, “Auto-calibrated parallel spiral imaging,” Magn. Reson.
Med., vol. 55, no. 3, pp. 619–625, 2006.

[33] A. Arunachalam, A. Lu, E. K. Brodsky, and W. F. Block, “GRAPPA for the 3D

radial trajectory (VIPR),” in Proc. Thirteenth ISMRM, 2005, p. 2674.

[34] C. M. Tsai and D. G. Nishimura, “Reduced aliasing artifacts using variable-

density k-space sampling trajectories,” Magn. Reson. Med., vol. 43, no. 3, pp.
452–8, 2000.

[35] P. Beatty, D. Nishimura, and J. Pauly, “Rapid gridding reconstruction with a

minimal oversampling ratio,” IEEE Trans. Med. Imag., vol. 24, no. 6, pp. 799–
808, 2005.

[36] M. A. Griswold, S. Kannengiesser, R. M. Heidemann, J. Wang, and P. M. Jakob,

“Field-of-view limitations in parallel imaging,” Magn. Reson. Med., vol. 52, no. 5,
pp. 1118–26, 2004.
BIBLIOGRAPHY 157

[37] C. A. McKenzie, M. A. Ohliger, E. N. Yeh, M. D. Price, and D. K. Sodickson,

“Coil-by-coil image reconstruction with SMASH,” Magn. Reson. Med., vol. 46,
no. 3, pp. 619–23, 2001.

[38] A. A. Samsonov, W. F. Block, A. Arunachalam, and A. S. Field, “Advances

in locally constrained k-space-based parallel MRI,” Magn. Reson. Med., vol. 55,
no. 2, pp. 431–438, 2006.

[39] P. T. Gurney, B. A. Hargreaves, and D. G. Nishimura, “Design and analysis of a

practical 3D cones trajectory,” Magn. Reson. Med., vol. 55, no. 3, pp. 575–582,
2006.

[40] A. Macovski, “Noise in MRI,” Magn. Reson. Med., vol. 36, no. 3, pp. 494–7,
1996.

[41] P. B. Roemer, W. A. Edelstein, C. E. Hayes, S. P. Souza, and O. M. Mueller,

“The NMR phased array,” Magn. Reson. Med., vol. 16, no. 2, pp. 192–225, 1990.

[42] P. Qu, C. Wang, and G. X. Shen, “Discrepancy-based adaptive regularization for

GRAPPA reconstruction,” J Magn Reson Imaging, vol. 24, no. 1, pp. 248–55,
2006.

[43] H. Sedarat and D. G. Nishimura, “On the optimality of the gridding reconstruc-
tion algorithm,” Medical Imaging, IEEE Transactions on, vol. 19, no. 4, pp.
306–17, 2000.

[44] K. A. Heberlein and X. Hu, “Kriging and GRAPPA: A new perspective on

parallel imaging reconstruction,” in Proc. Fourteenth ISMRM, 2006, p. 2465.

[45] M. S. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret,

Second-Order Cone Programming (SOCP), 1998. [Online]. Available:
http://www.stanford.edu/˜boyd/SOCP.html

[46] J. F. Sturm, “Using SeDuMi 1.02, a MATLAB toolbox for optimization over
symmetric cones,” Optimization Methods and Software, no. 11-12, pp. 625–653,
1999.
158 BIBLIOGRAPHY

[47] ——, SeDuMi 1.05, 2003. [Online]. Available:

http://fewcal.kub.nl/sturm/software/sedumi.html

[48] J. Löfberg, YALMIP 3, 2004. [Online]. Available:

http://control.ee.ethz.ch/˜joloef/yalmip.msql