Download dynamic contrast-enhanced breast mri - Stacks

DYNAMIC CONTRAST-ENHANCED BREAST MRI
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
Misung Han
August 2010
© 2010 by Misung Han. All Rights Reserved.
Re-distributed by Stanford University under license with the author.
This work is licensed under a Creative Commons AttributionNoncommercial 3.0 United States License.
http://creativecommons.org/licenses/by-nc/3.0/us/
This dissertation is online at: http://purl.stanford.edu/tw005pr0680
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.
Brian Hargreaves, Primary 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.
Bruce Daniel
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
Approved for the Stanford University Committee on Graduate Studies.
Patricia J. Gumport, Vice Provost Graduate Education
This signature page was generated electronically upon submission of this dissertation in
electronic format. An original signed hard copy of the signature page is on file in
University Archives.
iii
Abstract
T1 -weighted dynamic contrast-enhanced (DCE) breast MRI has shown high sensitivity in
detecting tumors, but has suffered from low specificity. Accuracy of tumor detection and
characterization is enhanced by high-quality images with high temporal resolution. This
thesis presents imaging techniques for improving DCE breast MRI in terms of image quality and imaging speed.
Bilateral breast imaging is more cost-effective than unilateral breast imaging and can
detect bilateral cancers. However, imaging the two breasts increase scan time due to increased volume. Independent slab-phase modulation of the two breast slabs can reduce
scan time by removing the need to image unnecessary space between them. Parallel imaging further accelerates data acquisition while maintaining spatial resolution. However, parallel imaging can result in noise amplification and residual aliasing artifacts. By combining slab-phase modulation with parallel imaging, reconstructed image quality can be improved compared parallel imaging alone. Phantom and in vivo experiments are presented
to demonstrate this.
By using parallel imaging in bilateral DCE imaging, temporal resolution is increased,
allowing better characterization of tumor contrast uptake curves. We used a temporal acceleration scheme from time-interleaved acquisitions in combination with slab-phase modulation and 3D spiral imaging. The details of TSENSE reconstruction are presented, with
showing patient study results.
iv
Fat suppression is important because unsuppressed fatty tissue can yield a signal intensity similar to that of enhanced tumors. In bilateral imaging, fat suppression can be
improved by independently shimming each breast. We combined independent shims with
dual-band spectral-spatial excitation and found that independent shims provided better fat
suppression than standard shims through in vivo experiments.
RF-spoiled gradient-echo sequences are normally used for DCE imaging in order to
provide pure T1 contrast. However, unsaturated flowing spins can result in inflow enhancement and pulsatile ghost artifacts, which degrade DCE data quantification. By applying
partial saturation over the slab upstream, flowing spins can be driven to the steady state before entering the imaging slab without generating net signal. We demonstrated that partial
saturation can reduce flow artifacts and provide T1 contrast along the imaging slab through
simulations, and phantom and in vivo experiments.
To conclude, the research presented in this thesis has contributed to improving temporal resolution, fat suppression, and T1 quantification in DCE breast MRI, increasing the
potential for more accurate diagnosis.
v
Acknowledgments
I have been so fortunate to have a Ph.D research area that fits my interests, which has made
my life at Stanford very rewarding and enjoyable. This thesis summarizes five years of my
Ph.D. research activities, none of which would be possible without the support and help
from many colleagues, friends, and family.
Brian Hargreaves is a great advisor both academically and personally. He is very patient, supportive, and encouraging. He has taught me everything necessary for Ph.D. research including MRI physics, pulse sequence programming, writing papers, and presenting research results. I am extremely grateful to him not only for his supervision of all this
work, but also for his personal care. He has always emphasized my own learning process
over the immediate outcome. In addition, I have always been encouraged by his ceaseless
enthusiasm in the MRI field, great passion for his own research, and continuous interest in
new areas.
Dwight Nishimura and John Pauly are the directors of the MRSRL group in Electrical
Engineering. Their classes made me deeply attracted to the MRI field together with my
interest in medical sciences, and has been a tremendous resource for my PhD research.
John Pauly has been a very attentive co-advisor and I was always amazed by his insightful
comments on my work.
vi
I thank Bruce Daniel, who has taught me the clinical side of breast MRI and has given
me critical comments on my research. His extensive intuition in MRI physics and innovative ideas about improving breast MRI have always motivated me. I feel so fortunate to
work with such a great clinician.
I am also thankful to Howard Zebker for being my oral defense chair and helpful feedback.
When I first started my research in the Body MR (BMR) group, I was the only graduate
student for the first few months. But within five years, it has grown, consisting of many
more graduate students, postdoctoral scholars, and research associates. I have been greatly
benefited by collaboration and interaction among the BMR members, academically and
socially. Wenmiao Lu and Ernesto Staroswiecki joined the group within a few months
after me and we shared the earliest stage of BMR together. Kristin Granlund, the second
BMR lady joined afterwards, made me feel very comfortable in BMR, and has helped a
lot, particularly with revising papers and improving oral presentations. Younger students,
Anderson Nnewihe, Caroline Jordan, and Hochin Kim have kept me thinking outside of
my projects. The BMR group has had many amazing advanced researchers who really
care graduate students. Neal Bangerter and Anthongy Faranesh were great mentors when
I first started. Marcus Alley has helped me a lot with MR data reconstruction and paper
writing. Seungbum Kim used to be a research associate as a mechanical engineer, and I
was always impressed by his various interests and passion for research. Kyung Sung and
Pauline Worters came to the group as postdoctoral scholars one and a half to two years ago.
As they are very generous and enthusiastic, I have enjoyed discussing research problems
with them. Monoj Sanadaram, and Catherine Moran are the most newest to the group but
they have made BMR more energetic with their experience and expertise.
As a member of the Lucas group, I have received a lot of supports from many people as
well. Anne Sawyer taught me about the use of scanners and various coils, and even came
vii
to the lab at midnight when I had problems with the scanners. Tom Brosnan rescued me
many times when I had problems with my complicated dual-boot computer. I thank Donna
Cronister and Marleys LeSene for all their administration work. Chunlei Lui and Calvin
Lew had much to teach me about parallel imaging applications. Jason answered various
questions ranging from MR physics to perl scripts. Sunmi Kim, Daniel Kopeinigg, Matus
Straka, and Jarrett Rosenberg were also very helpful in various projects. Jonduk Baik has
been a good friend to be able to share all aspects of life. I also thank Rebecca RakowPenner, Sonal Josan, Priti Balchandani, Sam Mazin, Lena Kaye, Sungwon Yoon, Catie
Chang, Rachelle Bitton, Scott Hsieh, Sandra Rodriguez, John Mendoza, Jared Starman,
Tao Xu, Christine Law, Jian Zhang, Jae Mo Park, and Eun Soo Choi, and all the others,
who have made my Lucas life more fun and relaxing. I would not forget all our outside
activities as well.
Outside of the Lucas Center, I am thankful to Sandy Napel, with whom I first started research at Stanford. I learned how to proceed in research independently from him and I was
exposed to various medical imaging areas when working in his group. I am also grateful
to many of MRSRL people for organizing a study group and learning EPIC programming
together. I especially thank Jongho Lee, Daeho Lee, Joelle Barral, and Emine Saritas for
their friendship and encouragement.
I am also thankful to people at GE ASL-West. I especially appreciated the chance to
do a summer internship in 2009, through which I learned more about pulse sequence and
scanner programming. I worked closely with Ajit Shankranarayan and Eric Han, and I was
really impressed by the number of projects which they were involved with concurrently.
Philip Beatty coauthored my paper, and gave me a lot of insights about parallel imaging
techniques.
Besides many colleagues, I have met so many precious people at Stanford, making my
life more meaningful and enjoyable. I was one of starting members of the Korean Christian
viii
Fellowship at Stanford (KCF), and have been involved with this fellowship for the last
seven years. KCF has been a great source of support at every stage of my graduate life.
I especially thank Meeyoung Park, Hyunjin Kim, Hyunok Lee, Youngsun Park, Inseoung
Kim, Jeunghoon Noh, and Jungwoo Lee, who were great supporters during the stressful
time of the first few years at Stanford. Pastor Don Kim and Maria SMN have taken care
of me as their own daughter. After marriage, gathering and sharing activities with other
newly-married couples has been extremely fun. Heehyun Jung, Giyoung Park, Youngsu
Choi, Jin Yoo, and Janghwan Choi, Dongjun Shin, and Hyunjee Kim, Namkeun Kim, and
Sookyung Kim become new but good friends of mine .
Thankfully, I have many long-time female friends who are engineers or scientists. I
thank them for their keeping encouraging me and sharing many thoughts about life balance.
In particular, Tammy Kim, Bomi Nam, and I attended the same high school and college,
and came to the States within one after another within one year, and we are still hanging
out frequently in the Bay area. I am also thankful to Kyeoungyeon Doo, who used to be
a long-time partner in the EE department over college years. We have been missing each
other so much since I came to the States.
Finally, I want to express my sincere thanks to my family. My parents have always
been supportive of my career with tremendous love and encouragement. My father has
been working very hard with great passion as an engineer. He has always been and will
continue to be a role model for me. My mother is my best friend. Especially after I came to
the States, I have talked to her about everything including study, research, career plans, and
relationships. She have understood me instantly and deeply, and always suggested the right
things for me. My sister is only one year younger than I am, and we grew up as friends
and competitors. I am really sorry for not being able to meet her more often. I wish her
the best of luck in her current medical residency training. My brother, the youngest in my
family, has always enjoyed teasing his sisters, but I know he cares for us. My husband,
ix
Hyunmin Park, has given me wonderful support and care ever since we got married. I was
very fortunate to meet him, and to find that he had a very similar personality to mine. As
he had already gone through similar career steps, he has been extremely understanding and
patient, and has given me valuable advice during my last Ph.D period. I dedicate this thesis
to all of my family.
x
Contents
Abstract
v
Acknowledgments
vii
1 Introduction
1
2 Basics of Magnetic Resonance Imaging
5
2.1
2.2
Physics of Nuclear Magnetic Resonance . . . . . . . . . . . . . . . . . . .
5
2.1.1
Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
2.1.2
Precession . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.1.3
Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.1.4
Bloch Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
Magnetic Fields for MR Imaging . . . . . . . . . . . . . . . . . . . . . . .
9
2.2.1
Static Magnetic Field B0 . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.2
RF Field B1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.3
Linear Gradient Fields G . . . . . . . . . . . . . . . . . . . . . . . 11
xi
2.3
2.4
Imaging Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1
Selective Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2
Spatial Encoding and Data Acquisition . . . . . . . . . . . . . . . 13
2.3.3
Sampling Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Parallel Imaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4.1
SENSE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4.2
GRAPPA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Breast MRI
28
3.1
Breast Cancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2
Breast MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3
Dynamic Contrast-Enhanced MRI . . . . . . . . . . . . . . . . . . . . . . 31
4 Parallel Imaging with Slab-Phase Modulation
34
4.1
Dual-Band Excitation in Bilateral Breast MRI . . . . . . . . . . . . . . . . 35
4.2
Parallel Imaging Applications . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3
Phantom Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4
4.3.1
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.2
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
In Vivo Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4.1
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4.2
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
xii
4.5
4.6
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.5.1
Improvement of Reconstructed Image Quality . . . . . . . . . . . . 54
4.5.2
Experimental Data Analysis . . . . . . . . . . . . . . . . . . . . . 55
4.5.3
Image Acquisition Considerations . . . . . . . . . . . . . . . . . . 56
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Rapid Bilateral Breast DCE-MRI
58
5.1
Breast MRI at Stanford Hospital . . . . . . . . . . . . . . . . . . . . . . . 59
5.2
DCE Imaging Pulse Sequence . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3
Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.1
Unaccelerated Reconstruction . . . . . . . . . . . . . . . . . . . . 62
5.3.2
TSENSE Reconstruction . . . . . . . . . . . . . . . . . . . . . . . 63
5.4
Patient Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.5
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5.1
Self-Calibrated Sensitivity Estimation and Residual Aliasing Artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5.2
SENSE Regularization . . . . . . . . . . . . . . . . . . . . . . . . 76
5.6
Comparison between TSENSE and TGRAPPA . . . . . . . . . . . . . . . 77
5.7
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6 Fat Suppression with Independent Shims
82
6.1
Field Inhomogeneity in Breast MRI . . . . . . . . . . . . . . . . . . . . . 83
6.2
Independent Shims with Dual-Band Spectral-Spatial Excitation . . . . . . . 85
xiii
6.3
In Vivo Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
7 Flow Artifact Reduction Using Partial Saturation
92
7.1
RF-Spoiled Gradient-Echo Sequences . . . . . . . . . . . . . . . . . . . . 93
7.2
Partial Saturation of Flowing Spins . . . . . . . . . . . . . . . . . . . . . . 95
7.3
Magnetization Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
7.4
Scan Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.5
7.6
7.4.1
Dephasing of Partial Saturation Slab . . . . . . . . . . . . . . . . . 106
7.4.2
Flow Phantom Experiments . . . . . . . . . . . . . . . . . . . . . 106
7.4.3
In Vivo Experiments . . . . . . . . . . . . . . . . . . . . . . . . . 108
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
7.5.1
Partial Saturation for AIF measurements . . . . . . . . . . . . . . . 110
7.5.2
Signal Uniformity and T1 Contrast by Partial Saturation . . . . . . . 113
7.5.3
Partial Saturation Slab Thickness . . . . . . . . . . . . . . . . . . 114
7.5.4
Variations in RF Pulse . . . . . . . . . . . . . . . . . . . . . . . . 114
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
8 Summary and Recommendations
117
8.1
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
8.2
Recommendations for Future Work . . . . . . . . . . . . . . . . . . . . . . 118
Bibliography
123
xiv
List of Tables
4.1
Noise Coupling Coefficients for Breast Phased-Array Coil . . . . . . . . . 42
4.2
Linear Regression Model for SNR Comparison . . . . . . . . . . . . . . . 55
6.1
Optimal Center Frequency and Linear Shim Values . . . . . . . . . . . . . 88
xv
List of Figures
2.1
Polarization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
2.2
Precession . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
2.3
Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8
2.4
Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.5
Linear Gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6
Selective Excitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7
k-Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.8
Phase-Encode Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.9
k-Space Trajectory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.10 k-Space Interval and Field-of-View . . . . . . . . . . . . . . . . . . . . . . 18
2.11 Phased-Array Coils . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.12 Spatial Aliasing due to Undersampled k-Space Acquisition . . . . . . . . . 21
2.13 SENSE Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.14 g-Factor Calculation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.15 Coil Sensitivity Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
xvi
2.16 GRPPA Kernel Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.1
Breast Anatomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2
X-Ray Mammography and Breast MRI . . . . . . . . . . . . . . . . . . . 30
3.3
Breast MR Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.4
Signal Enhancement Curve from Tumor . . . . . . . . . . . . . . . . . . . 32
4.1
Standard Excitation and Dual-Band Excitation . . . . . . . . . . . . . . . . 35
4.2
Dual-Band Spectral-Spatial Excitation . . . . . . . . . . . . . . . . . . . . 36
4.3
Independent Slab-Phase Modulation . . . . . . . . . . . . . . . . . . . . . 37
4.4
Undersampling with and without ISPM . . . . . . . . . . . . . . . . . . . 39
4.5
Self-Calibrated Parallel Imaging . . . . . . . . . . . . . . . . . . . . . . . 41
4.6
Phase Encoding over Two Breast Slabs . . . . . . . . . . . . . . . . . . . . 43
4.7
Phantom Experiments: Reconstruction Artifacts . . . . . . . . . . . . . . . 46
4.8
Phantom Experiments: g-Factor Maps . . . . . . . . . . . . . . . . . . . . 48
4.9
g-Factor Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.10 In Vivo Experiments: Reconstructed Images . . . . . . . . . . . . . . . . . 52
4.11 In Vivo Experiments: Reconstruction Artifacts . . . . . . . . . . . . . . . . 53
4.12 SNR Efficiency Comparison . . . . . . . . . . . . . . . . . . . . . . . . . 54
5.1
Bilateral DCE Breast MRI Protocol . . . . . . . . . . . . . . . . . . . . . 60
5.2
TSENSE Acquisition Order for 3D Spiral Imaging . . . . . . . . . . . . . 62
5.3
Image Reconstruction from Time-Interleaved Acquisition . . . . . . . . . . 64
xvii
5.4
Sensitivity Map Estimation from DCE Imaging . . . . . . . . . . . . . . . 65
5.5
Measurement of Oscillation Using a Frequency Spectrum . . . . . . . . . . 66
5.6
Patient Breast Images from unaccelerated and TSENSE Reconstructions . . 69
5.7
Signal Enhancement Curves from Using Different Sensitivity Maps . . . . 70
5.8
Quantitative Analysis of Residual Aliasing Artifacts . . . . . . . . . . . . . 72
5.9
Signal Enhancement Curves from Tumors . . . . . . . . . . . . . . . . . . 73
5.10 Dynamic and High-Resolution Breast Images . . . . . . . . . . . . . . . . 74
5.11 Comparison between Unaccelerated, TSENSE, and TGRAPPA Images . . . 79
5.12 Comparison between TSENSE and TGRAPPA . . . . . . . . . . . . . . . 80
6.1
Failure in Fat Suppression . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2
Field Maps of Breasts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.3
Magnet Shim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.4
Independent Shims with Dual-band Spectral-Spatial Excitation . . . . . . . 87
6.5
Fat-Suppressed Breast Images from Different Shims . . . . . . . . . . . . . 89
6.6
Fat-Suppressed Breast Images from Another Volunteer . . . . . . . . . . . 91
7.1
Gradient-Echo Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
7.2
Steady-State Signal with Varying Phase Increment Angle . . . . . . . . . . 95
7.3
Flow Artifacts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
7.4
Partial Saturation of Flowing Spins . . . . . . . . . . . . . . . . . . . . . . 97
7.5
Sequence with Partial Saturation . . . . . . . . . . . . . . . . . . . . . . . 98
xviii
7.6
Phase Graphs for Gradient-Echo Sequences with Partial Saturation . . . . . 100
7.7
Magnetization Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7.8
Magnetization Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.9
Comparison of Simulated Signal Profiles . . . . . . . . . . . . . . . . . . . 105
7.10 Dephasing of Partial Saturation Slab . . . . . . . . . . . . . . . . . . . . . 107
7.11 Maximum Intensity Projection Images of Tube Phantom . . . . . . . . . . 108
7.12 Signal Intensity Change along a Tube . . . . . . . . . . . . . . . . . . . . 109
7.13 Inflow Enhancement Reduction in Arteries . . . . . . . . . . . . . . . . . . 111
7.14 Composite RF Pulse and Magnetization Profile . . . . . . . . . . . . . . . 116
xix
Chapter 1
Introduction
The nuclear magnetic resonance (NMR) phenomenon was first discovered by Felix Bloch [1]
and Edward Purcell [2] independently in 1946. Originally, NMR was developed for chemical and physical molecular analysis, but in 1973 Paul Lauterbur first applied NMR to
imaging using gradient magnetic fields [3]. Since then magnetic resonance imaging (MRI)
has rapidly developed due to its main features. MRI does not use ionizing radiation and
offers more flexible image contrast between different tissues than other imaging modalities.
Currently, MRI is clinically used for imaging many body parts for many different purposes, including assessing morphology, detecting tumors or diseases, and visualizing function. Breast cancer detection is one of the important applications. Breast MRI has higher
sensitivity in detecting tumors than commonly-used X-ray mammography for high-risk patients [4, 5].
T1 -weighted dynamic contrast-enhanced (DCE) MRI is the most important breast MRI
technique. It involves repeated imaging following intravenous injection of a gadoliniumbased contrast agent. Contrast uptake by tumor angiogenesis causes decrease in T1 of
tumors, thus leads to signal enhancement on T1 -weighted images. Signal enhancement
1
CHAPTER 1. INTRODUCTION
2
patterns can be used to characterize tumors and discriminate malignant and benign tumors
[6–8]. However, DCE-MRI suffers from low specificity since benign tumors can enhance
with contrast agent as well, which results in unnecessary scans or biopsies.
The accuracy of tumor detection and characterization with DCE-MRI is improved by
high-quality images, particularly with high temporal and spatial resolution, high signal-tonoise ratio (SNR), reliable fat suppression, bilateral imaging, and reduction of flow artifacts. This thesis focuses on improvement of DCE-MRI in these areas.
The major contribution of this thesis is improving imaging speed in bilateral imaging
by applying parallel imaging. Parallel imaging emerged with the use of a phased-array
coil, where each coil element has different spatial selectivity [9]. Parallel imaging accelerates data acquisition by supplementing gradient encoding by spatial encoding with
coil sensitivities. This technique has reduced scan time significantly and allowed for new
clinical applications. However, parallel imaging can result in aliasing artifacts and noise
amplification if coil sensitivity calibration is not correct or reconstruction methods are not
appropriate.
First, this thesis demonstrates that parallel imaging can be suitably applied to bilateral
breast imaging with a commercial eight-channel breast coil. In particular, it shows that
slab-phase modulation of the two slabs [10] can improve parallel imaging quality in terms
of residual aliasing artifacts and SNR. Second, this thesis presents temporal acceleration
by exploiting parallel imaging techniques in bilateral DCE imaging. Specifically, the adaptive sensitivity encoding incorporating temporal filtering (TSENSE) technique [11] was
combined with 3D spiral imaging, providing very high temporal resolution. These DCE
imaging methods, proved to be extremely robust, has been used in routine clinical protocols at Stanford Hospital for the last three years, yielding more than 3000 patient scans.
CHAPTER 1. INTRODUCTION
3
Fat suppression is important because fatty tissue can yield signal intensity similar to
that of enhanced tumors. This thesis also presents that fat suppression can be improved in
bilateral imaging by applying independent shims to each breast.
Flow artifacts such as inflow enhancement [12] and pulsatile ghost artifacts [13,14] can
degrade DCE data quantification for contrast kinetic parameters [15, 16]. This thesis also
shows that partial saturation can eliminate flow artifacts efficiently in RF-spoiled gradientecho sequences.
The structure of this thesis is as follows:
Chapter 2: Basics of Magnetic Resonance Imaging
This chapter briefly reviews the physical principles of the MR phenomenon, magnetic fields
in an MRI sanner, MR imaging methods, and parallel imaging. Much of this chapter was
adapted from “Principles of Magnetic Resonance Imaging” by Nishimura [17].
Chapter 3: Breast MRI
This chapter describes breast cancer, and introduces breast MRI.
Chapter 4: Parallel Imaging with Slab-Phase Modulation
This chapter demonstrates that slab-phase modulation of the two breast slabs can not only
reduce scan time but also improve image quality for parallel imaging applications. Thorough analysis of reconstructed images from phantom and in vivo experiments is presented.
Most of this work has been published in a journal article [18] and a conference abstract [19].
Chapter 5: Rapid Bilateral Breast DCE-MRI
This chapter presents TSENSE acceleration for bilateral breast DCE-MRI to achieve high
temporal resolution. Details of TSENSE reconstruction including sensitivity map estimation and regularization are described and patient study results are provided. Most of this
work has been published in a journal article [20] and two conference abstracts [21, 22].
CHAPTER 1. INTRODUCTION
4
Chapter 6: Fat Suppression with Independent Shims
This chapter describes improvement of fat suppression by incorporating independent shims
to bilateral excitations. The method of applying independent shims to bilateral excitation
is described and a comparison is made between standard shims and independent shims by
in vivo experiments. This work has been published in a conference abstract [23].
Chapter 7: Flow Artifact Reduction Using Partial Saturation
This chapter describes reduction of flow artifacts by using partial saturation in RF-spoiled
gradient-echo imaging. The results from magnetization simulation, phantom and in vivo
experiments are provided, demonstrating that the partial saturation can provide accurate T1
contrast of blood. This work has been published in two conference abstracts [24, 25] and
has been submitted for consideration as a journal article [26].
Chapter 8: Summary and Recommendations
This chapter summarizes contributions presented in this thesis and outlines possible future
work.
Chapter 2
Basics of Magnetic Resonance Imaging
This chapter provides the basics of MRI including MR physics, three magnetic fields for
MRI, imaging methods, and parallel imaging, which would help to understand research
presented in this thesis.
2.1 Physics of Nuclear Magnetic Resonance
2.1.1 Polarization
Atoms with an odd number of protons and/or odd number of neutrons possess a spin angular
momentum and an associated spin magnetic dipole moment. These nuclei can be thought
as spinning charged spheres that give rise to a small magnetic moment and usually called
“spins.” Hydrogen (1 H) is the most commonly studied nucleus as it is the most abundant
in biological specimens.
In the absence of an external magnetic field, spins are oriented randomly and the net
magnetic moment is zero. However, in the presence of an external magnetic field B0 , spins
5
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
No Magnetic Field
a
6
Magnetic Field
b
B0
M0
Figure 2.1: (a) Without an external magnetic field, spins are oriented randomly and the net
magnetization is zero. (b) In the presence of an external magnetic field B0 , spins align in
the direction of B0 with the net magnetization proportional to B0 .
tend to align in the direction of B0 and create a net magnetization M (magnetic moment
per unit volume) (Fig. 2.1). The equilibrium magnetization magnitude M0 is proportional
to the magnetic field strength B0 and spins exhibit resonance at a well-defined frequency
called the Larmor frequency. The Larmor frequency is proportional to B0 as expressed by
f0 =
γ
B0 ,
2π
where gyromagnetic ratio γ is a unique constant for each nuclear species. For 1 H,
(2.1)
γ
2π
=
42.58 MHz/Tesla.
2.1.2 Precession
At thermal equilibrium, the net magnetization M and the applied magnetic field B0 are
pointed in the same direction. However, if M is made to point in a different direction to B0 ,
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
7
Precession
B0
M
x
y
z
Figure 2.2: If the magnetization M is perturbed to point a different direction than the B0
direction, magnetization precesses around the B0 direction.
(by RF excitation for MRI, Section 2.4), then M precesses about B0 . This relationship can
be written as
dM
= M × γ B,
dt
(2.2)
where B is the total magnetic field applied. When only B0 exists, M precesses about B0 at
the Larmor frequency (Fig. 2.2).
2.1.3 Relaxation
The perturbed magnetization precesses about the applied magnetic field, and at the same
time, it returns to its thermal equilibrium state by a process called relaxation (Fig. 2.3).
If we divide the magnetization into two components, the longitudinal component in the
direction of B0 and the transverse component located in the xy plane, relaxation of each
component can be separately described.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
b
Relaxation
a
8
Transverse Component (Mxy )
M0
z
y’
c
M
x’
Time
0
B0
Longitudinal Component (Mz)
M0
0
Time
Figure 2.3: (a) The perturbed magnetization M returns back to its thermal equilibrium state
exponentially. (b) The transverse component decays to zero by a time constant T1 . (c) The
longitudinal component returns to the thermal equilibrium value M0 by a time constant T2 .
Longitudinal Relaxation
The longitudinal magnetization (Mz ) returns to equilibrium along the z direction exponentially with a time constant T1 (spin-lattice relaxation time constant). With this time constant,
the longitudinal magnetization behaves as
Mz(t) = M0 + (Mz (0) − M0) exp(−t/T1).
(2.3)
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
9
Transverse Relaxation
The transverse magnetization (Mxy ) decays to zero exponentially with a time constant T2
(spin-spin time constant) as follows:
Mxy (t) = Mxy (0) exp(−t/T2).
(2.4)
T2 is always less than T1 . T1 and T2 values vary depending on tissue types and field strength.
2.1.4 Bloch Equation
The Bloch equation (Eq. 2.5) describes the dynamics of magnetization M in the presence
of the applied magnetic field B, with incorporating both precession and relaxation.
Mx i + My j (Mz − M0 )k
dM
= M × γB −
−
.
dt
T2
T1
(2.5)
Here, i, j, and k are unit vectors along the x, y, and z axes, and Mx and My are the x and y
components of M. The cross-product relation describes a precession behavior while terms
containing T1 and T2 describe the exponential behavior of the longitudinal and transverse
components, respectively. The equation solution represents the magnetization evolution as
a function of time.
2.2 Magnetic Fields for MR Imaging
An MRI scanner mainly consists of three main hardware components: a main magnet, an
RF system, and a magnetic field gradient system. The main magnet generates a strong static
field (B0 field) to polarize spins; the RF system is capable of generating a rotating magnetic
field in the transverse plane (B1 field); the magnetic field gradient system is designed to
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
10
produce time-varying linear gradient fields (G field). MR signals can be generated from
change of magnetization caused by interaction with the three different magnetic fields.
2.2.1 Static Magnetic Field B0
In an MRI scanner, a strong static magnetic field polarizes the spins, usually along the superior/inferior direction of the subject. Clinical magnets generally maintain a field strength
in the range of 0.1 - 3.0T with research systems available up to 9.4T. Inhomogeneities of
the field within the scan region are normally 1-3 parts per million (ppm).
2.2.2 RF Field B1
To obtain an MR signal, an RF magnetic field, B1 , tuned to the resonant frequency of the
spins is applied in the transverse plane to excite these spins out of equilibrium. The B1 field
rotates the magnetization vectors by a prescribed angle dependent on the amplitude and the
duration of the B1 (RF) pulse. Figure 2.4 shows magnetization excitation in the laboratory
frame of reference ((x, y, z) coordinate) and in the frame rotating at the Larmor frequency
f0 about the z axis ((x# , y# , z# ) coordinate). The two coordinates are related by
x# = x cos(2π f0t) − y sin(2π f0t),
y# = y sin(2π f0t) + x cos(2π f0t),
(2.6)
z# = z.
From Faraday’s law of induction, rotating magnetization vectors induce an electromotive
force (EMF) and generate signal in an RF receiver coil.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
11
Excitation
a
b
z
B1
M
y’
B0
y
f0
z’
x
Reference Frame
B0
x’
B1
M
Rotating Frame
Figure 2.4: RF excitation in the laboratory frame (a) and the rotating frame at the Larmor
frequency about the z axis.
2.2.3 Linear Gradient Fields G
If only B0 exists, it is not possible to distinguish the signals from different spatial locations
because all spins would resonate at the same frequency. In the MRI scanner, quasi-linear
gradient fields can be applied independently along the x, y and z axes in addition to the B0
field using specific linear gradient coils. For example, by applying a gradient Gx in the x
direction after excitation, the applied field is B0 + Gx x, and the frequency of spins becomes
a function of x location: f (x) =
γ
2π (B0 + Gx x)
(Fig. 2.5). The generated signal in the re-
ceiver coil combines magnetization at a specific x location modulated by a corresponding
resonant frequency. Then the Fourier transform of signal determines the contribution of
each frequency component, which maps to a particular x position. Normally, spatial localization (spatial encoding) in 2D or 3D domains is performed by applying linear gradients
along two or three axes.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
a
b
f (x) =
γ
(B0 + Gx x)
2π
Magnetic Field
Resonant Frequency
B(x) = B0 + Gx x
12
x Position
x Position
Figure 2.5: By applying a linear gradient (a), the resonant frequency of the excited spins
becomes linearly varying (b).
2.3 Imaging Methods
Conventionally, MR imaging consists of two parts: excitation and data acquisition. Excitation involves the use of RF pulses to tip the spins into the transverse magnetization,
thus magnetization signal can be generated in the receiver coil. Data acquisition involves
receiving signal with spatial encoding. Because of relaxation, the transverse magnetization
decays along the data acquisition. A sequence including both excitation and acquisition
parts is typically repeated many times during a scan and a set of acquired data is used
together to reconstruct images.
When exciting spins, it is more desirable to excite a limited volume than to excite the
entire volume. This selective excitation is normally used for breast MRI as well. This
section describes selective excitation, data acquisition, and then data sampling criteria.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
13
2.3.1 Selective Excitation
A transverse RF magnetic field, tuned to the Larmor frequency of the nuclear spins, tips the
magnetization vector over the entire volume. However, by turning on Gz during excitation,
the resonant frequency varies linearly along the z axis and only the spins that have resonant
frequencies within the bandwidth of B1 (t) are excited. To attain a rectangular slice profile,
the waveform of B1 (t) is usually like a truncated sinc function (Fig. 2.6a-b), but there are
a great variety of B1 (t) waveforms depending on specific applications. The excited slab
thickness ($z) can vary by changing the magnitude of Gz (Fig. 2.6c). For a 2D imaging
sequence, a thin slice is excited and 2D spatial encoding over the xy plane is performed;
for a 3D imaging sequence, a thicker slab is excited and z spatial encoding is applied in
addition to 2D spatial encoding to distinguish different z locations. The following sections
more focus on 2D imaging methods for simplicity.
2.3.2 Spatial Encoding and Data Acquisition
Assume that the transverse magnetization at the position (x, y) is m(x, y) following selective
excitation. Then the received signal at time t in the presence of linear gradient field Gx (t)i+
Gy (t)j would be
sr (t) =
! !
y x
m(x, y) exp{−iγ [B0t +
! t
0
Gx ( τ ) d τ x +
! t
0
Gy (τ ) d τ y]} dx dy.
(2.7)
Then the signal equation after demodulation at γ B0 can be simplified as
s(t) =
where [kx (t), ky (t)] = [ 2γπ
! !
y x
"t
m(x, y) exp{−i2π [kx (t)x + ky (t)y]} dx dy,
γ "t
0 Gx (τ ) d τ , 2π 0 Gy (τ ) d τ ].
(2.8)
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
14
RF Amplitude
a
c
b
z Position
Time
Magnitude
1
γG z
Slope =
Selective
Excitation
Δz
BW
Frequency
BW
Frequency
Figure 2.6: (a) The RF waveform for selective excitation and (b) its frequency spectrum.
(c) By turning on the z gradient during excitation, only the spins that have resonant frequencies within the bandwidth of B1 (t) would be excited. The excited slab thickness $z is
determined by the Gz magnitude.
If the 2D Fourier transform of m(x, y) is M(kx , ky ), then s(t) = M(kx (t), ky(t)). If
M(kx , ky ) over an adequate range of (kx , ky ) is obtained through s(t), then m(x, y) can
be imaged by the inverse Fourier transform of the obtained signal (Fig. 2.7). In MR, the
Fourier transform space is called k-space, where k represents the spatial-frequency variable
with a typical unit of cycles/cm.
The path trace in the k-space, which is determined by the applied linear gradients, is
called a k-space trajectory. For conventional 2D imaging, one k-space line is acquired per
one sequence repetition by applying x and y gradients as shown in Fig. 2.8. During readout,
constant x gradient, and no y gradient are applied. We usually refer to the x direction as the
readout direction, the y direction as the phase-encode direction, and each k-space line as
the phase-encode line. The time duration over one sequence is called sequence repetition
time (TR), and the time between tipping of the magnetization and acquisition of the k-space
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
15
ky
2D Fourier
Transform
kx
Spatial Frequency Domain
(k-space)
Image Domain
Figure 2.7: With spatial encoding, the acquired MR signal is the spatial frequency component of the object.
origin (kx = 0) is called echo time (TE). Over multiple sequence repetitions, entire phaseencode lines on different ky positions are acquired. As k-space data already reside on a
Cartesian grid, simply applying a 2D inverse Fourier transform can reconstruct the image.
This conventional k-space trajectory is called the 2DFT trajectory (Fig. 2.9a). The SNR in
the reconstructed image linearly increase with the square-root of the number of acquired
phase-encode lines if the maximum k-space coverage remains the same. The maximum
k-space coverage determines the image resolution.
Many other k-space trajectories can be used for acquisition because they are fully controlled by gradient waveforms. In particular, an echo-planar trajectory [27] and a spiral
trajectory [28, 29] are popular for fast data acquisition; they collect a much larger fraction
of k-space on each RF excitation than the conventional 2DFT trajectory.
The echo-planar trajectory acquires multiple phase-encode lines per excitation in a
raster-like manner (Fig. 2.9b). The entire phase encode lines can be acquired with a single excitation, but segmented k-space acquisition is possible with multiple excitations. By
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
RF Excitation
16
Imaging Gradients
RF
ky
TE
Gz
Gx
kx
Gy
Readout
TR
Figure 2.8: With the conventional 2DFT trajectory, for each excitation one k-space line is
acquired by applying a phase-encode lobe to determine the ky location and applying the x
gradient during readout.
acquiring data on the Cartesian grid, the data also can be reconstructed with a 2D inverse
Fourier transform.
In contrast, the spiral trajectory acquires data on a non-Cartesian grid, starting acquisition at the k-space origin and spiraling outward in 2D k-space with time-varying gradients
(Fig. 2.9c). Like the echo-planar trajectory, spiral imaging can acquire the entire k-space
data with a single excitation or with a multiple excitations using different spiral interleaves,
which are rotated with respect to one another. The spiral trajectory is more time-efficient
than the echo-planar trajectory because k-space is sampled continuously during readout.
For reconstruction, it requires gridding, resampling of acquired data to the 2D grid [30,31],
prior to applying a 2D inverse transform.
These rapid imaging trajectories are more prone to artifacts from B0 field inhomogeneities, local susceptibility variations, system eddy currents, and gradient delays. For the
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
a
2DFT
ky
b
kx
Echo-Planar
ky
c
17
Spiral
ky
kx
kx
Figure 2.9: (a) Conventional 2DFT imaging acquires one k-space line per excitation. The
echo-planar trajectory (b) and spiral trajectory (c) are widely used for fast acquisition. The
echo-planar trajectory acquires multiple phase-encode lines per excitation, whereas the
spiral trajectory acquires data from the k-space origin to outer k-space radially.
echo-planar trajectory, ghost artifacts along the phase-encode directions can be generated
due to the reverse direction of data acquisition between even and odd phase-encode lines
and geometrical distortion can occur in the image. For the spiral trajectory, image blurring
can be caused by off-resonant spins. Several techniques are used to correct off-resonance
effects after estimating the off-resonant frequency on each spatial location [32–34].
2.3.3 Sampling Criteria
For MR data acquisition, sampling occurs in the k-space domain along both the readout
direction and the phase-encode direction. By a Fourier transform property, sampling in one
domain results in replication in the corresponding transform domain (Fig. 2.10). If the kspace is sampled with an interval of $kx in the x direction and $ky in the y direction, then
the replication is with an interval of 1/$kx in the x direction and 1/$ky in the y direction.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
a
k-space Domain
18
Image Domain
b
ky
1/∆ky
FOVy
kx
∆ky
X
Y
1/∆kx
x
FOVx
∆kx
Figure 2.10: (a) Sampling in k-space and (b) the replication in the image domain. The kspace sampling intervals ($kx and $ky ) determine the imaging FOV of 1/$kx and 1/$ky
along the x and y directions. To avoid aliasing, $kx and $ky should be smaller than the
inverse of the object size, 1/X or 1/Y .
This determines the effective FOV of one replication as follows.
FOVx =
1
1
, and FOVy =
.
$kx
$ky
(2.9)
If the FOVx or FOVy is smaller than the object size in each x or y direction, then aliasing
arises.
Usually aliasing in the readout direction can be avoided by applying an anti-aliasing
filter. However, aliasing in the phase-encode direction always occurs if the k-space interval
is not sufficiently small. To avoid aliasing artifacts, the sampling interval $ky should be
determined based on the object size along the y direction.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
a
Single Coil
b
19
Phased-Array Coil
Figure 2.11: (a) Volume coil versus (b) phased-array coil. A phased-array coil consists of
small surface coils with different spatial selectivity to receive signal simultaneously over
the excited regions.
2.4 Parallel Imaging
Imaging speed is one of the most important considerations in clinical MRI. For the 2DFT
imaging, the scan time can be decreased by reducing the TR for given k-space sampling
patterns. However, hardware limits and patient safety restrict the possible minimum TR
value. Besides that, a long TR is sometimes required to generate proper image contrast.
Recently, parallel imaging has been widely used to decrease scan time by acquiring a
reduced number of phase-encode lines. It uses phased-array coils, where each coil element
has different spatial sensitivities [9] (Fig. 2.11). By supplementing gradient encoding by
spatial encoding from coil sensitivities, unaliased images can be reconstructed from undersampled data, with combining the signals from the coil elements.
There are many different parallel imaging methods, but can be divided into two classes,
“model-driven” reconstructions and “data-driven” reconstructions. Model-driven methods
require explicit knowledge of coil sensitivities to separate aliased signals and reconstruct
data by closely modeling the underlying physical process during image acquisition. This
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
20
class includes simultaneous acquisition of spatial harmonics (SMASH) [35], sensitivity encoding (SENSE) [36], and modified sensitivity encoding (mSENSE) [37]. In contrast, datadriven methods do not require explicit coil sensitivity information but rather reconstruct
unacquired data directly in k-space by using neighboring acquire data. This class includes
AUTO-SMASH [38], variable density AUTO-SMASH (VD-AUTO-SAMSH) [39], generalized autocalibrating partially parallel acquisitions (GRAPPA) [40], and autocalibrated
reconstruction for Cartesian sampling (ARC) [41]. The next subsections review SENSE
and GRAPPA, which are the most well-known techniques for each class.
2.4.1 SENSE
With an acceleration factor of two (R = 2), which means that every other phase-encode
line is not acquired, each coil image has aliasing artifacts due to the reduced FOV in the
phase-encode direction. For each pixel, signals from two positions that are originally one
half FOV away are superimposed, with each weighted by local coil sensitivity values. Figure 2.12 shows the superposition of the two pixels in the i’th aliased coil images with
denoting m as the original signal, and si as the spatial sensitivity of coil i.
SENSE reconstructs the full-FOV image from aliased coil images using the “unaliasing” process assuming that coil sensitivity information is known. For each position (x, y)
in aliased coil images, signal superposition can be represented as



 

n1 (x, y)
S1 (x, y) S1 (x, y + FOV /2) 
c1 (x, y)



 

m(x, y)

 .. 
..
 ..   ..

+


 . = .
. 
.

 m(x, y + FOV /2)

 

nN (x, y)
SN (x, y) SN (x, y + FOV /2)
cN (x, y)
(2.10)

CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
Full k-space data
21
Undersampled k-space data
si(x,y)m(x,y)
FOV/2
FOV
si(x,y)m(x,y) +
si(x,y-FOV/2)m(x,y-FOV/2)
si(x,y-FOV/2)m(x,y-FOV/2)
Figure 2.12: Spatial aliasing due to undersampled k-space acquisition. With an acceleration
factor of two, signals from two pixels that are originally half FOV away are superimposed.
where ci and ni denote the aliased coil image and the noise signal for the i’th coil. This
equation can be simplified as a matrix equation:
c = Sm + n
(2.11)
Then, the solution of m, argminm %Sm − c% can be determined by the least square fit:
m̂ = (SH Ψ−1 S)−1 SH Ψ−1 c,
(2.12)
where Ψ is the noise correlation matrix between coil elements [9]. Using the unaliasing process on a pixel-by-pixel basis, the full-FOV original image can be reconstructed.
Figure 2.13 illustrates SENSE reconstruction with an acceleration factor of two when a
four-channel phased-array coil is used.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
22
Aliased Coil Images
SENSE Image
Multi-Coil
Undersampled
k-Space Data
FFT
SENSE
Calculate
Unaliasing Matrix
Sensitivity Maps
Figure 2.13: Illustration of SENSE reconstruction for the case with an acceleration factor
of two (R = 2) with four receiver coils. The SENSE unaliasing matrix is determined from
the coil sensitivity information.
Accelerated images from parallel imaging reconstructions have lower SNR than unaccelerated images due to the reduced number of phase-encode lines and suboptimal coil
geometry. The coil geometry, the acceleration factor, and the acceleration direction determine spatially varying geometry factors (g-factors), which represent ill-conditioning of the
matrix inversion during SENSE unaliasing process. The SNR for accelerated images can
be calculated as
SNRaccel =
SNRfull
√ ,
g R
(2.13)
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
23
Sensitivity Maps
a
1.8
SENSE g-Factor Maps
b
c
d
0.5
R=3
R=2
R=4
Figure 2.14: For given sensitivity maps of four coil elements (a), g-factors at various acceleration factors are shown in (b-d).
where SNRfull is the optimal SNR on the composite of all coil images from full k-space
data [9].
Analytical calculation of g-factors for SENSE is driven in [36] as follows:
gSENSE,φ =
)
[(SH Ψ−1 S)−1 ]φ ,φ (SH Ψ−1 S)φ ,φ ,
(2.14)
where φ denotes the index of the pixel within the set of pixels to be separated. The g-factor
is always at least equal to one. Figure 2.14 shows g-factor maps at various acceleration
factors with a four-channel phased-array. With the increase of the acceleration factor, gfactors generally increase, which means noise is more amplified.
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
24
To eliminate aliasing artifacts perfectly, coil sensitivity maps should be accurately measured. Errors in sensitivity map estimation can generate residual aliasing artifacts and
further noise amplification [36]. Conventionally, coil sensitivity information is acquired by
an extra scan, but several variations of SENSE techniques exist, which use self-calibration.
Modified SENSE (mSENSE) [37] acquires a few additional lines at the center to estimate
coil sensitivity maps from the central k-space and applies SENSE for reconstruction. Adaptive sensitivity encoding incorporating temporal filtering (TSENSE) [11] is a temporal acceleration technique, which estimates sensitivity maps from multiple temporal frames and
applies SENSE for each temporal frame.
2.4.2 GRAPPA
Unlike SENSE, GRAPPA reconstructs data in k-space without explicit knowledge of coil
sensitivity information. GRAPPA exploits the correlation between neighboring k-space
data caused by the spread coil sensitivity spectrum (Fig. 2.15). The underlying assumption
of GRAPPA is that each k-space data point on a single coil can be represented as a linear
combination of its neighboring data on all coils and the set of linear combination kernel
weights is shift invariant in k-space.
To calculate the kernel weights, extra phase-encode lines (autocalibration signal (ACS))
are acquired in addition to the reduced acquisition in the central k-space. The central calibration data are then used to calculate kernel weights (Fig. 2.16). For each coil, a kernel
weight is calculated by a data-fitting approach, and is used to estimate unacquired phaseencode lines. Then reconstructed coil images can be combined into a final full-FOV image
using a sum-of-squares method or any array combination method [9, 42].
The analytical derivation of GRAPPA g-factors has been regarded as quite difficult, but
recent papers show that it can be derived by formulating GRAPPA in the image domain
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
a Uniform Sensitivity
25
b Vertical Shading
Figure 2.15: Spectrum of coil sensitivity. (a) With uniform sensitivity, the spectrum of the
coil is point-like, but (b) with sensitivity variation, the spectrum of coil sensitivity is spread
out and k-space data is blurred.
if ACS lines are only used for determining the kernels but not used for the final reconstruction [43, 44]. Synthesizing unacquired data in k-space by using kernel weights can be
expressed as convolution of undersampled k-space data with convolution kernel weights.
This is identical to the multiplication of aliased images with the Fourier transform of the
convolutional kernel as follows.
Mmacc =
N
∑ Mlred⊗wm, l
l=1
⇒
Imacc =
N
∑ IlredWm,l ,
(2.15)
l=1
where wm,l is a convolution kernel that applies to coil l to reconstruct coil m in k-space,
Mlred and Ilred are undersampled k-space data and the corresponding aliased image for coil
l, W is the inverse Fourier transform of the convolution kernel w, and N is the number of
coils. If reconstructed coil images are combined by using weights, pm , then
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
26
kx
ky
Acquired
Unacquired
ACS
Coil #
Figure 2.16: With the assumption that each k-space data point on a single coil can be
represented as a linear combination of its neighboring data on all coils, GRAPPA kernel
weights are calculated from calibration data.
acc
Icomb
N
=
N
∑ pm ∑
m=1
l=1
IlredWm,l
N
=
N
∑ ( ∑ pmWm,l )Ilred.
(2.16)
l=1 m=1
By formulating GRAPPA in the image domain, we can analyze how the noise in acquired coil images propagates into the final reconstructed image on each pixel. Thus, gfactors can be measured by comparing noise variance of unaccelerated images and accelerated images using Eq. 2.13 (SNR is inversely proportional to the noise variance). Let us
denote ql = ∑N
m=1 pmWm,l . Then the GRAPPA g-factor at (x, y) can be defined as
*
q̃Ψq̃H
,
gGRAPPA (x, y) = *
p̃Ψp̃H
where p̃ and q̃ are 1 x N vectors with q̃(l) = ql (x, y) and p̃(l) = pl (x, y).
(2.17)
CHAPTER 2. BASICS OF MAGNETIC RESONANCE IMAGING
27
GRAPPA is beneficial when the image includes areas where coil sensitivity maps are
difficult to obtain. In low-spin density regions, the image quality from SENSE might suffer
from inaccurate sensitivity maps, producing residual aliasing artifacts and noise amplification. With GRAPPA reconstruction kernel weights are estimated globally and are not affected by local signal void, and thus good reconstruction image quality might be achieved.
As a self-calibration technique where coil sensitivity information is extracted from acquired
data, GRAPPA is less sensitive to subject motion.
Chapter 3
Breast MRI
Breast cancer is the second leading cause of death due to cancer among American women
[45]. Approximately 180,000 women are diagnosed with breast cancer each year in the
United States. However, early diagnosis and treatment can reduce mortality rates [46].
The most common breast cancer screening method has been X-ray mammography, but
recently MRI has emerged as a powerful screening tool especially for high-risk patients.
This chapter presents a brief description of breast cancer and cancer detection using MRI.
3.1 Breast Cancer
The breast mainly consists of glandular and fatty tissue. The glandular tissue houses 15-20
lobes, which consist of lobules and ducts. The lobules produce milk; the ducts move milk
to the nipple. A layer of fatty tissue surrounds the breast glands and determines the shape
and size of the breast. Figure 3.1 depicts breast anatomy.
Breast cancer usually originates in the ducts or in the lobules. To meet high metabolic
demand for oxygen and nutrients, a malignant tumor requires angiogenesis, which induces
28
CHAPTER 3. BREAST MRI
29
Chest wall
Pectoralis muscle
Lobules
Nipple surface
Areola
Duct
Fatty tissue
Skin
http://www.turbosquid.com/3d-models/3dsmax-breast-anatomy/410670
Figure 3.1: Breast anatomy. The breast mainly consists of glandular tissues (lobules and
ducts) and fatty tissue.
sprouting and growth of pre-existing capillaries and formation of new vessels [47–49], thus
increasing perfusion. In addition, these capillaries exhibit a pathologic vessel wall architecture with leaky endothelial linings, thus tumors become hyperpermeable [50,51]. Different
types of tumors show different levels of angiogenesis with varying levels of vascularity and
permeability.
3.2 Breast MRI
Currently, the most common breast cancer screening method is X-ray mammography.
However, it has limited ability to detect tumors in young women, who have more dense
breasts than postmenopausal women [52]. Recently, T1 -weighted dynamic contrast-enhanced
(DCE) MRI has emerged as a powerful tool to detect cancer. Several studies show that
DCE-MRI has much higher sensitivity than mammography for detecting cancer in highrisk patients [4, 5, 53]. In 2007, the American Cancer Society recommended screening
CHAPTER 3. BREAST MRI
a
X-Ray Mammography
30
b
Contrast-Enhanced Breast MRI
Tumor
Figure 3.2: (a) X-ray mammography shows normal and moderately-dense glandular tissue
but there is no obvious sign of breast cancer. (b) Contrast-enhanced MRI shows an enhanced lesion, which is confirmed as an invasive ductal carcinoma by subsequent biopsy.
MRI for women at 20-25% or greater lifetime risk of breast cancer, based on six studies in which the sensitivity of MRI ranged from 71-100% compared to the sensitivity on
mammography of 16-40% [54]. Figure 3.2 compares images from X-ray mammography
and contrast-enhanced MRI. X-ray mammography depicts normal and moderately-dense
glandular tissue but there is no obvious sign of breast cancer. However, contrast-enhanced
MRI shows an enhanced mass. Subsequent MRI-guided biopsy confirmed the mass is a
malignant tumor of invasive ductal carcinoma.
While T1 -weighted DCE-MRI provides high sensitivity, a major drawback is poor specificity; benign tumors can enhance with contrast injection like malignant tumors. Various strategies are used to improve sensitivity, including quantifying signal enhancement
curves acquired with bolus intravenous injection and evaluating morphologic features of
lesions. Additionally, several non-contrast enhanced techniques can be used to improve
specificity. T2 -weighted MRI can be used to differentiate fibroadenomas and benign cysts
from tumors [55], and diffusion-weighted MRI has the potential to only detect malignant
CHAPTER 3. BREAST MRI
a
T2-Weighted
31
b
T1-Weighted, Pre-Contrast
c
T1-Weighted, Post-Contrast
Figure 3.3: Patient breast images from T2 -weighted imaging (a), pre-contrast T1 -weighted
imaging (b), and post-contrast T1 -weighted imaging (c). The tumor is visible on the T2 weighted image. With T1 -weighted contrast, the tumor is not visible on the pre-contrast
image but is highlighted on the post-contrast image.
tumors [56–58]. Figure 3.3 shows breast images with a tumor from T2 -weighted imaging,
and pre-contrast and post-contrast T1 -weighted imaging.
3.3 Dynamic Contrast-Enhanced MRI
DCE-MRI is currently the most important breast MRI technique for cancer detection [59].
When gadolinium-based contrast material is injected, lesion enhancement occurs mainly
for the following reasons: increased vascularity causes a focally increased inflow of contrast material, and increased vessel permeability induces an accelerated extravasation of
contrast material at the site of tumor. Thus, signal enhancement in MR images depends
on vessel density, vessel architecture, permeability, and tissue relaxation time. Figure 3.4
shows a signal intensity curve over time in the tumor (the same tumor in Fig. 3.2) measured
from DCE imaging. Rapid contrast uptake in the tumor results in signal enhancement following contrast injection.
CHAPTER 3. BREAST MRI
a
DCE Breast Image
32
b
Signal Enhancement Curves
Signal Intensity
1800
Tumor
600
0
Contrast Injection
Time
3.5 min
Figure 3.4: (a) Breast image at a post-contrast frame from DCE imaging. A tumor is
enhanced shown by an arrow. (b) Signal enhancement curve on the tumor. Rapid signal
enhancement can be seen following contrast injection.
It has been shown that signal enhancement patterns from lesions reflect tumor angiogenesis [60] and are helpful to discriminate between benign and malignant lesions [6–8].
In general, most malignant lesions enhance rapidly whereas benign lesions enhance at a
slower rate. Several studies show that quantitative measurement of contrast kinetic parameters from the contrast uptake curves can improve specificity in diagnosis [15, 16, 61].
To accurately characterize lesions from DCE data, DCE imaging requires both high
temporal and high spatial resolution. High temporal resolution provides accurate characterization of signal enhancement patterns, while high spatial resolution offers detailed visualization of lesion morphology such as spiculated margins, rim enhancement, and nonenhancing internal septations. In addition, DCE imaging needs to suppress fat signal for the
best possible detection of tumors, because fatty tissue can yield signal intensity similar to
that of enhanced tumors.
CHAPTER 3. BREAST MRI
33
Various fat suppression techniques have been used for breast MRI. Usually they exploit the resonant frequency difference between water and fat since the fat resonant frequency is 3.5 ppm lower. Fat-saturation techniques saturate fat before RF excitation [62],
and are time-efficient because fat saturation RF pulses do not need to be applied every
pulse sequence repetition. However, they are most sensitive to B0 and B1 field inhomogeneities. Water-only excitation (spectral-spatial excitation) techniques only excite water,
having reduced sensitivity to field inhomogeneities, but increase echo time and resultant
scan time [63]. Multi-point Dixon methods separate water and fat images from multi-echo
images, and are less sensitive to field inhomogeneities than fat-saturation and water-only
excitation methods [64–66]. However, they also increase scan time because of the need to
image at multiple echo times, and they are more sensitive to motion and other displacement
artifacts.
Bilateral imaging is more cost-effective and comfortable to patients than unilateral
imaging because both breasts can be imaged with one contrast injection on a single day.
Additionally, it helps to avoid diagnostic errors by comparing the two breasts and increases
the possibility of detecting bilateral tumors [67]. However, bilateral imaging is more difficult than unilateral imaging in achieving high temporal and spatial resolution due to the
increased imaging volume. It needs at least twice the scan time than unilateral imaging if
identical acquisition methods are used and spatial resolution is kept the same. Furthermore,
increased field inhomogeneities over the two breasts make fat suppression more challenging as well. However, with the recent development of hardware and imaging techniques,
bilateral MRI has been used for a clinical protocol in many outpatient centers.
Chapter 4
Parallel Imaging with Slab-Phase
Modulation
In sagittal bilateral breast MRI, two breasts of interest are separated in the slab direction.
Independent slab-phase modulation (ISPM) [10] incorporated to dual-band spectral-spatial
excitation [68, 69] eliminates the need to encode the space between the two breasts by
shifting the two breasts close together. Parallel imaging can also improve bilateral imaging
speed; however, it can result in noise amplification and residual aliasing artifacts. This
chapter investigates whether ISPM can improve parallel imaging quality, particularly when
self-calibrated parallel imaging methods are used. To compare model-driven reconstruction
and data-driven reconstruction, both mSENSE and GRAPPA methods were implemented
and applied. Phantom experiments verified that ISPM facilitates more homogenous noise
amplification factors with a lower average for both mSENSE and GRAPPA, and residual
aliasing artifacts were reduced for mSENSE. In vivo experiments demonstrated the improved SNR and reduced aliasing artifacts from using ISPM as well.
34
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
a
Single-Slab Excitation
b
35
Dual-Slab Excitation
R
L
Slab-Direction (z)
Figure 4.1: (a) Single-slab excitation excites a large volume including the right and left
breasts and the space between them. (b) A dual-band pulse excites a slab over each breast
simultaneously.
4.1 Dual-Band Excitation in Bilateral Breast MRI
In sagittal bilateral breast imaging, two breast volumes of interest are separated in the
slab direction (z direction). Conventionally, bilateral breast MRI used to be performed by
exciting and imaging the two separated breasts together using a volume-selective excitation
with a combination of a fat-suppression technique as shown in Fig. 4.1a. Recently, dualband spectral-spatial excitation has been suggested as an effective technique for bilateral
breast MRI because two separate slabs can be simultaneously excited without doubling the
RF excitation time [68, 69] (Fig. 4.1b). In addition, the dual-slab excitation can reduce
cardiac motion artifacts, because limited parts of the heart would be excited.
The dual-band pulse (Fig. 4.2) interleaves two flyback spectral-spatial pulses by playing
one on positive gradient lobes and the other on negative gradient lobes [70]. By exciting
each slab on a different gradient polarity, the two slabs can be excited independently. Each
slab can have each own slab location, spectral and spatial profiles, center frequency, linear
shims (see Chapter 6), and slab-phase modulation [10].
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
36
a
RF(t)
sGGGyGGGsGGGyGGGsGGGyGGGsGGGyGGGsGGGyGGGsGGGy
b
Gz(t)
12.4 ms
c
Spatial Profile
-12
-7
0
7
12 [cm]
d
Spectral Profile
-200 -100
0
100 200
[Hz]
Figure 4.2: (a-b) A dual-band spectral-spatial excitation consists of RF subpulses and alternating gradient lobes, where RF subpulses for each slab are applied on a different gradient
polarity. RF subpulses for the left breast (L) and RF subpulses for the right breast (R) are
applied alternatively in time. (c-d) Each RF subpulse determines spatial selectivity, while
the broad RF envelope determines spectral selectivity.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
a
Without ISPM
L
R
Slab-Direction FOV
b
37
With ISPM
L
R
Reduced FOV
Figure 4.3: Without ISPM, the two slabs and the empty space between them should be
encoded together. However, by using ISPM, the two slabs can be imaged as if they were
contiguous, and thus the necessary slab-direction FOV can be decreased.
For simultaneous imaging of the two breasts, 3D phase-encoding must be performed
over the two excited slabs and the empty space between them, which is unnecessary. ISPM
removes the need to encode the empty space by shifting the two slabs close together as if
they were contiguous (Fig. 4.3).
Shifting of a slab can be achieved by applying a linear phase modulation in kz phaseencode planes. By the increasing RF excitation phase by 2π m/(N$z) with kz phase-encode
number, where N is the number of slices within a slab and $z is a slice thickness, the slab
can be shifted by m$z. With the dual-band excitation, the slab-phase modulation can be
independently applied for each slab, thus the slabs can be shifted close together. Normally,
with ISPM the slab-direction FOV and the number of kz planes can be reduced for 2030% for bilateral breast imaging. ISPM can be considered a 3D extension of phase-offset
multiplanar imaging (POMP) [71].
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
38
4.2 Parallel Imaging Applications
By exciting the two slabs simultaneously, parallel imaging can be applied in the slabdirection (right/left direction) robustly by utilizing the sensitivity difference of coil elements in a phased-array. However, for a given data acquisition pattern, the appearance
of aliasing artifacts is different depending on whether ISPM is used or not and affects
reconstructed image quality. This section describes how aliasing artifacts occur with undersampling and how SENSE or self-calibrated parallel imaging methods can be applied
for either case.
Appearance of Aliasing Artifacts and SENSE Reconstruction
Here, we assume that two slabs with a thickness of a are separated by a distance of 0.5a
as shown in Fig. 4.4. Without ISPM the required slab-direction FOV is 2.5a including the
empty space of 0.5a between the two excited slabs. Conversely, with ISPM the required
slab-direction FOV can be reduced to 2a. If we undersample kz with an interval of 1/a,
the two slabs are superimposed with a slab-direction FOV of a. Whether ISPM is applied
or not, two slices, one from each slab, are superimposed at every z location. However, the
appearance of aliasing artifacts differ. When ISPM is not used, the slab centers are not
aligned, resulting in the medial region of one slab (M1 ) overlapping with the medial region
of the other slab (M2 ) and the lateral region of one slab (L1 ) overlapping with the lateral
region of the other slab (L2 ). Here, the boundary between the medial region and lateral
region is determined by the location of the edges of the aliased slab and thus depends on the
slab separation. When ISPM is used, the centers of each slab are aligned, always resulting
in the medial region of one slab overlapping with the lateral region of other slab. Therefore,
ISPM allows edge slices of each slab, where coil sensitivities are hard to estimate correctly,
for not overlapping with the inner region of the other slab.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
39
Aliased Slabs by
Undersampling in kz
Excited Slabs
no ISPM
Acquisition pattern
M1
L1
M2
a
L2
L1,L2 M1,M2
a
a/2
ISPM
a
1/a
M1
L1
L2
M2
a
a
L1,M2 M1,L2
a
*ISPM = indepedent
slab-phase modulation
Figure 4.4: If two slabs, each with thickness of a, are separated by a distance a/2, the total
slab-direction FOV is 2.5a. However, by incorporating ISPM, the required FOV can be
reduced to 2a. If undersampling is conducted in kz with a sampling interval 1/a, two slabs
are superimposed. Without ISPM, the centers of the two slabs are misaligned; with ISPM,
the centers of the two slabs are aligned.
For both cases, SENSE reconstruction with a factor of two unwraps each pair of two superimposed slices [72]. However, g-factors for SENSE reconstruction can vary because of
change in physical locations of the two superimposed slices. In particular, without ISPM,
the superimposed slices from the regions M1 and M2 have a physical separation of a while
the two superimposed slices from the regions L1 and L2 have a physical separation of 2a.
However, with ISPM the separation is constant (1.5a). In general, a higher g-factor is expected with less physical separation because the coil sensitivity difference between the two
superimposed slices would be smaller. Changing g-factors by relocating aliasing artifacts
is also used in the “controlled aliasing in parallel imaging results in higher acceleration”
(CAIPIRINHA) method [73].
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
40
Self-Calibrated Parallel Imaging Applications
When self-calibrated parallel imaging methods are used for acceleration, extra kz planes
need to be acquired in addition to the undersampled kz planes to provide coil sensitivity
information. However, the forms of the calibration line acquisition differ for the cases with
and without ISPM because the original slab-direction FOVs are different. As can be seen
in Fig. 4.5, when ISPM is not used, two additional kz planes should be acquired between
two consecutive undersampled kz planes to estimate alias-free coil sensitivity information.
If only one additional kz plane is acquired between the undersampled planes, the resulting
low-resolution images (with a slab-directon FOV of 2a) still have aliasing artifacts and
are inadequate for calculating coil sensitivity information of the two slabs. However, with
ISPM the acquisition of one additional kz plane between two consecutive undersampled kz
planes is sufficient to calculate coil sensitivity information within the two excited slabs.
From this undersampled data, both mSENSE and GRAPPA can be used to reconstruct
data. As introduced in Chapter 2, mSENSE uses SENSE after estimating coil sensitivity
maps from the central k-space data. For the cases with and without ISPM, mSENSE with
a reconstruction factor two can unwrap the two slabs. When ISPM is not used, estimated
low-resolution coil sensitivity maps have a slab-direction FOV of 3a and need to be rearranged by considering physical slab locations before unaliasing process; however, when
ISPM is used, the coil sensitivity maps have a slab-direction FOV of 2a and need not to
be rearranged. GRAPPA estimates unacquired data in k-space directly. When ISPM is
not used, GRAPPA estimates kz planes with an interval of 1/3a, resulting in a 3a slabdirection FOV in the image domain. When ISPM is used, GRAPPA estimates kz planes
with an interval of 1/2a, resulting in a 2a slab-direction FOV in the image domain.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
41
kz Acquisition Pattern
No ISPM
1/a
1/3a
ISPM
1/a
1/2a
Figure 4.5: For self-calibration, extra central kz planes need to be acquired in addition to
undersampled acquisition with an interval 1/a. When ISPM is not used, central kz planes
need to be acquired with an interval 1/3a. However, when ISPM is used, the central kz
planes with an interval 1/2a is sufficient for coil sensitivity calibration.
4.3 Phantom Experiments
We conducted a phantom experiment to compare parallel imaging quality between the cases
with and without ISPM for both mSENSE and GRAPPA reconstructions. Four different
protocols, without ISPM + mSENSE, without ISPM + GRAPPA, with ISPM + mSENSE,
and with ISPM + GRAPPA, were compared in terms of g-factors and reconstruction artifacts.
4.3.1 Methods
The phantom experiment was conducted with a liquid breast phantom (containing NaCl
and NiCl2 · 6H2O) at a GE 1.5T Excite scanner (GE Healthcare, Waukesha, WI) having a
40 mT/m maximum gradient amplitude and 150 mT/m/ms maximum gradient slew rate.
An eight-channel phased-array breast coil (GE Healthcare, Waukesha, WI) was used for a
signal receiver.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
42
Table 4.1: The magnitudes of coupling coefficients between eight coils in the breast phasedarray. ci,k represents the coupling coefficient between coil i and coil k. There exists coupling between coil 1 and coils 2-4, and coil 8 and coils 5-7, with more than 25% correlated
noise power. Individual couplings between coils 1-4 (located for the right breast) and coils
5-8 (located for the left breast) are minor.






c=





1.00
0.34
0.30
0.26
0.08
0.05
0.06
0.11
0.34
1.00
0.19
0.13
0.14
0.03
0.01
0.03
0.30
0.19
1.00
0.10
0.12
0.04
0.03
0.05
0.26
0.13
0.10
1.00
0.17
0.12
0.14
0.07
0.08
0.14
0.12
0.17
1.00
0.12
0.15
0.26
0.05
0.03
0.04
0.12
0.12
1.00
0.18
0.34
0.06
0.00
0.03
0.14
0.15
0.18
1.00
0.37
0.11
0.03
0.05
0.07
0.26
0.34
0.37
1.00












At first, we measured the noise correlation matrix from the receiver phased-array coil
using noise data acquired by scanning the phantom without RF excitation [9,74]. The noise
correlation matrix was used to combine multiple-coil images for optimal SNR [9,42] and to
synthesize the noise data for experimental g-factor measurement [75]. From the measured
noise correlation matrix (Ψ), coupling coefficients between coil i and coil k were calculated
*
as ci,k = Ψi,k / Ψi,i Ψk,k and the magnitudes are shown in Tab. 4.1. Given that coils 1-4
were located to image the right slabs and the others to image the left slabs, individual
couplings between coils 1-4 and coils 5-8 were minor. This actually means that the noise
correlation matrix might be ignored for optimal SNR when unfolding two pixels from each
slab.
From the localizer image, two slabs were located over the phantom separately, where
each slab was 9.6 cm thick and the slab separation was 5.7 cm. When ISPM was not used,
the two excited slabs, the space between the two slabs, and some outer space were phaseencoded with 192 sagittal slices and 1.5 mm slice thickness. Acquiring 192 slices along
the z axis was to reduce FOV to one slab thickness with an undersampling factor of three
as shown in Fig. 4.5. When ISPM was used, the two excited slabs were encoded with a
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
no ISPM
a
R
ISPM
b
R
L
43
L
Figure 4.6: After exciting two separate slabs over the breast phantom (as shown by yellow
boxes), slab-phase encoding is performed. (a) When ISPM is not applied, the two excited
slabs, the empty space between them, and some outer space are phase-encoded with 192
slices. (b) When ISPM is applied, the two slabs are shifted close together and only the two
excited slabs are encoded with 128 slices. The slice thickness is constant for both cases.
total of 128 sagittal slices and the same slice thickness (Fig. 4.6b). For both cases, a 3D
Cartesian readout was used with a 512 × 128 matrix over 20 × 20 cm2 FOV (in-plane),
30◦ flip angle, RF-spoiling, 22 ms TR and 7.6 ms TE. As the number of slices acquired for
data with and without ISPM is the ratio of 2/3, the expected image SNR from full k-space
*
data is the ratio of 2/3.
To obtain unaccelerated phantom images both with and without ISPM, fully-sampled
individual coil images were combined using the sum-of-squares (SOS) reconstruction by
incorporating the noise statistics in the receiver channel (Ψ) to provide optimal SNR [9].
This SOS reconstruction method can be expressed as
I(x, y, z) =
√
cH Ψ−1 c,
(4.1)
where c is a N x 1 vector where each element c(m) is the signal value for the m’th coil
image at (x, y, z), N is the number of coils, and I is the composite image.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
44
For parallel imaging applications, fully-sampled data were sub-sampled. Outer kz
planes were discarded with a kz interval set to the inverse of one slab thickness. For the data
without ISPM, kz planes were undersampled by a factor of 3 and 20 additional calibration
planes were remained, resulting in 31 central kz planes for estimating coil reconstruction
kernels. For the data with ISPM, the data were undersampled by a factor of 2 and 12 additional calibration planes were remained, resulting in 25 central kz planes. The numbers of
extra calibration planes were chosen to provide the same matrix sizes for fitting GRAPPA
kernels for both data sets.
Parallel imaging reconstruction was performed in each axially reformatted plane (kx ×
kz plane) after a 1D FFT in the y (superior/inferior) direction. For mSENSE reconstruction, low-resolution coil images from the Hamming-windowed central k-space data were
normalized by the square-root of SOS magnitudes and used as coil sensitivity maps. The
sensitivity matrix for each pixel of the aliased images was determined by using the coil
sensitivities at the physical locations of the two superimposed pixels. The SENSE unaliasing matrix was determined by integrating the estimated noise correlation matrix into
the sensitivity matrix [36]. When ISPM was not used, the non-reconstructed space outside
the excited slabs was set to zero. For GRAPPA reconstruction, kernels with size of 5 × 6
(kx × kz ) were determined from 121 × 31 (without ISPM) and 121 × 25 (with ISPM) central
full kx × kz data and used to reconstruct every missing line. Additional calibration kz planes
were used only to calculate reconstruction kernels and were not used for final reconstruction. No regularization was used for both mSENSE and GRAPPA. After reconstruction,
the two slab images were separated using the physical slab locations when ISPM was used.
The reconstructed image quality of the four protocols was compared in terms of reconstruction artifacts and SNR degradation. Reconstruction artifacts were assessed by subtracting fully-sampled SOS images from reconstructed images. The SNR degradation of
each protocol was quantified by estimating g-factors both analytically and experimentally.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
45
The analytical g-factors are calculated using Eqs. 2.14 and 2.17 in Section 2.4. Experimental g-factors were calculated using the “pseudo multiple replica” method [75]. This method
emulates acquisition of image replicas by adding synthesized noise generated from the
noise correlation matrix to the acquired k-space data and measures SNR from the pixel-bypixel evaluation of the mean signal and standard deviation through a stack of reconstructed
images [76]. One hundred sets of replicated k-space data were created and mSENSE and
GRAPPA reconstructions were applied. For each protocol, g-factor maps were calculated
as the ratio of SNR of unaccelerated images and accelerated images, divided by the squareroot of an acceleration factor [36].
4.3.2 Results
In Fig. 4.7, axially reformatted images of the phantom are shown. The unaccelerated images from full k-space data with and without ISPM are shown in Fig. 4.7a and d. When
ISPM was used, the reconstructed slabs were repositioned to the original location. Reconstruction artifacts from the four protocols are compared by subtracting accelerated images
from unaccelerated images as shown in Fig. 4.7b-c and e-f. For mSENSE reconstruction,
as low-resolution sensitivity maps were estimated, sensitivity values can be incorrectly estimated near the slab edges and residual aliasing artifacts can result. When ISPM was not
used, since the edges of one slab were located in inner regions of other slab after undersampling, artifacts were generated in inner regions of both slabs after reconstruction shown
by arrows. When ISPM was used, as the edges of both slabs were aligned, all residual
aliasing artifacts arising from sensitivity errors around the slab edges were located at the
edges of each slab, reducing aliasing artifacts in inner regions. For either case, imperfect
slab profiles with a finite transition width can also cause aliasing artifacts at the edges of
the two slabs. In GRAPPA reconstructed images, artifact levels were low as kernels were
estimated from sufficient calibration areas.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
46
no ISPM
b
a
c
max
max/16
ISPM
d
e
f
0
Full k-space
-max/16
Difference: mSENSE-Full
Difference: GRAPPA-Full
Figure 4.7: Axially reformatted phantom images from full k-space data without (a) and
with (d) ISPM (after reposition), and difference images between unaccelerated images
and accelerated images with mSENSE (b,e) and GRAPPA (c,f). Residual SENSE aliasing artifacts caused by incorrectly estimated sensitivity values are denoted by arrows. For
mSENSE, ISPM results in a more consistent artifact level across the slabs. Note that the
difference images alone cannot be used to compare SNR between different protocols because the images from full k-space data with and without ISPM have different SNR.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
47
Figure 4.8 shows estimated g-factor maps using the phantom images from the same
axial position as Fig. 4.7. Experimental g-factor maps calculated from the mean and standard deviation through 100 sets of reconstructed images show similar patterns to analytical
g-factor maps [75]. If we divide each slab into the medial region and the lateral region
corresponding to where the aliased slab is located as shown in Fig. 4.4, SENSE g-factors
in the medial regions (M1 , M2 ) are higher than in the lateral regions (L1 , L2 ) when ISPM
is not used. This is because of less physical separation between the two superimposed
slices from medial regions of each slab, M1 and M2 , than the two superimposed slices from
lateral regions of each slab, L1 and L2 (See page 39). Thus the coil sensitivity difference
between the two superimposed slices from the medial regions would be smaller and resultant g-factors would be higher. However, with ISPM, each pair of superimposed slices has
the same physical separation (1.6 times of one slab thickness), resulting in a more uniform
sensitivity difference and more homogenous g-factors across the volume. When comparing
the g-factors from SENSE and GRAPPA, some differences are apparent. With GRAPPA
g-factors, higher g-factors are resulted in the medial regions but g-factor variation across
the medial and lateral regions is more smooth than that of SENSE g-factors. In addition,
high g-factors are observed at the two edges of each slab, which might be attributed to
abrupt coil sensitivity changes that can lead to higher magnitude GRAPPA weights with
the limited size of kernels.
A SENSE reconstruction provided mean g-factors of 1.053 (without ISPM) and 1.030
(with ISPM), lower than mean GRAPPA g-factors of 1.351 (without ISPM) and 1.197 (with
ISPM). Higher mean g-factors of GRAPPA than SENSE would be due high g-factors at the
edges of the slab shown in Fig. 4.8. By incorporating ISPM, more homogenous and lower
g-factors were achieved for both SENSE and GRAPPA reconstructions.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
48
Analytical g-Factor Maps
L 1 M1
M2 L2
a
b
c
d
no ISPM
ISPM
2.2
Experimental g-Factor Maps
e
f
0.7
no ISPM
g
h
ISPM
SENSE
GRAPPA
Figure 4.8: The g-factors from the four different protocols are measured using phantom
images analytically (a-d) and experimentally (e-h). Experimental measurements show similar patterns to analytical measurements. Without ISPM, g-factors in the regions M1 and
M2 are generally higher than in the regions L1 and L2 for both SENSE and GRAPPA. By
incorporating ISPM, g-factors are more homogenous.
g-Factors Over Phantom
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
2
49
no ISPM
ISPM
1.5
1
mSENSE
GRAPPA
Figure 4.9: The mean g-factors analytically calculated over the phantom volume are shown
for the four different protocols. Vertical lines represent ± standard deviation. ‘×’ denotes
the 95th percentile of the g-factors. For both mSENSE and GRAPPA, ISPM results in a
lower average g-factor, and lower variation of the g-factor.
4.4 In Vivo Experiments
4.4.1 Methods
Eight volunteers were scanned after acquiring informed consent based on a protocol approved by our institution’s investigational review board. Sixty-four slices were located on
each breast slab with a 1.5-1.9 mm slice thickness depending on the breast size. The slab
separation between the two slabs was between 0.3 and 0.7 times one slab thickness. Like
phantom scans, when ISPM was not used, 192 slices were acquired; however, when ISPM
was used 128 slices were acquired. For each subject, two sets of data, one without and one
with ISPM, were acquired twice to measure SNR using the “difference method” [76, 77].
In this method, the mean signal is estimated from the sum images from the two replicated
scans, and the noise variance is estimated from the difference images of the two scans.
For parallel imaging applications, full k-space data were subsampled in the same manner as phantom data, retaining 31 central kz planes for data without ISPM and 21 central
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
50
kz planes for data with ISPM. Instead of the GRAPPA reconstruction using 2D kernels,
autocalibrated reconstruction for Cartesian sampling (ARC) using 3D kernels [41] was
used. ARC is a GE protocol equivalent to GRAPPA but improves computational efficiency
in calculating kernel weights and synthesizing unacquired data. A kernel with a size of
3 × 3 × 4 (kx × ky × kz ) was determined from 64 × 64 × 31 (without ISPM) or 64 × 64 × 21
(with ISPM) central k-space data. The ARC reconstruction used all central kz planes directly for the final reconstruction.
Reconstruction artifacts and SNR of each protocol were also compared for in vivo data.
Reconstruction artifacts were assessed by subtracting full k-space images from accelerated
images as was done for the phantom images. SNR was measured in a total of 89 ROIs
on breast tissues over the eight subjects by using the difference method. The ROIs were
located at identical physical locations for all protocols. As the number of kz planes used
for the reconstruction including calibration depends on whether or not ISPM was used, the
final metric was SNR efficiency (SNR divided by the square-root of the number of kz planes
used). For each ROI, the percent difference of SNR efficiency between different protocols
was measured. In addition, further analysis was conducted to assess the effects of ISPM
and reconstruction methods on the SNR efficiency with accounting for subject variability
using a mixed-effects regression model [78]. A linear additive model was set up with “use
of ISPM” and “type of reconstruction method” as two fixed factors, and “subject” as a
random factor as follows:
log10 (SNRefficiency ) = β0 + β1 × P + β2 × R.
(4.2)
Here, P = 0 or 1 if ISPM is not used or is used, and R = 0 or 1 if mSENSE or ARC is
used. βi are parameters to be estimated. Statistical analysis was done with Stata Release
9.2 (StataCorp LP, College Station, TX). A p-value of 0.05 was used to indicate statistical
significance.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
51
4.4.2 Results
Figure 4.10 shows axially reformatted breast images from full k-space data and from acceleration using the four protocols. Without acceleration, noise is uniformly distributed across
the image (Fig. 4.10a and d); however, using acceleration, noise level is spatially varying
(Fig. 4.10b-c and e-f). When comparing the unaccelerated images, the image without ISPM
(Fig. 4.10a) shows higher SNR than the image with ISPM (Fig. 4.10d) because it uses more
kz planes. For ARC, noise amplification at the two edges of each slab (Fig. 4.10c and f)
is noticeable as investigated with the phantom images. SENSE artifacts due to incorrectly
estimated sensitivity maps at the edges of the slabs are seen in Fig. 4.10b. Figure 4.11
shows axially reformatted images from another volunteer. The difference images between
accelerated images and unaccelerated images are shown with the unaccelerated images. In
the difference images, artifacts and noise amplification similar to those shown in Fig. 4.10
are again observed.
Figure 4.12 shows the averages of the percent differences in SNR efficiency from 89
ROIs between different pairs of two protocols. The SNR efficiency difference between
ISPM incorporation and no incorporation was calculated separately for mSENSE and ARC
reconstruction. In addition, the SNR efficiency difference between mSENSE and ARC was
calculated separately for cases with and without ISPM. The 95% confidence intervals for
each mean of the difference are depicted. By incorporating ISPM, SNR efficiency was increased by 5% when mSENSE reconstruction was used and 8% when ARC reconstruction
was used. When comparing different reconstruction methods, ARC provided 9% higher
SNR efficiency than mSENSE when ISPM was not used and 12% higher SNR efficiency
when ISPM was used. Higher SNR efficiency with ARC was mainly because central calibration planes were directly used for reconstruction, while they were only used to estimate
sensitivity maps for mSENSE.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
52
no ISPM
a
b
c
e
f
ISPM
d
Full k-space
mSENSE
ARC
Figure 4.10: In vivo images from full k-space data (a,d), mSENSE reconstruction (b,e)
and ARC reconstruction (c,f) are shown. Noise amplification at the edges of the two slabs
from using the ARC reconstruction are denoted by arrows (c,f). Note that signal to noise
level from these images does not reflect SNR efficiency of each method as the numbers of
kz planes directly used for reconstructing these images are not the same. In (b), SENSE
residual artifacts are indicated by arrows. In all images, cardiac motion artifacts are seen
inside the chest wall.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
53
no ISPM
a
b
c
max/4
max
d
ISPM
e
f
0
Full k-space
-max/4
Difference: mSENSE-Full
Difference: ARC-Full
Figure 4.11: In vivo images from another volunteer are shown. The reconstructed images
from full k-space data without and with ISPM (a,d) and difference images between unaccelerated images and accelerated images with mSENSE (b,e) and ARC (c,f) are shown.
Reconstruction artifacts and noise amplification denoted in Fig. 4.10 are observed.
The estimates of the linear mixed-regression model about SNR efficiency are shown in
Tab. 4.2. The 95% confidence intervals of the coefficients corresponding to ISPM application and reconstruction method do not include zero, which indicates that both elements
provide significant effects on SNR efficiency. Incorporating ISPM provides higher SNR
efficiency than not incorporating ISPM (p < 0.01), and ARC provides higher SNR efficiency than mSENSE (p < 0.0001), as shown with the previous analysis. There was high
between-subject variability, accounting for 64% of the remaining variance (95% confidence
interval: 40-89%).
SNR efficiency difference (%)
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
ISPM−no ISPM
no ISPM
54
ARC−mSENSE
mSENSE
16
12
8
4
0
mSENSE ARC
no ISPM ISPM
Figure 4.12: The SNR efficiency difference between ISPM incorporation and no incorporation is calculated separately for the mSENSE and ARC reconstructions. And, the SNR
efficiency difference between mSENSE and ARC is calculated separately for ISPM incorporation and no incorporation. The 95% confidence intervals of the mean differences are
shown.
4.5 Discussion
4.5.1 Improvement of Reconstructed Image Quality
Based on the results, the ISPM technique reduces residual aliasing artifacts and increases
SNR efficiency when self-calibrated parallel imaging methods are used for acceleration.
The ISPM application reduces the number of calibration lines and resulting acquisition
time, and simplifies SENSE reconstructions. In addition, it facilitates more homogenous
g-factors with lower average and peak g-factors for both SENSE and GRAPPA reconstructions. The g-factor improvement is based on the experiments with the eight-channel
phased-array coil, thus some different results could occur with a different coil arrangement.
The SENSE g-factor change achieved by applying ISPM depends on the geometry of
the breast, slab thickness, and slab separation because these affect the appearance of aliased
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
55
Table 4.2: Estimates of the coefficients in the linear regression model to assess effects of
ISPM (P) and reconstruction methods (R) on SNR efficiency.
Variable
Intercept
ISPM (P)
Reconstruction (R)
Parameter
β0
β1
β2
Estimate
-0.144
0.024
0.037
Standard Error
0.043
0.009
0.009
z
-2.81
-2.59
4.02
p > |z|
0.005
0.010
0.000
95% C.I.
[-0.227 -0.060]
[ 0.006 0.042]
[ 0.019 0.055]
C.I. = Confidence Interval
slabs. If ISPM is not used, the centers of the two slabs are aligned, and the g-factors are
similar to those achieved with ISPM only when the slab separation is an integer multiple
of the slab thickness. Otherwise, the medial regions of each slab overlap and the lateral
regions of each slab overlap, and higher g-factors are generated in the medial regions due
to less coil sensitivity difference. In such cases, incorporation of ISPM can be beneficial as
it provides perfect overlap and homogeneous g-factors.
When low-resolution sensitivity maps are used for mSENSE reconstruction, sensitivity
values can be incorrectly estimated around object edges or the air/tissue boundaries, and
residual aliasing artifacts can result. With ISPM those artifacts appear at the edges of each
slab, whereas without ISPM they can be located anywhere in inner regions of the imaging
slabs.
4.5.2 Experimental Data Analysis
To compare the SNR of the different protocols from in vivo data, the difference method
was used. This method assumes that there is no motion or signal change between the
two repeated scans. However, respiratory motion can cause misregistration of the two
images, thus the subtracted images from the two scans may contain artifacts. To reduce any
measurement errors, ROIs were located in artifact-free regions on the subtracted images.
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
56
For data-driven reconstruction, 2D kernels were used for the phantom experiment after the 1D FFT in the y direction, but 3D kernels were used for the in vivo experiments
by exploiting the ARC technique. Computing and applying 3D kernels is computationally
intensive with normal GRAPPA reconstruction, but the ARC technique allows faster reconstruction by synthesizing data in hybrid space. The applications of 2D kernels and 3D
kernels may not yield a large difference in reconstructed image quality if the calibration
area is sufficient. However, calibration data can be exploited more efficiently when estimating 3D kernels in kx × ky × kz space than when estimating 2D kernels in kx × kz space
for each y location.
An SNR comparison between ARC and mSENSE shows that ARC provides higher
SNR efficiency than mSENSE for in vivo measurement. This does not agree with phantom
analysis, which shows lower average g-factors with mSENSE than with GRAPPA. The is
for two reasons: the in vivo analysis from the limited number of ROIs cannot represent the
overall noise amplification of each protocol and can result in some discrepancy between
the phantom and the in vivo analysis; for in vivo ARC reconstruction, including extra calibration planes for final reconstruction provides higher SNR efficiency than is provided by
mSENSE, which uses extra calibration planes only to estimate sensitivity maps.
4.5.3 Image Acquisition Considerations
When applying ISPM, excitation RF phases should be linearly increased with the kz number. At the same time, there are constraints on RF excitation phases to maintain a steadystate of spin magnetization [79]. Only when RF excitation phases are constant, or linearly
or quadratically increasing can the steady state be maintained. Thus, the acquisition order
of kz planes should be properly determined for ISPM to be applied appropriately [10]. For
a normal acquisition scheme, the simple sequential ordering of kz planes can satisfy the
CHAPTER 4. PARALLEL IMAGING WITH SLAB-PHASE MODULATION
57
two constraints; however, for a variable density acquisition scheme, the acquisition order
should be carefully considered.
The dual-band pulse was used to excite only the two breasts without including the
region between them. However, small volumes of breast tissues and lymph nodes might
be located between the two breast slabs, and these may be relevant to breast diagnosis. In
our institution, dual-band excitation and ISPM are only used for dynamic imaging where
high temporal resolution is critical, whereas the two breast volumes in addition to the space
between them are imaged for other T1 and T2 -weighted imaging sequences.
4.6 Summary
In bilateral breast MRI, incorporating ISPM into dual-band excitation can improve image
quality when self-calibrated parallel imaging methods are used for acceleration. ISPM
modifies aliasing artifacts and generates g-factors that are more homogeneous and lower
on average for both mSENSE and GRAPPA reconstructions. For in vivo experiments, the
measured SNR efficiency increased in mSENSE by 5% and in ARC by 8% using ISPM
based on ROI analysis. Residual aliasing artifacts from mSENSE reconstruction were also
reduced with ISPM. Overall, the addition of ISPM to either model-driven or data-driven
parallel imaging methods simplifies reconstruction, increases image SNR efficiency, and
reduces residual artifact levels.
Chapter 5
Rapid Bilateral Breast DCE-MRI
In breast DCE-MRI, attaining both high temporal and spatial resolution is very important for accurate diagnosis. This chapter presents rapid bilateral breast DCE-MRI methods, which combine 3D spiral imaging [80, 81] and parallel imaging. For parallel imaging acceleration, we used adaptive sensitivity encoding incorporating temporal filtering
(TSENSE) [11]. TSENSE allows self-calibrated sensitivity map estimation by using timeinterleaved phase-encode acquisition. To develop TSENSE methods, we applied four different methods of sensitivity map calculation to clinical DCE data and compared residual
aliasing artifacts by analyzing resultant contrast uptake curves qualitatively and quantitatively. Signal intensity curves from the tumors verified temporal resolution improvement
by TSENSE over unaccelerated reconstruction. By incorporating regularization, image
quality degradation in regions where sensitivities are difficult to estimate was minimal.
Overall, TSENSE combined with spiral imaging is very time-efficient, providing 11 s temporal resolution, 1.1 x 1.1 x 3 mm3 spatial resolution, and 184 x 184 x 64 matrix size, and
generates minimal aliasing artifacts. These imaging and reconstruction methods were tailored to bilateral DCE protocols at Stanford Hospital and have been used to scan over 3000
58
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
59
patients. The last section in this chapter presents results with using temporal GRAPPA
(TGRAPPA) [82], which is an alternative reconstruction method to TSENSE.
5.1 Breast MRI at Stanford Hospital
At Stanford Hospital, breast exams have been mainly conducted on two GE 1.5T Excite
scanners (GE Healthcare, Waukesha, WI), one having a 40 mT/m maximum gradient amplitude and 180 mT/m/ms maximum gradient slew rate and the other having a 33 mT/m
maximum gradient amplitude and 120 mT/m/ms maximum gradient slew rate. Either an
eight-channel phased-array breast coil (GE Healthcare, Milwaukee, WI) or a seven-channel
phased-array breast coil (Invivo Corp, Waukesha, WI) was used for each scan. The bilateral
breast MRI protocol includes T1 -weighted high-spatial resolution and high temporal resolution imaging as shown in Fig. 5.1. Usually, pre-contrast high spatial resolution imaging is
conducted for the first six minutes. Afterwards, dynamic imaging during the wash-in period
is conducted for three and half minutes with the intravenous injection of 0.1 mmol/kg GdDTPA at 40 s after starting dynamic imaging. Then post-contrast high-resolution imaging
is performed, followed by the resumption of three and a half minutes of dynamic imaging
during the wash-out period. Post-contrast high-resolution images allow for localization of
possible lesions, and DCE images are used to measure signal enhancement curves. Suspicious lesions are later identified by the biopsy and pathology.
5.2 DCE Imaging Pulse Sequence
This section describes a bilateral DCE pulse sequence, which has been used at Stanford
Hospital for the last three years. Bilateral DCE imaging was performed by simultaneous
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
60
Contrast
Injection
Pre-contrast
High-res (6 min)
170 slices
0.4 x 1 x 1.5 mm3
Wash-in
Dynamic (3.5 min)
Post-contrast
High-res (6 min)
Wash-out
Dynamic (3.5 min)
32 slices 32 slices
1.1 x 1.1 x 3 mm3
170 slices
0.4 x 1 x 1.5 mm3
32 slices 32 slices
1.1 x 1.1 x 3 mm3
Figure 5.1: At Stanford Hospital, T1 -weighted MRI is performed for 19 minutes: 6 minutes of pre-contrast high-resolution imaging, 3.5 minutes of wash-in dynamic imaging, 6
minutes of post-contrast high-resolution imaging, and 3.5 minutes of wash-out dynamic
imaging.
excitation of the two slabs using a dual-band spectral-spatial pulse [68, 69] (Fig. 4.2) incorporating ISPM [10] (see Section 4.1). For each slab, 32 sagittal slices were prescribed
with 3-4 mm slice thickness. A flip angle of 40◦ was used to maximize the signal intensity
difference between normal tissues and enhanced lesions after contrast injection. T1 contrast
was achieved by using RF-spoiling with a gradient-echo sequence [83, 84].
To rapidly acquire data, a 3D “stack-of-spirals” imaging trajectory [80, 81] was used,
which employs phase encoding in the slab direction and spiral trajectories in each phaseencode plane (kz plane). Spiral trajectories consisted of nine spiral interleaves, each with
a 14 ms readout time. For off-resonance correction, low-resolution dual-echo single-shot
spiral acquisitions were performed every full temporal frame (full k-space acquisition) to
calculate field maps. The echo time difference was 4.5 ms, which allows frequencies from
-110 Hz to 110 Hz to be unambiguously resolved without errors due to residual fat [85].
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
61
In total, ten excitations, one for calculating field maps and nine for imaging, were applied
for each kz plane. The in-plane field-of-view (FOV) was set to 20 x 20 cm2 and in-plane
resolution was 1.1 x 1.1 mm2.
There are two constraints in slab-excitation phases when ISPM is employed, as discussed in [10]. First, to maintain a steady state of spins, the RF excitation phases for the
n’th sequence repetition must satisfy the equation: φn = an2 + bn + c. Second, to modulate
the slab, the RF excitation phases must increase linearly with the acquired slab-phase encode number, kz . To meet these constraints with an interleaved stack-of-spirals trajectory,
kz planes formed the innermost loop; i.e., all kz planes for the first spiral interleaf were acquired before moving to the second spiral interleaf (Fig. 5.2). The phase increment angle,
$φ , between the two consecutive kz values satisfied N$φ = 2π m (N: the total number
sections in each slab, m: an arbitrary integer), constraining that shift was a multiple of the
slice thickness. This ensured that the acquisition phases for all spiral interleaves within
a kz plane are the same. Quadratic RF phase increments were added in addition to linear
increments for RF-spoiling.
With a time-interleaved acquisition with a factor of two for TSENSE acceleration, half
of the kz planes were acquired for each temporal frame. Specifically, even kz planes were
acquired for the odd-number temporal frames and odd kz planes were acquired for the evennumber temporal frames. Even though the slab-excitation phases were linearly incremented
within a temporal frame, the linear phase increment constraint broke at the start of the next
temporal frame. Therefore, 16 dummy excitations were applied between different frames
for spins to attain a new steady state before data acquisition. With a 31 ms TR, the scan
time for a temporal frame (half of the kz planes) was 10 s. For the wash-in period, four
temporal frames were acquired before the contrast injection, and total 20 temporal frames
were imaged for 200 s imaging time. Hereafter, this chapter more focus on reconstruction
and data analysis for the wash-in period dynamic imaging.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
Temporal Frame 1
(Even kz planes)
Field Map Imaging Imaging
Echoes Interleaf 1 Interleaf 2
…
62
Temporal Frame 2
(Odd kz planes)
Imaging
Interleaf 9
Field Map Imaging Imaging
Echoes Interleaf 1 Interleaf 2
…
Imaging
Interleaf 9
…
16 Dummy Excitations
kz number: -32 -30 -28 … 26
28
30
-31 -29 -27 … 27 29
31
Figure 5.2: With a TSENSE acceleration factor of two, one half of the kz planes are acquired for each temporal frame, alternating between even and odd planes. To satisfy phase
modulation constraints and to provide an equal phase to each of the spiral interleaves, kz
planes were acquired as the innermost loop. Low-resolution dual-echo spiral acquisitions
for field map calculation are performed before imaging acquisition, i.e. as “interleaf 0.”
Sixteen dummy excitations were applied before acquiring data for the next temporal frame.
5.3 Reconstruction
From the acquired data using 3D spiral imaging and time-interleaved kz acquisition, we performed both unaccelerated and accelerated reconstructions and compared resultant image
quality and signal enhancement curves. This section describes the reconstruction methods.
5.3.1 Unaccelerated Reconstruction
From data undersampled by a factor of two in a time-interleaved manner, unaccelerated
full-FOV images were reconstructed by combining two contiguous temporal frames at ten
time points (Fig. 5.3). Data from each coil was Fourier-transformed in the slab direction
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
63
and then off-resonance was corrected slice-by-slice using linear field maps [33] calculated
from low-resolution dual-echo spiral images. After gridding off-resonance corrected data
to a 256 x 256 matrix, a 2D Fourier transform was applied in the xy plane, yielding a
full-FOV coil image. Multiple coil images were combined by using a square root sum-ofsquares method [9].
5.3.2 TSENSE Reconstruction
Images accelerated by a factor of two were reconstructed from the off-resonance corrected
and gridded k-space data at 20 temporal frames. With a half of the kz planes, aliasing occurs
through the slab direction in the image domain. SENSE reconstruction was applied to each
axially reformatted plane (256 × 64 matrix size) using estimated sensitivity maps.
Sensitivity Map Estimation
A major benefit of the time-interleaved acquisition is that coil sensitivities can be estimated
from the acquired data without any separate calibration scan. Averaging data over multiple
temporal frames can eliminate aliasing artifacts and generate full-FOV images to calculate
coil sensitivity maps. However, the use of data during rapid signal change might not suppress aliasing artifacts perfectly and might yield errors in estimated sensitivity values. In
this work, four different methods of acquiring full-FOV images (Fig. 5.4) were compared,
including (1) averaging the four temporal frames prior to contrast injection, (2) averaging
the last ten frames (late post contrast frames), (3) averaging all temporal frames, and (4)
averaging eight temporal frames in a sliding window fashion. For each approach, the fullFOV coil images divided by the square-root sum-of-squares magnitude of those images
were used as coil sensitivity maps [36]. Static sensitivity maps were calculated for the first
three methods, while time-adaptive sensitivity maps were calculated for the final method.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
64
Unaccelerated Images
Time-Intereaved
kz Acquisition
kz
Frame n-1
Frame n
Frame n+1
Frame n+2
…
P.I.
P.I.
SENSE
* P.I.=Prior Information
P.I.
SENSE
P.I.
SENSE
SENSE
TSENSE Accelerated Images
Figure 5.3: From data with a time-interleaved acquisition with an acceleration factor of two,
unaccelerated full-FOV images were reconstructed by combining two sequential frames.
For TSENSE reconstruction, a regularized SENSE reconstruction was applied for each
temporal frame using a view-share reconstructed image (by combining current frame and
previous frame) as prior information.
For the three static coil sensitivity maps, the residual aliasing artifact level was measured quantitatively from the signal intensity curves over time. Because of the alternating
sign of aliasing artifacts due to the interleaved k-space acquisition, SENSE residual artifacts caused by imperfect sensitivity maps resulted in oscillation in the signal intensity
curves. Assuming the signal intensity curves were smooth aside from the sharp uptake, the
amount of oscillation was measured from the Nyquist frequency. Even though other motion
(cardiac, respiratory or random patient motion) might produce a Nyquist frequency signal,
we also assumed that the predominant signal at the Nyquist frequency arised from residual aliasing TSENSE artifacts. To obtain a frequency spectrum, a Fourier transform was
applied to signal values of the corresponding intensity curve concatenated by the reverse
Signal Intensity
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
Contrast
Injection
0
1) Pre-contrast
65
5
10
15
20
Frame Number
4 Frames
10 Frames
2) Post-contrast
20 Frames
3) All Frames
8 Frames
4) Sliding Window
…
Figure 5.4: Sensitivity maps were estimated from full-FOV image data using four different
methods: (1) averaging the four temporal frames before contrast injection, (2) averaging
the last ten post-contrast frames, (3) averaging all temporal frames and (4) averaging eight
temporal frames in a sliding window fashion (adaptive sensitivity maps). Resultant reconstructed images were compared in terms of residual aliasing artifacts.
of the curve to eliminate the large discontinuity between the first and last temporal points.
The ratio of the mean oscillation amplitude to the mean signal over time was measured as
the ratio of the magnitude of the spectrum at the Nyquist frequency to the DC component
(Fig. 5.5).
SENSE Regularization
Coil sensitivities cannot be estimated correctly in regions where there is low or no signal
such as in fat-suppressed regions or inside the chest wall. In addition, cardiac motion
and the rapid signal change inside the heart after contrast injection make it difficult to
obtain correct sensitivity values in this region. Misestimated coil sensitivities can cause the
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
t
66
t
FFT
DC Component
Nyquist Component
0
Frequency
Figure 5.5: Oscillation in signal intensity curves can be measured using a frequency spectrum. The frequency spectrum is determined using the Fourier transform (FFT) of the
enhancement curve concatenated by the reverse of the curve. The mean oscillation amplitude and the mean signal over time are measured at the Nyquist and DC frequencies,
respectively.
reconstruction process to be ill-conditioned, yielding residual aliasing artifacts and noise
amplification. To improve tolerance to erroneous sensitivities, Tikonov regularization [86]
was used with incorporating a prior knowledge of the solution (prior information). When S
is a sensitivity matrix, c is a vector from aliased coil images, and m is a vector of original
signal (as used in Section 2.4.1), and m0 is a prior information of m, the solution can be
written as
mλ = argmin %Sm − c%22 + λ 2 %m − m0 %22 .
(5.1)
m
It minimizes the residual norm %Sm − c%2 with the side constraint of minimizing %m − m0 %2 .
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
67
As a prior information, a full-FOV image from averaging the current frame and previous frame (view sharing) [87] was used. This view-sharing reconstruction provided interpolated frames at twice the rate of the unaccelerated reconstruction but has the same
temporal resolution as unaccelerated images. Thus the regularization parameter, λ , acted
to trade off between improved image quality (reduction of noise amplification and residual
aliasing artifacts) and high temporal resolution. λ 2 was chosen to be 0.1 so that the prior
information would only influence regions where the sensitivity matrix is ill-conditioned.
Temporal Resolution Decrease by Regularization
The explicit solution of mλ in Eq. 5.1 is
λ
m =
R
∑ (w j
j=1
uHj c
σj
+ (1 − w j )vHj m0 )v j ,
wj =
σ 2j
σ 2j + λ 2
.
(5.2)
Here u j , v j and σ j are the left singular vectors, right singular vectors, and singular value of
S, generated by singular value decomposition (SVD), indexed by j. With an acceleration
factor of two, there are two singular values and the corresponding singular vectors. The
solution of the unregularized SENSE reconstruction, which minimizes the first term in
Eq. 5.1, is written as:
R
munreg =
uHj c
∑( σj
)v j .
(5.3)
j=1
The first term of the SENSE regularization solution is the unregularized solution scaled by
the weighting factor, w j (0 ≤ w j ≤ 1). Similarly, the second term of the SENSE regularized solution is the projection of the prior information onto the j’th right singular vector
scaled by (1 - w j ). The weighting factor w j represents the contribution of the unregularized
solution (Eq. 5.3) to the regularized solution (Eq. 5.2).
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
68
To determine the approximate temporal resolution decrease by regularization, we calculated the mean of w1 and w2 using a phantom consisting of two water bottles, where each
bottle represented a breast. After estimating sensitivity maps, the singular values from the
sensitivity matrices were calculated pixel-by-pixel within the bottles. The mean of two
singular values was 1.14 and 0.83, and the resultant mean of w1 and w2 was 0.93 and 0.87
when the regularization parameter of λ 2 was 0.1. Because the average of the mean w1
and w2 values was 0.9, the influence of prior information in the regularized SENSE reconstruction would be approximately 10%. Then, with 10 s temporal resolution from the
unregularized SENSE reconstruction and 20 s temporal resolution of the view-shared reconstructed images used for prior information, the effective temporal resolution from the
regularized SENSE reconstruction would be approximately 11 s.
5.4 Patient Study
Over 3000 patients have been scanned with this bilateral DCE sequence. All reconstructions were implemented in C and applied automatically after data acquisition. In this section, results from ten patient data are presented, including comparison between unaccelerated and accelerated images and comparison between TSENSE accelerated images reconstructed using the four different sensitivity maps.
Axially reformatted images of a patient from the unaccelerated and TSENSE reconstructions are shown in Fig. 5.6. Sensitivity maps were estimated by averaging the late postcontrast frames. No visible aliasing artifacts are seen in breast regions after the TSENSE
reconstructions; however, the regularization reduces noise amplification and aliasing artifacts inside the chest wall where sensitivity values were hard to estimate correctly. The
SNR between unaccelerated and accelerated images was not directly compared; however,
g-factors measured by phantom images were less than 1.1 in most of the regions where
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
69
Invasive Ductal Carcinoma
a
b
c
Figure 5.6: From a patient data with a large invasive ductal carcinoma, axially reformatted
images from the unaccelerated (a), unregularized TSENSE (b) and regularized TSENSE (c)
reconstructions are shown. With low g-factor values, there is no noticeable SNR decrease
and no visible aliasing artifacts from TSENSE acceleration in most regions except where
sensitivities cannot be estimated. The regularization reduces artifacts and noise amplification inside the chest wall (arrows).
breasts would be located. This indicates that SNR loss resulting from TSENSE acceleration is less than 15% by Eq. 2.13 if estimated coil sensitivity maps are free of artifacts.
In particular, TSENSE reconstruction with the four different sensitivity calculation
techniques was performed on selected ten patient data sets for the wash-in period. Figure 5.7 shows results from one patient data. Figure. 5.7a-b shows two slices reconstructed
from one aliased slice, one slice having an invasive ductal carcinoma and the slice from the
contralateral breast. Signal intensity curves were measured for two ROIs in tumor and one
ROI in normal breast tissue. We can see differences between the curves from the different
sensitivity calibration methods (Fig. 5.7c-f). When comparing those from the three static
sensitivity maps, the curve shapes are similar, while the oscillation amplitudes differ. The
signal intensity curves from the fourth method (adaptive sensitivity maps) have different
shapes than those from the first three methods (static sensitivity maps) during rapid contrast uptake. In the adaptive sensitivity maps, errors occur during the time of rapid signal
change due to imperfect cancellation of aliased regions. These errors in the sensitivity maps
propagate to the reconstructed images and thus to the shape of the resulting signal intensity
curves, notably around the onset of initial uptake and the end of rapid contrast uptake.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
70
ROI 2
a
b
ROI 3
ROI 1
c
Signal Intensity
200
0
d
Pre-contrast
1500
Post-contrast
1500
1
1
2
e
2
3
f
All Frames
1500
1
3
3
1
2
2
3
3
200
200
1
2
Time (min)
Sliding Window
1500
0
1
2
3
Time (min)
2000
1
2
Time (min)
3
0
1
2
Time (min)
3
Figure 5.7: (a, b) show two unwrapped slices from TSENSE reconstruction, which are
FOV/2 apart. A tumor (invasive ductal carcinoma) is shown in the slice (a). (c-f) show
signal enhancement curves on three ROIs (two located on tumor, one located on normal
breast tissue) from TSENSE accelerated images using different coil sensitivities (estimated
by averaging the four pre-contrast frames (c), the last ten post contrast frames (d), all
temporal frames (e), and eight temporal frames in a sliding window fashion (f)). The
arrows in (f) indicate where curve shapes differ from those acquired with other sensitivity
estimation methods if oscillations are ignored.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
71
Quantitative analysis of residual aliasing artifacts was performed on the TSENSE reconstructed images from the three different static sensitivity maps. The ratio of the mean
oscillation amplitude to the mean signal over time was measured using the frequency spectrum of the concatenated signal intensity curves. The ratio was measured in 35 tumor ROIs
and 57 normal breast tissue ROIs using the data from ten patients. The mean values over the
ROIs were calculated in tumors and normal breast tissues separately as shown in Fig. 5.8.
In tumor ROIs, the sensitivity maps calculated from the pre-contrast frames (method 1)
yield the lowest Nyquist residual aliasing artifact ratio (0.50%), followed by those from the
late post-contrast frames (method 2, 0.56%) and those from all frames (method 3, 1.73%).
In normal breast tissue ROIs, the sensitivity maps calculated from the post-contrast frames
(method 2) yield the lowest Nyquist residual aliasing artifact ratio (0.52%), followed by
those from all frames (method 3, 0.63%) and those from the pre-contrast frames (method
1, 1.31%).
Signal intensity curves of various tumors from the unaccelerated images and the TSENSE
accelerated images were compared in several verified lesions. For the TSENSE reconstruction, the sensitivity maps from the late post-contrast frames were used. Figure 5.9a shows
a slice where an enhanced infiltrating lobular carcinoma was detected and the signal intensity curves from the unaccelerated and accelerated images. Figure 5.9b-d shows slices
and curves from patients with an enhancing fibroadenoma, diffuse ductal carcinoma in situ,
and invasive ductal carcinoma, respectively. In all the tumors, higher temporal resolution
from the TSENSE reconstruction allows better characterization of the steep slope during
contrast uptake.
Figure 5.10 shows signal intensity curves from the TSENSE accelerated images, measured for two ROIs located in a large mass of invasive ductal carcinoma, the same lesion
shown in Fig. 5.6. Different enhancement rates from different ROIs are measured: rapid
enhancement in the periphery and slower enhancement in the interior. Although the static
Oscillation Amplitude to Mean Signal
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
72
0.03
Tumor
Breast
0.015
0
Pre−Contrast Post−Contrast
All Frames
Figure 5.8: Comparison of residual aliasing artifacts using the static sensitivity maps from
averaging the four pre-contrast frames, the last ten post-contrast frames and all temporal
frames. The ratio of the mean oscillation amplitude to the mean signal over time was measured from the frequency spectrum of each signal intensity curve. The mean values over 35
tumor ROIs and 57 normal breast tissue ROIs were calculated separately and shown. Error
bars represent ± standard deviation.
post-contrast image shows more detailed morphologic features in the lesion, the dynamic
images provide very detailed spatial information with a factor of 60 improvement in temporal resolution.
5.5 Discussion
By using dual-band spectral-spatial excitation and 3D spiral imaging, the two breast slabs
can be imaged efficiently with fat suppression. In addition, applying TSENSE in the slab
direction allows the two breast slabs to be imaged more rapidly, without much degradation
in image quality. This should provide better quantification of DCE data [88] and characterization of breast lesions. When compared to the VIBRANT sequence (GE Healthcare),
which is a specialized breast sequence by GE, our sequence is four and half times faster
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
ILC
1500
Signal Intensity
a
1100
3
Signal Intensity
1300
Unaccelerated
TSENSE
800
1
2
Time (min)
3
2100
Unaccelerated
TSENSE
1300
500
0
1
2
Time (min)
3
4600
IDC
Signal Intensity
d
2
Time (min)
2900
Signal Intensity
Diffuse
DCIS
1
1800
FA
300
0
c
Unaccelerated
TSENSE
700
300
0
b
73
3400
Unaccelerated
TSENSE
2200
1000
0
1
2
Time (min)
3
Figure 5.9: Comparison of signal intensity curves from unaccelerated images and TSENSE
accelerated images in enhanced lesions of (a) infiltrating lobular carcinoma, (b) fibroadenoma, (c) diffuse ductal carcinoma in situ, and (d) invasive ductal carcinoma. Due to
increased temporal resolution by TSENSE, the steep up-slope of curves can be better depicted.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
a
74
b
ROI 1
ROI 2
c
Signal Intensity
3000
2000
ROI1
1000
0
ROI2
0
1
2
3
Time (min)
Figure 5.10: (a) Dynamic TSENSE slice showing invasive ductal carcinoma and (b) the
corresponding slice from post-contrast high-resolution imaging. (c) Signal enhancement
curves from the two ROIs shown in (a). With our protocol, different enhancement curves
within one lesion can be measured with dynamic images and even more detailed morphologic features of the enhanced lesion can be assessed with high-resolution images.
given the same spatial resolution and imaging FOV. Although we have used spiral imaging,
spectral-spatial excitation and independent slab modulation, this work pertains specifically
to the application of TSENSE, which can be equally well applied to other imaging methods
(Cartesian, radial, etc.) and with other forms of fat suppression and excitation. This section
discusses more about residual aliasing artifacts and effects of SENSE regularization.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
75
5.5.1 Self-Calibrated Sensitivity Estimation and Residual Aliasing Artifacts
The time-interleaved acquisition used in TSENSE provides full-FOV images for estimating
coil sensitivities without any separate calibration scan. However, obtaining artifact-free
full-FOV images from acquired DCE data is challenging due to the signal intensity change
over time. A benefit of TSENSE acquisition is its ability to tolerate subject motion by
using adaptive sensitivity maps from a sliding window average. However, in our DCE
imaging studies, the sliding window method yields errors on sensitivity maps when they
are calculated from a window that includes frames of rapid signal change. As a result, rapid
tumor uptake curves cannot be measured accurately. As the breast coil array is static and
breast motion is not severe, using static coil sensitivity maps might be appropriate. To this
end, we applied several different sensitivity maps, comparing the resulting signal intensity
curves qualitatively and quantitatively.
Three different methods of calculating static sensitivity maps were compared to find the
method that generates the lowest level of aliasing artifacts. In averaging all frames or the
late post-contrast frames, aliasing artifacts from the rapid signal change can generate errors
in estimated sensitivity values. If only pre-contrast frames are averaged, sensitivity maps
are not affected by signal change. However, we found that using pre-contrast frames also
generates residual aliasing artifacts when post-contrast frames are reconstructed. In spiral imaging, cardiac motion manifests as swirling artifacts in each plane, whose intensity
increases with the fast wash-in of the heart; these motion artifacts make it difficult to estimate correct coil sensitivity values. However, to unwrap two superimposed pixels that are
corrupted by motion artifacts, it is actually helpful to use sensitivity values estimated from
images with artifacts. The calculated sensitivity values depend on the real coil sensitivities
at the location of interest and at the source of the motion artifact, and an signal amplitude
of the object at the location of interest and the motion artifact. As motion artifacts become
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
76
intense after contrast injection, appropriate sensitivity matrices to unalias artifact-disturbed
pixels may be different for pre-contrast and post-contrast frames. Therefore, sensitivity
maps from pre-contrast frames can yield residual aliasing artifacts for post-contrast frames.
Quantitative analysis of residual aliasing artifacts using the frequency spectrum of the
concatenated signal intensity curves shows that use of the sensitivity maps from the late
post-contrast frames (method 2) generates the lowest level of aliasing artifacts over all
ROIs (Fig. 5.8). In tumor ROIs, this method yields a much lower aliasing artifact level
than using the sensitivity maps from all temporal frames (method 3) because it does not
include temporal frames of rapid initial enhancement after injection. When compared to
using sensitivity maps from the pre-contrast frames (method 1), using sensitivity maps from
the late post-contrast frames (method 2) gives a similar level of aliasing artifacts in tumor
ROIs but a much lower level in normal breast tissue ROIs. This implies that appropriate
values for sensitivity matrices are more different for normal breast tissue than for tumors
between pre-contrast and post-contrast frames. This is reasonable as the signal level from
normal breast tissue is lower than that from tumors after contrast injection, thus relative
signal change by swirling motion artifacts becomes more significant. Overall, the signal
oscillation amplitude observed by using sensitivity maps from the late post-contrast frames
is very small compared to the mean signal value over time (0.56%), and therefore, it may
not affect characterization of contrast kinetics at all.
5.5.2 SENSE Regularization
Sensitivity maps are difficult to estimate correctly in regions inside the chest wall and as
well as in regions with suppressed fat, which can lead to residual aliasing artifacts or noise
amplification. In the heart, the signal changes very rapidly with fast wash-in and wash-out,
so even combining the late post-contrast frames may yield large errors in estimated sensitivity values. Cardiac motion can also cause errors. Applying background thresholding and
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
77
excluding low-signal regions from the unaliasing process, as suggested in [36], is problematic with our images because thresholding the breasts in images with fat suppression and
motion artifacts is difficult to do accurately and reliably. By incorporating regularization,
SENSE reconstruction can be more tolerant to misestimated coil sensitivities. Tikonov regularization using prior information provides a compromise between an inversion with no
conditioning and an inversion with prior information. The images used for a prior information have the temporal resolution of unaccelerated images, so the regularization trade-off
is between reduced aliasing artifacts and high temporal resolution. We calculated that the
decrease in temporal resolution by the regularization is around 10% compared to true 2x
accelerated images; this decrease occurs when sensitivity values are estimated correctly.
In areas of breast tissue sensitivity values can typically be accurately estimated. However,
inside the chest wall, sensitivity values may be greatly misestimated and the contribution
of prior information would be increased to stabilize SENSE reconstruction.
5.6 Comparison between TSENSE and TGRAPPA
For clinical exams, TSENSE has been used to accelerate dynamic images. However,
TGRAPPA [82] can be used instead of TSENSE to reconstruct undersampled data with
time-interleaved acquisition. TGRAPPA estimates kernel weights instead of sensitivity
maps from acquired data, and applies GRAPPA for each temporal frame to get full-FOV
images. This section compares TSENSE and TGRAPPA in terms of signal enhancement
curves and reconstructed image quality.
On fifteen patient data sets acquired with pulse sequence described in Section 5.2, both
TSENSE and TGRAPPA were applied. After 2D gridding and linear off-resonance correction of spiral data, SENSE and GRAPPA reconstructions were separately performed. Full
k-space reference data for calibration was determined differently for pre-contrast frames
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
78
and post-contrast frames. For the first four frames before contrast injection, reference data
was attained by averaging the first four temporal frames; for the post contrast frames, reference data was attained by averaging the last ten post-contrast frames. For SENSE, full
k-space coil images were normalized by the square-root sum-of-squares magnitudes and
used as sensitivity maps. For GRAPPA, the reference data set was used as a complete set
of auto-calibration signal (ACS) lines to estimate kernel weights. No regularization was
used for SENSE and GRAPPA reconstructions.
Figure 5.11 shows two bilateral unfolded sagittal slices from the TSENSE (a-b) and
TGRAPPA (d-e) reconstructions. Signal intensity curves from two ROIs within a tumor
from the reconstructed images and unaccelerated images are plotted in Fig. 5.11e, f. For
both acceleration methods, the increase in temporal resolution is noticeable by the increased enhancement rate. However, TGRAPPA generates less oscillation in the signal
intensity curves due to a fewer residual aliasing artifacts.
Figure 5.12 shows signal intensity curves from another patient having an infiltrating
lobular carcinoma, as well as axially reformatted images from TSENSE and TGRAPPA reconstructions. Aliasing artifacts from TSENSE can be seen inside the chest wall because of
errors in sensitivity maps. These artifacts are not visible in the GRAPPA reformatted slice.
When comparing the signal intensity curves from the tumor, TGRAPPA again generates
less oscillation than TSENSE.
For all of the 15 patients, TGRAPPA generated less oscillation in DCE curves and fewer
aliasing artifacts inside the chest wall than TSENSE. For TSENSE, correct sensitivityvalue estimation from dynamic images is sometimes challenging due to low signal regions
from fat-suppression, signal intensity changes over time, and spiral motion artifacts. These
issues can all result in localized residual aliasing artifacts. However, for GRAPPA artifacts
due to errors in coil kernel weights are rather spread out over the entire FOV and are less
noticeable.
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
79
TSENSE
b
c 3000
ROI 1
Signal Intensity
a
ROI 1
ROI 2
150 0
ROI 2
Unaccelerated
TSENSE
1
2
3
Time (min)
d
e
f
ROI 1
3000
Signal Intensity
TGRAPPA
150 0
ROI 2
Unaccelerated
TGRAPPA
1
2
3
Time (min)
Figure 5.11: (a-b) Two unwrapped slices using TSENSE from the left breast (a) and the
right breast. (c) Signal intensity curves measured in two ROIs located in the tumor. (d-e)
Two unwrapped slices using TGRAPPA. Signal intensity curves are shown in (f). Curves
from the unaccelerated images are also plotted in (c) and (f).
5.7 Summary
Temporal acceleration methods such as TSENSE and TGRAPPA can be used to improve
temporal resolution in dynamic imaging. By using time-interleaved phase-encode acquisition, full spatial-resolution sensitivity maps can be estimated for TSENSE or full k-space
ACS lines can be made to estimate coil kernel weights for TGRAPPA. In addition, these
self-calibration techniques are less sensitive to subject motion than the techniques that need
extra calibration scans.
For breast DCE-MRI, high temporal resolution is essential for providing accurate characterization of contrast kinetics. First, this chapter has presented rapid bilateral breast
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
b 1500
Signal Intensity
a
Tumor
500
Unaccelerated
TSENSE
TGRAPPA
0
TSENSE
c
80
1
2
Time (min)
3
TGRAPPA
d
Figure 5.12: (a) A sagittal slice reconstructed using TGRAPPA from a second patient. (bc) Axially reformatted slices from TSENSE (b) and TGRAPPA (c) corresponding to the
dashed line in (a). (d) Signal intensity curves from the ROI in the slice (a) from unacceleration, TSENSE and TGRAPPA. The yellow arrow in (c) indicates aliasing artifacts that are
not present in (d).
acquisition methods achieved by TSENSE acceleration with the combination of 3D spiral
imaging and dual-band spectral-spatial excitation. The coil sensitivity difference between
the two breast regions is sufficient to provide very low g-factors with an acceleration factor
of two in the left/right direction. However, accurate sensitivity estimation from acquired
data needs careful consideration due to swirling artifacts in spiral images, rapid signal
change in tumors, blood vessels and the heart after contrast injection, and low or no signal regions of fat and inside the chest wall. Based on our analysis, sensitivity maps by
averaging the late post-contrast frames yield the lowest level of residual aliasing artifacts
and oscillation in signal enhancement curves. SENSE regularization helps to stabilize the
CHAPTER 5. RAPID BILATERAL BREAST DCE-MRI
81
SENSE reconstruction in the regions where accurate sensitivity values are hard to estimate. From these bilateral DCE methods, 11 s temporal resolution and 1.1 x 1.1 x 3.5 mm3
spatial resolution can be achieved over the two breasts. Second, this chapter has also presented patient study results, comparing TGRAPPA with TSENSE. The study demonstrated
that TGRAPPA can generate fewer localized aliasing artifacts and less oscillation in signal
intensity curves than TSENSE.
Chapter 6
Fat Suppression with Independent
Shims
In breast MRI, fat suppression is important for detection of breast tumors because unsuppressed fat can mimic enhanced lesions. However, there exist large field inhomogeneities
within a breast due to susceptibility differences at the air/tissue boundary, the shape of the
breast [89], and the existence of lungs close to the breasts, and these can cause fat suppression failure. Furthermore, the amount of field variation is more significant across the two
breasts. Shimming the magnet with conventional linear shims may not reduce the field inhomogeneity over the two breasts sufficiently. We hypothesizes that if separate shims can
be applied to each breast, then fat suppression over the two breasts would be preformed
equally well as if either one of the breast is imaged. This chapter compars fat suppression
with standard shims and independent shims with incorporating to the dual-band spectralspatial excitation [68] at 1.5T.
82
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
83
6.1 Field Inhomogeneity in Breast MRI
Many fat suppression techniques used in breast MRI exploit resonant frequency differences
between fat and water spins (see Section 3.3). Therefore, when the field inhomogeneity is
large, fat suppression techniques can fail. Figure 6.1 shows regions of fat suppression
failure when the fat-saturation technique [62] (a) and the water-only excitation technique
[68] (b) are used respectively. In both images, the tumor is enhanced, but unsuppressed fat
in the inferior edge of the breast generates signal intensity similar to the bright tumor.
Post-Contrast T1-Weighted Image
T2-Weighted Image
a
b
Tumor
Tumor
Figure 6.1: Fat-suppressed breast images from a patient with tumor. (a) T2 -weighted image with fat saturation. (b) Post-contrast T1 -weighted image with water-selective spectralspatial excitation. The tumor is depicted by a yellow dashed arrow on each image. A field
inhomogeneity causes unsaturated fat in (a) or excited fat in (b), as denoted by purple solid
arrows.
Figure 6.2 shows field maps from a healthy volunteer calculated using a multi-echo
sequence with three different echo times [66]. The experiment was performed using a
GE 3T scanner (GE Healthcare, Waukesha, WI) and an eight-channel phased-array coil;
automated first-order shimming provided by the GE scanner was turned off. Based on
these field maps, the field variation within breasts ranged from -4 ppm to 1.6 ppm. There
existed a dominant linear field variation along the anterior-posterior direction due to the
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
a
84
b
1.6 ppm
1
2
3
-4 ppm
c
Axial Slice at 1
d
Axial Slice at 2
e
Axial Slice at 3
Figure 6.2: Field maps measured from a normal volunteer. One slice from the right breast
(a), and one from the left breast (b) are shown. (c-e) show axially reformatted images
1 +
2 and +
3 in (a-b). There are linear field variations along
corresponding to the positions +,
the anterior-to posterior direction and a strong field inhomogeneity around the inferior edge
of the breasts and nipples. We can expect that the optimal shim values for the left breast
and the right breast would be different.
hemisphere shape of breasts, and a strong field inhomogeneity around the inferior edge of
the breasts and nipples.
Normally scanners provide automated first-order shimming after determining the center frequency and linear shim values for the volume of interest (Fig. 6.3a), and thus reduce
field inhomogeneity. However, due to asymmetry of breasts optimal the center frequency
and linear shim values may be different to each other. Thus, the standard first-order shimming may not reduce field inhomogeneity sufficiently. If the first-order shimming can be
independently applied to each slab (Fig. 6.3b), the field inhomogeneity can be eliminated
to a greater extent, optimally for each slab.
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
Standard Shims
Standard
a
R
b
L
Bx,shim
85
Independent
IndependentShims
Shims
R
L
Bx,shim
x
x
Figure 6.3: Field inhomogeneity can be reduced by shimming the magnet field. With a
conventional first-order shimming (a), the shim magnetic field vary linearly all over the
two excited slabs, while with an independent first-order shimming to each breast (b), the
shim magnetic field can vary arbitrarily linearly within each slab.
6.2 Independent Shims with Dual-Band Spectral-Spatial
Excitation
As described in Section 4.1, the dual-band spectral-spatial pulse [68, 69] interleaves two
flyback spectral-spatial excitation pulses [70], one on positive gradient lobes and the other
on negative gradient lobes. With this time-interleaved excitation, each slab can have its
own spectral and spatial profiles, slab phase, center frequency, and linear shims.
We applied independent shims to each breast slab as presented in [68]. After measuring
the optimal center frequency and linear shim values for each slab from the scanner prescan
routine, the average center frequency and linear shims were applied using the standard firstorder shims provided on the scanner, and difference values were applied by using extra
frequency and gradient waveforms as shown in Fig. 6.4. These frequency and gradient
waveforms provided a constant amplitude on the RF subpulses for one slab (shown by
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
86
solid arrows), and the negative of this value on the RF subpulses for the other slab (shown
by dashed arrows), thus the center frequency and applied gradients alternated between each
slab excitation. Blips between the two different amplitudes ensured that the integral of the
waveform is linearly increasing over time during each slab excitation (Fig. 6.4e).
6.3 In Vivo Experiments
Eleven healthy volunteers were scanned at 1.5T with an eight-channel phased-array breast
coil after acquiring informed consent. For acquisition, flyback EPI [90] with an echo train
length (ETL) of four was used with a 20◦ flip angle, a 20 x 20 cm2 FOV and 256 x 192
matrix (in-plane), half-Fourier, 1.6 mm slice thickness with 64 sagittal slices per breast
(total 128 slices), and RF-spoiling. Independent slab-phase modulation [10] was incorporated into the dual-band pulse to avoid encoding the empty space between the breasts
(Section 4.1).
For each subject, the optimal center frequency and linear shim values along the x, y,
and z axes for each breast were measured separately from the prescan routine after locating
targeted shim volumes. Each shim volume was around 12 x 12 x 12 cm3 . Table 6.3 shows
the optimal center frequency and optimal shim gradient amplitudes of the two slabs for
all eleven subjects. The optimal values varied over the two breasts and between different
subjects.
For each subject, four different scans were conducted by varying shim values as follows:
(1) apply the left breast shim values to both breasts, (2) apply the right breast shim values to
both breasts, (3) apply the average shim values to both breasts (standard shims), (4) apply
the left breast shim values to the left breast and right breast shim values to the right breast
(independent shims).
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
87
a
RF(t)
b
G(t)
(slab select)
c
Gx,y,z(t)
(shim)
d
w(t)
(center frequency)
e
Integration of
Gx,y,z (t) or w(t)
Pauly, ISMRM 2003
Figure 6.4: Using time-interleaved excitation of the two slabs on different polarity of gradient lobes (a-b), different center frequency and linear shims can be applied to each slab
by using extra frequency and gradient waveforms (c-d). By locating blips between the two
different amplitudes, the integral of the waveform can be linearly increasing over time for
each slab excitation (e).
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
88
Table 6.1: The optimal center frequency and linear shim gradient amplitudes along all x, y,
and z axes, measured from the scanner, are shown for each left and right breast. The center
frequency is in Hz and the linear shim gradient amplitudes are in Hz/cm. The x axis is in
the right/left direction, the y axis is in the anterior/posterior direction, and the z axis is in
the superior/inferior direction.
Subject
1
2
3
4
5
6
7
8
9
10
11
Center Freq
60486
60471
60577
60477
60527
60630
60482
60419
60474
60510
60508
Left Slab
x Grad y Grad
0.65
16.25
-1.3
9.1
-6.5
12.35
1.3
13.65
-5.2
13.65
-7.15
12.35
1.3
16.25
2.6
18.2
2.6
9.1
-1.95
14.95
-3.25
18.2
z Grad
3.90
4.55
-5.85
4.55
0.65
-4.55
3.90
2.6
0.65
-3.9
-9.1
Center Freq
60455
60456
60529
60438
60500
60578
60447
60439
60465
60478
60462
Right Slab
x Grad y Grad
-4.55
18.2
-3.9
13.65
-1.95
13
-4.55
20.8
-1.95
16.25
0.65
12.35
-5.2
21.45
-3.9
20.15
-5.2
11.05
-3.25
15.6
-4.55
16.25
z Grad
0
2.6
-7.15
4.55
-5.85
-4.55
3.25
0
2.6
-1.95
-5.2
Figure 6.5 shows axially reformatted images close to the inferior edge of the breasts
from one volunteer with the four different shim methods. When the optimal left shim
values or the optimal right shim values were used for both breasts (a, b), the ipsilateral
breast achieved excellent fat suppression, while the contralateral breast had regions with
unsuppressed fat or suppressed water shown by arrows. When average shim values were
used for both breasts (c), failure of fat suppression were still seen in the right breast. With
independent shims (d), the fat suppression in each breast was as good as the ipsilateral
breast when optimal shim values for the left or the right breasts were used. Figure 6.5e-g
shows sagittal slices from the right and left breast. When standard shims were used, fat
suppression failed in the inferior edge of the breast (circles) and the images show partiallysuppressed water and unsuppressed fat signal; in contrast, the independent shims provided
excellent fat suppression in those regions.
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
a
Left Breast Shims
R
L
b
L
d
Right Breast Shims
R
L
Standard Shims
c
R
R
f
L
L
Independent Shims
Standard Shims
e
89
R
L
Independent Shims
g
R
h
L
Figure 6.5: Volunteer breast images from the four different shim methods. Axially reformatted images from the left breast shims (a), right breast shims (b), standard shims (c), and
independent shims (d) are shown. The regions of fat suppression failure are denoted by
arrows (solid purple: unsuppressed fat, dashed blue: suppressed water). Sagittal images
from the left and right breasts (corresponding to the dashed lines in (a)) demonstrate that
independent shims (g-h) achieve better fat suppression in the inferior edges of the breasts
(circles) than the standard shims (e-f). The dashed line in (e) corresponds to the axial
position of the images (a-d).
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
90
Images from another volunteer are presented in Fig. 6.6. Similarly, independent shims
improved fat suppression in regions of large field inhomogeneity. In eight out of the eleven
subjects, fat suppression was visually better with independent shims than with standard
shims over all 128 slices, and only three subjects showed no difference.
6.4 Summary
For bilateral breast MRI, uniform fat suppression over the two breasts is difficult due to
large field inhomogeneities. Many scanners provide automatic first-order shimming to reduce field inhomogeneities; however, the optimal shim values for the right and left breasts
can be different. This chapter has presented how the first order shimming can be applied
independently to each breast during dual-band spectral-spatial excitation and showed in
vivo images, comparing fat suppression with different shim values. Our results demonstrated that independent shims allow each breast to be optimally shimmed and provide
more uniform fat suppression over all 3D volume than standard shims. With the reduction
of artifacts such as unsuppressed fat or suppressed water, breast tissue can be delineated
better. This independent shim method does not need extra time to apply other than time to
measure shim values for two slabs, and thus is clinically useful.
CHAPTER 6. FAT SUPPRESSION WITH INDEPENDENT SHIMS
Right Breast Shims
Left Breast Shims
a R
L
b R
Standard Shims
c R
R
L
f
L
L
Independent Shims
d R
Standard Shims
e
91
L
Independent Shims
g
R
hL
Figure 6.6: Breast images from another volunteer. Axially reformatted images from the
four different shim methods (a-d) and sagittal images from the standard shims (e-f) and
independent shims (g-h) are shown. Here again, unsuppressed fat is denoted by the solid
purple arrows, and the independent shims provide more uniform fat suppression than the
standard shims in the inferior edges of the breasts (circles). Dashed lines indicate the two
sagittal positions and one axial position of these images.
Chapter 7
Flow Artifact Reduction Using Partial
Saturation
In RF-spoiled gradient-echo imaging, which is normally used for DCE imaging to provide
T1 contrast, unsaturated flowing spins can generate inflow enhancement and pulsatile ghost
artifacts. For T1 quantification of blood or tissue, these flow artifacts can degrade accuracy.
This chapter provides background information about RF-spoiled gradient-echo sequences
and demonstrates that flow artifacts can be reduced by applying partial saturation. Partial
saturation can drive flowing spins to a steady state before they enter the imaging slab, and
thus flow artifacts can be diminished. The ability to generate pure T1 contrast from flowing
spins was verified using magnetization simulations, and phantom and in vivo experiments.
92
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
93
7.1 RF-Spoiled Gradient-Echo Sequences
A gradient-echo sequence is consisted of a single RF excitation pulses, spoiler gradients,
and imaging gradients (Fig. 7.1). Steady-state magnetization can be achieved if each sequence repetition meets the following conditions: constant flip angle; constant TR; constant, or linearly or quadratically increasing, RF phase; and constant dephasing by gradients [79]. The steady-state signal contrast is determined by T1 , T2 , TR, TE, and flip angle.
When the TR is much longer than T2 (TR , T2 ), the transverse magnetization decays to
zero before the next excitation pulse; therefore, the steady state signal does not depend on
T2 , and T1 -weighted contrast is achieved. This steady-state signal amplitude immediately
after excitation is described by the Ernst formula [91] (Eq. 7.1):
Mss = sin(θ )
1 − exp(−T R/T1 )
,
1 − cos(θ ) exp(−T R/T1 )
(7.1)
where θ is a flip angle.
RF
TR
Gradients
Imaging
Gradients
Spoiler Gradient
Figure 7.1: Gradient echo sequences primarily consist of a single RF excitation pulse,
imaging gradients, and spoiler gradients.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
94
When the TR is comparable to or shorter than T2 , the transverse magnetization can no
longer be neglected, even when strong spoiler gradients are used. Sequential RF pulses
partly refocus transverse magnetization, leading to spin echoes, stimulated echoes, and
higher-order echoes. Therefore, the resultant steady-state signal is a composite of free
induction decay (FID) and echoes [92], and the steady-state signal depends on both T1 and
T2 .
Several strategies have been developed to spoil transverse magnetization and to provide
pure T1 contrast for short TR sequences [83, 84, 93–95]. Among them, RF-spoiling, which
combines RF phase cycling with spoiler gradients [83, 84], has emerged as the most commonly used solution. RF-spoiling increases the phases of sequential RF pulses quadratically, where the phase difference between two consecutive RF pulses is linearly incremented. This alters the contribution of each echo to the steady-state signal. With certain
phase difference increments, the steady-state signal is dominated by the FID because other
echoes add destructively and mostly cancel. Conventionally, a phase increment angle of
117◦ has been used since it effectively minimizes residual transverse magnetization for a
wide range of T1 , T2 , and flip angles [83], but other angles such as 84◦ [96], 96.5◦ [97],
50◦ [98], or 150◦ [99] are also suggested to provide better convergence properties or more
accurate T1 estimation. Figure 7.2 shows the magnitude of the steady-state signal as a
function of phase increment angle, simulated using the Bloch equations.
RF-spoiled gradient-echo sequences are commonly used for DCE imaging, where repetitive rapid T1 imaging is necessary to measure passage of the contrast agent in tissue. Signal
intensity can be directly used for estimating contrast concentration because contrast concentration is proportional to 1/T1 of tissue. For perfusion quantification, the arterial input
function (AIF), which is a time history of contrast concentration in the artery of interest,
should be known. However, fresh and unsaturated blood that enters the imaging slab can
generate flow artifacts and cause errors in the AIF measurement. Two detrimental flow
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
95
Steay-State Signal
0.16
0.12
0.08
0.04
0
30
60
90
120
150
Phase Increment Angle (degree)
180
Figure 7.2: Simulations of steady-state signal as a function of phase increment for RFspoiled gradient-echo sequence. Parameters used are T1 /TR = 50, T2 /TR = 10, and flip angle
= 30◦ . The constant dashed line represents the amplitude with ideal spoiling (determined
by the Ernst formula).
artifacts are inflow enhancement [12] and pulsatile ghost artifacts [13,14]. Inflow enhancement occurs at entry slices, generating signal intensity higher than pure T1 contrast, and
can lead to overestimation of contrast concentration. Pulsatile ghost artifacts arise from
view-to-view signal fluctuations caused by entrance of different amounts of unsaturated
magnetization over a cardiac cycle, hiding anatomical details.
7.2 Partial Saturation of Flowing Spins
Flow artifacts can be greatly reduced by partially saturating a slab upstream to the imaging
slab [100] as shown in Fig. 7.4. If identical flip angles and phase cycling are used for both
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
a
Inflow Enhancement
MRA, ed. Potchen et el., 1993
b
96
Pulsatile Ghost Artifacts
Felmlee et el., Radiology 1987
Figure 7.3: When fresh, unsaturated blood enters the imaging slab, it can cause artifacts
such as inflow enhancement (a) or pulsatile ghost artifacts (b). These artifacts can degrade
data quantification.
the partial saturation slab and the imaging slab, flowing spins can be driven to the steady
state while in the partial saturation slab without generating net signal.
When applying partial saturation to eliminate flow artifacts, there are several key considerations. First, the partial saturation slab should be thick enough for flowing spins to
experience a sufficient number of excitations to reach a steady state. Thus the partial saturation slab thickness should be determined with considering the flow velocity. Second,
no signal should arise from the partial saturation slab during data acquisition. Therefore,
spoiler gradients should be applied not only to provide RF-spoiling but also to avoid signal
generation from the partial saturation slab during data acquisition.
The pulse sequence we used is shown in Fig. 7.5. A minimum-phase RF pulse with
a 3.2 ms duration and time-bandwidth product (TBW) of 20 was used for both the partial
saturation RF pulse and the imaging RF pulse. Identical flip angles were used for the partial
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
97
Flow
Partial Saturation Slab
Imaging Slab
Figure 7.4: A partial saturation slab is located upstream exterior to the imaging slab to
drive flowing spins to the steady state.
saturation RF pulse and the imaging RF pulse; and identical phases were used for both RF
pulses, with a quadratic phase increase (117◦ phase increment angle) over each sequence
repetition.
Spoiler gradients were located at two different times for each sequence repetition, at the
end of the partial saturation RF pulse and at the end of the readout gradient. Immediately
after the partial saturation RF pulse, spoiler gradients were applied along the x, y, and z axes
simultaneously to maximize dephasing in the presence of a possibly nonuniform slab phase
profile and high spatial frequency components in the objects along any axis. Following the
readout gradient, a spoiler gradient was applied along the x axis to exploit the unbalanced
part of the preceding readout gradient for spoiling.
Signal generation and dephasing over time by this sequence was plotted using phase
graphs [101, 102] with ignoring imaging gradients. To simplify the phase graphs, we assumed that the spoiler gradient is applied only along one axis instead of all three axes at the
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
Partial Saturation RF
98
Imaging RF
RF
Gz
Acquisition
Gx
Gy
Spoiler Gradient
Spoiler Gradient
TR
Figure 7.5: For each sequence repetition, a partial saturation pulse and spoiler gradients are
applied prior to the standard RF-spoiled gradient-echo sequence. The partial saturation RF
pulse and the imaging RF pulse are phase-cycled equivalently. Spoiler gradients located
immediately after the partial saturation RF pulse and after acquisition to ensure that no
echoes are formed from the partial saturation slab during acquisition.
end of the partial saturation RF pulse, the spoiler gradients at the partial saturation RF pulse
and at the end of the readout gradient are applied for TR/2 - TRF and generate a phase twist
of −π to π radians across each voxel respectively, and data acquisition is performed instantaneously at TR/2 + TRF (Fig. 7.6a). Denoting Fn as a dephased transverse magnetization
state with n cycles of phase twist (i.e., 2π n) within a voxel, Fig. 7.6b and c present phase
graphs for the imaging slab and partial saturation slab respectively. The graphs demonstrate
that at the acquisition time, the magnetization in the imaging slab consists of magnetization states with an even number of n (..., F−4, F−2 , F0 , F2 , F4 , ...) where nonzero net signal
is produced due to the F0 state; however, the magnetization in the partial saturation slab
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION
99
consists of magnetization states with an odd number of n (..., F−3, F−1 , F1 , F3 , ...), resulting
in no signal. These magnetization pathways do not depend on the RF phase, and RF phase
cycling only alters the amplitude of each magnetization substate. As far as each spoiler
gradient generates an integer cycle of phase twist within a voxel, the magnetization in the
partial saturation slab only consists of dephased states at the acquisition time, generating
no net signal.
Figure 7.7 shows voxel signal evolution calculated using the Bloch equations from
both the imaging slab and partial saturation slab over two hundred sequence repetitions
a at the end of partial saturation RF pulse, +
b at the end
at three different time points (+
c at the end of the sequence). Tissue or scan parameters of
of imaging RF pulse, and +
T1 /T2 = 1000/200 ms, flip angle = 20◦ , and TR/TE = 20/4 ms were used. The left and right
columns present signal evolution without and with RF phase cycling (117◦ phase increment
angle), respectively. At the end of the imaging RF pulse, no signal occurs from the partial
saturation slab for both cases because the transverse magnetization only consists dephased
magnetization states at the point between spoiler gradients. In comparison, it is important
to note that some nonzero steady-state signal remains at the end of the sequence from the
partial saturation slab even with RF phase cycling (dashed arrow). This indicates why a
spoiler gradient after data acquisition is usually necessary to avoid signal generation from
the partial saturation slab.
7.3 Magnetization Simulation
We first calculated theoretical magnetization evolution due to the pulse sequence in Fig. 7.5
using the Bloch equations for stationary spins and for flowing spins with and without partial
saturation RF pulses. Simulation ranges included an imaging slab with a 9.6 cm thickness
(32 slices), a partial saturation slab with 5-15 cm thickness when partial saturation is used,
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 100
a
Partial Saturation RF
Imaging RF
Acquisition
b
TR
a
TRF
b
c
TR/2-TRF
Imaging Slab
F6
F4
F2
0 t
F-2
F-4
F-6
I.RF1
t=0
c
I.RF2
t=TR
I.RF3
t=2TR
I.RF4
t=3TR
t=4TR
Partial Saturation Slab
F8
F6
F4
F2
0 t
F-2
F-4
F-6
P.RF1
P.RF2
t=0
t=TR
P.RF3
t=2TR
P.RF4
t=3TR
t=4TR
* I.RF = Imaging RF, P.RF = Partial saturation RF
Figure 7.6: (a) Simplified pulse sequence using RF pulses and spoiler gradients. Spoiler
gradients are located at two different times. The phase evolution from a voxel for the
imaging slab (b) and the partial saturation slab (c) is presented over the first four sequence
repetitions. The acquisition points are denoted as dashed vertical lines. At those points, the
imaging slab generates a signal consisting of FID and refocused echoes, but the partial saturation slab does not generate any signal because magnetization consists only of dephased
states with an odd number of n (..., F−3, F−1 , F1 , F3 , ...).
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 101
No Phase Cycling
Quadratic Phase Cycling
Signal Intensity
a After Partial Saturation RF
0.3
0.2
0.1
0
−0.1
Imaging Slab
Partial saturation slab
0
50
100
150
200
0.3
0.2
0.1
0
−0.1
0
Imaging slab
Partial saturation slab
50
100
150 200
Signal Intensity
0.3
0.2
0.1
0
−0.1
Signal Intensity
b After Imaging RF (At aquisition)
0.3
0.2
0.1
0
−0.1
Imaging slab
Partial saturation slab
0
50
100
150
200
0.3
0.2
0.1
0
−0.1
0
Imaging slab
Partial saturation slab
50
100
150 200
c End of Sequence
Imaging slab
Partial saturation slab
0
50
100
150 200
Sequence Repetition
0.3
0.2
0.1
0
−0.1
Imaging slab
Partial saturation slab
0
50
100 150 200
Sequence Repetition
Figure 7.7: Signal evolution from a voxel at three different places over the sequence (dea +
b and +
c in Fig. 7.6a). The left and right columns show signal evolution both
noted as +,
without and with RF phase cycling, respectively. The solid arrows indicate that no signal
occurs from the partial saturation slab during data acquisition. Note that some nonzero
c from the partial saturation
signal is generated at the end of the sequence (at the point +)
slab even with RF spoiling (dashed arrow).
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 102
and extra regions with 4 cm thickness upstream and downstream to those slabs. For each
voxel, 40 x 40 isochromats were located and simulated over the xz plane, which excludes
the y dimension. Phase-encode and readout gradients were not incorporated, and the x and
z spoiler gradients immediately after the partial saturation RF pulse were applied whether
the partial saturation RF pulse was turned on or not.
Motion due to flow was simulated by spatially shifting the magnetization vectors at the
end of each TR based on the velocity, and filling empty spots with full magnetization. With
this simplification, flow effects during RF excitation [103,104] and second order dephasing
due to gradient spoilers [12] were neglected. Other parameters were T1 /T2 = 1000/200 ms
(close to those of blood at 1.5T), TR/TE = 20/4 ms, and flip angle = 20◦ .
Magnetization evolution is shown in Fig 7.8 for stationary spins and for flowing spins
with and without partial saturation of a 10 cm thick slab. For flowing spins, magnetization
of steady and periodic flow was compared for both with a mean velocity of 10 cm/s. Figure 7.8a, c, e, h, and j shows magnetization of isochromats immediately after applying the
imaging RF pulse for the first 200 sequence repetitions. From 2D magnetization simulations, magnetization along the z axis at a particular x location is displayed. The images on
the right of those images show the net magnetization obtained by averaging isochromats in
each voxel.
For stationary spins (Fig. 7.8a), individual spins do not attain a single steady state due
to quadratic RF phase cycling; instead, the magnetization becomes periodic and reaches a
“pseudo steady state” [105]. The net magnetization provides a uniform signal profile along
the z axis (Fig. 7.8b).
For steady flow without partial saturation, unsaturated spins at entry slices generate
high magnitude magnetization, leading to inflow enhancement at entry slices, and approach
a steady-state level as they move into the imaging slab (Fig. 7.8c-d). By applying partial
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 103
Net Magnetization
Magnetization
No Flow
b
z Location (cm)
−15
Velocity (cm/s)
a
g
−10
−5
Pulsatile Flow
Flow Waveform
30
20
10
0
0
40
80
120 160 200
Sequence Repetition
0
5
40
80
120
160
200
0
Sequence Repetition
40
80
120 160 200
Sequence Repetition
Net Magnetization
i
−15
z Location (cm)
0
Magnetization
h
−10
Steady Flow
d
z Location (cm)
−15
−10
−5
0
5
k
−15
5
Flow
e
f
−15
−10
z Location (cm)
−5
j
0
z Location (cm)
c
−10
−5
0
5
−5
0
0
40
80
120
160
200
0
40
Sequence Repetition
5
0
0
40
80
120
160
200
Sequence Repetition
0
40
80
120
0.8
80
120
160 200
Sequence Repetition
0
0.32
160 200
Sequence Repetition
Figure 7.8: Magnetization and voxel signal evolution over 200 sequence repetitions (over
4 s with TR = 20 ms) are shown for stationary spins (a-b), for steady flow without partial
saturation (c-d) and with partial saturation (e-f), and for pulsatile flow without partial saturation (h-i) and with partial saturation (j-k). The flow is in the positive z direction, and
the pulsatile flow waveform used is shown in (g). For stationary spins, each spin attains
a periodic pseudo steady state (a); however, the net magnetization of each voxel reaches a
steady state (b). Inflow enhancement is greatly reduced by partial saturation for both steady
and pulsatile flow (f,k). Slight signal fluctuation from sequence to sequence remains with
pulsatile flow (k).
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 104
saturation, spins are partially saturated before entering the imaging slab so that inflow enhancement is almost eliminated (Fig. 7.8e-f). No signal results from the partial saturation
slab. Small transient effects occur at the entry regions of the imaging slab, caused by the
RF magnetic field change at the intersection of the partial saturation slab and the imaging
slab.
For periodic flow, pulsatile flow was mimicked using a half-cosine velocity waveform
for systole and zero velocity for diastole over each period (Fig. 7.8g), assuming a heart rate
of 60 beats per minute with a mean velocity of 10 cm/s. All of the spins were assumed to
have the same velocity at certain time points. The number of unsaturated spins entering the
imaging slab varies with the flow velocity, resulting in significant signal variation over time
(Fig. 7.8h-i). However, when partial saturation is used, the inflow enhancement and signal
variation over time are greatly reduced (Fig. 7.8j-k) even though some signal fluctuation
from repetition to repetition remains (Fig. 7.8k).
In Fig. 7.9a, signal profiles from simulations in Fig. 7.8 are plotted together. For no
flow and steady flow, the signal profiles from the last sequence repetition are used because
the net signal from each voxel reaches a single value. However, for pulsatile flow, averaged
profiles over the fourth cardiac cycle are plotted to show the mean signal level from signal
fluctuation. The signal profiles show that inflow enhancement is greatly reduced by using
partial saturation for both steady flow and pulsatile flow. With steady flow, slight signal
fluctuation can be seen in the entry slices due to a change in the experienced RF field. With
pulsatile flow, the oscillation effects and fluctuation are averaged out within a cardiac cycle,
and a more uniform flow profile can be achieved, resulting in signal intensity close to that of
stationary spins. In Fig. 7.9b, the signal profiles from pulsatile flow by varying the partial
saturation slab thickness are compared with the profiles from no flow and pulsatile flow
without partial saturation. By using a partial saturation thickness of at least 10 cm, inflow
enhancement is almost completely eliminated for flow with the mean velocity of 10 cm/s.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 105
b
Signal by Varying Flow
0.3
Stationary
Steady flow without p.s.
Steady flow with p.s.
Pulsatile flow without p.s.
Pulsatile flow with p.s.
Net Magnetization
0.25
0.2
0.15
0.1
Stationary
Flow without p.s.
Flow with 5cm thick p.s.
Flow with 10cm thick p.s.
Flow with 15cm thick p.s.
0.25
0.2
0.15
0.1
0.05
0.05
0
−10
Signal by Varying Saturation Slab Thickness
0.3
Net Magnetization
a
-5
0
5
z Location (cm)
10
0
−10
−5
0
5
10
z Location (cm)
* p.s. = partial saturation
Figure 7.9: (a) Signal profiles for no flow, steady and pulsatile flow without partial saturation, and steady and pulsatile flow with a partial saturation slab with 10 cm thickness. For
pulsatile flow, signal fluctuation is averaged over one cardiac cycle. (b) Signal profiles for
no flow, pulsatile flow without partial saturation, and pulsatile flow with partial saturation
slab thickness of 5 cm, 10 cm and 15 cm. By partially-saturating a slab at least 10 cm thick,
the profile from flowing spins becomes similar to that of stationary spins.
7.4 Scan Experiments
To validate simulations and applicability to in vivo imaging, we conducted experiments
on a 1.5T scanner. First, dephasing of the partial saturation slab by spoiler gradients was
checked using a phantom. And then, flow phantom and in vivo experiments were conducted. Flow phantom experiments enabled us to verify partial saturation by varying T1 of
the fluid and velocity. For in vivo experiments, arteries from various body parts including
the neck, leg and abdomen were tested.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 106
7.4.1 Dephasing of Partial Saturation Slab
By using a standard GE birdcage head coil (GE Healthcare, Waukesha, WI) and a resolution
phantom, the imaging slab and partial saturation slab (both 10 cm-thick) were prescribed
and encoded together by extending the slab-direction imaged FOV to 30 cm (Fig. 7.10a).
The coronal reformatted image from the large-slab FOV scan to encode both the imaging
slab and partial saturation slab is shown in (Fig. 7.10b). The signal profile measured along
the dotted line indicates that spins from the partial saturation slab generated no net signal
(Fig. 7.10c).
7.4.2 Flow Phantom Experiments
Flow phantom experiments were conducted by using an electric pump that generates pulsatile flow in a connected pliable tube with a 6 mm inner diameter. The long tube penetrated
a bottle filled with agar and the immobilized part of the tube within the bottle was used for
flow analysis. The mean flow velocity was set at 10 cm/s, and the temporal portion of systole to diastole for periodic flow was set at 1:1. Liquid relaxation properties were varied by
doping pure water with Gadoteridol (Gd-DOPA). The first mixture had a Gd-DOPA concentration of 0.09 mmol/kg with measured T1 /T2 of 1160/550 ms, and the second mixture
had a concentration of 0.25 mmol/kg with measured T1 /T2 of 520/360 ms.
For each of two Gd-DOPA mixtures, three different flip angles of 10◦ , 20◦ , and 30◦
were tested, and for each case, three different scans were conducted: (1) with no flow and
no partial saturation; (2) with flow and no partial saturation; and (3) with both flow and
partial saturation. Other acquisition parameters were TR/TE = 20/4 ms, 20 x 20 cm2 FOV,
192 x 192 in-plane matrix size, a 10.8 cm-thick imaging slab throughout 36 slices, and a
16 cm thick partial saturation slab.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 107
Partial Saturation Slab
a
Imaging Slab
b
Signal Intensity
c 800
600
400
200
0
5
10
15
20
z Location (cm)
25
Figure 7.10: Phantom experiment to examine dephasing of the partial saturation slab. (a)
denotes the locations of the imaging slab and partial saturation slab with each 10 cm thick
on a localizer image. (b) Coronal reformatted image from the 30 cm slab-direction imaged FOV after applying partial saturation. (c) draws a signal profile along the dashed
line denoted in (b), demonstrating that the partial saturation slab is dephased completely,
generating no net signal.
Figure 7.11 shows coronal maximum intensity projection (MIP) reformatted images of
the tube when Gd-DOPA doped water (T1 /T2 = 520/360 ms) and a 30◦ flip angle were
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 108
a
b
c
d
Tube
Flow
Figure 7.11: MIP images of the tubes with flow. Here phase encoding along two axes is
performed by acquiring all ky lines before moving to the next kz plane. (a) Axial phantom
image with a tube for flow. (b-d) show maximum intensity projection (MIP) coronal reformatted images of the tube for stationary water, flowing water without and with partial
saturation. Partial saturation reduces inflow enhancement as well as ghost artifacts that are
visible in (c).
used. MIP images demonstrate that partial saturation greatly reduced inflow enhancement
and pulsatile ghost artifacts.
Figure 7.12 compares signal intensity variations over the 32 slices (after discarding 4
slices at the 2 edges of the slab) along the tube for no flow, flow without partial saturation,
and flow with partial saturation. Signal intensity was measured in an ROI consisting of
26 pixels, with an area of 0.16 cm2 . From the two different Gd-DOPA mixtures and three
different flip angles, the measured signal levels were different; however, for all six cases,
partial saturation provided uniform signal intensity across the slab for the flowing spins,
achieving a signal level similar to that of stationary spins.
7.4.3 In Vivo Experiments
After acquiring informed consent, we scanned healthy volunteers to test partial saturation
in the carotid arteries, abdomen aorta, and femoral arteries. Specifically, we performed
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 109
Gd-DOPA Doped Water with T1/T2 = 1160/550ms
Signal Intensity
Signal Intensity
400
Flip Angle:10˚
300
200
100
0
c
500
400
Signal Intensity
b
a 500
Flip Angle:20˚
300
200
100
10
20
30
0
500
400
200
100
10
Slice Number
20
0
30
300
200
Flip Angle:20˚
300
200
100
100
0
400
10
20
30
Slice Number
0
Signal Intensity
Flip Angle:10˚
20
30
Slice Number
g 500
500
Signal Intensity
Signal Intensity
f
400
10
Slice Number
Gd-DOPA Doped Water with T1/T2 = 520/360ms
e 500
Flip Angle:30˚
300
No flow
Flow without partial saturation
Flow with partial saturation
Flip Angle:30˚
400
300
200
100
10
20
30
Slice Number
0
10
20
30
Slice Number
Figure 7.12: The signal intensity change along the tube was quantified using manuallyplaced ROIs. For Gd-DOPA doped water with T1 /T2 = 1160/550 ms (a-c) and 520/360 ms
(d-f), flip angles of 10◦ , 20◦ , and 30◦ were examined. In all six cases, partial saturation reduces inflow enhancement almost completely and generates uniform signal intensity across
the slab with a signal level similar to when there is no flow.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 110
five neck scans, one abdomen scan, and four leg scans using a four-channel phased-array
neurovascular coil (MRI Devices Corp, Waukesha, WI), eight-channel phased-array cardiac coil (GE Healthcare, Waukesha, WI), and eight-channel phased-array knee coil (MRI
Devices Corp, Waukesha, WI), respectively. Common parameters for all acquisitions were
TR = 17-20 ms, TE = 4 ms, flip angle = 20◦ , and 36 slices with 0.3-0.4 cm slice thickness.
For the neck and leg scans, in-plane matrix sizes of 256 x 256 and 256 x 192 were used,
respectively, and for the abdomen scan, a 64 x 64 matrix size was used, yielding a 40 s scan
time to enable a single breathhold.
Figure 7.13 shows axial images from one of the neck, abdomen, and leg scans (a-c) and
signal intensity variations along the carotid artery (d), abdominal aorta (e), and femoral
artery (f) for the cases with and without partial saturation. For all three arteries, inflow enhancement was greatly eliminated by partial saturation and signal intensity was almost uniform across the slabs. For the carotid artery and femoral artery, the intensity inhomogeneity
from coil sensitivity variations was corrected using reference body coil images [106]; however, the coil sensitivity variation was not compensated for the abdominal aorta because
cardiac motion made it difficult to estimate accurate correction factors. Wrap-around artifacts through the slab direction added some errors in signal intensity measurement of the
arteries at the edge slices.
7.5 Discussion
7.5.1 Partial Saturation for AIF measurements
Partial saturation effectively eliminates flow artifacts, including inflow enhancement and
pulsatile ghost artifacts for the RF-spoiled gradient-echo sequence. By applying partial
saturation, flowing spins can achieve signal intensity similar to that of stationary spins,
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 111
Carotid Artery
a
b
Signal Intensity
d
Abdominal Aorta
e
200
150
Without partial saturation
With partial saturation
150
Femoral Artery
c
f
Without partial saturation
With partial saturation
500
Without partial saturation
With partial saturation
400
100
300
100
200
50
50
0
100
10
20
30
Slice Number (Inferior to Superior)
0
10
20
30
Slice Number (Superior to Inferior)
0
10
20
30
Slice Number (Superior to Inferior)
Figure 7.13: In vivo images and signal intensity variation across the arteries. Axial neck
(a), abdomen (b), and leg (c) images, and signal intensity changes over ROIs in the carotid
artery (d), abdominal aorta (e), and femoral artery (f) are shown. Using partial saturation,
inflow enhancement is eliminated for all three arteries.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 112
generating similar T1 maps. Accurate T1 contrast is critical in certain applications such as
DCE-MRI for the AIF measurement. Several calibration methods have been proposed to
determine the accurate AIF from images affected by flow artifacts; however, they require
extra flow velocity measurement [107, 108]. In contrast to inflow-corrected calibration,
partial saturation can be used to determine AIF from the acquired signal intensity itself.
Applying a saturation pulse and then waiting for a prescribed recovery period can quickly
drive spins to the steady state for both stationary and flowing spins [97]. However, due to
a single recovery period, the steady state can only be achieved for a limited range of T1
values.
In 3D time-of-flight (3DTOF) MR angiography [109], a ramp RF pulse [110, 111], also
called a tilted optimized nonsaturating excitation (TONE) pulse, can be used to equalize
the signal produced by flowing blood while it traverses through a thick slab. This RF pulse
produces a spatially varying flip angle profile; the flip angle is set to a lower value where the
blood first enters the slab and is progressively increased until reaching its maximal value
where the blood exits the slab. The resulting blood signal profile across the slab is more
uniform, and vascular details can be depicted better than with standard imaging. However,
this technique is useful for vessel visualization and would be difficult to combine with T1
quantification
A spatial 90◦ saturation of an upstream slab is also often used to reduce the blood signal
and pulsatile ghost artifacts [112, 113]. However, this method is also not appropriate for T1
quantification. Furthermore, compared to the partial saturation RF pulse, the full saturation
RF pulse either induces higher SAR for a given time-bandwidth product or yields less
spatial selectivity for a given pulse duration. Finally, full saturation is much more sensitive
to B1 inhomogeneity than partial saturation.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 113
7.5.2 Signal Uniformity and T1 Contrast by Partial Saturation
By applying partial saturation to flowing spins, their longitudinal magnetization can approach the steady state before entering the imaging slab, depending on the thickness of the
partial saturation slab. However, when spins pass through the intersection of the imaging
slab and partial saturation slab, their steady state may be perturbed. However, the consequent signal fluctuation is relatively small compared to inflow-related artifacts, and edges
slices are normally discarded due to aliasing from the slab transition band.
As described in Section 7.1, RF-spoiling eliminates coherent transverse magnetization through the destructive addition of transverse magnetization components for stationary
spins. When spins move in the direction of spoiler gradients, additional phase accumulation occurs for those spins [114]. If flow is steady, spins actually experience an additional
quadratic phase increase over sequence repetitions [115], and thus some transverse magnetization can be refocused again. However, when flow is pulsatile, spin phase changes
by spoiler gradients are incoherent, and the formation of coherent transverse magnetization can be prevented similarly to applying a random RF phase for each sequence repetition [116, 117]. Our phantom experiments actually demonstrate that the signal level from
flowing spins with partial saturation is similar to that from stationary spins, indicating that
there is no coherent refocusing of transverse magnetization by pulsatile flow.
The magnetization simulation with pulsatile flow is a simplified case to illustrate practical situations. Realistically, spins do not simply move in one direction. In addition, at a
given cardiac phase, spins at different locations have varying velocities since the flow velocity pattern is normally laminar within a cross-section of vessel. Therefore, there exists
incoherence in both the flow direction and velocity for each spin within a voxel. However,
the magnetization variations from different spins tend to average out within a voxel, reducing view-to-view signal fluctuations in practical situations. Our phantom and in vivo
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 114
experiments support this hypothesis because ghost artifacts occurred due to view-to-view
signal fluctuation are minimal with partial saturation.
7.5.3 Partial Saturation Slab Thickness
The thickness of the partial saturation slab determines the number of repetitions for which
inflowing spins experience the RF pulses before entering the imaging slab, and thus the
degree to which inflow enhancement is reduced. For complete elimination of inflow enhancement, the partial saturation slab should be thick enough for spins to experience a
sufficient number of RF pulses. Theoretically, when the TR is 20 ms and the flip angle is
20◦ , approximately 60 RF pulses are needed for blood to attain the steady state in longitudinal components. For blood with a mean velocity of 25 cm/s, a 30 cm thick partial saturation
slab is necessary. This is achievable with the body coil, which provides a homogenous excitation range of 40-50 cm. However, a very thick partial saturation is not desirable because
it also widens the transition band. Other approach might be using a spatially varying flip
angle profile within the partial saturation slab. If a larger flip angle is used at the entry of
the partial saturation slab and the flip angle is smoothly decreased to the imaging flip angle,
then spins can be driven to the steady state more quickly.
7.5.4 Variations in RF Pulse
Partial saturation increases the scan time for the pulse duration of one RF pulse and one
spoiler gradient. When the scan time is critical, the TR can be reduced by using different
types of RF pulses. For example, using a maximum-phase RF pulse instead of a minimumphase RF pulse for the partial saturation can decrease the spoiler gradient area after the
partial saturation RF pulse since the presaturation slab-select gradient itself generates gradient spoiling. For further reduction, two RF pulses, a maximum-phase pulse for partial
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 115
saturation and a minimum-phase pulse for imaging slab excitation, can be combined as one
RF pulse [25]. In that case, dephasing of the partial saturation slab is achieved by only
the slab-select gradient, thus an RF pulse with a high time-bandwidth product is necessary
to provide a sufficient spoiling gradient area. The RF pulse duration can be greatly minimized with the variable rate selective excitation (VERSE) design [118]. Figure 7.14 shows
the composite RF pulses (20◦ flip angle, time-bandwidth product of 96) and slab-select
gradient waveforms with and without the VERSE algorithm, and resultant magnetization
profiles from the VERSE excitation. The Imaging slab thickness and partial saturation slab
thickness were 12.8 cm and 13.8 cm.
7.6 Summary
In RF-spoiled gradient-echo imaging, flow artifacts can degrade image quality and make
data quantification difficult. This chapter has presented that partial saturation can effectively eliminate flow artifacts with increasing TR for the duration of one additional RF
pulse and one spoiler gradient. The magnetization simulation verified that the flowing
spins can be driven to the steady state longitudinal magnetization level without generating
a net signal until they enter the imaging slab. Phantom and in vivo experiments demonstrated that inflow enhancement and pulsatile ghost artifacts can be significantly reduced
in RF-spoiled gradient-echo sequences and that a signal level of pure T1 contrast can be
achieved for flowing spins over the entire slab.
CHAPTER 7. FLOW ARTIFACT REDUCTION USING PARTIAL SATURATION 116
B1 (G)
a
b
Original RF and Gradient
VERSE RF and Gradient
0.15
0.15
0.1
0.1
0.05
0.05
Gradient (G/cm)
0
4
0
4
3
3
2
2
1
1
0
2
0
4
2
6
4
6
6.9
0.25
0
6.9
0.25
Time (ms)
c
0.5
0.75
1
0.5
0.75
1
Time (ms)
Magnetization
d
0.2
0.1
0
−30
−20
−10
0
10
Phase/!
1
0.5
0
0.5
−1
−30
Signal Intensity
Magnitude
0.3
Slab Profile
0.3
0.2
0.1
Saturation Imaging
0
−30
−20
−10
0
10
z Location (cm)
−20
−10
0
10
z Location (cm)
Figure 7.14: A minimum-phase RF pulse for the partial saturation and a maximum-phase
RF pulse for the imaging slab excitation can be combined into one RF pulse. (a) and
(b) show the composite RF pulses and slab-select gradients before (a) and after (b) the
VERSE algorithm with a constraint of the peak RF amplitude to 0.17 gauss and a maximum
gradient strength to 3.9 gauss/cm. With VERSE, the pulse duration was reduced from
6.9 ms to 1 ms. (c-d) shows the magnitude and phase of the magnetization profile, and net
magnetization over each slice along the z direction from the VERSE excitation.
Chapter 8
Summary and Recommendations
8.1 Summary
DCE-MRI has very high sensitivity in detecting breast cancers but suffers from low specificity. The major contributions of this thesis are improving DCE-MRI in various areas to
improve imaging speed and image quality. The contributions are summarized as follows.
• Self-calibrated parallel imaging methods such as mSENSE and GRAPPA were applied to bilateral imaging with and without slab-phase modulation. We verified that
slab-phase modulation can improve image quality in terms of aliasing artifacts and
SNR through phantom and in vivo experiments. g-Factors were measured experimentally and analytically from phantom images for data with and without slab-phase
modulation. On average they were decreased with the use of slab-phase modulation
for both mSENSE and GRAPPA.
• A TSENSE reconstruction was designed and implemented, and used to accelerate
clinical DCE imaging. To find a good method in sensitivity estimation, different
sensitivity maps were used for SENSE reconstruction. Resultant reconstructed data
117
CHAPTER 8. SUMMARY AND RECOMMENDATIONS
118
were compared in terms of oscillation (residual aliasing artifacts) in tumor contrast
uptake curves. A SENSE regularization method, which specifically can be used with
dynamic imaging, was proposed. TSENSE and TGRAPPA reconstructions in DCE
imaging were compared in terms of residual aliasing artifacts. From our bilateral
DCE imaging methods, 11 s temporal resolution and 1.1 x 1.1 x 3.5 mm3 spatial
resolution were achieved over the two breasts.
• Fat suppression using dual-band spectral-spatial excitation was compared for standard shims and independent shims. Our in vivo experiments demonstrated that independent shims improve fat suppression, especially in the regions of large field
inhomogeneity.
• Partial saturation was added to standard RF-spoiled gradient-echo sequences. It was
verified that partial saturation can provide accurate T1 contrast due to reduction of
flow artifacts by thorough magnetization simulation, and phantom and in vivo experiments.
8.2 Recommendations for Future Work
Even though breast DCE-MRI has been used as a screening tool adjunct to mammography
in many hospitals, low specificity and lack of standardization regarding image acquisition
and interpretation guidelines remain as obstacles for widespread use of breast MRI. Further
development of imaging techniques to provide higher temporal and spatial resolution and
better image quality will allow widespread use in clinical practice. Strong gradients, high
magnetic field strength, and improved phased-array coils are certainly beneficial to meet
various demands for breast DCE-MRI.
Possible extensions of the work presented in this thesis are as follows:
CHAPTER 8. SUMMARY AND RECOMMENDATIONS
119
• Even though we used slab-phase modulation for bilateral breast imaging, it can be
also used for hip or leg imaging where two volumes of interest are separated. It would
be interesting to investigate whether slab-phase modulation also improves parallel
imaging quality in those body parts.
• 3D stack-of-spiral trajectories have been used for DCE acquisition. Blurring is one
of the major artifacts with spiral trajectories, and in this thesis simple linear offresonance correction was used to reduce blurring. However, this simple method
has several limitations. First, it can be hard to obtain an accurate field map from
breasts with fat suppression. Second, the field variation is not necessarily linear
within breasts. Use of an autofocus technique [119] or a multi-frequency interpolation method [34] could improve the spiral images with computational complexity.
• A fat-water separation technique using multiple echo images is less sensitive to field
inhomogeneities than spectral-spatial excitation. From spiral acquisitions made at
three different echo times, off-resonance correction and fat-water separation can be
simultaneously performed [120]. This technique was already applied for spiral breast
images, demonstrating less blurring and a more uniform water image [121]. However, this technique is computationally demanding. An efficient reconstruction algorithm is required for clinical applications.
• For spiral imaging, swirling motion artifacts from the heart can degrade image quality, and furthermore, the motion artifacts enhance after contrast injection. Applying heart saturation can significantly reduce motion artifacts but it increases scan
time [122]. Designing a saturation pulse with a short RF duration and optimizing the
frequency with which heart saturation pulses are applied are possible future work.
• Flyback EPI [90] (used in Chapter 6) is also an attractive alternative to spiral trajectories because it does not produce significant artifacts due to motion or off-resonance,
CHAPTER 8. SUMMARY AND RECOMMENDATIONS
120
and reconstruction is simpler. In-plane parallel imaging would be easier to apply as
well.
• Partial Fourier acquisition and reconstruction can be combined with parallel imaging
for further acceleration [123, 124]. Comparison between homodyne and projection
onto convex sets (POCS) reconstructions combined with TSENSE for clinical DCE
imaging has been already presented in [125]. A single-step reconstruction that combines partial Fourier reconstruction as a regularization parameter in parallel imaging [126] could be a robust reconstruction method.
• The standard 8-channel industrial coil only allows acceleration in one direction (right/left
direction). However, increasing the number of coil elements, higher SNR and twodimensional acceleration are possible. The design of a 16-channel phased-array coil
and improved parallel imaging capability are presented in [127]. This coil could
provide much higher temporal and spatial resolution.
• Compressed sensing (CS) is the most recent technique to accelerate data with random
undersampling and to reconstruct data exploiting sparsity of MR images [128]. For
DCE imaging, we can exploit temporal sparsity for CS application with a proper
spatiotemporal sampling pattern [129]. CS can be combined with parallel imaging
[130, 131].
• This thesis only qualitatively compares signal enhancement curves from unaccelerated reconstruction and TSENSE reconstruction. We could quantify contrast kinetic
parameters from these two sets of reconstructed images to compare diagnostic accuracy. We could also investigate the effect of SNR decrease by applying parallel
imaging on data quantification.
• As this thesis demonstrated, independent shims can improve fat suppression for dualband spectral-spatial excitation without scan time increase. However, due to some
CHAPTER 8. SUMMARY AND RECOMMENDATIONS
121
tediousness of measuring shim values for two volumes and prescribing them, this
method has not been used for clinical scans. Developing automated measurement
and implementation of independent shims would allow clinical applications.
• Even though the dual-band spectral-spatial excitation has been used for water-only
excitation, it could also be used as a fat-saturation technique by changing of the RF
frequency. In this case, independent shims can also provide more homogenous fat
saturation over the two breasts.
• From our simulation and phantom experiments, pulsatile flow provided a mean signal
level close to T1 contrast when partial saturation was used. It is reasonable because
spin phase changes by spoiler gradients are incoherent with pulsatile flow, thus formation of coherent transverse magnetization and thus increase in signal level would
be difficult. We could do more experiments by varying the velocity, the velocity
waveform, and T1 of the fluid for further generalization. We could test whether T1
contrast from pulsatile flow can be still achieved even turning off RF phase cycling.
• In this thesis, the improvement of AIF measurement by applying partial saturation
has not been investigated. Comparing measured AIFs and quantified contrast kinetic
parameters for the cases with and without partial saturation would demonstrate the
practical usefulness of applying partial saturation.
• A spatially varying flip angle profile for the partial saturation slab could be tested to
find out whether it can drive flowing spins to the steady state. Depending on T1 of the
flowing spins, the optimized partial saturation slab profile would vary, but a profile
that works for a wide range of T1 values could be proposed.
• The partial saturation technique could be used to measure gadolinium extraction fraction in the kidney to evaluate renal function. To quantify this, the gadolinium concentration in both the renal artery and the renal vain needs to be estimated based on
CHAPTER 8. SUMMARY AND RECOMMENDATIONS
122
T1 of the blood. The partial saturation technique would allow accurate quantification
of T1 .
• The partial saturation technique could also be used to measure contrast arrival time in
MR angiography. MR angiography needs to be performed immediately after arrival
of the contrast agent to achieve the highest SNR. By repeated imaging of the blood
with using the partial saturation technique, contrast arrival time could be determined
more accurately by monitoring blood signal change that is not affected by inflow
enhancement.
• Many techniques presented here were developed for 1.5T, but currently many breast
protocols are moving toward 3T. Increased SNR due to the increased field strength
would provide improved diagnostic accuracy. However, increased B0 and B1 field inhomogeneities [132] and increased RF power deposition are major challenges. These
effects should be considered when designing pulse sequence or analyzing DCE data.
Bibliography
[1] F. Bloch, W. W. Hansen, and M. Packard, “Nuclear Induction,” Phys Rev, vol. 69,
no. 3-4, pp. 127, Feb 1946.
[2] E. M. Purcell, H. C. Torrey, and R. V. Pound, “Resonance absorption by nuclear
magnetic moments in a solid,” Phys Rev, vol. 69, no. 1-2, pp. 37–38, 1946.
[3] P. C. Lauterbur, “Image formation by induced local interactions: examples employing nuclear magnetic resonance,” Nature, vol. 242, no. 5394, pp. 190–191, 1973.
[4] C. K. Kuhl, R. K. Schmutzler, C. C. Leutner, A. Kempe, E. Wardelmann, A. Hocke,
M. Maringa, U. Pfeifer, D. Krebs, and H. H. Schild, “Breast MR Imaging Screening
in 192 Women Proved or Suspected to Be Carriers of a Breast Cancer Susceptibility
Gene: Preliminary Results 1,” 2000.
[5] M. Kriege, C. T. Brekelmans, C. Boetes, P. E. Besnard, H. M. Zonderland, I. M.
Obdeijn, R. A. Manoliu, T. Kok, H. Peterse, M. M. Tilanus-Linthorst, S. H. Muller,
S. Meijer, J. C. Oosterwijk, L. V. Beex, R. A. Tollenaar, H. J. de Koning, E. J.
Rutgers, and J. G. Klijn, “Efficacy of MRI and mammography for breast-cancer
screening in women with a familial or genetic predisposition,” N Engl J Med, vol.
351, no. 5, pp. 427–437, 2004.
123
BIBLIOGRAPHY
124
[6] W. A. Kaiser and E. Zeitler, “MR imaging of the breast: fast imaging sequences
with and without Gd-DTPA. Preliminary observations.,” Radiology, vol. 170, no. 3,
pp. 681–686, 1989.
[7] E. S. Fobben, C. Z. Rubin, L. Kalisher, A. G. Dembner, M. H. Seltzer, and E. J.
Santoro, “Breast MR imaging with commercially available techniques: radiologicpathologic correlation,” Radiology, vol. 196, no. 1, pp. 143–52, 1995.
[8] T. J. Turkat, B. D. Klein, R. L. Polan, and R. H. Richman, “Dynamic MR mammography: a technique for potentially reducing the biopsy rate for benign breast
disease,” J Magn Reson Imaging, vol. 4, pp. 563–563, 1994.
[9] 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.
[10] B. A. Hargreaves, C. H. Cunningham, J. M. Pauly, and B. L. Daniel, “Independent
phase modulation for efficient dual-band 3D imaging.,” Magn Reson Med, vol. 57,
no. 4, pp. 798–802, 2007.
[11] P. Kellman, F. H. Epstein, and E. R. McVeigh, “Adaptive sensitivity encoding incorporating temporal filtering (TSENSE),” Magn Reson Med, vol. 45, no. 5, pp.
846–852, 2001.
[12] L. Axel, “Blood flow effects in magnetic resonance imaging,” AJR Am J Roentgenol,
vol. 143, no. 6, pp. 1157–1166, 1984.
[13] W. H. Perman, P. R. Moran, R. A. Moran, and M. A. Bernstein, “Artifacts from
pulsatile flow in MR imaging,” J Comput Assist Tomogr, vol. 10, no. 3, pp. 473–
483, 1986.
[14] R. G. De Graaf and J. P. Groen, “MR angiography with pulsatile flow,” Magn Reson
Imaging, vol. 10, no. 1, pp. 25–34, 1992.
BIBLIOGRAPHY
125
[15] U. Hoffmann, G. Brix, M. V. Knopp, T. Hess, and W. J. Lorenz, “Pharmacokinetic
mapping of the breast: a new method for dynamic MR mammography,” Magn Reson
Med, vol. 33, no. 4, pp. 506–514, 1995.
[16] P. S. Tofts, B. Berkowitz, and M. D. Schnall, “Quantitative analysis of dynamic
Gd-DTPA enhancement in breast tumors using a permeability model,” Magn Reson
Med, vol. 33, no. 4, pp. 564–568, 1995.
[17] D. G. Nishimura, Principles of Magnetic Resonance Imaging, Stanford University,
1996.
[18] M. Han, P. J. Beatty, B. L. Daniel, and B. A. Hargreaves, “Independent slab-phase
modulation combined with parallel imaging in bilateral breast MRI,” Magn Reson
Med, vol. 62, no. 5, pp. 1221–1231, 2009.
[19] M. Han and B. A. Hargreaves, “Slab-phase modulation combined with parallel
imaging in bilateral breast imaging,” in Proceedings of the 16th Annual Meeting
of ISMRM, Toronto, Canada, 2008, p. 2730.
[20] M. Han, B. L. Daniel, and B. A. Hargreaves,
“Accelerated bilateral dynamic
contrast-enhanced 3D spiral breast MRI using TSENSE,” J Magn Reson Imaging,
vol. 28, no. 6, pp. 1425–1434, 2008.
[21] M. Han, B. A. Daniel, and B. A. Hargreaves, “Rapid DCE breast MRI using
TSENSE accelerated 3D spiral imaging,” in Proceedings of the 15th Annual Meeting
of ISMRM, Berlin, Germany, 2007, p. 140.
[22] M. Han, B. L. Daniel, and B. A. Hargreaves, “Accelerated bilateral DCE 3D spiral
breast imaging: comparison between TSENSE and TGRAPPA,” in Proceedings of
the 16th Annual Meeting of ISMRM, Toronto, Canada, 2008, p. 2729.
BIBLIOGRAPHY
126
[23] M. Han, S. Rodriguez, A. M. Sawyer, B. L. Daniel, C. Cunningham, J. M. Pauly,
and B. A. Hargreaves, “Fat suppression with independent shims for bilateral breast
MRI,” in Proceedings of the 17th Annual Meeting of ISMRM, Honolulu, Hawaii,
USA, 2009, p. 580.
[24] M. Han and B. A. Hargreaves, “Elimination of inflow enhancement by partial presaturation in RF spoiled imaging,” in Proceedings of the 17th Annual Meeting of
ISMRM, Honolulu, Hawaii, USA, 2009, p. 2699.
[25] M. Han and B. A. Hargreaves, “Combined excitaiton and partial saturation to reduce inflow enhancement.,” in Proceedings of the 18th Annual Meeting of ISMRM,
Stockholm, Sweden, 2010, p. 4947.
[26] M. Han and B. A. Hargreaves, “Reduction of flow artifacts by using partial saturation
in RF-spoiled gradient-echo imaging,” Submitted to Magn. Reson. Med., 2010.
[27] P. Mansfield, “Multi-planar image formation using NMR spin echoes,” J. Phys. C.,
vol. 10, pp. L55–L58, 1977.
[28] C. B. Ahn, J. H. Kim, and Z. H. Cho, “High-speed spiral-scan echo planar NMR
imaging-I,” IEEE Trans Med Imaging, vol. 5, no. 1, pp. 2–7, 1986.
[29] C. H. Meyer, B. S. Hu, D. G. Nishimura, and A. Macovski, “Fast spiral coronary
artery imaging,” Magn Reson Med, vol. 28, no. 2, pp. 202–213, 1992.
[30] J. D. O’sullivan, “Fast sinc function gridding algorithm for Fourier inversion in
computer tomography.,” IEEE Trans Med Imaging, vol. 4, no. 4, pp. 200–207, 1985.
[31] J. I. Jackson, C. H. Meyer, D. G. Nishimura, and A. Macovski, “Selection of a convolution function for Fourier inversion using gridding,” IEEE Trans Med Imaging,
vol. 10, no. 3, pp. 473–478, 1991.
BIBLIOGRAPHY
127
[32] D. C. Noll, C. H. Meyer, J. M. Pauly, D. G. Nishimura, and A. Macovski, “A
homogeneity correction method for magnetic resonance imaging withtime-varying
gradients,” IEEE Trans Med Imaging, vol. 10, no. 4, pp. 629–637, 1991.
[33] P. Irarrazabal, C. H. Meyer, D. G. Nishimura, and A. Macovski, “Inhomogeneity
correction using an estimated linear field map,” Magn Reson Med, vol. 35, no. 2, pp.
278–282, 1996.
[34] L. C. Man, J. M. Pauly, and A. Macovski, “Multifrequency interpolation for fast
off-resonance correction,” Magn Reson Med, vol. 37, no. 5, pp. 785–792, 1997.
[35] D. K. Sodickson and W. J. Manning, “Simultaneous acquisition of spatial harmonics
(SMASH): fast imaging with radiofrequency coil arrays,” Magn Reson Med, vol. 38,
no. 4, pp. 591–603, 1997.
[36] K. P. Pruessmann, M. Weiger, M. B. Scheidegger, and P. Boesiger, “SENSE: sensitivity encoding for fast MRI,” Magn Reson Med, vol. 42, pp. 952–962, 1999.
[37] J. Wang, T. Kluge, M. Nittka, V. Jellus, B. Kühn, and B. Kiefer, “Parallel acquisition
techniques with modified SENSE reconstruction (mSENSE),” in Proceedings of the
First Würzburg Workshop on Parallel Imaging Basics and Clinical Applications,
Würzburg, 2001, p. 89.
[38] P. M. Jakob, M. A. Grisowld, R. R. Edelman, and D. K. Sodickson, “AUTOSMASH: a self-calibrating technique for SMASH imaging,” MAGMA, vol. 7, no. 1,
pp. 42–54, 1998.
[39] R. M. Heidemann, M. A. Griswold, A. Haase, and P. M. Jakob, “VD-AUTOSMASH imaging,” Magn Reson Med, vol. 45, no. 6, pp. 1066–1074, 2001.
BIBLIOGRAPHY
128
[40] M. A. Griswold, P. M. Jakob, R. M. Heidemann, M. Nittka, V. Jellus, J. Wang,
B. Kiefer, and A. Haase, “Generalized autocalibrating partially parallel acquisitions
(GRAPPA),” Magn Reson Med, vol. 47, no. 6, pp. 1202–1210, 2002.
[41] P. J. Beatty, A. C. Brau, S. Chang, S. M. Joshi, C. R. Michelich, E. Bayram, T. E.
Nelson, R. J. Herfkens, and J. H. Brittain, “A method for autocalibrating 2-D accelerated volumetric parallel imaging with clinically practical reconstruction times,”
in Proceedings of the 15th Annual Meeting of ISMRM, Berlin, Germany, 2007, p.
1749.
[42] D. O. Walsh, A. F. Gmitro, and M. W. Marcellin, “Adaptive reconstruction of phased
array MR imagery,” Magn Reson Med, vol. 43, pp. 682–690, 2000.
[43] F. A. Breuer, M. Blaimer, N. Seiberlich, P. M. Jakob, and M. A. Griswold, “A
general formation for quantitative g-factor calculation in GRAPPA reconstructions,”
in Proceedings of the 16th Annual Meeting of ISMRM, Toronto, Canada, 2008, p. 10.
[44] P. J. Beatty and A. C. Brau, “Analytical computation of g-factor maps for autocalibrated parallel imaging,” in Proceedings of the 16th Annual Meeting of ISMRM,
Toronto, Canada, 2008, p. 1297.
[45] A. Jemal, R. Siegel, E. Ward, Y. Hao, J. Xu, T. Murray, and M. J. Thun, “Cancer
statistics, 2008,” CA Cancer J Clin, vol. 58, no. 2, pp. 71–96, 2008.
[46] D. A. Berry, K. A. Cronin, S. K. Plevritis, D. G. Fryback, L. Clarke, M. Zelen, J. S.
Mandelblatt, A. Y. Yakovlev, J. D. Habbema, and E. J. Feuer, “Effect of screening
and adjuvant therapy on mortality from breast cancer.,” N Engl J Med, vol. 353, no.
17, pp. 1784–1792, 2005.
[47] J. Folkman and M. Klagsbrun, “Angiogenic factors,” Science, vol. 235, no. 4787,
pp. 442–447, 1987.
BIBLIOGRAPHY
129
[48] J. Folkman, K. Watson, D. Ingber, and D. Hanahan, “Induction of angiogenesis
during the transition from hyperplasia to neoplasia,” Nature, vol. 339, no. 6219, pp.
58–61, 1989.
[49] C. D. Pham, T. P. L. Roberts, N. Van Bruggen, O. Melnyk, J. Mann, N. Ferrara,
R. L. Cohen, and R. C. Brasch, “Magnetic resonance imaging detects suppression of
tumor vascular permeability after administration of antibody to vascular endothelial
growth factor,” Cancer Invest, vol. 16, no. 4, pp. 225–230, 1998.
[50] D. R. Senger, L. Water, L. F. Brown, J. A. Nagy, K. T. Yeo, T. K. Yeo, B. Berse,
R. W. Jackman, A. M. Dvorak, and H. F. Dvorak, “Vascular permeability factor
(VPF, VEGF) in tumor biology,” Cancer Metastasis Rev, vol. 12, no. 3, pp. 303–
324, 1993.
[51] W. Kolch, G. Martiny-Baron, A. Kieser, and D. Marmé, “Regulation of the expression of the VEGF/VPS and its receptors: role in tumor angiogenesis,” Breast Cacner
Res Treat, vol. 36, no. 2, pp. 139–155, 1995.
[52] T. M. Kolb, J. Lichy, and J. H. Newhouse, “Comparison of the performance of
screening mammography, physical examination, and breast US and evaluation of
factors that influence them: an analysis of 27,825 patient evaluations,” Radiology,
vol. 225, no. 1, pp. 165–175, 2002.
[53] M. O. Leach, C. R. Boggis, A. K. Dixon, D. F. Easton, R. A. Eeles, D. G. Evans, F. J.
Gilbert, I. Griebsch, R. J. Hoff, P. Kessar, S. R. Lakhani, S. M. Moss, A. Nerurkar,
A. R. Padhani, L. J. Pointon, D. Thompson, and R. M. Warren, “Screening with
magnetic resonance imaging and mammography of a UK population at high familial
risk of breast cancer: a prospective multicentre cohort study (MARIBS),” Lancet,
vol. 365, no. 9473, pp. 1769–1778, 2005.
BIBLIOGRAPHY
130
[54] D. Saslow, C. Boetes, W. Burke, S. Harms, M. O. Leach, C. D. Lehman, E. Morris,
E. Pisano, M. Schnall, S. Sener, R. A. Smith, E. Warner, M. Yaffe, K. S. Andrews,
and C. A. Russell, “American cancer society guidelines for breast screening with
MRI as an adjunct to mammography,” CA Cancer J Clin, vol. 57, no. 2, pp. 75–89,
2007.
[55] C. K. Kuhl, S. Klaschik, P. Mielcarek, J. Gieseke, E. Wardelmann, and H. H. Schild,
“Do T2-weighted pulse sequences help with the differential diagnosis of enhancing
lesions in dynamic breast MRI?,” J Magn Reson Imaging, vol. 9, no. 2, pp. 187–196,
1999.
[56] S. Sinha, F. A. Lucas-Quesada, U. Sinha, N. DeBruhl, and L. W. Bassett, “In vivo
diffusion-weighted MRI of the breast: potential for lesion characterization,” J Magn
Reson Imaging, vol. 15, no. 6, pp. 693–704, 2002.
[57] Y. Guo, Y.Q. Cai, Z.L. Cai, Y.G. Gao, N.Y. An, L. Ma, S. Mahankali, and J.H. Gao,
“Differentiation of clinically benign and malignant breast lesions using diffusionweighted imaging,” J Magn Reson Imaging, vol. 16, no. 2, pp. 172–178, 2002.
[58] Y. Kuroki, K. Nasu, S. Kuroki, K. Murakami, T. Hayashi, R. Sekiguchi, and
S. Nawano, “Diffusion-weighted imaging of breast cancer with the sensitivity encoding technique: analysis of the apparent diffusion coefficient value,” Magn Reson
Med Sci, vol. 3, no. 2, pp. 79–85, 2004.
[59] C. Hayes, A. R. Padhani, M. O. Leach, et al., “Assessing changes in tumour vascular function using dynamic contrast-enhanced magnetic resonance imaging,” NMR
Biomed, vol. 15, no. 2, pp. 154–163, 2002.
[60] M. V. Knopp, E. Weiss, H. P. Sinn, J. Mattern, H. Junkermann, J. Radeleff, A. Magener, G. Brix, S. Delorme, I. Zuna, et al., “Pathophysiologic basis of contrast
BIBLIOGRAPHY
131
enhancement in breast tumors,” J Magn Reson Imaging, vol. 10, no. 3, pp. 260–266,
1999.
[61] S. Mussurakis, D. L. Buckley, S. J. Bowsley, P. J. Carleton, J. N. Fox, L. W. Turnbull,
and A. Horsman, “Dynamic contrast-enhanced magnetic resonance imaging of the
breast combined with pharmacokinetic analysis of gadolinium-DTPA uptake in the
diagnosis of local recurrence of early stage breast carcinoma,” Invest Radiol, vol.
30, no. 11, pp. 650–662, 1995.
[62] B. R. Rosen, V. J. Wedeen, and T. J. Brady, “Selective saturation NMR imaging,” J
Comput Assist Tomogr, vol. 8, no. 5, pp. 813, 1984.
[63] C. H. Meyer, J. M. Pauly, A. Macovskiand, and D. G. Nishimura, “Simultaneous
spatial and spectral selective excitation,” Magn Reson Med, vol. 15, no. 2, pp. 287–
304, 1990.
[64] W. T. Dixon, “Simple proton spectroscopic imaging,” Radiology, vol. 153, no. 1,
pp. 189, 1984.
[65] G. H. Glover, “Multipoint Dixon technique for water and fat proton and susceptibility imaging,” J Magn Reson Imaging, vol. 1, no. 5, pp. 521–530, 1991.
[66] S. B. Reeder, Z. Wen, H. Yu, A. R. Pineda, G. E. Gold, M. Markl, and N. J. Pelc,
“Multicoil Dixon chemical species separation with an iterative least-squares estimation method,” Magn Reson Med, vol. 51, no. 1, pp. 35–45, 2004.
[67] C. D. Lehman, C. Gatsonis, C. K. Kuhl, R. E. Hendrick, E. D. Pisano, L. Hanna,
S. Peacock, S. F. Smazal, D. D. Maki, T. B. Julian, E. R. DePeri, D. A. Bluemke, and
M. D. Schnall, “MRI evaluation of the contralateral breast in women with recently
diagnosed breast cancer,” N Engl J Med, vol. 356, no. 13, pp. 1295–1303, 2007.
BIBLIOGRAPHY
132
[68] J. M. Pauly, C. H. Cunningham, and B. L. Daniel, “Independent dual-band spectralspatial pulses,” in Proceedings of the 11th Annual Meeting of ISMRM, Toronto,
Canada, 2003, p. 966.
[69] C. H. Cunningham, B. L. Daniel, R. Rakow-Penner, and J. M. Pauly, “Independent
dual-band spectral-spatial pulses: implementation for bilateral breast MRI,” in Proceedings of the 13th Annual Meeting of ISMRM, Miami Beach, FL, USA, 2005, p.
1849.
[70] J. O. Fredrickson, C. H. Meyer, and N. J. Pelc, “Flow effects of spectral spatial
excitation,” in Proceedings of the 5th Annual Meeting of ISMRM, Vancouver, 1997,
p. 113.
[71] G. H. Glover, “Phase-offset multiplanar (POMP) volume imaging: a new technique.,” J Magn Reson Imaging, vol. 1, no. 4, pp. 457–461, 1991.
[72] L. Dougherty, G. Isaac, M. A. Rosen, L. W. Nunes, P. J. Moate, R. C. Boston, M. D.
Schnall, and H. K. Song, “High frame-rate simultaneous bilateral breast DCE-MRI,”
Magn Reson Med, vol. 57, no. 1, pp. 220–225, 2007.
[73] F. A. Breuer, M. Blaimer, R. M. Heidemann, M. F. Mueller, M. A. Griswold, and
P. M. Jakob, “Controlled aliasing in parallel imaging results in higher acceleration
(CAIPIRINHA) for multi-slice imaging,” Magn Reson Med, vol. 53, no. 3, pp. 684–
691, 2005.
[74] P. Kellman and E. R. McVeigh, “Image reconstruction in SNR units: a general
method for SNR measurement,” Magn Reson Med, vol. 54, no. 6, pp. 1439–1447,
2005.
BIBLIOGRAPHY
133
[75] P. M. Robson, A. K. Grant, A. J. Madhuranthakam, R. Lattanzi, D. K. Sodickson,
and C. A. McKenzie, “Comprehensive quantification of signal-to-noise ratio and gfactor for image-based and k-space-based parallel imaging reconstructions.,” Magn
Reson Med, vol. 60, pp. 895–907, 2008.
[76] S. B. Reeder, B. J. Wintersperger, O. Dietrich, T. Lanz, A. Greiser, M. F. Reiser,
G. M. Glazer, and S. O. Schoenberg, “Practical approaches to the evaluation of
signal-to-noise ratio performance with parallel imaging: application with cardiac
imaging and a 32-channel cardiac coil,” Magn Reson Med, vol. 54, no. 3, pp. 748–
754, 2005.
[77] M. J. Firbank, A. Coulthard, R. M. Harrison, and E. D. Williams, “A comparison of
two methods for measuring the signal to noise ratio on MR images,” Phys Med Biol,
vol. 44, no. 12, pp. N261–N264, 1999.
[78] B. T. West, K. B. Welch, and A. T. Galecki, Linear mixed models: a practical guide
using statistical software, Chapman & Hall/CRC, 2006.
[79] W. T. Sobol and D. M. Gauntt, “On the stationary states in gradient echo imaging,”
J Magn Reson Imaging, vol. 6, no. 2, pp. 384–398, 1996.
[80] P. Irarrazabal and D. G. Nishimura, “Fast three dimensional magnetic resonance
imaging,” Magn Reson Med, vol. 33, no. 5, pp. 656–662, 1995.
[81] Y. F. Yen, K. F. Han, B. L. Daniel, S. Heiss, R. L. Birdwell, R. J. Herfkens, A. M.
Sawyer-Glover, and G. H. Glover, “Dynamic breast MRI with spiral trajectories: 3D
versus 2D,” J Magn Reson Imaging, vol. 11, no. 4, pp. 351–359, 2000.
[82] F. A. Breuer, P. Kellman, M. A. Griswold, and P. M. Jakob, “Dynamic autocalibrated
parallel imaging using temporal GRAPPA (TGRAPPA),” Magn Reson Med, vol. 53,
no. 4, pp. 981–985, 2005.
BIBLIOGRAPHY
134
[83] Y. Zur, M. L Wood, and L. J. Neuringer, “Spoiling of transverse magnetization in
steady-state sequences,” Magn Reson Med, vol. 21, no. 2, 1991.
[84] A. P. Crawley, M. L. Wood, and R. M. Henkelman, “Elimination of transverse
coherences in FLASH MRI,” Magn Reson Med, vol. 8, no. 3, 1988.
[85] G. H. Glover and E. Schneider, “Three-point Dixon technique for true water/fat
decomposition with B0 inhomogeneity correction,” Magn Reson Med, vol. 18, no.
2, pp. 371–383, 1991.
[86] A. N. Tikhonov, “Solution of incorrectly formulated problems and the regularization
method,” Soviet Math. Dokl, vol. 4, no. 4, pp. 1035–1038, 1963.
[87] S. J. Riederer, T. Tasciyan, F. Farzaneh, J. N. Lee, R. C. Wright, and R. J. Herfkens,
“MR fluoroscopy: technical feasibility,” Magn Reson Med, vol. 8, no. 1, pp. 1–15,
1988.
[88] E. Henderson, B. K. Rutt, and T. Y. Lee, “Temporal sampling requirements for the
tracer kinetics modeling of breast disease.,” Magn Reson Imaging, vol. 16, no. 9, pp.
1057–1073, 1998.
[89] N. Maril, C. M. Collins, R. L. Greenman, and R. E. Lenkinski, “Strategies for
shimming the breast,” Magn Reson Med, vol. 54, no. 5, pp. 1139–1145, 2005.
[90] D. A. Feinberg, R. Turner, P. D. Jakab, and M. Von Kienlin, “Echo-planar imaging
with asymmetric gradient modulation and inner-volume excitation,” Magn Reson
Med, vol. 13, no. 1, pp. 162–169, 1990.
[91] R. R. Ernst, G. Bodenhausen, and A. Wokaun, Principles of nuclear magnetic resonance in one and two dimensions, Oxford University Press, USA, 1987.
BIBLIOGRAPHY
135
[92] M. L. Gyngell, “The steady-state signals in short-repetition-time sequences,” J
Magn Reson, vol. 81, no. 3, pp. 474–483, 1989.
[93] M. L. Wood, M. Silver, and V. M. Runge, “Optimization of spoiler gradients in
FLASH MRI,” Magn Reson Imaging, vol. 5, no. 6, pp. 455–463, 1987.
[94] J. Frahm, W. Hänicke, and K. D. Merboldt, “Transverse coherence in rapid FLASH
NMR imaging,” J Magn Reson, vol. 72, no. 2, pp. 307–314, 1987.
[95] H. Z. Wang and S. J. Riederer, “A spoiling sequence for suppression of residual
transverse magnetization,” Magn Reson Med, vol. 15, pp. 175–191, 1990.
[96] F. H. Epstein, J. P. Mugler III, and J. R. Brookeman, “Spoiling of transverse magnetization in gradient-echo (GRE) imaging during the approach to steady state,” Magn
Reson Med, vol. 35, no. 2, 1996.
[97] R. F. Busse and S. J. Riederer, “Steady-state preparation for spoiled gradient echo
imaging,” Magn Reson Med, vol. 45, no. 4, pp. 653–661, 2001.
[98] C. Preibisch and R. Deichmann, “Influence of RF spoiling on the stability and accuracy of T1 mapping based on spoiled FLASH with varying flip angles,” Magn Reson
Med, vol. 61, no. 1, pp. 125–135, 2009.
[99] R. Treier, A. Steingoetter, M. Fried, W. Schwizer, and P. Boesiger, “Optimized and
combined T1 and B1 mapping technique for fast and accurate T1 quantification in
contrast-enhanced abdominal MRI,” Magn Reson Med, vol. 57, no. 3, pp. 568–576,
2007.
[100] Y. Pang, M. Bernardo, B. Turkbey, G. Ravizzini, D. Thomasson, and P. Choyke,
“Minimizing inflow effects in measured arterial input function for prostate DCEMRI,” in Proceedings of the 16th Annual Meeting of ISMRM, Toronto, Canada,
2008, p. 2775.
BIBLIOGRAPHY
136
[101] R. Kaiser, E. Bartholdi, and R. R. Ernst, “Diffusion and field-gradient effects in
NMR Fourier spectroscopy,” The Journal of Chemical Physics, vol. 60, pp. 2966–
2979, 1974.
[102] J. Henning, “Multiecho imaging sequences with low refocusing flip angles,” J Magn
Reson, vol. 78, no. 3, pp. 397–407, 1988.
[103] C. Yuan, G. T. Gullberg, and D. L. Parker, “The solution of Bloch equations for
flowing spins during a selective pulse using a finite difference method,” Med Phys,
vol. 14, pp. 914–921, 1987.
[104] J. H. Gao and J. C. Gore, “Effects of slow flow on slice profile and NMR signal in
fast imaging sequences,” Phys Med Biol, vol. 39, no. 9, pp. 1489–1500, 1994.
[105] V. Denolin, C. Azizieh, and T. Metens, “New insights into the mechanisms of signal
formation in RF-spoiled gradient echo sequences.,” Magn Reson Med, vol. 54, no.
4, pp. 937–954, Oct 2005.
[106] W. W. Brey and P. A. Narayana, “Correction for intensity falloff in surface coil
magnetic resonance imaging,” Med Phys, vol. 15, pp. 241–245, 1988.
[107] M. K. Ivancevic, I. Zimine, X. Montet, H. Jean-Noel, F. Lazeyras, D. Foxall, and J. P.
Vallee, “Inflow effect correction in fast gradient-echo perfusion imaging.,” Magn
Reson Med, vol. 50, no. 5, pp. 885–891, 2003.
[108] F. Peeters, L. Annet, L. Hermoye, and B. E. Van Beers, “Inflow correction of hepatic
perfusion measurements using T1-weighted, fast gradient-echo, contrast-enhanced
MRI.,” Magn Reson Med, vol. 51, no. 4, pp. 710–717, Apr 2004.
[109] T. J. Masaryk, M. T. Modic, J. S. Ross, P. M. Ruggieri, G. A. Laub, G. W. Lenz,
E. M. Haacke, W. R. Selman, M. Wiznitzer, and S. I. Harik, “Intracranial circulation:
BIBLIOGRAPHY
137
preliminary clinical results with three-dimensional (volume) MR angiography.,” Radiology, vol. 171, no. 3, pp. 793–799, 1989.
[110] D. Atkinson, M. Brant-Zawadzki, G. Gillan, D. Purdy, and G. Laub, “Improved MR
angiography: magnetization transfer suppression with variable flip angle excitation
and increased resolution.,” Radiology, vol. 190, no. 3, pp. 890–894, 1994.
[111] A. Priatna and C. B. Paschal, “Variable-angle uniform signal excitation (VUSE) for
three-dimensional time-of-flight MR angiography,” J Magn Reson Imaging, vol. 5,
no. 4, pp. 421–427, 1995.
[112] J. P. Felmlee and R. L. Ehman, “Spatial presaturation: a method for suppressing flow
artifacts and improving depiction of vascular anatomy in MR imaging.,” Radiology,
vol. 164, no. 2, pp. 559–564, Aug 1987.
[113] J. Frahm, K. D. Merboldt, W. Hänicke, and A. Haase, “Flow suppression in rapid
FLASH NMR images,” Magn Reson Med, vol. 4, no. 4, pp. 372–377, 1987.
[114] S. Patz, “Some factors that influence the steady state in steady-state free precession,”
Magn Reson Imaging, vol. 6, no. 4, pp. 405–413, 1988.
[115] J. H. Duyn, “Steady state effects in fast gradient echo magnetic resonance imaging,”
Magn Reson Med, vol. 37, no. 4, pp. 559–568, Apr 1997.
[116] R. Freeman and H. D. W. Hill, “Phase and intensity anomalies in Fourier transform
NMR,” J Magn Reson, vol. 4, pp. 366–383, 1971.
[117] W. Lin and H. K. Song, “Improved accuracy in T1 mapping and flip angle correction
with random spoiling in radial gradient echo imaging,” in Proceedings of the 17th
Annual Meeting of ISMRM, Honolulu, Hawaii, USA, 2009, p. 4440.
BIBLIOGRAPHY
138
[118] S. Conolly, D. Nishimura, A. Macovski, and G. Glover, “Variable-rate selective
excitation,” J Magn Reson, vol. 78, no. 3, pp. 440–458, 1988.
[119] D. C. Noll, J. M. Pauly, C. H. Meyer, D. G. Nishimura, and A. Macovskj, “Deblurring for non-2D Fourier transform magnetic resonance imaging,” Magn Reson Med,
vol. 25, no. 2, pp. 319–333, 1992.
[120] P. T. Gurney, B. A. Hargreaves, and D. G. Nishimura, “Simultaneous off-resonance
correction and fat/water separation for noncartesian trajectories using a multifrequency least-squares approach,” in Proceedings of the 15th Annual Meeting of
ISMRM, Berlin, Germany, 2007, p. 1635.
[121] K. Granlund and B. Hargreaves, “Multi-frequency Off-resonance Correction and
Water/Fat Separation for Spiral Breast Imaging,” in Proceedings of the 16th Annual
Meeting of ISMRM, Toronto, Canada, 2008, p. 587.
[122] R. Rakow-Penner, B. Hargreaves, L. Pisani, B. Daniel, and G. H. Glover, “Heart
saturation for bilateral breast MRI: reduction of cardiac motion artifacts,” in Proceedings of the 16th Annual Meeting of ISMRM, Toronto, Canada, 2008, p. 3765.
[123] K. F. King and L. Angelos, “SENSE with partial Fourier homodyne reconstruction,”
in Proceedings of the 8th Annual Meeting of ISMRM, Denver, 2000, p. 153.
[124] D. B. Clayton and R. Bammer, “Methods for SENSE reconstruction with partial
k-space acquisition,” in Proceedings of the 13th Annual Meeting of ISMRM, Miami
Beach, FL, USA, 2005, vol. 13, p. 686.
[125] M. Han, B. L. Daniel, and B. A. Hargreaves, “Accelerated bilateral DCE 3D spiral
breast imaging using TSENSE and partial fourier methods,” in Proceedings of the
15th Annual Meeting of ISMRM, Miami Beach, FL, USA, 2005, p. 3359.
BIBLIOGRAPHY
139
[126] M. Bydder and M. D. Robson, “Partial Fourier partially parallel imaging,” Magn
Reson Med, vol. 53, no. 6, pp. 1393–1401, 2005.
[127] N. A. Nnewihe, T. Grafendorfer, B. L. Daniel, P. Calderon, M. T. Alley, F. Robb,
and B. A. Hargreaves, “16-channel custom-fitted bilateral breast coil for parallel
imaging in two directions,” in Proceedings of the 18th Annual Meeting of ISMRM,
Stockholm, Sweden, 2010, p. 644.
[128] M. Lustig, D. Donoho, and J. M. Pauly, “Sparse MRI: The application of compressed
sensing for rapid MR imaging,” Magn Reson Med, vol. 58, no. 6, pp. 1182–1195,
2007.
[129] U. Gamper, P. Boesiger, and S. Kozerke, “Compressed sensing in dynamic MRI,”
Magn Reson Med, vol. 59, no. 2, pp. 365–373, 2008.
[130] K. F. King, “Combining compressed sensing and parallel imaging,” in Proceedings
of the 16th Annual Meeting of ISMRM, Toronto, Canada, 2008, p. 1488.
[131] R. Otzao and D. Sodickson, “Distributed compressed sensing for accelerated MRI,”
in Proceedings of the 17th Annual Meeting of ISMRM, Honolulu, Hawaii, USA,
2009, p. 378.
[132] C. K. Kuhl, H. Kooijman, J. Gieseke, and H. H. Schild, “Effect of B1 inhomogeneity
on breast MR imaging at 3.0 T,” Radiology, vol. 244, no. 3, pp. 929–930, 2007.