University of Southern California Dissertations and Theses

Measurement of human brain perfusion with arterial spin labeling magnetic resonance imaging at ultra-high field

(USC Thesis Other)

Measurement of human brain perfusion with arterial spin labeling magnetic resonance imaging at ultra-high field

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Measurement of Human Brain Perfusion with Arterial Spin Labeling Magnetic Resonance
Imaging at Ultra-high Field

by
Kai Wang

A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BIOMEDICAL ENGINEERING)

August 2021

Copyright 2021 Kai Wang
ii
DEDICATION
This dissertation is dedicated to my mom and dad, and my fiancée Xin.

iii
ACKNOWLEDGEMENTS
I would like to thank my advisor, Dr. Danny Wang, for his support and guidance throughout
the whole journey. His vision and knowledge of the research field have guided me so many times
when I encountered obstacles. His passion and devotion to science have inspired me to be more
hardworking. I would never forget his encouragement and motivation when I was faced up with
challenges, nor would I ever forget the Thanksgiving dinners at his house, which let me feel the
warmth of a family despite being far away from home. I am grateful for what he has offered me. I
feel so lucky to have him as my supervisor. I am also immensely grateful for the expert insight and
additional mentorship I have received from my committee members: Drs. Krishna Nayak, John
Wood, Lirong Yan, and Hosung Kim. Thank you for contributing your valuable time and energy
to my scientific progress.
I wish to thank Siemens scientists, Drs. Jin Jin, Omer Oran, Patrick Liebig, Bernd Stoeckel.
From knowing little about our new pTx system to being able to proficiently using it to complete
my thesis, it took dozens of emails with my endless questions, dozens of pages of your detailed
answers, and your extreme patience, kindness, and selflessness. You have taught me not only about
the knowledge of the system but also about the quality of being a better person.
Thank you to my present and former lab mates Samantha Ma, Xingfeng Shao, Kay Jann,
Chenyang Zhao, Ziwei Zhao, Qinyang Shou, Soroush Heidari Pahlahvian, and Mayank Jog, for
offering your help in research and in everyday life. All the hotpot, Korean barbecue, ramen, curry,
pho, fish taco, and boba that we had together have made this journey so much more enjoyable and
weight-gaining. Thank you to our MRI technologist Katherine Martin for the help of countless
hours of scans. Thank you to our office staffs for all the assistant and support.
iv
To my dearest family and my loveliest fiancée, Xin Teng, thank you from the deepest depth
of my heart for your unconditional love and emotional support. I would not be able to finish this
journey without your constant encouragement and your faith in me. Despite we are thousands of
miles apart, just knowing you are there for me has always given me the strength to move forward.
Thank you for everything.

v
Table of Contents
DEDICATION ................................................................................................................................. ii
ACKNOWLEDGEMENTS ........................................................................................................... iii
LIST OF TABLES ......................................................................................................................... vii
LIST OF FIGURES ..................................................................................................................... viii
LIST OF ABBREVIATIONS ....................................................................................................... xv
ABSTRACT ................................................................................................................................ xvii
Chapter 1. General Introduction to the Dissertation ...................................................................... 1
1.1. Significance of Measuring Human Brain Perfusion ...................................... 1
1.2. Principles of ASL sequences ......................................................................... 2
1.3. UHF: opportunities and challenges ............................................................... 6
1.4. Parallel RF transmission (pTx) ...................................................................... 7
1.5. Overview of Studies .................................................................................... 10
Chapter 2. Optimization of adiabatic pulses for PASL at 7T: Comparison with pCASL ........... 12
2.1. Abstract ........................................................................................................ 12
2.2. Introduction ................................................................................................. 12
2.3. Materials and Methods ................................................................................ 14
2.4. Results ......................................................................................................... 23
2.4.1 Adiabatic Inversion Pulse Simulation ....................................................... 23
2.4.2 Inversion Profile Evaluation ...................................................................... 24
2.4.3 In-vivo pCASL optimization .................................................................... 27
2.4.4 In-vivo PASL evaluation ........................................................................... 28
2.5. Discussion .................................................................................................... 32
2.6. Conclusion ................................................................................................... 39
Chapter 3. Optimization of Pseudo-continuous Arterial Spin Labeling at 7T ............................. 40
3.1. Abstract ........................................................................................................ 40
3.2. Introduction ................................................................................................. 41
3.3. Methods ....................................................................................................... 42
3.4. Results ......................................................................................................... 47
3.5. Discussion .................................................................................................... 52
3.6. Conclusion ................................................................................................... 58
Chapter 4. Increased Labeling Efficiency of Pseudo-Continuous Arterial Spin Labeling with
Parallel Transmission B1 shimming ........................................................................... 59
4.1. Abstract ........................................................................................................ 59
4.2. Introduction ................................................................................................. 60
4.3. Materials and Methods ................................................................................ 61
4.4. Results ......................................................................................................... 67
vi
4.5. Discussion .................................................................................................... 75
4.6. Conclusion ................................................................................................... 79
Chapter 5. Exploration of Continuous Arterial Spin Labeling (CASL) with Dynamic Parallel
Transmission (pTx) .................................................................................................... 80
5.1. Abstract ........................................................................................................ 80
5.2. Introduction ................................................................................................. 81
5.3. Theory .......................................................................................................... 83
5.4. Methods ....................................................................................................... 85
5.5. Results ......................................................................................................... 89
5.6. Discussion .................................................................................................... 95
5.7. Conclusion ................................................................................................... 98
Chapter 6. Conclusion and Ongoing Work .................................................................................. 99
6.1. Optimization of the preparation pulses ........................................................ 99
6.2. Expansion of the LE simulation models .................................................... 101
6.3. Optimization of conventional CASL ......................................................... 102
References .................................................................................................................................. 105
APPENDIX A .............................................................................................................................. 111

vii

LIST OF TABLES

Table 2.1. Parameters and losses of the optimized pulses. For the trFOCI pulse, the parameter
vector is [Amax ,w,r1,r2,r3,r4,r5,µ,β,τ1,τ2]. FOCI has the lowest labeling efficiency
loss and trFOCI has the lowest uniformity loss, while WURST has the lowest combined
loss with a good balance of both labeling efficiency and the inversion band uniformity. 23
Table 2.2 Quantitative metrics of four PASL sequences and pCASL sequence. Metrics were
measured per subject, and the average value were shown here. LE, labeling efficiency; RS,
residual tissue signal; rLE, relative labeling efficiency. .................................................... 31
Table 3.1 The scan protocol for the in vivo experiments. TOF images were obtained for
angiography, after which pCASL and PASL sequences were acquired with the same
acquisition parameters. ...................................................................................................... 47
Table 3.2 The parameters of the pCASL sequences for the in-vivo evaluation as well as their
perfusion. Labeling 3 achieved the highest perfusion and thus was chosen as the optimal
7T pCASL. ......................................................................................................................... 49
Table 4.1 The scan protocol for the in vivo experiments. TOF images were acquired for
angiography, channel-specific 𝐵1+ maps were obtained for pTx 𝐵1+ shimming, and
pCASL and PASL sequences were acquired with the same acquisition parameters. (*)
TR=6000ms for pCASL-CP, and 9000ms for pCASL-univ and pCASL-indv due to more
constraining SAR limit (see section 4.5.4) ........................................................................ 66
Table 4.2 Quantitative metrics of pCASL sequences with three shimming modes. pCASL with
indv-shim outperforms the other two with the highest rMeanB1 and lowest rAsym, highest
perfusion, ICC, wsCV, and LE. Univ-shim also achieved considerable increase of
rMeanB1, perfusion, and LE compared with CP-shim. .................................................... 75

viii

LIST OF FIGURES
Figure 1.1 The position of the labeling slab and the inversion band. Depending on the different
relative location of the labeling and inversion slabs, three categories of PASL are
(A)EPISTAR, (B)PICORE, and (C)FAIR. .......................................................................... 2
Figure 1.2 The implementation of CASL label (first column) and the two methods of control. In
order to ensure identical MT effects, labeling plane is flipped above the imaging slab with
the same offset as label, which only works with single slice acquisition (second column).
Amplitude modulation of the labeling pulse splits the labeling plane into two, which has
identical MT effects as label and has nearly no tagging (third column). SS: single slice;
AM: amplitude modulation ................................................................................................. 4
Figure 1.3 Principle of the flow driven adiabatic inversion The effective field rotates from positive
to negative, thus inverts the flow magnetization. Adapted from: Detre JA, Alsop DC. Eur
J Radiol. 1999; 30(2):115-124 ............................................................................................. 4
Figure 1.4 The RF and gradient of the pCASL (A) label block and (B) control block. The long RF
pulse of CASL was separated into many small RF pulses, and the corresponding gradient
was increased for a more selective profile. Adapted from: Dai W., et al. Magn Reson Med.
2008 Dec: 60(6): 1488-1497 ................................................................................................ 5
Figure 1.5 The simulated labeling efficiency of a pCASL sequence with the presence of 𝐵1+ and
𝐵0 inhomogeneities. Parameters of the pulse train were adopted from Wu WC, et al., Magn
Reson Med, 2007 ................................................................................................................. 6
Figure 1.6 The intrinsic SNR is roughly proportional to 𝐵0
1.65
, which means at 7T the SNR is
nearly three times higher than the 3T. Adapted from: Pohmann R., Speck O., and Scheffler
K. Magn Reson Med. 2016 Feb: 75(2): 801-9 .................................................................... 7
Figure 1.7 Demonstration of the combined 𝐵1+ field of pTx. Relative uniform 𝐵1+ map is
achieved by overlapping the peak and valley of the standing waves of different transmit
channels. Adapted from mriquestions.com. ........................................................................ 7
ix
Figure 1.8 A The channel-specific 𝐵1+ map acquired on a gel phantom with an 8-channel pTx
system at 7T. Unique excitation pattern of each channel can be observed. B The channel-
specific 𝐵1+ map acquired on the same phantom with a 2-channel pTx system at 3T. In
comparison to 7T, the channel-specific excitation was more homogeneous, and the inter-
channel coherence was high. ............................................................................................... 9
Figure 2.1 The definition of the three bands in the inversion profile ........................................... 15
Figure 2.2 The parameter optimization pipeline for the (A) HS, WURST, FOCI pulses, and (B)
trFOCI pulse. A. For HS, WURST, and FOCI pulses, parameter set was obtained by
traversing the 2D parameter space, based on which the pulse waveform was generated, the
inversion profile was simulated with Bloch simulation, and the loss was calculated. B. For
trFOCI, the genetic algorithm was adopted to optimize the 11 parameters which consists
of Iteration and Greedy Hill-climbing. .............................................................................. 17
Figure 2.3 B0 and B1 map in a young healthy subject at 7T. A) B0 offset appeared to be linearly
shifting along z-direction for intracranial region. B) The highest B1 amplitude appeared at
the center of the brain while it dropped dramatically (<30%) below the brainstem. ........ 18
Figure 2.4 RF and gradient waveform and inversion profile with scaled B1 of HS, FOCI, trFOCI
and WURST pulses. The peak B1 amplitude of HS was set to 20uT, then the other three
pulses were scaled to have the same SAR as HS. Inversion profile was simulated with
targeted B1 scaled by 25%, 50%, 75% and 100% for each pulse. When scaled by 50%, the
uniform inversion band no longer exists with FOCI. trFOCI has a more uniform inversion
band but lower labeling efficiency compared with HS and WURST. ............................... 24
Figure 2.5 Inversion profile measured with phantom experiment. Row 1, the phantom image of
four inversion pulses with inversion thickness=100mm, red line indicating where the
inversion profile was plotted; Row 2, the inversion profile with SAR Normalized pulses;
Row 3, the inversion profile with pulses of Equal peak RF amplitude. ............................ 25
Figure 2.6 Inversion profile measured with in-vivo experiment. Same layout was used as phantom
experiments. Highest labeling efficiency was achieved with WURST, which was consistent
with phantom experiments. ................................................................................................ 26
Figure 2.7 A) The labeling plane overlain on the coronal view of TOF angiography. The red dash
line indicates the labeling plane. B) The B1 map at the labeling plane. Compared with the
desired B1 amplitude, the B1 amplitude at the targeted arteries (red dots, upper two: ICAs,
lower two: VAs) is scaled to about 50%. C) The B0 map at the labeling plane. The B0
offset at the targeted arteries is about 90Hz. D) Labeling efficiency vs. B1 and B0. Ideally
(B1 amplitude = 100% of desired 25°, B0 offset=0), the labeling efficiency is as high as
0.8; with B1=50% and B0 offset=0, the labeling efficiency=70%; with B1=50% and B0
offset=90Hz, the labeling efficiency=60%. ....................................................................... 27
x
Figure 2.9 Perfusion map acquired using the pulses with the same peak RF amplitude. The 𝐵1+
amplitude of WURST, FOCI, and trFOCI pulses were increased, leading to increased
labeling efficiency. For HS and FOCI, the residual tissue signal is more dominant at bottom
slices. Red box indicates the central slice, of which the mean gray matter perfusion is
0.42%, 0.60%, 0.51%, 0.46%, and 0.38% for HS, WURST, FOCI, trFOCI and pCASL,
respectively. ....................................................................................................................... 29
Figure 2.8 Perfusion map acquired using the pulses with the same SAR. HS and FOCI had higher
signal intensity especially at bottom slices, likely originated from residual tissue signal.
Compared with WURST, trFOCI had lower signal intensity, suggesting lower labeling
efficiency. Red box indicates the central slice, of which the mean gray matter perfusion is
0.42%, 0.42%, 0.39%, 0.33%, and 0.38% for HS, WURST, FOCI, trFOCI and pCASL,
respectively. ....................................................................................................................... 29
Figure 2.10 Comparison of the rLE term across 4 adiabatic pulses. The WURST pulse had the
highest rLE under both constraints of SAR Normalized and Equal peak RF amplitude
(P<0.01), which indicated a higher perfusion signal ratio against residual tissue signal of
WURST. rLE, relative labeling efficiency. ....................................................................... 30
Figure 2.11 CBF maps of three representative subjects. A) CBF acquired with PASL of peak B1
amplitude WURST pulse. B) CBF maps of the same subjects acquired with pCASL.
Consistent pattern of perfusion maps were obtained between PASL and pCASL, while
WURST-PASL achieved higher perfusion amplitude compared with pCASL. Note that the
display range is different from panel A. ............................................................................ 32
Figure 2.12 Simulated inversion profile of original trFOCI and the corresponding optimized
WURST, with and without 𝛥𝐵0. Apparent distortion of inversion profile was observed for
trFOCI with the presence of 𝛥𝐵0, while for WURST only a slight transition was introduced
which suggests improved robustness of WURST to B0 inhomogeneity compared with
trFOCI. ............................................................................................................................... 34
Figure 2.13 Inversion profile measured with a gel phantom with original trFOCI pulses from
literature and the corresponding WURST pulse. When SAR was normalized, higher
inversion efficiency was observed for WURST. For WURST/trFOCI2010 pair, the pulses
were optimized with inversion thickness=50mm and pulse duration=13ms, thus trFOCI
parameter of [5.98, 0.32, 0.73, 0.18, 0.33, 0.81, 0.04, 4.9, 4.9, 0, 0.83] (Hurley et al., 2010)
was used; for WURST/trFOCI2016 pair, the pulses were optimized with inversion
thickness=50mm and pulse duration=10ms, thus trFOCI parameter of [31.52, 0.28, 0.49,
0.13, 0.32, 0.83, 1, 0.8, 5.34, 0.46, 0.32] (Bause et al., 2016) was used. .......................... 34
Figure 3.1 A. The red dash line shows the labeling plane on top of the Maximum Intensity
Projection on the coronal plane of TOF; B. The B1 amplitude map at the labeling plane
from one representative subject, normalized by the designed value; C. The B0 map at the
xi
labeling plane from the same representative subject; D&E. The distribution of the B1
amplitude and B0 offset with the 44 inflow arteries. ........................................................ 43
Figure 3.2 Demonstration of one pCASL labeling block. The labeling block can be precisely
described by the timing parameters of RF duration and RF gap, and the gradient parameters
𝐺𝑚𝑎𝑥 and 𝐺𝑎𝑣𝑒. 𝐺𝑚𝑎𝑥 can be replaced by 𝑔𝑅𝑎𝑡𝑖𝑜: 𝐺𝑚𝑎𝑥 = 𝐺𝑎𝑣𝑒⋅𝑔𝑅𝑎𝑡𝑖𝑜. .............. 44
Figure 3.3 A, B&C. The contour map of the simulated LE for RF duration=300us, 500us, and
800us, respectively. These three RF duration values were chosen from all RF durations as
representative cases. The highest LE with each RF duration was marked with blue asterisks.
The highest LE in the (RF duration, 𝐺𝑎𝑣𝑒, 𝑔𝑅𝑎𝑡𝑖𝑜) 3D parameter space was achieved with
(300us, 0.3 𝑚𝑇/𝑚, 11). Labeling 0 was marked with red asterisk. D&E LE vs. B1 and B0
with Labeling 1 and Labeling 0, respectively. ................................................................... 48
Figure 3.5 The perfusion maps of the remaining three subjects with Labeling 1. Severe
contamination by the pCASL labeling side band was observed in the imaging volume. The
seemingly different location of the side band was due to the different labeling offset which
was adjusted for each subject. ........................................................................................... 50
Figure 3.4Perfusion map of one representative subject. Perfusion map of Labeling 1 was
contaminated by the aliasing of the labeling plane, while the difference between the other
three pCASL labeling was visually unnoticeable which was settled with quantitative
analysis. ............................................................................................................................. 50
Figure 3.6 The fractional perfusion map of the six subjects with visit 1 (A) and visit 2 (B). Robust
perfusion was achieved for each subject and each visit, although with variation across
subjects. ............................................................................................................................. 51
Figure 3.7 Simulated excitation profile of the pCASL labeling parameters. A. With Labeling 0,
the first aliasing appeared at 21mm, well below typical labeling offset value (~80mm); B.
With Labeling 1, the first aliasing appeared at 89mm, which likely will affect imaging
volume. C&D. Compared with Labeling 1, Labeling 2 (C) and 3 (D) had a higher gave
resulting in a smaller aliasing location (44mm and 48mm, respectively), and also a higher
𝐺𝑚𝑎𝑥 resulting in a lower aliasing signal amplitude. ....................................................... 54
Figure 3.8 Simulation of LE with v=25cm/m for RF duration=300us (A), 500us (B), and 800us
(C). The optimal parameter set was RF duration=300us, 𝐺𝑎𝑣𝑒 = 0.5mT/m, 𝑔𝑅𝑎𝑡𝑖𝑜=12.
Compared with results from v=40cm/s (Figure 3.3), the optimal 𝐺𝑎𝑣𝑒 increased from
0.3mT/m to 0.5mT/m due to lower flow velocity. ............................................................ 55
Figure 3.9 The excitation profile, RF and gradient waveform of the optimized pCASL labeling,
VERSE modified pCASL labeling, and VERSE modified with RF duration=500us.
Without VERSE modification, the labeling block parameters are: RF duration=300us,
Gmax=6mT/m, gradient ramp time=30us, duration of the rewinding gradient=190us, RF
period=550us, and RF duty cycle=300/550=54.6%; With VERSE modification, RF
xii
duration=300us, gradient ramp time=70us, duration of the rewinding gradient=220us, RF
period=660us, and RF duty cycle=300/660=45.6%. Clear aliasing pattern was observed in
B. Shown in C, when the modified RF duration was increased to 500us, the aliasing pattern
disappeared. ....................................................................................................................... 57
Figure 4.1 The flow chart of indv-shim. TOF images were acquired for localizing the inflowing
arteries at the labeling plane, channel-specific 𝐵1+ map was acquired and fed to the
optimization problem, which was solved with fmincon solver. ........................................ 63
Figure 4.3 Same layout was used as Figure 4.2. For all subjects, w=20 yielded consistent increase
of rMeanB1 and a comparable if not smaller rAsym compared to CP-shim. ................... 68
Figure 4.2 A&D The ROI maps on the TOF images. B&E The rMeanB1 ratio between indv-shim
and CP-shim with 𝑤 ∈[0,50]. C&F The rAsym Ratio with ∈[0,50]. For both subjects,
w=20 yielded a rMeanB1ratio of nearly 1.2 and an rAsym ratio smaller than 1. Therefore,
w=20 was chosen. .............................................................................................................. 68
Figure 4.4 A&B The rMeanB1 ratio and rAsym ratio between univ-shim and CP-shim w ∈[0,50].
w = 40 was chosen for univ-shim. C&D The rMeanB1 and rAsym curve of three shimming
modes, respectively. The univ-shim achieved considerable mean B1 increase compared to
CP-shim for all subjects, which was comparable to indv-shim, although the asymmetry was
higher than indv-shim. ....................................................................................................... 69
Figure 4.5 A&B The rMeanB1 curve of the three shimming modes. Indv-shim had consistently
higher rMeanB1 than CP-shim for each subject, and univ-shim had similar performance as
indv-shim except for measurement 9. B The rAsym curve of the three shimming modes.
Indv-shim had the lowest rAsym while the rAsym of CP-shim and univ-shim were
comparable. C The combined B1 map of one representative case of the three shimming
modes, first row amplitude, second row phase. Performance of indv-shim and univ-shim
were comparable, both of which achieved higher B1 amplitude within ROI (colored circles)
compared with CP-shim. A dark band appeared (yellow arrow) with indv-shim and univ-
shim, which would not affect the inflowing arteries. ........................................................ 71
Figure 4.6 The fractional perfusion maps of one representative case: visit 1 (A) and visit 2 (B).
Consistent and high-quality perfusion maps were obtained with all pCASL sequences,
while difference among three shimming modes were not observable visually. Good
repeatability was achieved. ................................................................................................ 72
Figure 4.7 A and B The fractional perfusion map for 2 other subjects, respectively. Top panel,
first visit; bottom panel, second visit. Perfusion maps of good quality and decent
repeatability were achieved with all pCASL sequences .................................................... 73
Figure 4.8 Comparison of the mean GM CBF of three shimming modes. Perfusion of pCASL with
indv-shim was significantly higher than that of CP-shim (increased by 9.5%, p<0.05).
xiii
Perfusion of pCASL with univ-shim was also increased compared with that of CP-shim by
5.3% although the difference was not significant probably due to limited sample size. ... 74
Figure 4.9 Bland-Altman plot of pCASL with CP-shim (A), indv-shim (B), and univ-shim (C).
All pCASL sequences showed decent repeatability. ......................................................... 75
Figure 5.1 Demonstration of DCASL principle. By two channel groups working in turn, flow-
driven adiabatic inversion of CASL can be achieved without exceeding RFPA duty cycle
limit. ................................................................................................................................... 85
Figure 5.2 (A&B) The label and control RF pulses of two groups of channels, with no mismatch
between the two groups. The combined RF showed a perfect constant pattern for label and
amplitude modulated pattern for control. (C&D) The label and control RF pulses of the
two group with the RF amplitude of Group2 scaled by 0.8 compared with Group1. The
combined RF pulses showed abrupt jumps when channel group changed. ....................... 88
Figure 5.3 Loss vs. iteration for Group1 (A) and Group2 (B) during the optimization process.
Following the definition in Section 5.4.1, Loss1 and Loss2 were loss calculated with real
part and imaginary part, respectively. Loss for both Group1 and Group2 converged to close
to 0 after about iteration 20, and early stop was achieved for both groups. ...................... 90
Figure 5.4 A The global SAR vs. 𝛼, calculated as the ratio against the conventional CASL. Global
SAR dropped monotonically with 𝛼. B The local SAR vs. 𝛼, calculated as the ratio against
the conventional CASL. Local SAR first dropped then plateaued with 𝛼, 𝛼=0.1 was the
turning point. C The amplitude mismatch between the two groups vs. 𝛼. D The phase
mismatch between the two groups vs. 𝛼. .......................................................................... 91
Figure 5.5 The simulated LE with 𝐵1+ amplitude and phase mismatch between the two groups
for A) Label condition, B) Control condition, and C) the net LE (Label - Control). A higher
mismatch (amplitude or phase) led to decreased net LE, however, the LE was relatively
insensitive to the mismatch: LE>0.75 was achieved with most of the mismatch parameter
range. ................................................................................................................................. 92
Figure 5.6 The relationship between LE and 𝛼 generated from the LE vs. amplitude/phase
mismatch lookup table and the mismatch vs. 𝛼 relationship. ............................................ 92
Figure 5.7 The combined 𝐵1+ maps of the conventional CASL and Group1 and Group2 of
DCASL. The amplitude map is shown in first row, and phase map second row. Both
Group1 and Group2 of DCASL were able to achieve combined 𝐵1+ map well matching
that of CASL. ..................................................................................................................... 93
Figure 5.8 The external pulses loaded into the scanner. Pulses were successfully loaded for both
groups and both label and control conditions, but then the RFPA did not allow it. .......... 94
xiv
Figure 5.9 A The comparison of SAR between DCASL and CASL. For each RF channel, with
DCASL the working time is 50% but the amplitude is double compared with CASL, since
SAR is proportional to the square of RF amplitude, the total SAR of DCASL is
approximately double compared with CASL. B The comparison of SAR between pCASL
and CASL. With pCASL, the RF channels are turned on for only ~50% time and that the
Hanning pulse has a high peak amplitude, which leads to 2.76 times SAR of CASL. ..... 95
Figure 6.1 The inversion profile of the BS inversion pulse directly adopted from 3T with (A)
𝛥𝐵0=−100𝐻𝑧, (B) 𝛥𝐵0=0𝐻𝑧, and (C) 𝛥𝐵0=100𝐻𝑧. The inversion efficiency
dropped dramatically when 𝐵1+ was less than 50%. .................................................... 100
Figure 6.2 The fractional perfusion maps of three subjects, acquired with pCASL sequence with
and without BS, shown in top row and bottom row, respectively. Perfusion signal was
significantly reduced by BS. The ratio between the perfusion of pCASL with BS and that
without BS was 63.1%. .................................................................................................... 100
Figure 6.3 A The weighted average labeling efficiency of CASL with 𝐺𝑎𝑣𝑒, with weighting for
each 𝐵1+/𝐵0 condition as their corresponding relative frequency, the highest LE (0.83)
was achieved with 𝐺𝑎𝑣𝑒=0.6 mT/m. B The LE vs. 𝐵1+ and 𝐵0 inhomogeneities. The
highest LE was achieved at approximately 70% 𝐵1+ amplitude, while LE was insensitive
to 𝐵0 offset. ..................................................................................................................... 102
Figure 6.4 The perfusion map of the single-slice CASL sequence with different RF and gradient
amplitude. The combination of high RF amplitude (2.25uT) and low gradient amplitude
(0.8mT/m) yielded the highest perfusion signal. ............................................................. 103
Figure 6.5 The simulated labeling efficiency of label and control with AM. LE of label is
independent from 𝐵0 offset, while the control showed a period fluctuation of LE with 𝐵0
offset, shorter period for higher 𝐺𝑎𝑣𝑒. For 𝐺𝑎𝑣𝑒=0.6mT/m, the LE fluctuation was not
obvious, but LE of control was as high as 0.35 resulting in a net LE (label - control) of
0.55. ................................................................................................................................. 104

xv
LIST OF ABBREVIATIONS
AM . . . . . . . . Amplitude Modulation
ASL . . . . . . . . Arterial Spin Labeling
CASL . . . . . . . . Continuous Arterial Spin Labeling
CBF . . . . . . . . Cerebral Blood Flow
CP . . . . . . . . Circular Polarization
CSF . . . . . . . . Cerebral Spinal Fluid
DSC . . . . . . . . Dynamic Susceptibility Contrast
DCE . . . . . . . . Dynamic Contrast Enhanced
FAIR . . . . . . . . Flow-sensitive Alternating Inversion Recovery
FOCI . . . . . . . . Frequency-Offset-Correction-Inversion
GM . . . . . . . . Gray Matter
GRASE . . . . . . . . Gradient and Spin-Echo
HS . . . . . . . . Hyperbolic Secant
ICA . . . . . . . . Internal Carotid Artery
ICC . . . . . . . . Intraclass Correlation Coefficient
LE . . . . . . . . Labeling Efficiency
MT . . . . . . . . Magnetic Transfer
MRI . . . . . . . . Magnetic Resonance Imaging
PASL . . . . . . . . Pulsed Arterial Spin Labeling
PCASL . . . . . . . . Pseudo-continuous Arterial Spin Labeling
pTx . . . . . . . . Parallel RF Transmission
RFPA . . . . . . . . Radiofrequency Power Amplifier
RMSE . . . . . . . . Root Mean Square Error
SAR . . . . . . . . Specific Absorption Rate
SD . . . . . . . . Standard Deviation
SMS . . . . . . . . Simultaneous Multi Slice

xvi
SNR . . . . . . . . Signal-to-Noise Ratio
TFL . . . . . . . . turbo Fast Low Angle Shot
TOF . . . . . . . . Time of Flight
UHF . . . . . . . . Ultra-High Field
VA . . . . . . . . Vertebral Artery
VERSE . . . . . . . . Variable-rate Selective Excitation
VOP . . . . . . . . Virtual Observation Point
WM . . . . . . . . White Matter
wsCV . . . . . . . . Within-subject Coefficient of Variation
WURST . . . . . . . . Wideband-Uniform-Rate-Smooth-Truncation

xvii
ABSTRACT
Arterial spin labeling (ASL) is a perfusion magnetic resonance imaging (MRI) technique
that utilizes magnetically labeled arterial blood water as an endogenous tracer to measure cerebral
blood flow (CBF). The noninvasive nature and the ability to quantitatively measure tissue
perfusion make ASL ideal for research and clinical studies. The main limitation of ASL technique
is the low signal-to-noise (SNR) due to the intrinsically small fraction of the labeled arterial blood
(~1%) and T1 relaxation of the label.
Ultra-high field (UHF) benefits ASL with an increased intrinsic SNR of MRI signal (𝐵
N
1.65
)
due to the dielectric resonance of the magnetic field and a prolonged tracer half-life (blood T1).
However, the implementation of ASL at UHF is not straightforward. Due to the inhomogeneity of
the transmit 𝐵
O
(𝐵
O
P
) field and the 𝐵
N
field, it is challenging to achieve high labeling efficiency
with both pulsed ASL (PASL) and pseudo-continuous ASL (pCASL), which are the two
commonly used ASL techniques. Meanwhile, parallel transmission (pTx) provides previously
unavailable degree of freedom that allows full spatial and temporal control of the 𝐵
O
P
field. With a
transmit coil array that consists of several elements with spatially distinct 𝐵
O
P
field pattern, the
extra degrees of freedom can be exploited to overcome the effects of 𝐵
O
P
inhomogeneities, which
provides potential solutions to mitigate the issues in implementing ASL at UHF.
The overall objective of this work is to develop a suite of reliable ASL technologies at
UHF. First, the 7T PASL sequence was optimized by optimizing and evaluating the adiabatic
inversion pulses with the 𝐵
O
P
and 𝐵
N
inhomogeneities at 7T. The wide-band-uniform-rate-smooth-
truncation (WURST) pulse achieved the lowest loss in simulation and achieved a superior
performance compared with the other three pulses in the experiments. Second, the 7T pCASL
xviii
parameters were optimized to achieve a high labeling efficiency with the 𝐵
O
P
and 𝐵
N

inhomogeneities at 7T. The optimized pCASL sequence achieved robust labeling efficiency as
well as good repeatability. Third, the optimized pCASL was incorporated with pTx 𝐵
O
P
shimming
aiming at further improving the labeling efficiency by increasing the 𝐵
O
P
amplitude at the inflowing
arteries. Both “indv-shim” (shimming weights calculated for each individual subject) and “univ-
shim” (universal shimming weight calculated based on a group of subjects) successfully achieved
increased perfusion signal intensity compared with the circular polarized (CP) mode. Fourth, an
innovative continuous ASL (CASL) with Dynamic pTx pulse (noted as DCASL) was proposed.
By utilizing channel-specific RF pulses for each transmit channel, the proposed DCASL showed
intrinsic insensitivity to 𝐵
N
offset and relatively low Specific Absorption Rate (SAR) similar to
conventional CASL sequence without violating the duty cycle limit of each RF transmit channel.
In conclusion, reliable perfusion measurements were obtained with the PASL and pCASL
sequences optimized for 7T, pTx 𝐵
O
P
shimming methods further increased the LE of pCASL
sequence, and the theoretical framework of an innovative DCASL implemented with dynamic pTx
was proposed and evaluated.

1
Chapter 1. General Introduction to the Dissertation
1.1. Significance of Measuring Human Brain Perfusion
Cerebral Blood Flow (CBF) or brain perfusion is a key parameter for in-vivo assessment
of neurovascular function. Since perfusion is normally coupled with metabolism, perfusion can be
treated as a surrogate maker of brain metabolism and neuronal activity. There are three main
techniques for perfusion MRI: dynamic susceptibility contrast (DSC) MRI, which acquires a series
of T2 or T2*-weighted images to capture the image intensity drop caused by shortened T2/T2* due
to the injected Gadolinium-based contrast and therefore estimate the perfusion; dynamic contrast-
enhanced (DCE) MRI, which measures the T1 drop induced by the injected Gadolinium-based
contrast to estimate perfusion; and arterial spin labeling (ASL), which utilizes magnetically labeled
arterial blood water as an endogenous tracer to measure CBF. Unlike the DSC MRI and DCE MRI,
ASL does not require the injection of exogenous contrast, which eliminates contrast-related risk
(Power, Talbot, Kucharczyk, & Mandell, 2016) and therefore has been widely applied for clinical
and scientific studies.
ASL signal is close to the site of neural activation as most of the labeled arterial water
exchanges with tissue water in capillaries (Detre & Wang, 2002). It has been shown that ASL
perfusion fMRI is able to visualize orientation columns in the cat visual cortex with superior spatial
resolution compared to BOLD fMRI (Duong, Kim, Uğurbil, & Kim, 2001). The hemodynamic
response of perfusion signals has also been shown to arise ~1sec earlier than that of BOLD signals
(Liu et al., 2000). Unlike BOLD fMRI that detects relative signal changes between two conditions,
ASL provides quantitative perfusion measurements both at rest and during task activation. The
main limitation of ASL technique is the low signal-to-noise ratio (SNR) due to the intrinsically
2
small fraction of labeled arterial blood (~1%), and T1 relaxation of the label (Rempp et al., 1994),
which is usually compensated by repeated measurements at the cost of longer scan time.
1.2. Principles of ASL sequences
There are three categories of ASL sequences: pulsed ASL (PASL), continuous ASL
(CASL), and pseudo-continuous ASL (pCASL).
PASL uses inversion pulses
to label the inflowing arterial blood
(Kim & Tsekos, 1997). The
duration of the inversion pulses is
usually ~10ms, during which the
magnetization of the blood within a
certain segment of the inflow
arteries are inverted, which will
serve as the source of the perfusion
signal. Depending on the relative
locations of the inversion slab and
imaging slab, PASL can be further
separated into three categories. The
first PASL method is called Echo-
Planar Imaging-based Signal
Targeting by Alternating
Radiofrequency pulses
(EPISTAR), which uses an excitation pulse to saturate the imaging slab, an inversion pulse below
Figure 1.1 The position of the labeling slab and the inversion band.
Depending on the different relative location of the labeling and
inversion slabs, three categories of PASL are (A)EPISTAR, (B)PICORE,
and (C)FAIR.

3
the imaging slab for label and above the imaging slab for control (Figure 1.1A). The second
method is called Proximal Inversion with Control of Off-Resonance Effects (PICORE), which
utilizes identical labeling scheme for label but utilizes an off-resonance inversion pulse for control
(Figure 1.1B). The third method is called Flow-sensitive Alternating Inversion Recovery (FAIR),
which is a commonly used labeling scheme (Kim & Tsekos, 1997) with a selective inversion slab
limited to region near the imaging slab for the label images, and an on-resonance non-selective
inversion slab for control (Figure 1.1C).
Adiabatic inversion pulses are often utilized by PASL. Unlike traditional pulses which have
a flip angle proportional to the RF amplitude, the adiabatic pulses manipulate the magnetization
with specially modulated amplitude and phase, so that once the adiabatic condition is met, the
resultant magnetization is not related to the actual amplitude of the RF (Tannús & Garwood, 1997).
Commonly used adiabatic pulses include the Hyperbolic Secant (HS) pulse (Silver, Joseph, &
Hoult, 1969), the Wideband-Uniform-Rate-Smooth-Truncation (WURST) pulse (Kupce, 1995),
etc.
The subtraction of the label and control image pairs is hence the perfusion signal caused
by the inflowing labeled blood. For the purpose of quantification, an inferior saturation pulse can
be applied to spoil the labeled magnetization before image acquisition and therefore determines
the bolus length of the labeled blood (Wong, Buxton, & Frank, 1998).
4
Unlike PASL, CASL uses one long constant RF pulse (~1s) along with a positive gradient
in z-direction, which utilizes flow-driven
adiabatic inversion to tag the inflow blood
(Williams, Detre, Leigh, & Koretsky, 1992):
as the blood flowing up, the effective 𝐵
O
P

field, which is the combination of the RF
pulse and the gradient induced frequency
offset in the rotating frame, rotates from
positive to negative; The magnetization of
the blood follows the effective 𝐵
O
P
field and
thus is inverted (Figure 1.2). For
the control block, signal caused
by the magnetic transfer (MT) effect has to be identical to the label. As shown in Figure 1.3, one
way to construct a control block is to apply the inversion above the imaging slab with identical

Figure 1.2 Principle of the flow driven adiabatic inversion
The effective field rotates from positive to negative, thus
inverts the flow magnetization. Adapted from: Detre JA,
Alsop DC. Eur J Radiol. 1999; 30(2):115-124
Figure 1.3 The implementation of CASL label (first column) and the two methods of control. In order to ensure
identical MT effects, labeling plane is flipped above the imaging slab with the same offset as label, which only works
with single slice acquisition (second column). Amplitude modulation of the labeling pulse splits the labeling plane
into two, which has identical MT effects as label and has nearly no tagging (third column). SS: single slice; AM:
amplitude modulation
5
offset distance to produce identical MT effects, which only works with single slice (SS) acquisition
(Williams et al., 1992) (thus referred as SS method below). The other way is to apply an amplitude
modulation (AM) to the RF so that the effective labeling plane is split into two, which theoretically
has no net tagging of flow magnetization and has identical MT effects at imaging slab (Alsop &
Detre, 1998). The AM control block allows acquisition of multiple slices, but the labeling
efficiency is lower (~74%) compared with its SS alternative (J. Wang et al., 2005).
However, most clinically
available MR systems have a duty cycle
limit which forbids the long RF pulse.
To get around this problem, pCASL
separates the long RF pulse into as many
as 1000 or more shaped RF pulses as a
pulse train (Dai, Garcia, de Bazelaire, &
Alsop, 2008). As demonstrated in
Figure 1.4A, for the label block, the
long RF pulse is split, the gradient is
increased to avoid the effect of side
band of the inversion pulse, and a
“rewinding” gradient is used to ensure a comparable average gradient over time as CASL.; For the
control block (Figure 1.4B), the RF pulses have alternating polarity which preserves the
magnetization while controls for MT effects.
Compared with PASL, pCASL shows higher SNR and repeatability at 3T although at a
higher Specific Absorption Rate (SAR) and thus is recommended for 3T perfusion imaging (Alsop
Figure 1.4 The RF and gradient of the pCASL (A) label block
and (B) control block. The long RF pulse of CASL was
separated into many small RF pulses, and the corresponding
gradient was increased for a more selective profile. Adapted
from: Dai W., et al. Magn Reson Med. 2008 Dec: 60(6): 1488-
1497
6
et al., 2015). However, at ultra-high field (UHF) where the SAR is more limiting along with other
challenges, the choice between different ASL methods has to be carefully determined based on
specific applications.
1.3. UHF: opportunities and challenges
UHF provides dual benefits for ASL. First, the higher 𝐵
N
field provides an increased
intrinsic SNR of MRI signal, which is roughly proportional to 𝐵
N
1.65
(Pohmann, Speck, &
Scheffler, 2016) (Figure 1.5); Second, the higher B0 field provides a prolonged tracer half-life
(blood T1), which allows a larger perfusion-based contrast and a longer imaging time window.
However, UHF has its own intrinsic challenges. On one hand, since the resonance frequency is
proportional to the 𝐵
N
field strength, at UHF the wavelength of the RF pulse is shorter, leading to
a more inhomogeneous transmit 𝐵
O
field. On the other hand, the 𝐵
N
offset issue is also more severe
due to hardware manufacturing challenges and increased susceptibility effect.
The inhomogeneities of 𝐵
O
P
and 𝐵
N
field has posed challenges for implementing ASL
sequences at UHF. For pulsed PASL,
the inhomogeneous 𝐵
O
P
field leads to
spatially varied flip angle for
traditional inversion pulse, which leads
to reduced labeling efficiency and
residual tissue signal overwhelming
the perfusion signal. For pCASL, the
𝐵
O
P
drop and 𝐵
N
inhomogeneity at the
labeling plane (usually below
Figure 1.5 The simulated labeling efficiency of a pCASL
sequence with the presence of 𝐵
O
P
and 𝐵
N
inhomogeneities.
Parameters of the pulse train were adopted from Wu WC, et al.,
Magn Reson Med, 2007
7
cerebellum for brain perfusion imaging)
make it difficult to achieve a high
labeling efficiency at 7T (W. C. Wu,
Fernández-Seara, Detre, Wehrli, &
Wang, 2007). On the one hand, a lower
𝐵
O
P
amplitude allows less efficient
adiabatic inversion when the flow passes
the labeling plane; On the other hand,
the existence of the 𝐵
N
offset induces
additional phase increment between
each labeling pulse that cannot be accounted for, which further undermines the labeling efficiency.
Figure 1.6 shows the simulated labeling efficiency vs. the relative 𝐵
O
P
amplitude and 𝐵
N
offset.
Therefore, the inhomogeneity problem of 𝐵
O
P
and 𝐵
N
field has to be addressed properly for the
implementatin of ASL sequences at UHF.
1.4. Parallel RF transmission (pTx)
To mitigate the 𝐵
O
P
inhomogeneity problem at UHF, the concept of pTx was proposed
(Hoult & Phil, 2000; Ibrahim, Lee,
Baertlein, Kangarlu, & Robitaille,
2000). Compared with single
transmission (1Tx) system that utilizes
one channel to emit RF pulses, a pTx
system uses a transmit coil array, which
consists of multiple channels designed to
Figure 1.6 The intrinsic SNR is roughly proportional to 𝐵
N
1.65
,
which means at 7T the SNR is nearly three times higher than the
3T. Adapted from: Pohmann R., Speck O., and Scheffler K. Magn
Reson Med. 2016 Feb: 75(2): 801-9
Figure 1.7 Demonstration of the combined 𝐵
O
P
field of pTx.
Relative uniform 𝐵
O
P
map is achieved by overlapping the peak and
valley of the standing waves of different transmit channels.
Adapted from mriquestions.com.
8
produce spatially distinct 𝐵
O
P
field patterns. A simple way to understand the alleviation of 𝐵
O
P

inhomogeneity problem by pTx is shown in Figure 1.7: different transmit channels are arranged
in a way such that their peak and valley of the RF standing waves are cancelled with each other,
therefore resulting in a relative uniform combined 𝐵
O
P
field. Since the idea of pTx-facilitated RF
acceleration was proposed (Katscher, Börnert, Leussler, & van den Brink, 2003; Zhu, 2004), pTx
has gained great interest of the research community. Currently, dual-channel transmitters have
been installed for latest 3T MRI scanners, and 8 or 16-channel transmitters have been installed
with many 7T scanners.
With a pTx system, since each of the channel is driven by a separate RF amplifier, each
can independently control the magnitude and phase of their own RF pulse. According to the
principle of superposition, the 𝐵
O
P
field produced in the subject will be the sum of 𝐵
O
P
field
produced by each channel, hence Eq. 1-1a:
𝑩
𝟏
(𝒓,𝑡)=∑ 𝐁
𝟏,𝐢
(𝒓,𝑡)
Y
Z
[\O
(Eq. 1-1a)
where 𝑁
^
is the channel number, 𝒓 is the spatial vector. Note that for each channel, the 𝐵
O
P
field it
produced can be separated into a temporal pulse waveform function and a spatial sensitivity map,
as shown in Eq. 1-1b
𝑩
𝟏
(𝒓,𝑡)=∑ 𝑝
[
(𝑡)𝑆
[
(𝒓)
Y
Z
[\O
(Eq. 1-1b)
where 𝑝
[
is the RF pulse waveform of the i-th channel, and 𝑆
[
is the transmit spatial sensitivity
map of the i-th channel. This format of pTx implementation is called “dynamic pTx”, in which
each channel has a channel-specific pulse waveform 𝑝
[
(𝑡).
If the pulse waveforms are the same for all channels, then Eq.1-1b can be simplified as Eq.
1-2:
𝑩
𝟏
(𝒓,𝑡)=𝑝(𝑡)∑ 𝛼
[
𝑆
[
(𝒓)
Y
Z
[\O
(Eq. 1-2)
9
where 𝛼
[
is the channel-specific
complex weight that modifies
the amplitude and phase of the
input RF pulse for each channel.
This format of pTx
implementation is called “static
pTx”. The third form of pTx is
“multi-pulse pTx”, for which
the channel-specific weights can vary throughout an MRI sequence, e.g., different weights for
different pulses or slices. The channel-specific transmit sensitivity map is usually required for all
pTx methods, which can be acquired with a 𝐵
O
P
mapping sequence. Figure 1.8 shows the channel-
specific 𝐵
O
P
map of an 8-channel pTx system at 7T (acquired on Siemens 7T Terra system with
Nova 8Tx/32Rx coil) and the channel-specific 𝐵
O
P
map of a 2-channel pTx system at 3T (acquired
on Siemens 3T Prisma system with 2Tx/20Rx coil) for comparison. The much lower inter-channel
coherence of 7T 𝐵
O
P
maps compared with 3T provides more freedom of pTx at 7T.
Many pTx studies have focused on the design of channel inputs with two separate goals in
mind: 1) to mitigate the 𝐵
O
P
inhomogeneity problem, and 2) to achieve local excitation. Note that
𝐵
O
P
inhomogeneity compensation can be achieved by all three forms of pTx: The static pTx scales
the amplitude and shifts the phase of RF pulses to alter the distribution of the combined 𝐵
O
P
field
(Mao, Smith, & Collins, 2006; Setsompop, Wald, Alagappan, Gagoski, & Adalsteinsson, 2008;
van den Bergen, Van den Berg, Bartels, & Lagendijk, 2007), and the dynamic pTx can directly
incorporate 𝐵
O
P
and 𝐵
N
inhomogeneities into the pulse design (Cloos et al., 2012; Katscher et al.,
2003). Meanwhile, other goals of pTx design can be achieved with their corresponding objective
Figure 1.8 A The channel-specific 𝐵
O
P
map acquired on a gel phantom with
an 8-channel pTx system at 7T. Unique excitation pattern of each channel
can be observed. B The channel-specific 𝐵
O
P
map acquired on the same
phantom with a 2-channel pTx system at 3T. In comparison to 7T, the
channel-specific excitation was more homogeneous, and the inter-channel
coherence was high.
10
functions, such as “maxmin (maximize the minimum)” shimming for adiabatic conditions
(Balchandani et al., 2011).

1.5. Overview of Studies
The purpose of this dissertation is to develop a reliable ASL perfusion sequence at 7T by
optimizing the existing ASL techniques and developing innovative ASL sequences allowed by
pTx. First, the currently widely used PASL and pCASL sequences were optimized based on the
scenarios of 𝐵
O
P
and 𝐵
N
inhomogeneities at 7T; Second, pTx shimming was utilized to further
increase the labeling efficiency of the pCASL sequence with channel-specific weights optimized
to increase the 𝐵
O
P
amplitude at the inflowing arteries; Third, an innovative CASL technique with
dynamic pTx shimming was proposed to increase the labeling efficiency as well as the robustness
to 𝐵
O
P
and 𝐵
N
inhomogeneities.
Optimization of adiabatic pulses for PASL at 7T
In Chapter 2, the PASL sequences were optimized and evaluated at 7 Tesla. With a custom-
defined loss function specifically designed for Flow-sensitive Alternating Inversion Recovery
(FAIR) PASL sequence, four most commonly used adiabatic pulses were optimized in simulation
and evaluated with phantom and in-vivo experiments. The wide-band-uniform-rate-smooth-
truncation (WURST) pulse achieved the lowest loss in simulation and achieved a superior
performance compared with the other three pulses in the experiments.
Optimization of pCASL at 7T
In Chapter 3, the parameters of the pCASL labeling were optimized for 7T. In order to
achieve higher labeling efficiency with the 𝐵
O
P
and 𝐵
N
inhomogeneities at 7T, the pCASL labeling
parameters were optimized based on the 𝐵
O
P
and 𝐵
N
distributions at the inflowing arteries obtained
11
from the subject cohort of Chapter 2, and the optimized pCASL labeling was implemented for in-
vivo experiment. The optimized pCASL sequence achieved robust labeling efficiency as well as
good repeatability.
Increased Labeling Efficiency of pCASL with pTx 𝑩
𝟏
P
shimming
In Chapter 4, the optimized pCASL sequence was incorporated with pTx 𝐵
O
P
shimming to
further increase the labeling efficiency. The pTx 𝐵
O
P
shimming was used to increase the 𝐵
O
P

amplitude at the inflowing arteries. The “indv-shim”, which utilized shimming weights calculated
for each individual subject, and the “univ-shim”, which used a set of universal weights calculated
based on a group of subjects, were evaluated, with circular polarized (CP)-shim as benchmark.
The three 𝐵
O
P
shimming modes were implemented as three pCASL sequences (pCASL-indv,
pCASL-univ, and pCASL-CP, respectively) and were evaluated with in-vivo experiments.
Compared with pCASL-CP, pCASL-indv achieved 9.5% higher perfusion signal (p<0.05), and
pCASL-univ increased perfusion signal by 5.3% (p=0.35).
An Innovative CASL Sequence Implemented with Dynamic pTx
In Chapter 5, an innovative CASL with Dynamic pTx pulse (DCASL) was proposed. By
utilizing channel-specific RF pulses for each transmit channel, the proposed DCASL has the
intrinsically insensitivity to 𝐵
N
offset and relatively low Specific Absorption Rate (SAR) similar
to conventional CASL sequence without violating the duty cycle limit of each RF transmit channel.
The pTx weights were calculated with a gradient descent algorithm based on a custom-defined
loss function. In simulation, the DCASL achieved comparable (~2% lower) labeling efficiency
compared with conventional CASL. Although not supported by the current sequence programming
software, it is still beneficial to the perfusion community since future software or other MRI
vendors may support it without the need of hardware modification.
12
Chapter 2. Optimization of adiabatic pulses for PASL at 7T:
Comparison with pCASL
2.1. Abstract
Objectives To optimize and evaluate adiabatic pulses for PASL at ultra-high field 7T.
Materials and Methods Four common adiabatic inversion pulses, including Hyperbolic-Secant
(HS), Wideband-Uniform-Rate-Smooth-Truncation (WURST), Frequency-Offset-Correction-
Inversion (FOCI), and time-resampled FOCI (trFOCI) pulses, were optimized based on a custom-
defined loss function that included labeling efficiency and inversion-band uniformity. The
optimized pulses were implemented in FAIR sequences and tested on phantom and 11 healthy
volunteers with two constraints: 1) SAR normalized; and 2) Equal peak RF amplitude,
respectively. A pCASL sequence was implemented for comparison. Quantitative metrics such as
perfusion and relative labeling efficiency versus residual tissue signal (rLE) were calculated.
Results Among the four pulses, WURST yielded the lowest loss in simulation and achieved a good
balance between labeling efficiency and residual tissue signal from both phantom and in-vivo
experiments. WURST-PASL showed significantly higher rLE compared to the other sequences
(P<0.01), while the perfusion signal was increased by 40% when the highest 𝐵
O
P
amplitude was
used. The four PASL sequences yielded comparable perfusion signals compared to pCASL but
with less than half SAR.
Conclusion Optimized WURST with the highest 𝐵
O
P
amplitude allowed was recommended for 7T
PASL.
2.2. Introduction
ASL is a perfusion MRI technique that utilizes magnetically labeled arterial blood water
13
as an endogenous tracer to measure CBF. The main limitation of ASL technique is the low SNR
due to the intrinsically small fraction of labeled arterial blood (~1%) and T1 relaxation of the label
(Rempp et al., 1994). UHF benefits ASL with an increased intrinsic SNR of MRI signal (𝐵
N
1.65
(Pohmann et al., 2016)) and a prolonged tracer half-life (blood T1). A number of ASL studies have
been performed to exploit the potential benefits of UHF ASL (Bause et al., 2016; Ghariq,
Teeuwisse, Webb, & van Osch, 2012; Luh, Talagala, Li, & Bandettini, 2013; Y. Wang et al., 2015;
Zimmer et al., 2016; Zuo et al., 2013).
However, implementation of ASL on UHF is not straightforward. Although pCASL is
recommended for 3T(Alsop et al., 2015), the 𝐵
O
P
drop and 𝐵
N
inhomogeneity at the labeling plane
make it difficult to achieve high labeling efficiency at 7T (W. C. Wu et al., 2007). The pCASL
sequence is further limited by the SAR of RF pulses that increases approximately quadratically
with field strength. For PASL, although SAR is less limiting compared with pCASL, the
inhomogeneous 𝐵
O
P
field leads to a spatially varied flip angle for traditional inversion pulses. The
FAIR labeling scheme(Kim, 1995) is commonly used for PASL at UHF, since the inversion pulse
is applied at the center of the imaging slab where 𝐵
O
P
field is relatively high. However, the resultant
perfusion map can still be contaminated by residual tissue signals due to mismatched inversion
profiles of the label and control pulses, especially for slices close to the edge of the inversion-band.
Adiabatic pulses have been used as the inversion pulse for PASL to mitigate the effects of
𝐵
N
and 𝐵
O
P
inhomogeneities at UHF, including Hyperbolic Secant (HS) (Silver et al., 1969),
Wideband Uniform Rate Smooth Truncation (WURST) (Kupce, 1995), Frequency Offset
Correction Inversion (FOCI) (Ordidge, Wylezinska, Hugg, Butterworth, & Franconi, 1996), and
time-resampled FOCI (trFOCI) pulses(Hurley et al., 2010). Theoretically, once the adiabatic
condition is met, the inversion profile is insensitive to variations in 𝐵
O
P
amplitude. In practice,
14
however, the properties of the inversion profile depend on multiple factors including the 𝐵
O
P

amplitude, 𝐵
N
offset, and parameters in generating the adiabatic pulses. In a previous study
(Hurley et al., 2010), the parameters of the trFOCI pulses were optimized based on a custom-
defined loss function using a genetic algorithm for a range of 𝐵
N
and 𝐵
O
P
field variations at UHF.
However, the performance of the trFOCI pulse and other adiabatic pulses has not been rigorously
evaluated at UHF for PASL applications.
The goal of the present work was to optimize the four adiabatic pulses (HS, WURST,
FOCI, and trFOCI) for FAIR PASL based on a loss function that takes into account the labeling
efficiency and inversion-band uniformity. The performance of the four optimized adiabatic pulses
for FAIR PASL at 7T was evaluated with both phantom and in-vivo experiments, as well as
comparison with a pCASL sequence adjusted for 𝐵
N
offset.
2.3. Materials and Methods
2.3.1. Loss function of PASL simulation
The profile of an inversion pulse can be described with three bands: pass-band, transition-
band, and inversion-band (Figure 2.1). The definition of the three bands in the inversion profile is
as follows: The pass-band (30mm on each side) is where the magnetization is preserved, the
inversion-band (60mm) is where the magnetization is inverted, and the transition-band (40mm on
each side) is between the pass-band and inversion-band. The widths of the three bands were
designed to satisfy a typical PASL experiment, with a 100mm selective inversion (between the
two zero-crossing points), and an imaging slab thickness of at least 60mm.
15
For the FAIR sequence, perfusion
can be calculated by the subtraction of the
images acquired with the selective and non-
selective inversion pulses. On one hand,
higher labeling efficiency is key to a
typically low-SNR PASL experiment. On
the other hand, the uniformity of the
inversion-band is of vital importance since
the amplitude of the perfusion signal is on
the order of 1% of the background tissue signal. Thus, a 1% fluctuation of signal in the inversion-
band will lead to residual tissue signal overwhelming the perfusion signal. Therefore, we defined
a loss function comprising of two terms: the labeling efficiency term 𝐿𝑜𝑠𝑠
cdd
, defined as the Root
Mean Square Error (RMSE) compared with the ideal profile of the inversion, transition and pass
bands; and the uniformity term 𝐿𝑜𝑠𝑠
ef[d
, defined as the standard deviation (SD) of the inversion-
band (Eq. 2-1), which was defined based on the assumption that the profile of non-selective
inversion is a constant without presence of the gradient (if adiabatic condition is met), and any
variation of the selective inversion-band will result in residual tissue signal. A weight 𝑤 was used
to combine the two terms, which was determined by approximately equal contribution of 𝐿𝑜𝑠𝑠
cdd

and 𝐿𝑜𝑠𝑠
ef[d
in the loss function. We performed a pilot study of parameter optimization with a
larger parameter step size, and comparable 𝐿𝑜𝑠𝑠
cdd
and 𝐿𝑜𝑠𝑠
ef[d
were observed for each pulse
with 𝑤 =300, which was used in the loss function. Strictly speaking, the RMSE term also
included the SD of the inversion-band, however the signal contribution is very small (<1%).
𝐿𝑜𝑠𝑠 =𝐿𝑜𝑠𝑠
cdd
+𝑤∗𝐿𝑜𝑠𝑠
ef[d
(Eq.2-1)
Figure 2.1 The definition of the three bands in the inversion
profile
16
2.3.2 Parameter Optimization
The definition of the HS, WURST, FOCI and trFOCI pulses were described in literature
(Hurley et al., 2010; Kupce, 1995; Ordidge et al., 1996; Silver et al., 1969) and summarized in
Appendix A. According to their definition, two parameters must be optimized for HS, WURST,
and FOCI pulses, respectively, while 11 parameters need to be optimized for trFOCI.
A schematic plot of the optimization process is shown in Figure 2.2. For HS, WURST,
and FOCI, the pulses and corresponding gradients were calculated across the 2-dimensional
parameter space. The inversion profile was then generated with Bloch simulation, from which the
loss function defined above was calculated (Figure 2.2A). For trFOCI, since it requires an
astronomic amount of computation to traverse the 11-dimensional parameter space, a genetic
algorithm modified from (Hurley et al., 2010) was used (Figure 2.2B). Considering the parameter
combination as an 11-dimensional vector, the initial population was chosen as the 250 vectors with
the lowest loss from 5000 randomly generated vectors. Then, a two-stage procedure was performed
to find the optimal vector including Iterations and Greedy Hill-climbing. For each iteration, the
50 vectors with the lowest loss from the last iteration were chosen as the “parental vectors”. The
population were produced by the “parental vectors” in two ways: two-point crossover and
mutation. Two-point crossover produced 600 “offspring vectors”, while mutation produced 50.
Therefore, a total of 650 vectors were produced. Once the group loss reached a steady-state, the
top 10 vectors of the final iteration were sent through Greedy Hill-climbing, during which each of
the 11 elements of each vector was adjusted individually and iteratively until the loss reached a
steady-state. The parameter vector with the lowest loss after Greedy Hill-climbing was the final
winner. The code of genetic algorithm was shared at our lab website (http://loft-lab.org).
17
The peak amplitude for the HS pulse was set to 20uT which is the typical peak 𝐵
O
P

amplitude that can be achieved within hardware limits at 7T (verified with vendor software), and
the WURST, FOCI, and trFOCI pulses amplitude were scaled to match the same SAR as HS.
Figure 2.3 shows representative 𝐵
N
and 𝐵
O
P
maps acquired on a 7T Siemens Terra MR scanner on
a human subject (25yr, male). Accordingly, we considered three scenarios of 𝐵
N
offset in the
simulation: 0, -300 to 300Hz linear shift, and 300 to -300Hz linear shift from -Zmax to Zmax,
respectively. To incorporate the effect of 𝐵1
P
inhomogeneity, the actual pulse amplitude was
scaled by 25%, 50%, 75%, and 100% of the target amplitude, respectively. The loss of each 𝐵
N

offset and 𝐵
O
P
inhomogeneity scenario was summed up as the final loss for each parameter
Figure 2.2 The parameter optimization pipeline for the (A) HS, WURST, FOCI pulses, and (B) trFOCI
pulse. A. For HS, WURST, and FOCI pulses, parameter set was obtained by traversing the 2D
parameter space, based on which the pulse waveform was generated, the inversion profile was
simulated with Bloch simulation, and the loss was calculated. B. For trFOCI, the genetic algorithm
was adopted to optimize the 11 parameters which consists of Iteration and Greedy Hill-climbing.

18
combination of each pulse to represent the possible B0/ 𝐵1
P
variations at 7T. Lastly, the optimized
HS, WURST, FOCI, and trFOCI pulses were generated with the optimal parameter set for phantom
and in-vivo experiments.
2.3.3 Inversion Profile Evaluation
Phantom and in-vivo experiments were performed on 7T MRI MAGNETOM Terra
(Siemen Healthcare, Erlangen, Germany) with an 8Tx/32Rx head coil (Nova Medical, MA, USA).
The RF-shimming mode was True-Form (Nistler, Diehl, Renz, & Eberler, 2004), which uses 45°
phase increment for each adjacent transmit channel to mimic the single transmit circularly
polarized (CP) coil. To demonstrate the inversion profile, the optimized pulses were applied with
corresponding gradients in the readout direction, immediately followed by turbo Fast Low Angle
Shot (TFL) readout (FOV: 210mm×210mm, Matrix size: 256×256). The inversion-band thickness
was set as 70mm, 100mm, and 300mm, respectively. The sequences were tested on a gel phantom
(https://www.goldstandardphantoms.com/funstar) (diameter=18cm) with T1≈700𝑚𝑠 , T2≈
Figure 2.3 B0 and B1 map in a young healthy subject at 7T. A) B0 offset appeared to be linearly shifting
along z-direction for intracranial region. B) The highest B1 amplitude appeared at the center of the brain
while it dropped dramatically (<30%) below the brainstem.
19
80𝑚𝑠, and 𝐵
O
P
field distribution close to that of a human head. The inversion profile was measured
along the central line of the phase-encoding direction, normalized by M0 (images acquired with
the identical protocol without inversion). The inversion profile was further demonstrated with in-
vivo experiments on one healthy subject (29yr male) with the same sequences as phantom
experiments.
Based on the inversion profile of the 100mm inversion pulse, the transition width, the
inversion efficiency, and the inversion profile difference between label and control were evaluated.
The transition width was defined as Mz between (0.9Mz_min+0.1Mz_max) and
(0.1Mz_min+0.9Mz_max). The inversion efficiency was defined as the area of (1–profile) on the
inversion-band divided by that of the ideal profile (corrected for T1 recovery). The inversion
profile difference was calculated as the mean difference between the inversion band of the
normalized inversion profile of 100mm and 300mm.
Since SAR is a major limiting factor affecting the performance of inversion pulses at UHF,
two constraints for 𝐵
O
P
amplitude were implemented: 1) SAR Normalized, the transmission
voltage was manually set to maximum peak RF amplitude allowed by the hardware (570 Volt) for
HS since it has the highest RF amplitude with the same SAR and is the easiest to be constrained
by hardware limit among all inversion pulses, and the voltage was scaled proportionally for other
pulses to achieve the same SAR as HS; 2) Equal peak RF amplitude, the maximum peak RF
amplitude allowed by the system was employed for each of the four pulses.
2.3.4 In-vivo ASL Experiment Protocol
The optimized pulses were implemented in FAIR ASL sequences (noted as HS-PASL,
WURST-PASL, FOCI-PASL, and trFOCI-PASL, respectively), with a single-shot TFL readout.
The scan parameters were as follows: TR=4000ms, TE=1.38ms, flip angle= 8° ,
20
bandwidth=490Hz/Px, inversion thickness=100/300mm for label and control,
TI2/TI1=1600/700ms, 9 slices with 5mm thickness and 2.5mm interslice gap, each slice
acquisition duration=150ms, FOV=210x192.5mm, matrix size=96x60, partial Fourier=6/8,
sequential k-space ordering, 50 measurements including 2 M0 and 24 label/control pairs for
3min22s.
For pCASL, 𝐵
O
P
and 𝐵
N
maps were measured at the labeling plane of one representative
subject (25yr male). Bloch simulation was performed to evaluate pCASL labeling efficiency vs.
𝐵
O
P
and 𝐵
N
inhomogeneity. The existing parameters of the pCASL pulse train were: mean
Gz=0.6mT/m, slice-selective Gz=6mT/m, Hanning window‐shaped RF pulse with duration=500
μs, RF gap=360 μs, flip angle=25° assuming laminar flow with a mean velocity of 40cm/s (Dai et
al., 2008).
Since the 𝐵
O
P
amplitude was constrained by global SAR and temporal resolution, the
pCASL sequence was only optimized by compensating the 𝐵
N
offset induced phase increment
between pCASL labeling pulses (W. C. Wu et al., 2007). We first measured the 𝐵
N
map of the
labeling plane using a dual-echo GRE sequence. Then, the phase adjustment was calculated from
the mean 𝐵
N
offset of the four inflow arteries (left and right internal carotid arteries (ICAs) and
vertebral arteries (VAs)) offline, which was then manually incorporated into the pCASL sequence.
The parameters of the pCASL TFL sequence were: TR=6000ms, TE=1.38ms, flip angle=8°,
labeling duration=1500ms, post-labeling delay=1500ms. The imaging parameters of pCASL TFL
readout were the same as as PASL TFL except that 50 measurements including 2 M0 and 24
label/control pairs were acquired for 5min2s. The labeling plane was placed based on a Time-of-
Flight (TOF) angiography image (Resolution=0.25x0.25mm, FOV=181.25x200mm, 3min49s) to
21
ensure that the inflow arteries were relatively straight and parallel to the 𝐵
N
field at the labeling
plane.
2.3.5 In-vivo ASL Data Analysis
To calculate perfusion maps for the PASL and pCASL sequences, images were corrected
for rigid head motion using SPM12 (Wellcome Trust Centre for Neuroimaging, UCL), then
pairwise subtraction of label and control images was performed, followed by averaging across
measurements.
The perfusion weighted signal for PASL with TFL readout can be expressed as (Alsop et
al., 2015; Zuo et al., 2013):
Δ𝑀 =
no∙qr
s
∙^tu
vNNNw ∙xyz{
| }
~
| s

∙
(
s
)

(
s
)

P(O
s
)∙
s(
s
)

(s
s
)
(Eq.2-2)

Similarly, the TFL pCASL perfusion weighted signal can be modeled as:
Δ𝑀 =
no∙qO

∙^tu ∙(Oxyz (/qO

))
vNNNw ∙xyz{

| s

∙
(
s
)

(
s
)

P(O
s
)∙
s(
s
)

(s
s
)
(Eq.2-3)

where 𝜆 is the brain/blood partition coefficient, 𝛼 is the labeling efficiency, 𝑇𝐼
n
is the inversion
time, 𝑃𝐿𝐷 is the post-labeling delay, 𝐸
O
=exp (−
q
qO
), 𝜃 is the flip angle of RF excitation in the
TFL readout, and 𝑗 is the index of the center phase-encoding line of the TFL readout.
Two metrics were calculated for the evaluation of adiabatic pulses: labeling efficiency and
residual tissue signal (between label and control acquisitions). The following notation was used:
We assume the mean gray matter (GM) perfusion weighted signal for each slice is 𝑃
⃑
=
[𝑝
O
,𝑝
n
,…,𝑝
f
] where n is the slice number. The ratio of perfusion weighted signals measured by
PASL and pCASL, 𝑷
𝐏𝐀𝐒𝐋
(𝑖) and 𝑷
𝒑𝑪𝑨𝑺𝑳
(𝑖), can be derived from dividing Eq.2-2 by Eq.2-3

𝑷
𝑷𝑨𝑺𝑳
(𝑖) / 𝑷
𝒑𝑪𝑨𝑺𝑳
(𝑖)=𝐴∙
o
¬
o
®Z¬
(Eq.2-4a)

22
where 𝐴 =
qr
s
∙xyz (
| }
~
| s

)
qO

∙(Oxyz (/qO

))
is a constant, and 𝛼
¯°±²
and 𝛼
³^°±²
are the labeling
efficiency for PASL and pCASL, respectively. Both pCASL labeling efficiency 𝛼
³^°±²
and PASL
labeling efficiency 𝛼
¯°±²
should be the same for each slice; however, due to the imperfect
inversion profile, the measured PASL labeling efficiency would be contaminated by a slice-
dependent residual tissue signal. Therefore, the point division of 𝑃
¯°±²
and 𝑃
³^°±²
, noted as
𝑅, becomes:

𝑹= 𝑷
𝑷𝑨𝑺𝑳
./ 𝑷
𝒑𝑪𝑨𝑺𝑳
=
°
o
®Z¬
(𝛼
¯°±²
+ µ𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
O,¯°±²
,𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
n,¯°±²
,…,𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
f,¯°±²
¹)
(Eq.2-4b)

where 𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
[,¯°±²
is the correction term related to the residual tissue signal of the i-th slice.
Based on the principle of the FAIR sequence (Kim, 1995), there should be no residual tissue signal
in the perfusion map of the central slice. Therefore, let 𝑚 =
fPO
n
(𝑛 𝑖𝑠 𝑎𝑛 𝑜𝑑𝑑 𝑛𝑢𝑚𝑏𝑒𝑟), then
𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
½,¯°±²
=0, and 𝑹(𝑚) indicates the relative labeling efficiency of the PASL inversion
pulse (noted as labeling efficiency or LE metric) without contamination of residual tissue signal,
and the standard deviation of 𝑹 will be
°
o
®Z¬
std(𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
O,¯°±²
,𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
n,¯°±²
,…,𝑅𝑒𝑠𝑖𝑑𝑢𝑎𝑙
f,¯°±²
) , which indicates the measured
perfusion signal variation across slices caused by the residual tissue signal (noted as residual signal
or RS metric). We further defined a metric as the ratio between LE and RS: 𝑟𝐿𝐸 =
²
±
, which
indicates the relative labeling efficiency versus the residual tissue signal, and thus is directly
related to perfusion map quality and accuracy. Following the calculation of the mean GM perfusion
for each slice, the LE, RS, and rLE metrics were calculated for each subject from Eq. 2-4b, and
23
pair-wise t-test was performed to test if there was significant difference of the rLE metric across
the four adiabatic inversion pulses.
Quantitative CBF maps were calculated for the WURST-PASL with peak 𝐵
O
P
amplitude,
and for the pCASL sequence respectively. For PASL, labeling efficiency=0.95 was
assumed(Wong, 2005), and for pCASL, labeling efficiency was simulated based on 𝐵
O
P
/𝐵
N
at
labeling plane. To further quantify the image quality of the perfusion maps, temporal SNR (tSNR)
of GM and white matter (WM) perfusion maps were also calculated for the whole image volume
of each subject and each sequence, respectively.
2.4. Results
2.4.1 Adiabatic Inversion Pulse Simulation
Optimized Parameters Bandwidth Peak
pulse
amplitude
Efficiency
loss
Uniformity
loss
(weighted)
Combined
Loss
HS 𝜇 =14.9,𝛽 =6.14 3.79 𝑘𝐻𝑧 20 𝜇𝑇 4.84 9.23 14.07
WURST 𝑘 =1.58×10
Æ
,𝑛 =2 5.06 𝑘𝐻𝑧 13.5 𝜇𝑇 5.64 5.52 11.16
FOCI 𝜇 =6.08,𝛽 =4.34 1.09 𝑘𝐻𝑧 12.7 𝜇𝑇 3.01 156.83 159.84
trFOCI [21.23, 0.53, 0.62,
0.82,0.27, 0.96, 1.00,
4.58, 6.51, 0.64, 0.04]
1.24𝑘𝐻𝑧 15.4 𝜇𝑇 8.27 5.06 13.33
The optimized parameters, bandwidth, peak pulse amplitude and loss of the four optimized
adiabatic inversion pulses are shown in Table 2.1. In the simulation, the WURST pulse showed
the lowest loss among the four pulses, while FOCI showed the highest loss due to significantly
Table 2.1. Parameters and losses of the optimized pulses. For the trFOCI pulse, the parameter vector is
[𝐴
½ÇÈ
,𝑤,𝑟
O
,𝑟
n
,𝑟
É
,𝑟
Æ
,𝑟
Ê
,𝜇,𝛽,𝜏
O
,𝜏
n
] . FOCI has the lowest labeling efficiency loss and trFOCI has the lowest
uniformity loss, while WURST has the lowest combined loss with a good balance of both labeling efficiency and the
inversion band uniformity.
24
greater variations in the inversion-band than the other pulses. Among the HS, trFOCI, and WURST
pulses, the uniformity losses of trFOCI and WURST were lower than that of HS, while HS had a
slightly lower inversion efficiency loss. In Figure 2.4, the RF waveform (top row) of the optimized
pulses as well as the corresponding gradients (middle row) and their inversion profiles (bottom
row) are shown. The peak amplitude of the HS pulse was set as 20uT (100% 𝐵
O
P
amplitude), and
the rest pulses were scaled to have the same SAR as that of the HS pulse. The inversion profiles
were further simulated with 75%, 50%, and 25% of the maximum amplitude of the corresponding
pulses. The maximum gradient amplitudes and slew rates of all waveforms were well within the
hardware limitation.
2.4.2 Inversion Profile Evaluation
As shown in Figure 2.5, for each of the 4 pulses, the inversion profile was measured from
the central phase-encoding line (red dashed line) of the image and normalized by M0 of the gel
Figure 2.4 RF and gradient waveform and inversion profile with scaled B1 of HS, FOCI, trFOCI and WURST
pulses. The peak B1 amplitude of HS was set to 20uT, then the other three pulses were scaled to have the same SAR
as HS. Inversion profile was simulated with targeted B1 scaled by 25%, 50%, 75% and 100% for each pulse. When
scaled by 50%, the uniform inversion band no longer exists with FOCI. trFOCI has a more uniform inversion band
but lower labeling efficiency compared with HS and WURST.
25
phantom. The inversion-band Mz was around -0.45, which was due to the T1 recovery during the
time to reach k-space center. Evaluated with SAR Normalized setting (row 2), in terms of transition
width, FOCI had the widest (11.48mm) while HS (6.56mm) and WURST (6.57mm) pulses were
comparable, which is consistent with the simulation results. The trFOCI pulse had the narrowest
transition width(3.28mm). In terms of inversion efficiency, the HS and WURST pulses were
comparable (95% vs 95%), FOCI was slightly lower (88%) due to wide transition-band, while
trFOCI was much lower (76%), which was consistent with the highest efficiency loss in
simulation. In terms of inversion profile difference, WURST (0.0062) and FOCI (0.0059) were
lower than HS (0.0079), while trFOCI (0.0110) was the highest, which indicates the uniformity of
the respective inversion band given the 𝐵
O
P
and 𝐵
N
field at 7T. When the RF transmission voltage
Figure 2.5 Inversion profile measured with phantom experiment. Row 1, the phantom image of four inversion
pulses with inversion thickness=100mm, red line indicating where the inversion profile was plotted; Row 2, the
inversion profile with SAR Normalized pulses; Row 3, the inversion profile with pulses of Equal peak RF
amplitude.
26
was increased to hardware limit (row 3), the profile of HS, WURST, and FOCI did not change
dramatically, suggesting that the adiabatic condition was already met with normalized SAR (row
2). For trFOCI, the bowl-shaped distortion of the inversion-band disappeared with the maximum
RF voltage, suggesting that the inversion efficiency of trFOCI dropped with a lower B1.
As shown in Figure 2.6, the inversion profile of the 4 pulses with 100mm width was also
measured on a human head. Same layout was used as Figure 2.5. Although the quantification of
the transition width was confounded by the brain structure, sharper transition for trFOCI and wider
transition for FOCI compared with HS and WURST can still be observed; Assuming that the
average T1 is 1900ms(Wright et al., 2008) at 7T , the inversion efficiency for HS, WURST, FOCI,
and trFOCI were 91.3%, 94.7%, 92.4%, and 82.4% (calculated with SAR normalized images),
respectively, which is consistent with phantom experiment results.
Figure 2.6 Inversion profile measured with in-vivo experiment. Same layout was used as phantom experiments.
Highest labeling efficiency was achieved with WURST, which was consistent with phantom experiments.
27
2.4.3 In-vivo pCASL optimization
Figure 2.7 shows the position of the labeling plane (A), 𝐵
O
P
and 𝐵
N
maps of the labeling
plane (B&C) for one representative subject, and the simulated pCASL labeling efficiency vs. 𝐵
N

and 𝐵
O
P
magnitude values (D). Based on the acquired 𝐵
O
P
and 𝐵
N
map, left and right ICAs and VAs,
which were targeted for labeling, were associated with around 50% of desired 𝐵
O
P
amplitude and
Figure 2.7 A) The labeling plane overlain on the coronal view of TOF angiography. The red dash line indicates
the labeling plane. B) The B1 map at the labeling plane. Compared with the desired B1 amplitude, the B1
amplitude at the targeted arteries (red dots, upper two: ICAs, lower two: VAs) is scaled to about 50%. C) The
B0 map at the labeling plane. The B0 offset at the targeted arteries is about 90Hz. D) Labeling efficiency vs.
B1 and B0. Ideally (B1 amplitude = 100% of desired 25°, B0 offset=0), the labeling efficiency is as high as
0.8; with B1=50% and B0 offset=0, the labeling efficiency=70%; with B1=50% and B0 offset=90Hz, the
labeling efficiency=60%.
28
around -90Hz 𝐵
N
offset. From Figure 2.7D we estimated that adjusting the phase increment based
on the 𝐵
N
map of the labeling plane in individual subjects can increase the labeling efficiency from
around 60% to approximately 70%. Therefore, labeling efficiency of 70% was used for the pCASL
CBF calculation for in-vivo experiment.
2.4.4 In-vivo PASL evaluation
Figure 2.8 shows the fractional perfusion maps (perfusion signal divided by M0) of four
TFL PASL sequences with the same SAR and the pCASL sequence from a representative subject.
The central slice is marked with the red box. The HS, WURST, and FOCI-PASL had higher
fractional perfusion signals than trFOCI-PASL and pCASL. With PASL, the first two slices had
higher perfusion values than the rest of the slices. This artifact was caused by residual tissue signal
since the perfusion signal intensity should be consistent across all slices as suggested by the
pCASL result. Signal intensity fluctuation across slices was observed for trFOCI-PASL, e.g.,
lower perfusion signal on the third and sixth slices. Considering that during simulation the trFOCI
inversion-band was relatively uniform, the possible reasons for the discrepancy will be discussed
below.
Figure 2.9 shows the fractional perfusion maps calculated from four PASL sequences with
the same peak 𝐵
O
P
amplitude and the pCASL sequence for the same subject, following the same
layout as Figure 2.8. It can be observed that except for HS-PASL whose 𝐵
O
P
amplitude did not
29
change, all other PASL sequences showed a higher perfusion signal than their counterparts in
Figure 2.8. However, residual tissue signal also increased.
For quantitative analysis, the LE, RS, and rLE metrics were calculated for each sequence
and each subject under each of the 𝐵
O
P
amplitude constraints. Figure 2.10 shows the box plots of
Figure 2.8 Perfusion map acquired using the pulses with the same peak RF amplitude. The 𝐵
O
P
amplitude of WURST,
FOCI, and trFOCI pulses were increased, leading to increased labeling efficiency. For HS and FOCI, the residual
tissue signal is more dominant at bottom slices. Red box indicates the central slice, of which the mean gray matter
perfusion is 0.42%, 0.60%, 0.51%, 0.46%, and 0.38% for HS, WURST, FOCI, trFOCI and pCASL, respectively.
Figure 2.9 Perfusion map acquired using the pulses with the same SAR. HS and FOCI had higher signal intensity
especially at bottom slices, likely originated from residual tissue signal. Compared with WURST, trFOCI had lower
signal intensity, suggesting lower labeling efficiency. Red box indicates the central slice, of which the mean gray
matter perfusion is 0.42%, 0.42%, 0.39%, 0.33%, and 0.38% for HS, WURST, FOCI, trFOCI and pCASL,
respectively.
30
measured rLE values, and the WURST-PASL had significantly higher rLE compared with any of
the other three PASL sequences for both constraints (P<0.01 for all comparisons). As rLE indicates
the relative labeling efficiency compared with residual tissue signal, the highest rLE of WURST-
PASL suggests the perfusion map of WURST was superior in terms of quality and accuracy
compared with other pulses. This quantitative result is consistent with our visual observations
based on Figures 2.8 and 2.9, and also agrees with the phantom experiment and simulation. The
mean LE, RS, and rLE values across all subjects are listed in Table 2.2.
Other quantitative metrics are also shown in Table 2.2, including mean GM perfusion,
SNR of GM and WM perfusion map, and SAR percentage relative to the FDA limit (first level,
3.2W/kg) monitored during experiments by the vendor software. FOCI showed higher values of
labeling efficiency (i.e., perfusion, LE, and SNR) which was in agreement with the simulation
results; however, due to its wide transition-band, edge slices were severely contaminated by
residual tissue signals, as reflected by the highest RS.

Figure 2.10 Comparison of the rLE term across 4 adiabatic pulses. The WURST pulse had the highest rLE
under both constraints of SAR Normalized and Equal peak RF amplitude (P<0.01), which indicated a higher
perfusion signal ratio against residual tissue signal of WURST. rLE, relative labeling efficiency.
31
When the peak RF amplitude was increased to the maximum allowed (Constraint 2), the
perfusion weighted signal was increased for each pulse as well as LE, RS, and SNR terms. For
WURST-PASL, the fractional perfusion signal increased 40% from 0.45% to 0.63%.
Figure 2.11A shows CBF maps of three subjects using WURST-PASL with peak 𝐵
O
P

amplitude, which are consistent across slices and subjects. The average GM and WM CBF of all
the 11 subjects was 61.4±19.4 and 21.0±8.7 ml/100g/min respectively, well matching literature
values(Pantano et al., 1984; Parkes, Rashid, Chard, & Tofts, 2004). Figure 2.11B shows CBF
maps of the same 3 subjects using the pCASL sequence. The average GM and WM CBF for

GM perfusion
(%)
LE RS rLE GM
SNR
WM
SNR
SAR
(%)
Constraint1
SAR
normalized
HS 0.50 1.32 0.39 3.73 1.89 1.41 21
WURST 0.45 1.21 0.18 8.41 1.37 0.86 21
FOCI 0.59 1.54 0.43 4.15 2.00 1.56 21
trFOCI 0.37 0.99 0.29 3.73 1.08 0.77 21

Constraint2
Equal RF
amplitude
HS 0.50 1.32 0.39 3.73 1.89 1.41 21
WURST 0.63 1.68 0.30 7.40 1.86 1.21 38
FOCI 0.64 1.69 0.44 4.39 2.10 1.54 40
trFOCI 0.44 1.16 0.31 4.13 1.28 0.94 38

pCASL

0.41

-

-

-

1.63

0.86

88
Table 2.2 Quantitative metrics of four PASL sequences and pCASL sequence. Metrics were measured per subject,
and the average value were shown here. LE, labeling efficiency; RS, residual tissue signal; rLE, relative labeling
efficiency.
32
pCASL was 32.3±10.91 and 17.2±4.4 ml/100g/min, respectively. The reason for the lower
labeling efficiency of pCASL than the theoretical value will be discussed below.
2.5. Discussion
In this study, we systematically optimized and evaluated 4 common adiabatic pulses for
FAIR-based PASL at 7T. The optimized WURST pulse achieved a significantly higher rLE
compared to HS, FOCI, and trFOCI, and the perfusion signal was comparable to and even higher
than pCASL with an increased 𝐵
O
P
amplitude. Therefore, the optimized WURST with the highest
possible 𝐵
O
P
amplitude within hardware and SAR limitations is recommended for 7T perfusion
imaging.
Figure 2.11 CBF maps of three representative subjects. A) CBF acquired with PASL of peak B1 amplitude WURST
pulse. B) CBF maps of the same subjects acquired with pCASL. Consistent pattern of perfusion maps were obtained
between PASL and pCASL, while WURST-PASL achieved higher perfusion amplitude compared with pCASL. Note
that the display range is different from panel A.
33
2.5.1. Definition of Loss Function
In this study, the loss function was designed to incorporate both inversion efficiency and
inversion band uniformity, while earlier studies have mainly focused on inversion
efficiency(Hurley et al., 2010; Zimmer et al., 2016). We considered three 𝐵
N
and four 𝐵
O
P

conditions during RF pulse optimization to achieve the optimal balance between inversion
efficiency and inversion band uniformity for the four common adiabatic pulses. Note that the
resultant inversion profile may appear bowl-shaped due to the 𝐵
O
P
/𝐵
N
distribution at 7T (Figure
2.5, trFOCI). Since the 𝐵
O
P
/𝐵
N
field for one specific voxel should be the same during label and
control, and if the designed inversion band is uniform (𝐿𝑜𝑠𝑠
ef[d
=0) with given 𝐵
O
P
/𝐵
N
, the
subtraction of label and control will be zero, even though the measured profile shows a bowl-
shape. Therefore, as long as a low uniformity loss is achieved in simulation, although the actual
inversion profile may be seemingly ununiform, a clean subtraction of brain tissue signal in label
and control acquisitions can be achieved for perfusion imaging.
2.5.2. Comparison of WURST with FOCI and trFOCI
The FOCI pulse is typically used when a sharp transition-band is needed. In our
experiment, however, the FOCI pulse was associated with the widest transition. One possible
reason is that FOCI is more sensitive to the adiabatic condition with SAR constraints: the same
adiabatic condition should remain the same for FOCI as the original HS pulse it was modified
from, however due to the modulation function, the 𝐵
O
P
amplitude was increased along with SAR.
Therefore, only a smaller bandwidth of inversion could be achieved under the same SAR
constraint. This resulted in a lower gradient amplitude, and correspondingly a wider transition-
band. Moreover, due to the same reason, when B1 was scaled by 50% or less, the FOCI pulse
34
could no longer preserve a uniform inversion-band, resulting in a significantly higher uniformity
loss compared with other pulses (Table 2.1).

Figure 2.12 Simulated inversion profile of original trFOCI and the corresponding optimized WURST, with
and without 𝛥𝐵
N
. Apparent distortion of inversion profile was observed for trFOCI with the presence of 𝛥𝐵
N
,
while for WURST only a slight transition was introduced which suggests improved robustness of WURST to
B0 inhomogeneity compared with trFOCI.
Figure 2.13 Inversion profile measured with a gel phantom with original trFOCI pulses from literature and
the corresponding WURST pulse. When SAR was normalized, higher inversion efficiency was observed for
WURST. For WURST/trFOCI2010 pair, the pulses were optimized with inversion thickness=50mm and
pulse duration=13ms, thus trFOCI parameter of [5.98, 0.32, 0.73, 0.18, 0.33, 0.81, 0.04, 4.9, 4.9, 0, 0.83]
(Hurley et al., 2010) was used; for WURST/trFOCI2016 pair, the pulses were optimized with inversion
thickness=50mm and pulse duration=10ms, thus trFOCI parameter of [31.52, 0.28, 0.49, 0.13, 0.32, 0.83,
1, 0.8, 5.34, 0.46, 0.32] (Bause et al., 2016) was used.

35
The trFOCI pulse has drawn growing interest in the field due to its robustness to 𝐵
O
P

inhomogeneity, sharp transition-band, and uniform inversion-band. Hurley et al. (Hurley et al.,
2010) first introduced the concept of trFOCI, optimized the parameters for several inversion
thickness based on a metric for inversion profile accuracy, and compared the performance of
trFOCI with HS and FOCI pulses; O’Brien et al. (O'Brien et al., 2014) compared trFOCI and HS
pulses for anatomical imaging; later, Zimmer et al. (Zimmer et al., 2016) compared trFOCI with
FOCI for PICORE PASL experiment, and Bause et al. (Bause et al., 2016) optimized the trFOCI
parameters for 9.4T FAIR PASL. However, none of these studies quantified the inversion-band
uniformity, which is of vital importance for FAIR PASL as explained in Methods Section 2.3.1,
nor did they optimize and include WURST pulse in the comparison. We optimized WURST based
on our loss function with the same settings (i.e., pulse duration, B1 amplitude, inversion thickness,
etc.) used by Hurley et al. and Bause et al. The optimized WURST pulses had consistently lower
loss (10.4 vs. 62.6, and 15.8 vs. 127.3) in simulation (Figure 2.12) and higher inversion efficiency
(73.8 vs. 69.1 and 74.1 vs. 70.1) in gel phantom experiment (Figure 2.13) compared with the
corresponding trFOCI pulses, suggesting that WURST may provide a better option than trFOCI
for the application of FAIR PASL. As shown in Figure 2.12 comparing WURST with trFOCI
from Bause et at., low B1 amplitude in the presence of B0 offset leads to severe distortion of the
inversion profile for trFOCI, which likely contributed to the observed perfusion signal fluctuation
observed with in-vivo experiments (Figure 2.8). In contrast, B0 offset only causes small shift of
the profile for WURST, suggesting that WURST is more robust to B0 offset than trFOCI.
The optimized WURST pulse had a different waveform compared with the traditional pulse
(Kupce, 1995), which had a sausage-like shape and low peak pulse amplitude. The original
WURST pulse was designed for 1.5T MR experiments, for which the hardware peak 𝐵
O
P
amplitude
36
was more limiting instead of SAR. However, for 7T experiments, SAR limitation has to be
considered especially for perfusion imaging with high-energy adiabatic pulses. Therefore, in the
simulation, we limited all the pulses to have the same SAR, and the optimized WURST pulse had
an amplitude modulation factor of n=2 instead of the original n=40, which proved to be more SAR-
efficient. In practice, if SAR permits, an increased 𝐵
O
P
amplitude for the inversion pulse is always
recommended.
2.5.3. Comparison of PASL and pCASL
In the present study, the definition of the RS metric was based on the following
consideration: The measured perfusion image consists of both residual tissue signal and actual
perfusion signal. Since the pCASL perfusion image is free of residual tissue signal, the intrinsic
regional CBF variation can be controlled by dividing the PASL perfusion by pCASL perfusion
signal in a slice-by-slice manner (Eq.2-4). Note that the recommended WURST pulse did not have
the highest SNR compared with other pulses, which was not in conflict with the highest rLE term.
The custom-defined rLE term reflects the labeling efficiency relative to the residual tissue signal.
When the residual tissue signal is present, the measured SNR will be inflated due to the signal
contributed by the static background tissue although the true perfusion signal may not increase. In
fact, when the SNR was calculated for each slice, the variation across slices was quite small for
WURST, while HS and trFOCI had around 30% higher SNR for the two bottom slices compared
with the average of the rest 7 slices for WM which has low perfusion signal, suggesting inflated
SNR due to residual tissue signal. Therefore, the rLE term is a more accurate metric over SNR for
assessing the performance of adiabatic inversion pulses. Thus, the optimized WURST is
recommended based on its performance and robustness, as well as the ease of implementation.
37
When compared with pCASL, WURST-PASL showed advantages although the SNR was
lower. First, the WURST pulse had a much lower SAR even with peak 𝐵
O
P
amplitude (38% vs.
88% of first-level SAR limit). Second, WURST-PASL had a shorter TR (4s vs. 6s), which allowed
more measurement per unit time thereby increasing the SNR efficiency by Í
v
Æ
=1.22, assuming
negligible T1 effect. Considering that the TR may be further shortened as allowed by SAR in
PASL, the increase of SNR efficiency will be even higher. Third, pCASL required additional
scans, such as TOF angiography and 𝐵
O
P
/𝐵
N
mapping of the labeling plane, to improve labeling
efficiency, making it less straightforward for clinical usage. One potential drawback of PASL at
7T, however, is the limited imaging coverage allowed by the head coil (as a minimum of 10cm
thick labeling slab is required). For pCASL, whole brain coverage should be achievable given the
relatively thin labeling plane.
In this study, we used pCASL as a benchmark for the evaluation of PASL pulses, since
theoretically it has no residual tissue signal. However, the labeling efficiency for pCASL was
relatively low, even with an additional phase increment to correct for 𝐵
N
offset at the labeling
plane. The possible reasons are: First, the 𝐵
O
P
field was relatively low (~50% of brain center) at
the labeling plane which was not corrected, resulting in decrease of labeling efficiency from the
ideal case of ~90% to 70%, or even lower considering additional 𝐵
N
offset; Second, since the
arterial blood was labeled at four arteries (left and right ICAs and VAs), the phase increment for
the correction of 𝐵
N
offset could only be calculated from the mean 𝐵
N
offset of the four arteries.
Thus, for each artery, the 𝐵
N
-induced phase offset could not be completely corrected, leading to
compromised labeling efficiency. One potential pathway of further increasing the labeling
efficiency is to increase the 𝐵
O
P
amplitude of the labeling pulses. However, as SAR was already
88% of the FDA limit, this can only be achieved at the cost of even lower temporal resolution than
38
the current TR=6000ms. Several other methods have been proposed for the improvement of 7T
pCASL, including an extra neck labeling coil (Mora Álvarez, Stobbe, & Beaulieu, 2019) , using
dielectric pads to improve 𝐵
O
P
, and a phase-cycled prescan to correct phase offset for each artery
(Luh et al., 2013). Due to the focus on PASL optimization for this project, however, only correction
for the global 𝐵
N
offset was applied for pCASL. A systematic study on the optimization of pCASL
at 7T will be conducted and presented separately.
2.5.4. Limitations and future directions
It is worth noting that pTx has proved to be a powerful tool for UHF experiments(Deniz,
2019; X. Wu et al., 2018; X. Wu et al., 2019; Zhu, 2004). The advantage of pTx can be beneficial
to 7T ASL imaging. Take Ref (Tong, Jezzard, Okell, & Clarke, 2020) for example, the authors
improved 𝐵
O
P
amplitude at the labeling plane using static RF shimming, which in turn increased
the labeling efficiency for pCASL using VERSE pulses with reduced SAR. In the present study,
we only used the TrueForm method, making the current results easily generalized to the single
transmit CP coils. In practice, however, the imaging coverage with the 8Tx/32Rx coil is larger the
1Tx/32Rx coil. Based on our experience, only the upper brain can be imaged with our optimized
adiabatic pulses using the 1Tx/32Rx coil. Therefore, a logical next step for this project is to exploit
the full capability of the pTx system for maximizing the performance of PASL and pCASL
sequences while maintaining or even reducing SAR at 7T.
Several limitations have to be addressed in future studies. First, 2D TFL was used for image
acquisition in this project, which could not cover the whole brain during the imaging time window.
More efficiency acquisition schemes such as Simultaneous Multi-slice (SMS) (Y. Wang et al.,
2015) and 3D Gradient and Spin-Echo (GRASE) can be applied to increase imaging speed and
coverage (Spann et al., 2020). Second, background suppression can be used to increase the SNR
39
of the perfusion map, although at the expense of higher SAR and possibly lower labeling
efficiency(Garcia, Duhamel, & Alsop, 2005; Shao, Wang, Moeller, & Wang, 2018).

2.6. Conclusion
In this project, we systematically optimized and evaluated four adiabatic pulses using a
custom-defined loss function for FAIR PASL at 7T. The WURST pulse showed the lowest loss in
simulation and yielded superior performance compared with other inversion pulses for PASL
sequences and the pCASL sequence. Thus, the optimized WURST with highest peak 𝐵
O
P

amplitude allowed by hardware and/or SAR is recommended for 7T PASL imaging. Future work
will be performed on the joint design of adiabatic inversion pulses and pTx to further improve the
performance of 7T ASL. This work has been published in MRM (K. Wang et al., 2021) and a C2P
sequence has been disseminated to other 7T sites and is being implemented in Siemens WIP
sequence for 7T ASL.

40
Chapter 3. Optimization of Pseudo-continuous Arterial Spin
Labeling at 7T
3.1. Abstract
Objectives To optimize pseudo-continuous arterial spin labeling (pCASL) parameters for ultra-
high field 7T.
Methods pCASL labeling parameters including the RF duration, RF gap, mean gradient 𝐺
ÇÎc
, and
the ratio of maximum gradient and mean gradient 𝑔𝑅𝑎𝑡𝑖𝑜, were optimized based on the 𝐵
O
P
and
𝐵
N
field distribution at 7T. The labeling efficiency was estimated with Bloch equation simulation
over a wide range of pCASL parameters and 𝐵
O
P
/𝐵
N
conditions, and the weighting for each 𝐵
O
P
/𝐵
N

condition was their corresponding relative frequency of the occurrence estimated from a
previously published cohort. The candidate pCASL parameters were implemented and evaluated
with four subjects. The optimal pCASL was chosen as the one without distortion or contamination
of perfusion images by side lobes of labeling pulses that had the highest fractional perfusion signal.
The robustness and repeatability of the optimal pCASL sequence was then evaluated with in-vivo
experiments, along with a pulsed arterial spin labeling (PASL) sequence.
Results The optimal pCASL parameter set was: RF duration/gap=300/250us, 𝐺
ÇÎc
=0.6𝑚𝑇/
𝑚,𝑔𝑅𝑎𝑡𝑖𝑜 =10. The perfusion of the optimal pCASL sequence was 1.27±0.13%, and the
labeling efficiency was estimated to be 0.63. Fair repeatability was achieved with intra-class
correlation coefficient (ICC) and averaged within-subject Coefficient of Variation (wsCV) to be
0.66, 4.60%, respectively.
Conclusion The optimized pCASL was able to achieve robust and repeatable perfusion imaging at
7T
41
3.2. Introduction
In Chapter 2, the Pulsed Arterial Spin Labeling (PASL) sequence was optimized by
optimizing the parameters of the adiabatic inversion pulses for the 𝐵
O
P
and 𝐵
N
conditions at 7T.
However, for 3T perfusion, the pseudo-continuous ASL (pCASL) showed better repeatability and
reliability, and therefore was recommended by the perfusion research community (Alsop et al.,
2015). The advantage of pCASL over PASL is likely to hold true at 7T, but a fair comparison can
only be made with a pCASL sequence that is systematically optimized and adjusted for 7T.
Implementing pCASL at ultra-high field (UHF) is not straightforward, the largest challenge
being that the 𝐵
O
P
and 𝐵
N
field inhomogeneities at the labeling plane undermine the flow-driven
adiabatic inversion condition, resulting in low labeling efficiency (Ghariq et al., 2012; W. C. Wu
et al., 2007). A number of previous studies have been performed to address the challenges of
implementing pCASL at UHF. To address the 𝐵
O
P
amplitude drop caused by the 𝐵
O
P

inhomogeneity, the labeling plane was raised from neck region to the bottom of cerebrum to take
advantage of the higher 𝐵
O
P
amplitude (Zuo et al., 2013), and dielectric pads were used to improve
𝐵
O
P
efficiency and homogeneity (Y. Wang et al., 2015). To address the 𝐵
N
offset problem, a
prescan was used to correct for off-resonance effects at the labeling plane (Luh et al., 2013).
However, few studies have been performed to optimize the parameters for pCASL based on the
specific distributions of 𝐵
O
P
and 𝐵
N
field at UHF.
The goal of this study was to develop robust pCASL perfusion imaging at 7T. We first
optimized the pCASL labeling parameters based on the 𝐵
O
P
and 𝐵
N
distributions at the inflowing
arteries obtained from a previously published subject cohort (K. Wang et al., 2021), then the
optimized pCASL was implemented and evaluated with in-vivo experiments. We demonstrated
42
that the optimized pCASL is able to obtain reliable perfusion maps at 7T even with the 𝐵
O
P
and 𝐵
N

inhomogeneities.
3.3. Methods
3.3.1. Estimation of the 𝑩
𝟏
P
/𝑩
𝟎
distribution at 7T
In order to better evaluate the performance of the pCASL labeling at 7T, the conditions of
𝐵
O
P
and 𝐵
N
associated with pCASL labeling was estimated. The distribution of the 𝐵
O
P
and 𝐵
N

condition was estimated from a previously published cohort (K. Wang et al., 2021) with 11
subjects and thus 44 inflowing arteries (left and right internal carotid arteries (ICAs) and vertebral
arteries (VAs) ).
First, for each subject, a Time-of-Flight (TOF) sequence was used for angiography, the
labeling plane was placed at the C1 segment (Bouthillier classification) of the ICAs (Figure 3.1A),
and a 𝐵
O
P
(Figure 3.1B) and 𝐵
N
map (Figure 3.1C) were acquired at the labeling plane with turbo-
Fast-Low-Angle-Shot (TFL) and GRE sequences, respectively; Second, the ROIs of the four
inflowing arteries were manually drawn on the TOF image; Third, the ROIs was resampled
according to the 𝐵
O
P
and 𝐵
N
maps, respectively, and the mean 𝐵
O
P
amplitude and 𝐵
N
offset was
43
calculated for each artery. The histogram of the relative 𝐵
O
P
amplitude scaling and the 𝐵
N
offset
were shown in Figure 3.1D and Figure 3.1E, respectively.
3.3.2. pCASL labeling parameter optimization with numerical simulation
Based on the principle of pCASL (Dai et al., 2008), a pCASL labeling block can be
determined by the following parameters: the RF FA, RF duration, RF gap, mean gradient 𝐺
ÇÎc
,
and gradient amplitude during each RF pulse 𝐺
½ÇÈ
, as shown in Figure 3.2. 𝐺
½ÇÈ
can be expressed
Figure 3.1 A. The red dash line shows the labeling plane on top of the Maximum Intensity Projection on the
coronal plane of TOF; B. The B1 amplitude map at the labeling plane from one representative subject, normalized
by the designed value; C. The B0 map at the labeling plane from the same representative subject; D&E. The
distribution of the B1 amplitude and B0 offset with the 44 inflow arteries.

44
as 𝐺
ÇÎc
⋅𝑔𝑅𝑎𝑡𝑖𝑜. A commonly used set of pCASL parameters at 3T include RF FA=25°, RF
duration=500us, RF gap=420us 𝐺
ÇÎc
=0.6mT/m, and 𝑔𝑅𝑎𝑡𝑖𝑜=10 (Dai et al., 2008; Shao et al.,
2019), noted as “Labeling 0”.
To optimize pCASL at 7T, numerical simulation was performed with Bloch equation to
estimate the labeling efficiency (LE) for a wide range of pCASL parameters: Hanning-shaped RF
pulse, RF duration = [300:100:1000] us, 𝐺
ÇÎc
= [0.1:0.1:1.2] mT/m, and 𝑔𝑅𝑎𝑡𝑖𝑜 = [5:1:15]. The
RF gap was determined such that the ratio between RF gap and RF duration was the same as
Labeling 0 in order for the duty cycle of the RF power amplifier for playing out the pCASL pulses
to remain constant (54.35%), and the RF FA was determined to maintain the same time-averaged
SAR. The flow velocity was assumed to be 40cm/s (Zhao, Vidorreta, Soman, Detre, & Alsop,
2017). The LE was defined as the subtraction of the inversion efficiency of the Label and Control,
where for the Label the RF pulses are in phase while for Control a 180° phase increment was added
between adjacent RF pulses. The balanced pCASL scheme was used to avoid eddy current effect
(W. C. Wu et al., 2007).
The 𝐵
O
P
and 𝐵
N
inhomogeneities were incorporated in the optimization process by
considering 𝐵
O
P
amplitude variations with a scaling range of [20:10:120]% of the reference value,
Figure 3.2 Demonstration of one pCASL labeling block. The labeling block can be precisely described by the
timing parameters of RF duration and RF gap, and the gradient parameters 𝐺
½ÇÈ
and 𝐺
ÇÎc
. 𝐺
½ÇÈ
can be replaced
by 𝑔𝑅𝑎𝑡𝑖𝑜: 𝐺
½ÇÈ
= 𝐺
ÇÎc
⋅𝑔𝑅𝑎𝑡𝑖𝑜.

45
and 𝐵
N
offset with a range of [-200:20:200] Hz. The weighting for each 𝐵
O
P
/𝐵
N
condition was their
corresponding relative frequency. Therefore, the optimal set of pCASL parameters was chosen as
the one with the highest weighted sum of LE, noted as “Labeling 1”.
3.3.3. pCASL implementation and optimization
A pCASL sequence with a single-shot TFL readout was implemented with the following
parameters: FOV=210×192.5mm, matrix size=96×60, 9 slices with 5mm thickness and 2.5mm
interslice gap, each slice acquisition duration=150ms, partial Fourier=6/8, sequential k-space
ordering, TR=6000ms, TE=1.38ms, FA= 8° , labeling duration=1000ms, post-labeling
delay=1500ms, 40 measurements including 2 M0 scans and 19 label/control pairs were acquired
in 4min2sec.
Four sets of pCASL parameters were evaluated: 1. The benchmark parameters, Labeling
0; 2. The parameters optimized by simulation, Labeling 1; 3. RF (FA, duration, and RF gap) of
Labeling 1, and gradient (𝐺
ÇÎc
and 𝑔𝑅𝑎𝑡𝑖𝑜) of Labeling 0, noted as “Labeling 2”; 4. Same as
Labeling 2, except that the RF gap was reduced so that the gradient slew-rate reached hardware
limit, noted as “Labeling 3”. Labeling 2 was implemented to ameliorate the possible aliasing
problem (See 3.4 Results and 3.5 Discussion Section “pCASL labeling optimization in
simulation”), while Labeling 3 was to further minimize the RF gap with respect to Labeling 2.
Four subjects were recruited (2 Male, age=28.8±8.9 years) to evaluate the performance of
the four parameter sets. This study was approved by the Institutional Review Board of the
University of Southern California, and written informed consent was obtained prior to the
experiments from each subject. In-vivo experiments were performed on a 7T MRI MAGNETOM
Terra (Siemen Healthcare, Erlangen, Germany) with an 8Tx/32Rx head coil (Nova Medical, MA,
USA). The 𝐵
O
P
-shimming mode was True-Form(Nistler et al., 2004) which uses same amplitude
46
and 45° phase increment for each adjacent transmit channel to mimic the single transmit circularly
polarized (CP) coil . The scan protocol included a TOF to locate the labeling plane (same as Figure
3.1A), and four pCASL sequences with the same FOV and same labeling plane.
To calculate perfusion maps for the pCASL sequences, images were corrected for rigid
head motion using SPM12 (Wellcome Trust Centre for Neuroimaging, UCL), then pairwise
subtraction of label and control images was performed, followed by averaging across
measurements. The perfusion maps were then normalized by the M0 image resulting in fractional
perfusion maps. The four sets of labeling parameters were evaluated based on two criteria: 1)
perfusion images without distortion or contamination by the side lobes of labeling pulses, and 2)
labeling efficiency indicated perfusion signal intensity. The pCASL parameter set which yielded
the highest fractional perfusion signal without distortion or contamination would be considered as
the optimal parameter set for 7T pCASL.
3.3.4. In-vivo experiment
Six subjects were recruited (3 Male, age=25.8±3.2 years) for this validation study. Each
subject underwent two scans which were exactly 24-hour apart to control for daily physiological
fluctuation of brain perfusion.
Parameters TOF pCASL PASL
FOV (mm) 200×200 210×192.5 210×192.5
Matrix size 768×768 96×88 96×88
Slice number 136 9 9
Slice thickness (mm) 0.5 5 5
Voxel size(mm) 0.26×0.26 2.19×2.19 2.19×2.19
TR (ms) 19 6000 6000
TE (ms) 3.42 2.38 2.38
Measurements N/A 40 40
47
Duration 3min49sec 4min2sec 4min2sec

The scan protocol for each visit is listed in Table 3.1, which included: 1. A TOF sequence
for the localization of labeling plane and inflowing arteries; 2. A optimized pCASL sequence with
labeling parameters optimized in Methods Section 1&2; 3. A pulsed ASL sequence optimized in
a previous study(K. Wang et al., 2021) as a comparison for the pCASL sequence. The labeling
plane position was the same as shown in Figure 3.1A.
The process of preprocessing and perfusion map calculation was the same as the pCASL
parameter optimization study (Methods Section 2). The mean gray matter (GM) fractional
perfusion signal (normalized by M0) was calculated for each subject and each visit respectively.
The repeatability of the ASL sequences was evaluated by calculating the intraclass correlation
coefficient (ICC) and the average within-subject coefficient of variation (wsCV) of the perfusion
signal.
The LE of the pCASL sequences was also estimated. Assuming the LE of PASL is 0.95
(K. Wang et al., 2021; Wong, 2005), the mean GM CBF of the central slice was calculated(Alsop
et al., 2015; Zuo et al., 2013) (noted as CBFPASL). The mean GM CBF of the pCASL sequence
was calculated with LE=1(Alsop et al., 2015; Zuo et al., 2013) (noted as CBFpCASL), then the
estimated LE of the pCASL sequence was CBFpCASL/CBFPASL.
3.4. Results
3.4.1. pCASL parameter optimization in simulation
Figure 3.3 shows the contour map of the simulated LE of RF duration = 300us (A), 500us
(B), and 800us (C) w.r.t. 𝐺
ÇÎc
and 𝑔𝑅𝑎𝑡𝑖𝑜 . The contours are open-ended because the LE
calculation was skipped for those parameter sets that violated the hardware slew-rate limit
Table 3.1 The scan protocol for the in vivo experiments. TOF images were obtained for
angiography, after which pCASL and PASL sequences were acquired with the same
acquisition parameters.

48
(200T/m/s) to reduce simulation time. Due to the difficulty to display the 3D parameter space (RF
duration, 𝐺
ÇÎc
and 𝑔𝑅𝑎𝑡𝑖𝑜), results of three representative RF durations are shown. The highest
LE (weighted by 𝐵
O
P
/𝐵
N
conditions) achieved with the three RF durations were 0.81, 0.78, and
0.64, respectively, which showed a decreasing trend of LE w.r.t. RF duration. Indeed, the highest
LE (0.81) in the whole 3D parameter space was achieved with RF duration = 300us, 𝐺
ÇÎc
=
0.3 𝑚𝑇/𝑚, and 𝑔𝑅𝑎𝑡𝑖𝑜 =11, which served as Labeling 1.
The LE of the optimized parameter w.r.t. 𝐵
O
P
scaling and 𝐵
N
offset is shown in Figure
3.2D. For comparison, the same layout was used for the benchmark parameter (Labeling 0) in
Figure 3.2E. The LE of Labeling 1 showed small variation across the whole 𝐵
N
offset frequency
range, and LE>0.80 was achieved when 𝐵
O
P
scaling factor was larger than 0.6. Comparing to
Figure 3.3 A, B&C. The contour map of the simulated LE for RF duration=300us, 500us, and 800us, respectively.
These three RF duration values were chosen from all RF durations as representative cases. The highest LE with
each RF duration was marked with blue asterisks. The highest LE in the (RF duration, 𝐺
ÇÎc
, 𝑔𝑅𝑎𝑡𝑖𝑜) 3D parameter
space was achieved with (300us, 0.3 𝑚𝑇/𝑚, 11). Labeling 0 was marked with red asterisk. D&E LE vs. B1 and B0
with Labeling 1 and Labeling 0, respectively.

49
Labeling 1, although higher peak LE across the 𝐵
O
P
and 𝐵
N
parameter space was achieved(0.88 vs.
0.84), Labeling 0 showed larger LE variation w.r.t. 𝐵
N
offset suggesting less robustness to 𝐵
N

offset; moreover, Labeling 0 had a generally higher LE than Labeling 1 only when 𝐵
O
P
scaling was
large (greater than 0.8), while with 𝐵
O
P
scaling<0.6, LE was lower for Labeling 0.
3.4.2. In-vivo evaluation for pCASL parameter optimization
The parameters of the four sets of labeling parameters (Labeling 0, 1, 2, and 3) are reported
in Table 3.2. The fractional perfusion map of one representative subject is shown in Figure 3.4,
which exhibits perfusion maps possibly contaminated by the aliasing of the pCASL labeling pulses
when a smaller gradient was used (Labeling 1: 𝐺
ÇÎc
=0.3 𝑚𝑇/𝑚, 𝑔𝑅𝑎𝑡𝑖𝑜 =11, thus 𝐺
½ÇÈ
=
0.3×11=3.3 𝑚𝑇/𝑚). This pattern was consistent among other subjects (Figure 3.5). Therefore,
Labeling 1 was dropped from further evaluation. The mean GM fractional perfusion signal was
calculated for each labeling parameter set, also reported in Table 3.2. Labeling 3 had a higher
fractional perfusion signal (1.26±0.22%) compared to Labeling 0 (1.14±0.26%) and Labeling 2
(1.21±0.28%) although without statistical significance due to the limited sample size, therefore,
Labeling 3 was chosen as the optimal 7T pCASL labeling and used later for in-vivo experiments.

Name RF duration
(us)
RF FA
(°)
RF gap
(us)
𝐺
ÇÎc

(mT/m)
𝑔𝑅𝑎𝑡𝑖𝑜 Mean Perfusion
(%)
Labeling 0 500 25 420 0.6 10 1.14±0.26
Labeling 1 300 15 270 0.3 11 0.26±0.08
Labeling 2 300 15 270 0.6 10 1.21±0.28
Labeling 3 300 15 250 0.6 10 1.26±0.22
Table 3.2 The parameters of the pCASL sequences for the in-vivo evaluation as well as their perfusion. Labeling 3
achieved the highest perfusion and thus was chosen as the optimal 7T pCASL.

50

3.4.3. In-vivo experiment
The perfusion map of all six representative subject is shown in Figure 3.6. The optimal
pCASL sequence achieved perfusion maps of decent quality, with an average whole brain GM
perfusion of 1.27±0.13%. In terms of the repeatability, the ICC of the optimal pCASL was 0.66,
and the wsCV averaged across subjects was 4.60%, indicating a good level of repeatability.
Figure 3.5Perfusion map of one representative subject. Perfusion map of Labeling 1 was contaminated by the
aliasing of the labeling plane, while the difference between the other three pCASL labeling was visually
unnoticeable which was settled with quantitative analysis.
Figure 3.4 The perfusion maps of the remaining three subjects with Labeling 1. Severe contamination by the
pCASL labeling side band was observed in the imaging volume. The seemingly different location of the side band
was due to the different labeling offset which was adjusted for each subject.

51
Assuming the LE of the PASL is 0.95, CBFPASL was calculated to be 58.1mL/100g/min
well matching literature values (Pantano et al., 1984; Parkes et al., 2004), based on which the LE
of the optimal pCASL was estimated to be 0.62±0.10.

Figure 3.6 The fractional perfusion map of the six subjects with visit 1 (A) and visit
2 (B). Robust perfusion was achieved for each subject and each visit, although with
variation across subjects.
52
3.5. Discussion
In this study, we optimized the pCASL labeling based on the distribution of 𝐵
O
P
/𝐵
N
fields
at 7T. With in-vivo evaluation, the optimized pCASL labeling was adjusted, i.e., higher 𝐺
ÇÎc
(from
0.3mT/m to 0.6mT/m) and higher 𝐺
½ÇÈ
(from 3.3mT/m to 6mT/m), to generate the optimal
pCASL labeling. The optimized pCASL was able to achieve reliable and repeatable perfusion map.
3.5.1. pCASL labeling optimized in simulation
In this study, the pCASL parameters were optimized specifically for 7T applications, i.e.,
to provide LE with increased robustness to 𝐵
N
offset with a lower 𝐵
O
P
amplitude. Our simulation
yielded Labeling 1 (RF duration=300us, 𝐺
ÇÎc
=0.3mT/m, 𝑔𝑅𝑎𝑡𝑖𝑜=11) as the optimal set of pCASL
parameters which fits well with our expectation. Based on the principle of balanced pCASL, the
LE has a periodic fluctuation against Δ𝑝ℎ𝑎𝑠𝑒, which is the 𝐵
N
induced phase offset, which means
when Δ𝑝ℎ𝑎𝑠𝑒 =±𝜋, the Label and Control acquisitions will switch. Considering the phase error
accrued between adjacent labeling pulses is proportional to (RF (duration+gap)⋅𝐵
N
offset), a
shorter RF duration and gap would improve the robustness to 𝐵
N
offset. Meanwhile, according to
the flow-driven adiabatic inversion explanation of pCASL(Dai et al., 2008), 𝐺
ÇÎc
has to be reduced
accordingly when 𝐵
O
P
amplitude is lower, thus by simulation 0.3mT/m was the optimal 𝐺
ÇÎc

instead of the recommended value 0.6mT/m (Alsop et al., 2015).
However, the in-vivo evaluation suggested Labeling 3 as the optimal pCASL labeling
rather than Labeling 1. Possible reasons for the discrepancy are discussed as following.
3.5.2. Interference with imaging volume
The side lobes of the labeling pulses can interfere with the imaging volume, which was not
accounted for in the simulation. Note the RF pulse duration as 𝜏, and the pulse period as T. The
53
labeling pulse train can be viewed as the convolution of a Hanning windows and a comb function
of period T:
𝑓(𝑡)=
O
q
𝑐𝑜𝑚𝑏Ô
Õ
q
Ö∗𝐻𝑎𝑛(
Õ

) (Eq.3-1)
Then the main lobe width of the Hanning window is determined by the Fourier transform
of the Hanning function:
ℱ{𝐻𝑎𝑛Ô
Õ

Ö=𝜏𝑠𝑖𝑛𝑐(𝜋𝜏𝑓)⋅
O
n(OØ
~
d
~
)
(Eq.3-2)
Thus, the first zero-crossing is at 𝑓 =1/𝜏. Let 𝑓 ≡𝛾𝐺
½ÇÈ
⋅𝑧
O
, then 𝑧
O
=
O
ÛÜ
ÝÞß

, which
means a lower 𝐺
½ÇÈ
and a lower 𝜏 will lead to wider profile of the Hanning pulse.
Meanwhile, the location of the aliasing is determined by the Fourier transform of the comb
function:
ℱ{
O
q
𝑐𝑜𝑚𝑏Ô
Õ
q
Ö =𝑐𝑜𝑚𝑏(𝑓𝑇) (Eq.3-3)
Thus, the aliasing location is at 𝑓 =
à
q
,𝑘 =1,2,3,…. Let 𝑓 ≡𝛾𝐺
ÇÎc
⋅𝑧
n
, then 𝑧
n
=
à
ÛÜ
Þáâ
q
,
which means a lower 𝐺
ÇÎc
and a lower T will lead to farther distance of the first aliasing.
54
Therefore, Labeling 1 has wider side band of labeling profile due to two reasons: 1. The
lower 𝐺
½ÇÈ
(0.3×11=3.3 mT/m) and shorter RF duration widened the profile of the Hanning
pulse; 2. The shorter RF period
and the lower 𝐺
ÇÎc
pushed the
side band farther from the
labeling plane. To alleviate the
effect of labeling side band, one
must increase the RF period or
increase the 𝐺
½ÇÈ
. Since the RF
period was deliberately
decreased for improved 𝐵
N

offset robustness, the latter is
the only option left. Figure 3.7
analyzed the simulated excitation
profile of the labeling pulse on
static tissue for Labeling 0, 1, 2, and 3. Side band of Labeling 1 extended 89 mm away from the
labeling plane, which is likely in the imaging volume.
3.5.3. Labeling plane location
While seemingly unrelated to the LE simulation, it poses additional factor to be considering
when deciding the gradient parameters. Previous work has suggested three options: L1: on the
inferior border of the cerebellum, L2: above the carotid artery bifurcation and below V3 segment,
and L3: below the bifurcation. Although L2 was recommended at 3T (Zhao et al., 2017), at 7T the
very limited transmission coil coverage has to be considered, meaning the 𝐵
O
P
amplitude drops
Figure 3.7 Simulated excitation profile of the pCASL labeling parameters.
A. With Labeling 0, the first aliasing appeared at 21mm, well below typical
labeling offset value (~80mm); B. With Labeling 1, the first aliasing
appeared at 89mm, which likely will affect imaging volume. C&D.
Compared with Labeling 1, Labeling 2 (C) and 3 (D) had a higher gave
resulting in a smaller aliasing location (44mm and 48mm, respectively),
and also a higher 𝐺
½ÇÈ
resulting in a lower aliasing signal amplitude.
55
dramatically below cerebellum, hence L1 was adopted in this study. However, L1 is likely to suffer
from susceptibility effects due to air-bone and tissue-bone interfaces, and a wider labeling plane
would increase the possibility for interference of susceptibility effects. Therefore, a higher 𝐺
½ÇÈ

would be preferred to narrow the main lobe of the labeling plane.
3.5.4. Flow velocity
A constant blood velocity (40cm/s) was assumed in the simulation, which may be over-
simplified. In reality, the flow is temporally pulsatile and spatially laminar. Zhao et al. calculated
the weighted contribution of flow with different velocity (Figure 2 from (Zhao et al., 2017)), which
suggested a lower “effective velocity” (~25cm/s) may be more suitable for the pCASL labeling
simulation. Since the product of velocity and 𝐺
ÇÎc
determines the frequency offset experienced by
the flow, a lower velocity would require a higher optimal 𝐺
ÇÎc
.
Figure 3.8 shows the simulated LE with v=25cm/s, same layout was used as Figure 3.3A-
C, in which case the optimal parameter is (RF duration=300us, 𝐺
ÇÎc
=0.05mT/m, 𝑔𝑅𝑎𝑡𝑖𝑜=12).
With all factors above considered, we decided to keep the recommended gradient parameters
and only update the timing. Consistent with the analysis that shorter RF period and gap is more
robust to 𝐵
N
offset, Labeling 3 (i.e., the RF gap was shortened) achieved the highest LE and
therefore was chosen as the optimal pCASL scheme at 7T.
Figure 3.8 Simulation of LE with v=25cm/m for RF duration=300us (A), 500us (B), and 800us (C). The optimal
parameter set was RF duration=300us, 𝐺
ÇÎc
= 0.5mT/m, 𝑔𝑅𝑎𝑡𝑖𝑜=12. Compared with results from v=40cm/s (Figure
3.3), the optimal 𝐺
ÇÎc
increased from 0.3mT/m to 0.5mT/m due to lower flow velocity.
56
3.5.5. VERSE modification
It is worth noting that the VERSE RF modification can be used to reduce SAR for pCASL
labeling. However, based on our simulation, VERSE may not be a good fit for pCASL labeling
pulse. First, VERSE reduces SAR by stretching and compressing RF waveforms and
simultaneously adjusting gradient waveform. Since the pCASL labeling pulse is very short and the
gradient is limited by hardware slewrate, the SAR reduction effect is very limited (Teeuwisse,
Webb, & van Osch, 2010) (e.g., only 9.5% SAR reduction for RF duration=300us with 𝐺
½ÇÈ

increased from 6mT/m to the “VERSEed” value of 14mT/m and hardware slewrate 200Tm/ms).
Second, a longer gradient ramp time and a longer rewinding gradient of labeling pulse are needed
when 𝐺
½ÇÈ
is increased to the “VERSEed” value, leading to a lower RF amplifier duty cycle,
which, according to Section 3.5.2, worsens the aliasing problem. Longer RF duration and RF
period can ameliorate the aliasing problem, but the pCASL LE would be more sensitive to 𝐵
N

offset.
Figure 3.9 demonstrates the excitation profile (A, B, and C) as well as the pulse and
gradient waveform (D, E, and F) of the optimized pCASL labeling (A&D), its VERSE
modification (B&E), and VERSE with RF duration=500us (C&F), respectively. The VERSE
modification algorithm was adopted from (Gai & Zur, 2007).
57
3.5.6. Limitations and future directions
First, the simulation model needs to be expanded to include the continuous flow velocity
and an independent RF gap value. Second, the LE improvement at UHF needs to be verified with
comparison of ASL sequences at 3T and 7T. Third, the potential of the parallel RF transmission
(pTx) system can be explored to further increase the LE and robustness the pCASL sequence at
7T. Fourth, 2D TFL was used for image acquisition in this study, which could not cover the whole
brain during the image acquisition window. More efficient acquisition schemes such as
Simultaneous Multi-slice (SMS) (Y. Wang et al., 2015) and 3D Gradient and Spin-Echo (GRASE)
can be applied to increase imaging speed and coverage (Spann et al., 2020).
Figure 3.9 The excitation profile, RF and gradient waveform of the optimized pCASL labeling, VERSE modified
pCASL labeling, and VERSE modified with RF duration=500us. Without VERSE modification, the labeling block
parameters are: RF duration=300us, Gmax=6mT/m, gradient ramp time=30us, duration of the rewinding
gradient=190us, RF period=550us, and RF duty cycle=300/550=54.6%; With VERSE modification, RF
duration=300us, gradient ramp time=70us, duration of the rewinding gradient=220us, RF period=660us, and RF
duty cycle=300/660=45.6%. Clear aliasing pattern was observed in B. Shown in C, when the modified RF duration
was increased to 500us, the aliasing pattern disappeared.
58
3.6. Conclusion
We optimized the pCASL labeling parameters based on the 𝐵
O
P
/𝐵
N
field distribution at 7T.
The proposed pCASL sequence was able to achieve reliable and repeatable perfusion
measurements at 7T. This work is under revision by MRM and a C2P sequence has been
disseminated to other 7T sites.

59
Chapter 4. Increased Labeling Efficiency of Pseudo-Continuous
Arterial Spin Labeling with Parallel Transmission B1 shimming
4.1. Abstract
Objective To improve the labeling efficiency (LE) of the pseudo-continuous arterial spin labeling
(pCASL) with parallel RF transmission (pTx) transmit B1 (𝐵
O
P
) shimming.
Materials and Method In order to achieve increased LE, pTx 𝐵
O
P
shimming was used to
specifically increase the 𝐵
O
P
amplitude within the regions of inflowing arteries. Two 𝐵
O
P
shimming
modes were defined: The “indv-shim”, which utilized shimming weights calculated for each
individual subject, and the “univ-shim”, which used a set of universal weights calculated based on
a group of subjects and did not require any adjustment for individual subjects. Both indv-shim and
univ-shim were evaluated, with circular polarized (CP)-shim as the benchmark. The pCASL
sequences with the three 𝐵
O
P
shimming modes were evaluated with in-vivo experiments in 6
subjects who underwent 2 repeated scans one day apart, along with a pulsed ASL (PASL)
sequence. Quantitative metrics such as mean 𝐵
O
P
amplitude, perfusion, and ICC were calculated.
Results Both indv-shim and univ-shim achieved significantly increased 𝐵
O
P
amplitude compared
with CP-shim in simulation and in in-vivo experiment (p<0.01). Compared to CP-shim, perfusion
signal was increased by 9.5% with indv-shim (p<0.05) and by 5.3% with univ-shim (p=0.35). All
three pCASL sequences achieved fair to good repeatability with ICC>=0.5.
Conclusions By effectively increasing the 𝐵
O
P
amplitude at inflowing arteries, both indv-shim and
univ-shim can improve LE of the pCASL sequence at 7T.
60
4.2. Introduction
In Chapter 3, we specifically optimized the pseudo-continuous Arterial Spin Labeling
(pCASL) labeling parameters based on the 𝐵
O
P
and 𝐵
N
inhomogeneity condition at 7T, and robust
perfusion measurement was achieved. However, potential improvement of the labeling efficiency
(LE) may still be achieved with higher 𝐵
O
P
amplitude.
Parallel RF transmission (pTx) provides additional degrees of freedom that allows full
spatial and temporal control of the 𝐵
O
P
field. With a transmit coil array that consists of several
elements with spatially distinct 𝐵
O
P
field pattern and channel-specific RF pulse waveform, the extra
degrees of freedom can be exploited to overcome the effect of 𝐵
O
P
inhomogeneity for
implementing ASL at UHF. Recently, Tong et al. (Tong et al., 2020) utilized pTx 𝐵
O
P
shimming
to minimize the difference between the actual flip angle (FA) and the target value, and variable-
rate selective excitation (VERSE) was further used to decrease the Specific Absorption Rate (SAR)
of RF power. However, the labeling efficiency (LE) was very low possibly because that the pCASL
labeling parameters were directly adopted from 3T which may not be suited for 7T scans. Besides,
the shimming weight had to be calculated for each subject, which prolonged total scan time and
presented a challenge for clinical application.
For typical dynamic pTx applications, the pTx pulses need to be calculated based on the
𝐵
O
P
and 𝐵
N
maps acquired for individual subject, which complicates the workflow and thus
impedes the clinical translation. To address this problem, the idea of “universal pulse” was
proposed by Gras et al. (Gras, Vignaud, Amadon, Le Bihan, & Boulant, 2017). By calculating the
pTx pulse based on a group of subjects, the pulse was robust to the variation of 𝐵
O
P
and 𝐵
N
among
subjects, and therefore no calibration step was needed. However, the optimal design of “universal
61
pulse” for pCASL is challenging and has never been attempted, given the flowing spins that
confound the precise control of the excitation k-space.
The goal of this study was to increase the LE of pCASL sequence at 7T with pTx 𝐵
O
P

shimming technologies. Two types of pTx 𝐵
O
P
shimming modes were developed and evaluated:
the “indv-shim” which utilized shimming weights optimized for each individual subject, and the
“univ-shim” which utilized a set of universal weight optimized on a separate cohort and thus did
not require any calibration steps. The default True-Form mode mimicking the single channel
Circular-Polarized (CP) coil served as benchmark (CP-shim). We demonstrated that both indv-
shim and univ-shim provide significantly increased 𝐵
O
P
amplitude at inflowing arteries and
increased LE compared with CP-shim.
4.3. Materials and Methods
4.3.1. 𝑩
𝟏
P
shimming evaluation cohort
During the development of the 𝐵
O
P
shimming methods, data of a subject cohort was used
for evaluation purpose. This cohort (n=12, 11 male, age=29±6 years) includes subjects that were
recruited during the testing of the optimized pCASL sequences (Chapter 3). Specifically, the Time-
of-Flight (TOF) angiography was acquired to locate the labeling plane and the inflowing arteries,
and the channel-specific 𝐵
O
P
map was acquired at the labeling plane (Chapter 3, Figure 3.1A).
The study design was approved by the Institutional Review Board of the University of Southern
California, and written consent was obtained from each subject prior to the experiments. Note that
there is no overlapped subject between this cohort and the cohort participated in the in vivo
experiment in the current project.
62
4.3.2. Definition of 𝑩
𝟏
P
shimming cost function
The cost function of the optimization includes two terms: the efficiency term and the
asymmetry term. The efficiency term penalized low 𝐵
O
P
amplitude by taking the form of 𝐸𝑓𝑓 =
−∥𝑩
𝒄𝒉𝒏
⋅𝒖∥
n
, where 𝑩
𝒄𝒉𝒏
is the channel specific 𝐵
O
P
of the voxels within the ROIs, and 𝒖 is the
complex weight for each channel. The asymmetry term was specifically designed to penalize the
inter-hemispheric 𝐵
O
P
difference of the left and right arteries, which would cause imbalanced
perfusion signal of the left and right hemispheres. It was defined as the square of the amplitude of
the difference between the 𝐵
O
P
amplitude of the left and right ROIs: 𝐴𝑠𝑦𝑚 = ∥|𝑩
𝒄𝒉𝒏,𝒍𝒆𝒇𝒕
⋅𝒖| −
|𝑩
𝒄𝒉𝒏,𝒓𝒊𝒈𝒉𝒕
⋅𝒖|∥
n
. The cost function was then defined as a weighted sum of the two terms:
𝑐𝑜𝑠𝑡 =𝐸𝑓𝑓+𝑤⋅𝐴𝑠𝑦𝑚
=−∥𝑩
𝒄𝒉𝒏
⋅𝒖∥
n
+𝑤⋅∥|𝑩
𝒄𝒉𝒏,𝒍𝒆𝒇𝒕
⋅𝒖| −|𝑩
𝒄𝒉𝒏,𝒓𝒊𝒈𝒉𝒕
⋅𝒖|∥
n
(Eq. 4-1)
where w is the weighting for the asymmetry term. Therefore, u was optimized by
minimizing the cost function using the “fmincon” solver of MATLAB (The MathWorks, Natick,
MA), during which only the phases of u was updated.
63
4.3.3. Indv-shim and asymmetry weight evaluation
The 𝐵
O
P
shimming weight calculated per subject is noted as “individual weight”, and the
shimming method is noted as “indv-shim”. A flowchart of the optimization process for the indv-
shim was shown in Figure 4.1. First, the labeling plane location was determined with a TOF
angiography, and the four ROIs marking the inflowing arteries were manually drawn on the TOF
image; Second, the ROI maps were resampled based on the spatial resolution of the 𝐵
O
P
maps;
Third, the channel-specific complex 𝐵
O
P
values of the voxels within the four ROIs were fed into
the 𝐵
O
P
shimming optimization algorithm.
The True-Form mode served as the benchmark, noted as CP-shim. Similar to the definition
of the efficiency term and asymmetry term in the cost function, two metrics were defined to
evaluate the proper asymmetry term weighting w: 𝑟MeanB1, which is the mean 𝐵
O
P
amplitude
within ROIs; and 𝑟Asym, which is the mean absolute value of the left and right 𝐵
O
P
amplitude
difference divided by 𝑟MeanB1. These two metrics were calculated over 𝑤 ∈[0,50], and the ratio
Figure 4.1 The flow chart of indv-shim. TOF images were acquired for localizing
the inflowing arteries at the labeling plane, channel-specific 𝐵
O
P
map was acquired
and fed to the optimization problem, which was solved with fmincon solver.
64
of the indv-shim and the CP-shim was calculated for both 𝑟MeanB1 and 𝑟Asym. The optimal w
was manually selected based on the rMeanB1 ratio-w and rAsym ratio-w curves: a rMeanB1
ratio>1, which means increased 𝐵
O
P
amplitude compared with CP-shim, and a rAsym ratio<1,
which means decreased asymmetry relative to CP-shim, were required. The weighting selection
process was performed on two representative subjects from the evaluation cohort and the optimal
w was used for all subjects of the in-vivo experiments (Section 4.3.5).
4.3.4. Universal weight
Considering the extra time and manual operation needed to calculate the individual weight,
a “universal weight” was developed without adjustments for individual subjects during scan. This
shimming method is noted as “univ-shim”. The development and evaluation of universal weight
was performed on the evaluation cohort. The optimization process was similar to that of indv-
shim, except that the cost function was calculated with all ROIs of the group of subjects, meaning
that both the efficiency term and the asymmetry term were calculated with 4×12=48 ROIs.
Numerical simulation was utilized to evaluate the univ-shim with leave-one-out cross-
validation: for each of the 12 subjects, one universal weight was calculated with 11 subjects and
then evaluated on the remaining subject. Since the optimal asymmetry weighting w may be
different from that of indv-shim, the weighting evaluation process was repeated by calculating the
𝑟MeanB1 and 𝑟Asym metrics over 𝑤 ∈[0,50], averaged across 12 subjects. For comparison, the
𝑟MeanB1 and 𝑟Asym metrics were also calculated with the 𝐵
O
P
shimming results of CP-shim and
indv-shim. Once the method of univ-shim was verified, a final universal weight was calculated
with all 12 subjects and applied in the in-vivo experiments.
65
4.3.5. In-vivo experiment
Six subjects were recruited (3 Male, age=25.8±3.2 years. Each subject underwent two
scans which were exactly 24-hour apart to control for daily physiological fluctuation of brain
perfusion.
The three 𝐵
O
P
shimming modes of CP-shim, indv-shim, and univ-shim were implemented
with pCASL sequences (pCASL-CP, pCASL-indv, and pCASL-univ, respectively). Shimming
weights were only applied for the labeling pulses, while the default CP weight was used for the
pulses of acquisition. Based on the pCASL optimization results of Chapter 3, the labeling
parameter of the pCASL sequences are: RF duration=300us, RF gap=250us, RF flip angle=15°,
𝐺
ÇÎc
=0.6mT/m, 𝑔𝑅𝑎𝑡𝑖𝑜=10, labeling duration=1000ms, and post label delay=1500ms.
The scan protocol for each visit is listed in Table 4.1, which included: 1. A TOF sequence
for the localization of labeling plane and inflowing arteries; 2. A modified TFL sequence to acquire
channel-specific 𝐵
O
P
map; 3. Three pCASL sequences of pCASL-CP, pCASL-indv, and pCASL-
univ; 4. A pulsed ASL sequence optimized in a previous study (K. Wang et al., 2021) as a
comparison for the pCASL sequences. The labeling plane position was kept the same for all
pCASL sequences. The FOV and 𝐵
N
shimming volume were kept the same for all ASL sequences.
66
Parameters TOF Tfl_B1map pCASL PASL
FOV (mm) 200×200 200×200 210×192.5 210×192.5
Matrix size 768×768 128×128 96×88 96×88
Slice number 136 25 9 9
Slice thickness (mm) 0.5 3 5 5
Voxel size(mm) 0.26×0.26 1.56×1.56 2.19×2.19 2.19×2.19
TR (ms) 19 6480 6000 or 9000 (*) 6000
TE (ms) 3.42 1.75 2.38 2.38
Measurements N/A N/A 40 40
Duration 3min49sec 1min6sec 4min2sec or
6min2sec (*)
4min2sec
Table 4.1 The scan protocol for the in vivo experiments. TOF images were acquired for angiography, channel-specific
𝐵
O
P
maps were obtained for pTx 𝐵
O
P
shimming, and pCASL and PASL sequences were acquired with the same
acquisition parameters. (*) TR=6000ms for pCASL-CP, and 9000ms for pCASL-univ and pCASL-indv due to more
constraining SAR limit (see section 4.5.4)
The performance of 𝐵
O
P
shimming was evaluated by calculating the rMeanB1 and rAsym
metrics for each subject of each 𝐵
O
P
shimming mode. To calculate perfusion maps for the pCASL
sequences, images were corrected for rigid head motion using SPM12 (Wellcome Trust Centre for
Neuroimaging, UCL), then pairwise subtraction of label and control images was performed,
followed by averaging across measurements. The perfusion maps were then normalized by the M0
image resulting in fractional perfusion maps. The mean gray matter (GM) fractional perfusion
signal was calculated for each subject, each visit, and each ASL sequences respectively. The
repeatability of the ASL sequences was evaluated by calculating the intraclass correlation
coefficient (ICC) and the average within-subject coefficient of variation (wsCV) of the perfusion
signal. In order to evaluate the performance improvement with pTx 𝐵
O
P
shimming, two-way
ANOVA was used to compare the rMeanB1, rAsym, and perfusion of pCASL-univ, pCASL-indv,
and pCASL-CP, with shimming type and subject as the independent variables and visits as
repetitions.
67
The LE of the pCASL sequences was also estimated. Assuming the LE of PASL is 0.95
(K. Wang et al., 2021; Wong, 2005), the mean GM CBF of the central slice was calculated (Alsop
et al., 2015; Zuo et al., 2013) (noted as CBFPASL). The mean GM CBF of the pCASL sequences
were calculated with LE=1 following the method in (Alsop et al., 2015; Zuo et al., 2013) (noted
as CBFpCASL), then the estimated LE of the pCASL sequence was CBFpCASL/CBFPASL.
4.4. Results
4.4.1. Indv-shim and asymmetry weight evaluation
Figure 4.2 shows the asymmetry weight evaluation process with two representative
subjects. Figure 4.2 A&D shows the TOF image of the labeling plane as well as the ROI of the
inflow arteries. Figure 4.2 B&E shows the variation of rMeanB1 ratio (indv-shim versus CP-
shim) vs. the asymmetry weight w, and Figure 4.2 C&F shows the variation of rAsym ratio vs. w.
Generally, a higher asymmetry weight w led to lower rMeanB1 ratio and lower rAsym ratio since
68
mean 𝐵
O
P
amplitude was sacrificed to achieve lower asymmetry. As a balance between the two
Figure 4.3 A&D The ROI maps on the TOF images. B&E The rMeanB1 ratio between indv-
shim and CP-shim with 𝑤 ∈[0,50]. C&F The rAsym Ratio with ∈[0,50]. For both subjects,
w=20 yielded a rMeanB1ratio of nearly 1.2 and an rAsym ratio smaller than 1. Therefore,
w=20 was chosen.
Figure 4.2 Same layout was used as Figure 4.2. For all subjects, w=20 yielded consistent
increase of rMeanB1 and a comparable if not smaller rAsym compared to CP-shim.
69
terms, w = 20 was chosen because it yielded an rAsym ratio < 1 for both subjects while largely
preserving mean 𝐵
O
P
amplitude gain from 𝐵
O
P
shimming (rMeanB1 ratio = ~1.2). The results of
three other subjects are shown in Figure 4.3, which validated our choice of w = 20.

4.4.2. Universal Weight
The same process as used for indv-shim was performed to determine the optimal
asymmetry weight for univ-shim. The rMeanB1 ratio and the rAsym ratio (vs. CP-shim) are shown
Figure 4.4 A&B The rMeanB1 ratio and rAsym ratio between univ-shim and CP-shim w ∈[0,50].
w = 40 was chosen for univ-shim. C&D The rMeanB1 and rAsym curve of three shimming modes,
respectively. The univ-shim achieved considerable mean B1 increase compared to CP-shim for
all subjects, which was comparable to indv-shim, although the asymmetry was higher than indv-
shim.
70
in Figure 4.4A and 4.4B, respectively, and w =4 0 was chosen with rMeanB1 ratio ≈ 1.14 and
rAsym ratio < 1.
The simulated subject-wise rMeanB1 and rAsym for each shimming mode are shown in
Figure 4.4C and 4.4D, respectively. With univ-shim, the rMeanB1 was significantly increased
compared with CP-shim (0.72±0.06 vs. 0.64±0.07, p<0.001), although lower than the indv-shim
(0.75±0.08). In terms of the asymmetry term rAsym, the univ-shim was comparable to the CP-
shim (0.13± 0.07 and 0.14± 0.06, respectively), which were both higher than indv-shim
(0.06±0.03). Therefore, the univ-shim was able to increase the mean 𝐵
O
P
amplitude by nearly 14%
while maintaining a comparable asymmetry level as the CP-shim in our simulation.
4.4.3. In-vivo experiment
Calculated with the channel-specific 𝐵
O
P
map acquired for each subject, the rMeanB1 and
rAsym of three 𝐵
O
P
shimming modes are shown in Figure 4.5A and Figure 4.5B, respectively.
The two visits of each subject were treated as repeated measurements when comparing the 𝐵
O
P

shimming modes. The rMeanB1 was significantly increased with univ-shim compared to CP-shim
(0.83±0.10 vs. 0.75±0.11, p<0.001). Although the rMeanB1 of indv-shim was the highest
(0.86±0.09), it was not significantly different from that of univ-shim. In terms of rAsym, the univ-
shim (0.13±0.06) was slightly lower than the CP-shim (0.14±0.06), with indv-shim being the
lowest (0.06±0.03, p=0.001 and 0.002 against CP-shim and univ-shim, respectively). This result
was highly consistent with the simulation results in universal weight evaluation, suggesting that
univ-shim was able to increase the mean 𝐵
O
P
amplitude within ROIs of inflowing arteries without
increasing the asymmetry. Figure 4.5C shows the amplitude (top row) and phase (bottom row) of
the combined 𝐵
O
P
map of three shimming modes for one representative subject. Indv-shim and
univ-shim obtained similar combined 𝐵
O
P
map. Both indv-shim and univ-shim increased 𝐵
O
P

71
amplitude within the ROIs (colored circles) although at the cost of a dark band (yellow arrow),
which would not affect the inflowing arteries.

Figure 4.5 A&B The rMeanB1 curve of the three shimming modes. Indv-shim had consistently
higher rMeanB1 than CP-shim for each subject, and univ-shim had similar performance as indv-
shim except for measurement 9. B The rAsym curve of the three shimming modes. Indv-shim had the
lowest rAsym while the rAsym of CP-shim and univ-shim were comparable. C The combined B1
map of one representative case of the three shimming modes, first row amplitude, second row phase.
Performance of indv-shim and univ-shim were comparable, both of which achieved higher B1
amplitude within ROI (colored circles) compared with CP-shim. A dark band appeared (yellow
arrow) with indv-shim and univ-shim, which would not affect the inflowing arteries.
72
The perfusion map of one representative subject is shown in Figure 4.6. Perfusion maps
of 2 other subjects are shown in Figure 4.7. All pCASL sequences achieved perfusion maps of
decent quality and good repeatability. The average GM fractional perfusion values of three 𝐵
O
P

shimming modes are shown in Figure 4.8. Compared with perfusion of pCASL with CP-shim
(1.27± 0.13%), perfusion of pCASLwith indv-shim (1.39± 0.13%) was significantly higher
(P<0.05) which was increased by 9.5%, while perfusion of pCASL with univ-shim (1.34±0.14%)
was 5.3% higher than that of pCASL with CP-shim although the difference was not significant (P
=0.35).

Figure 4.6 The fractional perfusion maps of one representative case: visit 1 (A) and visit 2 (B).
Consistent and high-quality perfusion maps were obtained with all pCASL sequences, while
difference among three shimming modes were not observable visually. Good repeatability was
achieved.
73

Figure 4.7 A and B The fractional perfusion map for 2 other subjects,
respectively. Top panel, first visit; bottom panel, second visit. Perfusion maps of
good quality and decent repeatability were achieved with all pCASL sequences
74
The ICC of pCASL with CP-shim, indv-shim, and univ-shim was 0.66, 0.73, and 0.50,
respectively, and the wsCV averaged across subjects was 4.60%, 3.94%, 6.90%, respectively,
indicating a good level of repeatability of the three pCASL sequences. The corresponding Bland-
Altman plot is shown in Figure 4.9.
Assuming the LE of the PASL is 0.95, CBFPASL was calculated to be 58.1mL/100g/min,
based on which the LE of pCASL with CP-shim, indv-shim, and univ-shim was estimated to be
0.62, 0.69, and 0.66, respectively. Quantitative indices including rMeanB1, rAsym, perfusion,
ICC, wsCV, and LE are summarized in Table 4.2.

Figure 4.8 Comparison of the mean GM CBF of three shimming modes. Perfusion of pCASL
with indv-shim was significantly higher than that of CP-shim (increased by 9.5%, p<0.05).
Perfusion of pCASL with univ-shim was also increased compared with that of CP-shim by 5.3%
although the difference was not significant probably due to limited sample size.
75

CP-shim indv-shim univ-shim
rMeanB1 0.75±0.11 0.86±0.09 0.83±0.10
rAsym 0.14±0.06 0.06±0.03 0.13±0.06
Perfusion (%) 1.27±0.12 1.39±0.13 1.34±0.14
ICC 0.66 0.73 0.50
wsCV (%) 4.60±2.74 3.94±2.68 6.90±4.07
LE 0.62±0.10 0.69±0.15 0.66±0.14

4.5. Discussion
4.5.1. 𝑩
𝟏
P
shimming with multi-pulse pTx
The pTx offers a high degree of freedom to pulse design with the multiple transmit channels
driven by their individual RF amplifier. Three main categories of pTx pulse have been proposed:
the static pTx, where the same RF waveform is used for all RF channels and the relative 𝐵
O
P

ampitude and phase is adjusted by the channel-specific complex weight; the dynamic pTx, where
Figure 4.9 Bland-Altman plot of pCASL with CP-shim (A), indv-shim (B), and univ-shim (C). All pCASL
sequences showed decent repeatability.
Table 4.2 Quantitative metrics of pCASL sequences with three shimming modes. pCASL with indv-shim
outperforms the other two with the highest rMeanB1 and lowest rAsym, highest perfusion, ICC, wsCV, and LE.
Univ-shim also achieved considerable increase of rMeanB1, perfusion, and LE compared with CP-shim.
76
each channel is powered with different RF waveforms; and the multi-pulse pTx, where the
channel-specific weights can change throughout an MRI sequence.
Apparently, the dynamic pTx offers the highest degree of freedom with its channel-specific
RF waveform. However, dynamic pTx is challenging for pCASL given the flowing spins that
confound the precise control of the excitation k-space for optimal pTx pulse design (Katscher et
al., 2003). Moreover, the purpose of 𝐵
O
P
shimming for pCASL labeling is only to increase 𝐵
O
P

amplitude within the inflowing arteries, leading to possible destructive patterns elsewhere (e.g.,
imaging volume) and thus static pTx is not suitable. Therefore, we chose to use the multi-pulse
pTx in this study, where the adjusted weight was only used for pCASL labeling, and the CP-mode
was still used for image acquisition.
4.5.2. 𝑩
𝟏
P
shimming cost function
The cost function consists of two terms: the efficiency term 𝐸𝑓𝑓 =−∥𝑩
𝒄𝒉𝒏
⋅𝒖∥
n
, and
the asymmetry term 𝐴𝑠𝑦𝑚 = ∥|𝑩
𝒄𝒉𝒏,𝒍𝒆𝒇𝒕
⋅𝒖| −|𝑩
𝒄𝒉𝒏,𝒓𝒊𝒈𝒉𝒕
⋅𝒖|∥
n
. Intuitively, the aim of the
efficiency term was to increase the 𝐵
O
P
amplitude within the inflowing arteries, which has been
well accepted (Setsompop et al., 2008; Tong et al., 2020). On the other hand, the asymmetry term
has not been proposed before. Although the inhomogeneity of the combined 𝑩
𝟏
P
field was not
included in the cost function because the ROIs for the inflowing arteries are relatively small
considering the low resolution of the channel-specific 𝑩
𝟏
P
map, in our experience, the left-right
asymmetry can become a problem leading to considerable difference of the labeling efficiency for
the left and right hemispheres, resulting in left-right asymmetry on the perfusion map, hence the
left-right asymmetry term was added in the cost function. The anterior-posterior asymmetry has
been considered as well, but the ICAs are closer to the central hotspot of the combined 𝑩
𝟏
P
map
77
than the PCAs, therefore, it would sacrifice the 𝑩
𝟏
P
efficiency too much to reduce anterior-
posterior asymmetry, hence anterior-posterior asymmetry was not included.
The efficiency term and the asymmetry term in the cost function have a different form from
the rMeanB1 term and rAsym term in the asymmetry weight evaluation process. In the cost
function, the two terms take the form of the square of amplitude, so that 1. The two terms have the
same unit, therefore, the optimization result will not depend on the unit of the input images; The
convex function is easier for the optimizer to solve. Meanwhile, the mean 𝐵
O
P
amplitude within
ROI rMeanB1 and the left-right absolute 𝐵
O
P
amplitude difference over mean are more intuitive to
quantify the optimization results.
4.5.3. Universal Pulse
The computational time needed for the 𝐵
O
P
shimming weights takes a few seconds on a
MacBook Pro, but considering the time needed to acquire the 𝐵
O
P
map, transferring data to a
workstation, drawing ROI, and calculating and manually inputting the shimming weights to scan,
the extra time needed for the indv-shim is about 5min. This extra time and effort required make it
difficult for clinical applications.
To address this problem, inspired by the universal pulse(Gras et al., 2017), we further
proposed univ-shim to improve the workflow of pCASL with pTx 𝐵
O
P
shimming. This idea is
supported by the observation that despite the cross-subject variation, the location of the inflowing
arteries of different subjects are relatively consistent as well as the channel-specific 𝐵
O
P
map
pattern. Indeed, the universal weight common for different subjects was found, which was
evaluated with cross-validation in simulation and validated in all 6 subjects of the in-vivo
experiments without any adjustments. With univ-shim, the pCASL sequence can be used just like
78
any sequence without pTx 𝐵
O
P
shimming, with no need to acquire the 𝐵
O
P
map or adjust the
shimming weight during the scan.
We believe this universal weight can be generalized to other sites with the same scanner/RF
coil setup, although further validation is needed for a larger cohort of subjects and for subject with
an abnormal distribution of arteries (e.g., severely tortuous arteries). Therefore, the proposed univ-
shim provides a promising solution to the 𝐵
O
P
shimming implementation problem for pCASL at
7T.
4.5.4. SAR constraint
For the proposed pCASL 𝐵
O
P
shimming, only the phase of the channel-specific weights
were optimized, which is only an interim solution for SAR constraint. The best strategy would be
to include both global SAR and local SAR constraints into the weight optimization. However, this
is not feasible yet due to hardware limitation: theoretically, the global SAR is easily represented
by 𝑆𝐴𝑅
ô
=𝑘
O
Σ‖𝑢
[
‖
n
, where 𝑢
[
is the weight for i-th channel and 𝑘
O
is a constant to convert 𝑢
[
to
RF power amplifier voltage then to SAR, the local SAR is controlled with a virtual observation
point (VOP) method (Eichfelder & Gebhardt, 2011): 𝑆𝐴𝑅
÷
=max
ø
(𝒇
ù
⋅𝑨
𝒋
⋅𝒇), where 𝒇 is the RF
waveform, 𝑨
𝒋
is the system matrix of each observation point that is calculated based on the
excitation profile information of the RF coil used. However, because that information is not
provided by the coil vendor, the local SAR constraint is enforced by limiting the power (or, 𝐵
O
P

amplitude) of each RF channel very conservatively: 𝑆𝐴𝑅
÷
=max
[
(𝑘
n
⋅8⋅‖𝑢
[
‖
n
) where 𝑘
n
>𝑘
O
,
leading to overestimation of SAR when the 𝐵
O
P
shimming mode is turned on even if the shimming
weights are exactly the same as the True-Form mode. Specifically for the Siemens Terra scanner
system used in this study, 𝑘
n
=2.5𝑘
O
, indicating a possibly 1.5 times overestimation of SAR.
When 𝐵
O
P
amplitude is different for different channels, the overestimation is even higher.
79
Therefore, we decided to use the same amplitude for all channel weights and only optimize the
phase.
4.5.5. Limitations
First, 2D TFL was used for image acquisition in this study, which could not cover the
whole brain during the image acquisition window. More efficient acquisition schemes such as 3D
Gradient and Spin-Echo (GRASE) can be applied to increase imaging speed and coverage (Spann
et al., 2020). However, unlike the TFL acquisition which intrinsically suppresses long 𝑇
O
signal,
background suppression will be needed for GRASE acquisition which needs to be optimized for
7T. Second, more SAR-efficiency shimming weights can be obtained by directly including SAR
constraint in the optimization and releasing weight amplitude as subject of optimization once the
coil information for VOP is available.
4.6. Conclusion
We achieved increased labeling efficiency for 7T pCASL with pTx 𝐵
O
P
shimming of indv-
shim and univ-shim. pCASL with indv-shim achieved significantly higher perfusion signal
compared with that of pCASL with CP-shim (9.5%, P<0.05), and the perfusion signal was
increased by 5.3% with univ-shim that required no calibration steps. This work (Chapter 3&4) is
under revision by MRM and a C2P sequence has been disseminated to other 7T sites.

80
Chapter 5. Exploration of Continuous Arterial Spin Labeling
(CASL) with Dynamic Parallel Transmission (pTx)
5.1. Abstract
Objectives To propose an innovative Continuous Arterial Spin Labeling (CASL) sequence
implemented with Dynamic parallel transmission (pTx) noted as DCASL without violating RF
power amplifier (RFPA) duty cycle limit.
Theory The eight transmit channels were divided into two groups (Group1 and Group2), each with
a periodic rectangular waveform that is half period apart from each other. One group is turned on
and the other is off at any timepoint during CASL labeling, and the combined 𝐵
O
P
at the four
inflowing arteries (left and right Internal Carotid Arteries (ICAs) and Vertebral Arteries (VAs))
remains unchanged when the working status of the two groups shift. There exist the pTx weights
for each channel to satisfy the requirements above since the number of variables and equations are
the same.
Materials and Methods The pTx weights were solved using a gradient descent optimization
algorithm and a custom-defined cost function consisting of a fidelity term and a SAR penalty term
with weighting α. The proper α was determined based on the SAR and the labeling efficiency (LE)
over α∈[0.001:0.001:0.5]. The amplitude and phase mismatch between the two groups were
evaluated over the range of α, and the LE was estimated with Bloch simulation against the
amplitude/phase mismatch. The performance of the proposed DCASL was evaluated on one
subject and compared with the conventional CASL sequence with default True-Form mode.
Results The pTx weights were obtained with approximately 50 iterations of the gradient descent
optimization. The SAR penalty weighting term 𝛼=0.1 was chosen since it yielded the lowest SAR
81
at the cost of a small LE decrease (~2%). Compared with CSL, comparable performance of
DCASL was achieved at the cost of a higher SAR (~2.29 times), although not supported by current
scanner platform yet.
Conclusions By utilizing dynamic pTx, the proposed DCASL can achieve a LE close to CASL at
7T without violating the RFPA duty cycle limit for each transmit channel.
5.2. Introduction
Arterial spin labeling (ASL) is a perfusion MRI technique that utilizes magnetically labeled
arterial blood water as an endogenous tracer to measure cerebral blood flow (CBF). Due to its non-
ionizing and completely non-invasive nature, ASL is very suitable for perfusion studies in healthy
individuals, patients with renal insufficiency and those who need repetitive follow-ups. There are
three main categories of ASL sequences: the continuous ASL (CASL), the pulsed ASL (PASL),
and the pseudo-continuous ASL (pCASL).
Developed by Williams et al.(Williams et al., 1992), CASL was the first proposed ASL
technique. In 1992, the original CASL sequence was used to measure rat brain CBF. With CASL,
the spin of the inflowing arterial blood is continuously inverted at the neck region with a
continuous RF pulse and gradient waveform, then the signal was later measured at the imaging
volume with the subtraction of a control image. However, the long RF pulse leads to strong
magnetic transfer (MT) effect, which interferes with perfusion measurements. The original CASL
utilized a control that applies the labeling plane to the distal side of the imaging volume with the
same labeling offset as label, which only works with single slice acquisition, noted as CASL-SS;
Later, Detre et al. proposed the amplitude modulation (AM) method, noted as CASL-AM, which
modulated the labeling RF pulse with a sinusoidal function, leading to equivalently two effective
labeling planes in the frequency domain to preserve the magnetization by inverting it twice.
82
The limitations of CASL are obvious. The long continuous RF pulses are oftentimes not
supported by commercially available scanners due to RF power amplifier (RFPA) duty cycle
violation. Modified from CASL, the pseudo-continuous ASL (pCASL) was proposed, which
separates the long RF pulse into many short RF pulses (Dai et al., 2008). However, pCASL itself
faces challenges when implementing at ultra-high field (UHF). First, the phase increment between
two adjacent RF pulses needs to be carefully calculated so that the RF pulses are always in-phase
with the flow spin, but the severe 𝐵
N
inhomogeneity at UHF greatly undermines the phase tracking,
leading to decreased LE (Figure 3.3D&E). Second, the Specific Absorption Rate (SAR) increases
approximately quadratically with the field strength, leading to ~4 times the SAR at 7T compared
with 3T, while the SAR of pCASL sequence is normally high.
On the other hand, CASL suits well with UHF. Because it utilizes a continuous long RF
pulse for flow-driven adiabatic inversion, the 𝐵
N
inhomogeneity only causes a slight shift of the
labeling plane along the z-direction without affecting the labeling efficiency, therefore, CASL is
extremely robust against 𝐵
N
inhomogeneity. Also, since CASL has a much longer and flatter RF
pulse shape, the SAR is much lower compared with pCASL assuming the time-average flip angle
(FA) is the same, which helps ameliorate SAR constraint problem.
Parallel RF transmission (pTx) provides previously unavailable degrees of freedom for RF
pulse design. With dynamic pTx, the RF waveform of each individual transmit channel can be
manipulated independently. Theoretically, since each transmit channel is powered by its own
RFPA, the duty cycle calculation for each channel should be independent from each other.
In this study, we proposed an innovative CASL sequence implemented with Dynamic pTx
(DCASL). By separating the transmit channels into two groups and controlling them to work in
83
turn, the DCASL sequence can be achieved with a duty cycle of 50% for each transmit channel.
The proposed DCASL sequence can be used for both CASL-SS and CASL-AM.
5.3. Theory
In conventional CASL, the flowing spin is inverted by the constant RF and gradient
following adiabatic inversion when passing by the labeling plane. With DCASL, although a
constant RF is unachievable due to RFPA duty cycle limit, the RF pulses of each transmit channel
can be manipulated so that the combined RF at each of the four inflowing arteries (left and right
internal carotid arteries (ICAs) and vertebral arteries (VAs)) is constant.
Considering the spatially slow-varying channel-specific 𝐵
O
P
field and the relatively small
ROIs of the inflowing arteries, the 𝐵
O
P
field of one artery can be approximated with the mean 𝐵
O
P

field within the ROI for each channel. Therefore, for an 8-channel pTx system, the channel-specific
𝐵
O
P
field of the four ROIs can be represented by:
𝑿= ÿ
𝑎
OO
⋯ 𝑎
Oø
⋮ ⋱ ⋮
𝑎
[O
⋯ 𝑎
[ø
$,
where 𝑎
[ø
is the mean 𝐵
O
P
field of the i-th ROI and the j-th channel, 𝑖 = 4 is the number of ROIs,
and 𝑗 = 8 is the number of transmit channels.
Assume the weight of each transmit channels is 𝒖=[𝑢
O
;𝑢
n
; …;𝑢
ø
], then the combined
𝐵
O
P
field at each ROI is:
𝒀=𝑿⋅𝒖= ÿ
𝑎
OO
⋯ 𝑎
Oø
⋮ ⋱ ⋮
𝑎
[O
⋯ 𝑎
[ø
$⋅'
𝑢
O
𝑢
n
…
𝑢
ø
( (Eq.5-1)
The eight channels are divided into two groups, both of which utilizes a periodic
rectangular RF waveform with duty cycle of 50%. Channels within the same group have the same
84
timing of RF but their amplitude and phase are adjusted by channel-specific complex weights, and
the RF waveforms of one group is lagged by half a period than the other group. Group division
should be done in a way that tries to maximize the intra-group spatial incoherence of all coils to
facilitate the channel weight optimization. One possible solution is odd number channels for
Group1 (channel 1, 3, 5, 7) and even number channels for Group2 (channel 2, 4, 6, 8). Suppose at
Time1, Group1 is on and Group2 is off, and at Time2, Group1 is off and Group2 is on, then the
combined 𝐵
O
P
field at each ROI at Time1 and Time2 is:
𝒀
𝒈𝟏
=𝑿⋅𝒖
𝒈𝟏
= ÿ
𝑎
OO
⋯ 𝑎
Oø
⋮ ⋱ ⋮
𝑎
[O
⋯ 𝑎
[ø
$⋅
⎣
⎢
⎢
⎢
⎢
⎢
⎢
⎡
𝑢
O
0
𝑢
É
0
𝑢
Ê
0
𝑢
,
0
⎦
⎥
⎥
⎥
⎥
⎥
⎥
⎤
(Eq.5-2)
and
𝒀
𝒈𝟐
=𝑿⋅𝒖
𝒈𝟐
= ÿ
𝑎
OO
⋯ 𝑎
Oø
⋮ ⋱ ⋮
𝑎
[O
⋯ 𝑎
[ø
$⋅
⎣
⎢
⎢
⎢
⎢
⎢
⎢
⎡
0
𝑢
n
0
𝑢
Æ
0
𝑢
v
0
𝑢
1
⎦
⎥
⎥
⎥
⎥
⎥
⎥
⎤
(Eq.5-3)
If 𝒖=𝒖
𝒈𝟏
+𝒖
𝒈𝟐
can be found such that 𝒀
𝒈𝟏
=𝒀
𝒈𝟐
= 𝒀
2
, where 𝒀
2
is the desired 𝐵
O
P
field at each
ROI, then the DCASL can be implemented. Figure 5.1 shows a simple demonstration of the
DCASL principle explained above.
85

For any 𝒀
2
, the 𝒖 can be found that satisfies
𝒀
𝒈𝟏
= 𝒀
2
(Eq.5-4a)
𝒀
𝒈𝟐
= 𝒀
2
(Eq.5-4b)
since for each of the two equations there are four unknown variables and four equations.
5.4. Methods
5.4.1. pTx Weight optimization
The pTx weight optimization process was the same for solving Eq.5-4a and Eq.5-4b,
therefore only the process of solving Eq.5-4a was described below. The optimization was
performed using a gradient descent optimizer, with the cost function defined as:
𝐿𝑜𝑠𝑠 = 3𝑿⋅𝒖
𝒈𝟏
−𝒀
2
3
n
+𝛼⋅3𝒖
𝒈𝟏
3
n
(Eq.5-5)
which includes the fidelity term 3𝑿⋅𝒖
𝒈𝟏
−𝒀
2
3
n
and a specific absorption rate (SAR) penalty term
3𝒖
𝒈𝟏
3
n
, combined with a weighting term 𝛼. Since the variables in Eq.5-5 are complex, Eq.5-5
was further separated into two equations, one for the real part and the other for the imaginary part:
Figure 5.1 Demonstration of DCASL principle. By two channel groups working in turn, flow-driven
adiabatic inversion of CASL can be achieved without exceeding RFPA duty cycle limit.
86
𝐿𝑜𝑠𝑠1= 3𝑿
𝒓𝒆𝒂𝒍
⋅𝒖
𝒈𝟏𝒓𝒆𝒂𝒍
−𝑿
𝒊𝒎𝒂𝒈
⋅𝒖
𝒈𝟏𝒊𝒎𝒂𝒈
−𝒀
𝒓𝒆𝒂𝒍
6
3
n
+𝛼⋅3𝒖
𝒈𝟏𝒓𝒆𝒂𝒍
3
n
(Eq.5-6a)

𝐿𝑜𝑠𝑠2= 3𝑿
𝒓𝒆𝒂𝒍
⋅𝒖
𝒈𝟏𝒊𝒎𝒂𝒈
+𝑿
𝒊𝒎𝒂𝒈
⋅𝒖
𝒈𝟏𝒓𝒆𝒂𝒍
−𝒀
7𝒎𝒂𝒈
6
3
n
+𝛼⋅3𝒖
𝒈𝟏𝒊𝒎𝒂𝒈
3
n
(Eq.5-6b)
Define 𝒀
2
=[1;1;1;1], then 𝒀
𝒓𝒆𝒂𝒍
6
=[1;1;1;1] and 𝒀
7𝒎𝒂𝒈
6
=[0;0;0;0]. During each
iteration, the derivative of Loss1 and Loss2 with regard to 𝒖
𝒈𝟏𝒓𝒆𝒂𝒍
and 𝒖
𝒈𝟏𝒊𝒎𝒂𝒈
were calculated
respectively, and both 𝒖
𝒈𝟏𝒓𝒆𝒂𝒍
and 𝒖
𝒈𝟏𝒊𝒎𝒂𝒈
were updated based on their derivatives of Loss1 and
Loss2. The step size for the gradient was 0.5 which decayed to 0.05 when iteration number was
above 50 and to 0.005 when iteration number was above 100. The optimization process stopped
when either the maximum iteration number 200 or a steady state of the loss function was reached
(difference compared with last iteration < 1E-4).
The weighting term of the SAR penalty was evaluated with one representative subject. A
higher 𝛼 would reduce the SAR but perturb the optimization away from the exact solution of the
fidelity term, and the mismatch of the amplitude and phase between Group1 and Group2 would
undermine the flow-driven adiabatic inversion process and therefore reduce the labeling efficiency
(LE). Therefore, the optimal 𝛼 was determined jointly by the resultant LE and SAR compared with
conventional CASL sequence. Ideally, the optimal 𝛼 would lead to the lowest SAR that preserves
most LE, i.e., at most 5% decrease compared with the highest LE.
To obtain the relationship between 𝛼 and the amplitude and phase mismatch, the gradient
descent optimization was performed over 𝛼 ∈[0.001:0.001:0.5], and the amplitude and phase
mismatch of each 𝛼 was recorded. The amplitude mismatch was calculated as the ratio between
the amplitude difference and the amplitude of Group1, and the phase mismatch was calculated as
the mean phase difference between Group1 and Group2. To obtain the relationship between the
amplitude/phase mismatch and LE, LE was first simulated over a range of amplitude and phase
87
mismatch between the two groups to generate a lookup table of LE vs. amplitude/phase mismatch
(see Section 5.4.2 for details), and the LE of each 𝛼 was estimated based on their resultant
amplitude/phase mismatch from the lookup table.
Meanwhile, to monitor the SAR with 𝛼 ∈[0.001:0.001:0.5], both global SAR and local
SAR were calculated. With a higher 𝛼, the resultant pTx weights would result in a lower combined
𝐵
O
P
amplitude, which was corrected so that the mean 𝐵
O
P
amplitude within the four inflowing
arteries were the same as the conventional CASL. The global SAR was calculated as the sum of
square of the channel weight amplitude, and the local SAR was calculated as the maximum square
of the channel weight amplitude of all channels. A more precise estimation of the local SAR is by
Virtual Observation Point (VOP) (Eichfelder & Gebhardt, 2011), but the coil information needed
was not available, as discussed in Section 5.6.2. The global SAR and local SAR were calculated
for both DCASL and CASL, and the ratio between the two was calculated.
5.4.2. LE simulation vs. 𝑩
𝟏
P
amplitude/phase mismatch
Numerical simulation was performed with Bloch equation to estimate the labeling
efficiency (LE) with amplitude and phase mismatch. The amplitude mismatch was simulated with
a ratio of 𝑩
𝟏
P
amplitude variation from -20% to 20% between Group2 and Group1 with a step size
of 2%, and the phase mismatch was simulated with phase difference from -11° to +11° with a step
size of 1°. The gradient amplitude was 1.6mT/m, and the RF amplitude for Group1 was fixed at
2.25uT (J. Wang et al., 2005) with phase=0. This setup was because the optimized pTx weights
can be scaled to achieve any desired RF amplitude without changing the mismatch between the
two groups.
The simulation was performed over time from -400ms to 400ms with a step size of 0.01ms,
assuming flow velocity = 40cm/s, blood T1/T2 = 2100ms/60ms, 𝐵
O
P
and 𝐵
N
inhomogeneities
88
ignored. The switching frequency of Group1 and Group2 was 20Hz. LE was simulated for both
label and control conditions, where the control was implemented with amplitude modulation (AM)
with a modulation frequency of 100Hz (J. Wang et al., 2005). The construction of control condition
RF block follows the same principle as label, except that the rectangular-shape RF pulse was
replaced with sinusoidal-shape pulses, as shown in Figure 5.2. The final LE w.r.t. the 𝑩
𝟏
P

amplitude/phase mismatch was calculated as the subtraction of the label and the control conditions.
5.4.3. Sequence implementation
Experiments were performed on a 7T MRI MAGNETOM Terra (Siemens Healthcare,
Erlangen, Germany) with an 8Tx/32Rx head coil (Nova Medical, MA, USA) on one subject. The
Figure 5.2 (A&B) The label and control RF pulses of two groups of channels, with no mismatch between
the two groups. The combined RF showed a perfect constant pattern for label and amplitude modulated
pattern for control. (C&D) The label and control RF pulses of the two group with the RF amplitude of
Group2 scaled by 0.8 compared with Group1. The combined RF pulses showed abrupt jumps when
channel group changed.
89
study was approved by the Institutional Review Board of the University of Southern California,
and written informed consent was obtained prior to the experiments.
To implement the proposed DCASL sequence, the 𝐵
O
P
of the four inflowing arteries were
require. First, Time-of-Flight (TOF) images were acquired to obtain angiography, the labeling
plane was placed at the C1 segment (Bouthillier classification) of the ICAs, and a 𝐵
O
P
map was
acquired at the labeling plane with a turbo-Fast-Low-Angle-Shot (TFL) sequence; Second, the
ROIs of the four inflowing arteries were manually drawn on the TOF image; Third, the ROIs were
resampled to match the voxel size of 𝐵
O
P
map, and the mean 𝐵
O
P
was calculated for each artery and
fed to the optimization algorithm.
Using the conventional CASL with default True-Form mode as reference, once the pTx
weights were calculated, they were scaled so that the mean 𝐵
O
P
amplitude within the four inflowing
arteries was the same as the CASL sequence. The weights were then converted to the RFPA
voltage by referring to the scanner reference voltage, and the RF waveforms in voltage of the eight
transmit channels were provided as an external RF pulse for the scanner.
5.5. Results
5.5.1. pTx weight optimization
The simulation was performed with data from a previously acquired cohort described in
Section 4.3.5. The four ROIs for the four inflowing arteries were manually drawn on the TOF
image, the ROIs were resampled based on the voxel size of the channel-specific 𝐵
O
P
maps, and the
mean 𝐵
O
P
for each ROI was calculated and fed to the optimization algorithm.
The optimization algorithm was verified with one representative subject, with SAR penalty
weighting 𝛼 =0.02 . Figure 5.3 shows the loss (only the fidelity term) tracked during the
optimization for Group1 and Group2, respectively. Optimization stopped at iteration number 37
90
and 52 for Group1 and Group2, respectively, which shows the optimization algorithm was
effective and efficient. The resultant combined 𝐵
O
P
for Group1 and Group2 were [0.85-0.01i; 0.88-
0.01i; 1.04-0.05i; 1.02-0.03i] and [0.84+0.05i; 0.92+0.01i; 1.08-0.04i;0.93-0.02i], resulting in a
5.4% and 5.2% Root Mean Square Error (RMSE) compared with the ground truth [1;1;1;1].
The optimization process was repeated with 𝛼 ∈[0.001:0.001:0.5], during which the
global SAR, the local SAR, the amplitude mismatch between the two channel groups, and the
phase mismatch between the two channel groups were calculated. Figure 5.4 shows that the global
SAR decreased, from 2.25 to 2.02, monotonically with 𝛼, and the local SAR first decreased from
2.6 to 2.2 then plateaued, with turning point at 𝛼 =0.1. Meanwhile, with 𝛼 increasing from 0.001
to 0.5, the amplitude mismatch increased from 4.5% to 7.5%, and the phase mismatch increased
from 2° to 5°.

Figure 5.3 Loss vs. iteration for Group1 (A) and Group2 (B) during the optimization process. Following the definition
in Section 5.4.1, Loss1 and Loss2 were loss calculated with real part and imaginary part, respectively. Loss for both
Group1 and Group2 converged to close to 0 after about iteration 20, and early stop was achieved for both groups.
91
5.5.2. LE simulation and SAR weighting evaluation
Figure 5.5 shows the simulated LE of the label, control, and label-control w.r.t. the 𝐵
O
P

amplitude and phase mismatch. Apparently, a higher amplitude or phase mismatch led to decreased
net LE (Label - Control), however, the LE was relatively insensitive to the mismatch. In the round-
shape region with a radius of approximately amplitude mismatch =15% or the phase mismatch =
9°, the net LE was above 0.75, comparing to the peak value LE=0.80 where amplitude and phase
mismatch were 0.

Figure 5.4 A The global SAR vs. 𝛼, calculated as the ratio against the conventional CASL.
Global SAR dropped monotonically with 𝛼. B The local SAR vs. 𝛼, calculated as the ratio
against the conventional CASL. Local SAR first dropped then plateaued with 𝛼, 𝛼=0.1 was
the turning point. C The amplitude mismatch between the two groups vs. 𝛼. D The phase
mismatch between the two groups vs. 𝛼.
92

Figure 5.6 shows the relationship between LE and 𝛼 , generated from the LE vs.
amplitude/phase mismatch relationship
and the mismatch vs. 𝛼 relationship.
Although the lookup table resolution
was relatively low resulting in a stair-
shape LE vs. 𝛼 function, it was obvious
that LE above 0.775 was achieved for all
𝛼, the peak LE being 0.80.
Since the local SAR is more
limiting than the global SAR at 7T even
Figure 5.5 The simulated LE with 𝐵
O
P
amplitude and phase mismatch between the two groups for A)
Label condition, B) Control condition, and C) the net LE (Label - Control). A higher mismatch
(amplitude or phase) led to decreased net LE, however, the LE was relatively insensitive to the
mismatch: LE>0.75 was achieved with most of the mismatch parameter range.
Figure 5.6 The relationship between LE and 𝛼 generated from
the LE vs. amplitude/phase mismatch lookup table and the
mismatch vs. 𝛼 relationship.
93
for the conventional CASL, and that the local SAR ratio of DCASL and CASL was higher than
that of global SAR (Section 5.5.1), local SAR was used to determine the most “SAR-efficient” 𝛼,
which led to 𝛼 = 0.1. Based on the simulation of LE with amplitude and phase mismatch, 𝛼 = 0.1
led to amplitude mismatch = 7.5%, phase mismatch = 5°, resulting in the LE only dropped from
80% to 79.3%. Therefore, 𝛼=0.1 was chosen as the weighting of the SAR penalty term.
5.5.3. Sequence implementation
Following the procedure described in Section 5.5.1, the channel-specific 𝐵
O
P
maps were
acquired at the labeling plane, and mean 𝐵
O
P
was calculated for each of the four inflowing arteries
with the ROIs manually drawn on the TOF image which was then fed into the optimization
algorithm, with 𝛼=0.1.
The pTx weights were calculated, and the performance of the DCASL was compared with
conventional CASL. Figure 5.7 shows the combined 𝐵
O
P
maps of the conventional CASL (True-
Figure 5.7 The combined 𝐵
O
P
maps of the conventional CASL and Group1 and Group2 of
DCASL. The amplitude map is shown in first row, and phase map second row. Both Group1
and Group2 of DCASL were able to achieve combined 𝐵
O
P
map well matching that of CASL.
94
Form mode), the channel Group1 of DCASL, and the channel Group2 of DCASL, first row
amplitude, second row phase. Although the combined 𝐵
O
P
maps of both Group1 and Group2 were
inhomogeneous, the 𝐵
O
P
(both amplitude and phase) at the four inflowing arteries matched well
with CASL. The amplitude and phase mismatch between the two groups were 5.5% and 2.84°,
respectively, which yielded a LE of approximately 0.794 according to the simulation results in
Section 5.5.2. The ratio between the global SAR of DCASL and that of CASL was 1.58, and the
ratio of local SAR was 2.29.
The pTx weights were used to generate the external RF pulses for the scanner, which was
successfully loaded as shown in Figure 5.8 from the scanner. Unfortunately, it was not allowed
by the RFPA possibly due to duty cycle limits, which was discussed in Section 5.6.3.
Figure 5.8 The external pulses loaded into the scanner. Pulses were successfully loaded for both
groups and both label and control conditions, but then the RFPA did not allow it.
95
5.6. Discussion
In this project, we developed the theoretical framework of an innovative DCASL sequence
implemented with dynamic pTx. The proposed DCASL sequence was able to achieve comparable
LE compared with the conventional CASL without violating the RFPA duty cycle limit. Although
currently it is not supported by our scanner software, it may be supported by scanners of other
vendors and by future software versions since there is no hardware limit violated.
5.6.1. SAR constraint

The proposed DCASL sequence has approximately twice the SAR of the convention CASL
sequence (global SAR ratio: 1.58, local SAR ratio: 2.29). The reason was only half the channels
are turned on during the RF pulse. Assuming equal contribution of each channel to the mean 𝐵
O
P

amplitude, to achieve the similar 𝐵
O
P
amplitude as conventional CASL where all eight channels
are turned at any time, the amplitude of each channel must approximately double for DCASL.
Figure 5.9 A The comparison of SAR between DCASL and CASL. For each RF channel, with DCASL the working
time is 50% but the amplitude is double compared with CASL, since SAR is proportional to the square of RF amplitude,
the total SAR of DCASL is approximately double compared with CASL. B The comparison of SAR between pCASL
and CASL. With pCASL, the RF channels are turned on for only ~50% time and that the Hanning pulse has a high
peak amplitude, which leads to 2.76 times SAR of CASL.
96
Therefore, for each channel, the working time (and thus the duty cycle) is reduced by half, but the
amplitude doubles leading to a quadruple power, which combined with the 50% working time
leads to a double SAR compared with CASL. A simple demonstration is shown in Figure 5.9A.
However, the SAR reduction of CASL compared with pCASL must be considered.
Following the same principle as the SAR comparison between CASL and DCASL, since with
CASL the RF is turned on at any time during labeling while the duty cycle of pCASL is
approximately 50%, much lower SAR can be obtained with CASL due to a longer RF pulse with
lower peak amplitude. Assuming the same time-average flip angle is achieved with pCASL and
CASL, and for pCASL the parameters are: RF duration = 500us, RF gap = 420us, Hanning pulse
with flip angle of each RF pulse = 25° (commonly accepted parameter used by (Dai et al., 2008;
Shao et al., 2019), etc.), then the SAR of pCASL is 2.76 times that of CASL, shown in Figure
5.9B. Therefore, the ratio between SAR of DCASL and pCASL will be 2.29/2.76 = 83%. Even if
we consider the increased RF amplitude (2.25uT, recommended by (J. Wang et al., 2005)) for
CASL used in the LE simulation, the SAR of DCASL will still be comparable to that of pCASL
(133%, which is overestimated because of overestimation of local SAR, discussed in Section 5.6.2).
Note that the parameters (RF amplitude = 2.25uT, gradient = 1.6mT/m) was proposed for 3T scans,
possible SAR reduction can be achieved if these parameters are optimized for 7T conditions.
Although the exact number may vary depending on individual subject, a comparable SAR of
DCASL should be achieved compared with pCASL.
5.6.2. SAR penalty in the cost function
The SAR penalty was added in the pTx weight optimization cost function in the format of
the l2-norm of the pTx weight, which penalized global SAR but did not explicitly include local
SAR. The best strategy would be to also include local SAR in the cost function. However, detailed
97
information about the RF coil field is unavailable for the system matrix required by the VOP
(Eichfelder & Gebhardt, 2011) method which controls the local SAR, and currently the local SAR
is only constrained conservatively by limiting the power of each RF channel. Therefore, no explicit
term regarding local SAR was added in the cost function. This may be the reason that the global
SAR decreased monotonically with 𝛼 while local SAR plateaued after 𝛼 =0.1. Once the VOP-
formatted local SAR term is included in the cost function, we expect that the local SAR will also
monotonically decrease with 𝛼, and a lower local SAR can be achieved if more amplitude/phase
mismatch is acceptable.
5.6.3. Implementation obstacles
Unfortunately, the proposed DCASL sequence was still not allowed by the scanner
platform. This was because currently, the sequence programming software only uses one object to
control the eight RF channels and gradients in three directions which does not provide the
flexibility to manipulate the timing of each RF channels independently. Therefore, at Figure 5.8
when the RF amplitude of one transmit channel is 0, it is still “on” but with a 0 value. However,
hardware-wise the eight RF channels are powered by eight independent RFPA, so it is theoretically
feasible to have independent control of the timing of each RF channel. Therefore, we believe that
the framework of the DCASL still benefits the ultra-high field ASL community, for it may be used
with future software of sequence programming or by other scanner vendors.
5.6.4. Limitations and future directions
First, the SAR penalty in the cost function did not explicitly include local SAR, which can
be included once the needed information for VOP is available. Second, the simulation of LE vs.
amplitude/phase mismatch did not consider the 𝐵
O
P
and 𝐵
N
inhomogeneity problem at 7T. The
sensitivity of LE to the amplitude/phase mismatch may be different with different 𝐵
O
P
and 𝐵
N

98
inhomogeneities. Third, since the labeling efficiency of DCASL should be very similar to
conventional CASL, labeling parameters including RF amplitude and gradient amplitude need to
be optimized for CASL with the 𝐵
O
P
and 𝐵
N
conditions at 7T for the optimal performance of
DCASL.
5.7. Conclusion
We proposed an innovative CASL with Dynamic pTx (DCASL) calculated with a gradient
descent algorithm and a custom-defined cost function. The proposed DCASL was able to achieve
comparable performance as the conventional CASL without violating the RFPA duty cycle limit
and had comparable or even lower SAR compared with commonly used pCASL sequence.

99
Chapter 6. Conclusion and Ongoing Work
In this dissertation, we systematically optimized and evaluated the four commonly used
adiabatic inversion pulses for PASL based on 7T 𝐵
O
P
and 𝐵
N
conditions, and the proposed
WURST-PASL sequence was able to achieve reliable perfusion measurements. We also optimized
the pCASL labeling parameters based on 7T 𝐵
O
P
and 𝐵
N
conditions, and reliable perfusion maps
were acquired with the optimized pCASL sequence with good repeatability. In addition, we
utilized the pTx 𝐵
O
P
shimming to increase the labeling efficiency (LE) by specifically increase the
𝐵
O
P
amplitude at the inflowing arteries. Compared with the default True-Form mode, the indv-shim
achieved significantly higher perfusion intensity with shimming weights calculated for individual
subject, and the univ-shim with weights calculated on a subject cohort also increased perfusion
without any calibration steps for individual subject. Furthermore, we proposed an innovative
CASL implemented with dynamic pTx (DCASL) that achieved comparable performance with
conventional CASL without violating the RFPA duty cycle limit of each transmit channel.
To further increase the reliability, repeatability, and accuracy of the perfusion
measurements at 7T, there still exist several improvements that are worth exploring in the future.
6.1. Optimization of the preparation pulses
For both PASL and pCASL, the pulses used for the ASL preparation have been optimized
based on the 𝐵
O
P
and 𝐵
N
conditions at 7T. However, some preparation pulses, such as the inferior
saturation pulse in PASL for CBF quantification, are still using their 3T parameters, which may
interfere with the perfusion measurement. The inferior saturation pulses in QUIPSS II PASL
(Wong et al., 1998) is supposed to suppress any perfusion signal after a delay time. However,
given the 𝐵
O
P
and 𝐵
N
inhomogeneities at 7T especially at the target region of the saturation where
it is below the imaging volume, the saturation may not be ideal, leading to residual contribution of
100
the lagging inflowing blood that is not accounted for in the quantification, which may lead to
overestimation of CBF.
The background suppression (BS) was not used by either PASL or pCASL in this
dissertation due to SAR limitation. Fortunately, the CSF signal was suppressed by the turbo Fast
Low Angle Shot (TFL) readout due to its long T1. However, in the future if more efficient
Figure 6.1 The inversion profile of the BS inversion pulse directly adopted from 3T with (A) 𝛥𝐵
N
=−100𝐻𝑧, (B)
𝛥𝐵
N
=0𝐻𝑧, and (C) 𝛥𝐵
N
=100𝐻𝑧. The inversion efficiency dropped dramatically when 𝐵
O
P
was less than 50%.
Figure 6.2 The fractional perfusion maps of three subjects, acquired with pCASL sequence with and without
BS, shown in top row and bottom row, respectively. Perfusion signal was significantly reduced by BS. The
ratio between the perfusion of pCASL with BS and that without BS was 63.1%.
101
acquisition schemes such as Turbo Gradient Spin Echo (TGSE) (Spann et al., 2020) are used, BS
may be necessary to suppress CSF signal and to further increase the SNR. Figure 6.1 shows the
simulated profile of the original BS inversion pulse adopted from 3T with 𝐵
O
P
amplitude scaled by
[30:10:100] % and 𝐵
N
offset = -100Hz, 0, and 100Hz, respectively. The inversion efficiency was
not sensitive to 𝐵
N
offset but dropped fast with a lower 𝐵
O
P
amplitude. Considering a BS block
typically has 2 inversion pulses, the reduction of the perfusion signal intensity would be even
worse. Figure 6.2 shows perfusion maps of three subjects with pCASL, acquired with (top row)
and without (bottom row) BS, with every other parameter controlled. The ratio of perfusion
between with and without BS was 63.1%. Therefore, BS inversion pulses also need to be
optimized.
To optimize the saturation and the inversion pulses, the proper weight for each 𝐵
O
P
/𝐵
N
is
needed. Unlike the pCASL labeling pulse which only applies on the labeling plane, the target
region of the saturation pulse and the inversion pulse are a volume, therefore, the 𝐵
O
P
/𝐵
N

distribution in the 3D volume will need to be considered.
6.2. Expansion of the LE simulation models
In the pCASL LE simulation, we optimized three pCASL labeling parameters including
RF duration, 𝐺
ÇÎc
, and 𝑔𝑅𝑎𝑡𝑖𝑜, assuming the mean flow velocity is 40cm/s. In practice, however,
the flow velocity is temporally pulsatile and spatially laminar. The simplified assumption of the
flow velocity may lead to bias of the optimal labeling parameters. In the future, a more complex
flow model will be used to in the pCASL parameter optimization, as well as more parameter
dimensions including the RF gap.
In the DCASL LE simulation, we simulated the LE vs. the amplitude and phase mismatch
between the two channel groups, assuming the ideal CASL labeling with RF amplitude = 2.25uT,
102
gradient = 1.6mT/m, no 𝐵
O
P
and 𝐵
N
inhomogeneities. Considering the 𝐵
O
P
and 𝐵
N
inhomogeneities
at 7T, the accurate LE should be simulated with the 4D parameter space of amplitude and phase
mismatch and 𝐵
O
P
and 𝐵
N
inhomogeneities. Therefore, in the future an expanded model with 𝐵
O
P

and 𝐵
N
inhomogeneities will be used.
6.3. Optimization of conventional CASL
The performance of the proposed DCASL sequence is limited by the performance of its
corresponding convention CASL sequence, while currently the CASL parameters (RF amplitude
= 2.25uT, gradient=1.6mT/m) are directly adopted from 3T (J. Wang et al., 2005), which may not
be suitable for 7T given the 𝐵
O
P
and 𝐵
N
inhomogeneities. Figure 6.3 shows the simulated LE of
CASL labeling with 𝐵
O
P
scaled by [20:10:120] % from the nominal 2.25uT and 𝐵
N
offset frequency
= [-200:20:200] Hz, assuming mean flow velocity = 40cm/s. Figure 6.3A is the weighted average
of LE vs. gradient, with each 𝐵
O
P
/𝐵
N
condition weighted by their relative frequency (see Chapter
3 Section 3.3.2), which shows the optimal gradient given the 𝐵
O
P
/𝐵
N
inhomogeneities at 7T was
Figure 6.3 A The weighted average labeling efficiency of CASL with 𝐺
ÇÎc
, with weighting for each 𝐵
O
P
/𝐵
N
condition
as their corresponding relative frequency, the highest LE (0.83) was achieved with 𝐺
ÇÎc
=0.6 mT/m. B The LE vs. 𝐵
O
P

and 𝐵
N
inhomogeneities. The highest LE was achieved at approximately 70% 𝐵
O
P
amplitude, while LE was insensitive
to 𝐵
N
offset.
103
0.6mT/m. Figure 6.3B is the LE vs. 𝐵
O
P
/𝐵
N
inhomogeneities with the optimal gradient 0.6mT/m,
which shows that CASL labeling is indeed insensitive to 𝐵
N
offset but the LE changes with 𝐵
O
P

amplitude.
Taking advantage of the Continuous Waveform mode of the Siemens Terra system, we
were able to implement a conventional single-slice CASL sequence and performed an exploratory
in-vivo experiment evaluating the different amplitudes of RF (1.5, 2.25uT) and gradient (0.8, 1.6,
2.4mT/m), values adopted from (J. Wang et al., 2005). Results of one subject were shown in
Figure 6.4, which shows consistent results with simulation. In the future, more complex flow
model will need to be included in the simulation, and a larger subject cohort will be used for in-
vivo experiments to verify simulation results.

Figure 6.4 The perfusion map of the single-slice CASL sequence with different RF and
gradient amplitude. The combination of high RF amplitude (2.25uT) and low gradient
amplitude (0.8mT/m) yielded the highest perfusion signal.
104
Meanwhile, the control condition for CASL will also be optimized at 7T. Single slice
CASL (CASL-SS) only works with single slice that is parallel to the labeling plane. The amplitude
modulated CASL (CASL-AM) suffers from reduced labeling efficiency (LE) especially when the
RF field is reduced to overcome the SAR limit at higher field (Gach & Dai, 2004; J. Wang et al.,
2005; Werner, Norris, Alfke, Mehdorn, & Jansen, 2005), and the AM is sensitive to 𝐵
N
offset as
shown in Figure 6.5. Therefore, in the future the modulation frequency will be optimized based
on LE with different 𝐵
N
offset. Alternatives such as advanced 𝐵
N
offset correction (Berry, Jezzard,
& Okell, 2019; Sengupta et al., 2011) will also be explored.

Figure 6.5 The simulated labeling efficiency of label and control with AM. LE of label is independent from 𝐵
N
offset,
while the control showed a period fluctuation of LE with 𝐵
N
offset, shorter period for higher 𝐺
ÇÎc
. For
𝐺
ÇÎc
=0.6mT/m, the LE fluctuation was not obvious, but LE of control was as high as 0.35 resulting in a net LE
(label - control) of 0.55.
105
References
Alsop, D. C., & Detre, J. A. (1998). Multisection cerebral blood flow MR imaging with
continuous arterial spin labeling. Radiology, 208(2), 410-416.
doi:10.1148/radiology.208.2.9680569
Alsop, D. C., Detre, J. A., Golay, X., Günther, M., Hendrikse, J., Hernandez-Garcia, L., . . .
Zaharchuk, G. (2015). Recommended implementation of arterial spin-labeled perfusion
MRI for clinical applications: A consensus of the ISMRM perfusion study group and the
European consortium for ASL in dementia. Magn Reson Med, 73(1), 102-116.
doi:10.1002/mrm.25197
Balchandani, P., Khalighi, M., Hsieh, S., Setsompop, K., Pauly, J., & Spielman, D. (2011).
Adiabatic B1 shimming algorithm for multiple channel transmit at 7T. Paper presented at
the Proceedings of the 19th Scientific Meeting, ISMRM. Montreal.
Bause, J., Ehses, P., Mirkes, C., Shajan, G., Scheffler, K., & Pohmann, R. (2016). Quantitative
and functional pulsed arterial spin labeling in the human brain at 9.4 T. Magn Reson Med,
75(3), 1054-1063. doi:10.1002/mrm.25671
Berry, E. S. K., Jezzard, P., & Okell, T. W. (2019). Off-resonance correction for pseudo-
continuous arterial spin labeling using the optimized encoding scheme. Neuroimage, 199,
304-312. doi:10.1016/j.neuroimage.2019.05.083
Cloos, M. A., Boulant, N., Luong, M., Ferrand, G., Giacomini, E., Le Bihan, D., & Amadon, A.
(2012). kT -points: short three-dimensional tailored RF pulses for flip-angle
homogenization over an extended volume. Magn Reson Med, 67(1), 72-80.
doi:10.1002/mrm.22978
Dai, W., Garcia, D., de Bazelaire, C., & Alsop, D. C. (2008). Continuous flow-driven inversion
for arterial spin labeling using pulsed radio frequency and gradient fields. Magn Reson
Med, 60(6), 1488-1497. doi:10.1002/mrm.21790
Deniz, C. M. (2019). Parallel Transmission for Ultrahigh Field MRI. Top Magn Reson Imaging,
28(3), 159-171. doi:10.1097/RMR.0000000000000204
Detre, J. A., & Wang, J. (2002). Technical aspects and utility of fMRI using BOLD and ASL.
Clin Neurophysiol, 113(5), 621-634. doi:10.1016/s1388-2457(02)00038-x
Duong, T. Q., Kim, D. S., Uğurbil, K., & Kim, S. G. (2001). Localized cerebral blood flow
response at submillimeter columnar resolution. Proc Natl Acad Sci U S A, 98(19), 10904-
10909. doi:10.1073/pnas.191101098
Eichfelder, G., & Gebhardt, M. (2011). Local specific absorption rate control for parallel
transmission by virtual observation points. Magn Reson Med, 66(5), 1468-1476.
doi:10.1002/mrm.22927
106
Gach, H. M., & Dai, W. (2004). Simple model of double adiabatic inversion (DAI) efficiency.
Magn Reson Med, 52(4), 941-946. doi:10.1002/mrm.20240
Gai, N. D., & Zur, Y. (2007). Design and optimization for variable rate selective excitation using
an analytic RF scaling function. J Magn Reson, 189(1), 78-89.
doi:10.1016/j.jmr.2007.08.017
Garcia, D. M., Duhamel, G., & Alsop, D. C. (2005). Efficiency of inversion pulses for
background suppressed arterial spin labeling. Magn Reson Med, 54(2), 366-372.
doi:10.1002/mrm.20556
Ghariq, E., Teeuwisse, W. M., Webb, A. G., & van Osch, M. J. (2012). Feasibility of
pseudocontinuous arterial spin labeling at 7 T with whole-brain coverage. MAGMA,
25(2), 83-93. doi:10.1007/s10334-011-0297-0
Gras, V., Vignaud, A., Amadon, A., Le Bihan, D., & Boulant, N. (2017). Universal pulses: A
new concept for calibration-free parallel transmission. Magn Reson Med, 77(2), 635-643.
doi:10.1002/mrm.26148
Hoult, D. I., & Phil, D. (2000). Sensitivity and power deposition in a high-field imaging
experiment. J Magn Reson Imaging, 12(1), 46-67. doi:10.1002/1522-
2586(200007)12:1<46::aid-jmri6>3.0.co;2-d
Hurley, A. C., Al-Radaideh, A., Bai, L., Aickelin, U., Coxon, R., Glover, P., & Gowland, P. A.
(2010). Tailored RF pulse for magnetization inversion at ultrahigh field. Magn Reson
Med, 63(1), 51-58. doi:10.1002/mrm.22167
Ibrahim, T. S., Lee, R., Baertlein, B. A., Kangarlu, A., & Robitaille, P. L. (2000). Application of
finite difference time domain method for the design of birdcage RF head coils using
multi-port excitations. Magn Reson Imaging, 18(6), 733-742. doi:10.1016/s0730-
725x(00)00143-0
Katscher, U., Börnert, P., Leussler, C., & van den Brink, J. S. (2003). Transmit SENSE. Magn
Reson Med, 49(1), 144-150. doi:10.1002/mrm.10353
Kim, S. G. (1995). Quantification of relative cerebral blood flow change by flow-sensitive
alternating inversion recovery (FAIR) technique: application to functional mapping.
Magn Reson Med, 34(3), 293-301. doi:10.1002/mrm.1910340303
Kim, S. G., & Tsekos, N. V. (1997). Perfusion imaging by a flow-sensitive alternating inversion
recovery (FAIR) technique: application to functional brain imaging. Magn Reson Med,
37(3), 425-435. doi:10.1002/mrm.1910370321
Kupce, E. (1995). Adiabatic pulses for wideband inversion and broadband decoupling. Journal
of magnetic resonance, Series A(115), 4.

107
Liu, H. L., Pu, Y., Nickerson, L. D., Liu, Y., Fox, P. T., & Gao, J. H. (2000). Comparison of the
temporal response in perfusion and BOLD-based event-related functional MRI. Magn
Reson Med, 43(5), 768-772. doi:10.1002/(sici)1522-2594(200005)43:5<768::aid-
mrm22>3.0.co;2-8
Luh, W. M., Talagala, S. L., Li, T. Q., & Bandettini, P. A. (2013). Pseudo-continuous arterial
spin labeling at 7 T for human brain: estimation and correction for off-resonance effects
using a Prescan. Magn Reson Med, 69(2), 402-410. doi:10.1002/mrm.24266
Mao, W., Smith, M. B., & Collins, C. M. (2006). Exploring the limits of RF shimming for high-
field MRI of the human head. Magn Reson Med, 56(4), 918-922. doi:10.1002/mrm.21013
Mora Álvarez, M. G., Stobbe, R. W., & Beaulieu, C. (2019). High resolution continuous arterial
spin labeling of human cerebral perfusion using a separate neck tagging RF coil. PLoS
One, 14(4), e0215998. doi:10.1371/journal.pone.0215998
Nistler, J., Diehl, D., Renz, W., & Eberler, L. (2004). Homogeneity Improvement Using A 2 Port
Birdcage Coil. Paper presented at the International Society for Medicine Resonance in
Medicine, Kyoto.
O'Brien, K. R., Magill, A. W., Delacoste, J., Marques, J. P., Kober, T., Fautz, H. P., . . . Krueger,
G. (2014). Dielectric pads and low- B1+ adiabatic pulses: complementary techniques to
optimize structural T1 w whole-brain MP2RAGE scans at 7 tesla. J Magn Reson
Imaging, 40(4), 804-812. doi:10.1002/jmri.24435
Ordidge, R. J., Wylezinska, M., Hugg, J. W., Butterworth, E., & Franconi, F. (1996). Frequency
offset corrected inversion (FOCI) pulses for use in localized spectroscopy. Magn Reson
Med, 36(4), 562-566. doi:10.1002/mrm.1910360410
Pantano, P., Baron, J. C., Lebrun-Grandié, P., Duquesnoy, N., Bousser, M. G., & Comar, D.
(1984). Regional cerebral blood flow and oxygen consumption in human aging. Stroke,
15(4), 635-641. doi:10.1161/01.str.15.4.635
Parkes, L. M., Rashid, W., Chard, D. T., & Tofts, P. S. (2004). Normal cerebral perfusion
measurements using arterial spin labeling: reproducibility, stability, and age and gender
effects. Magn Reson Med, 51(4), 736-743. doi:10.1002/mrm.20023
Pohmann, R., Speck, O., & Scheffler, K. (2016). Signal-to-noise ratio and MR tissue parameters
in human brain imaging at 3, 7, and 9.4 tesla using current receive coil arrays. Magn
Reson Med, 75(2), 801-809. doi:10.1002/mrm.25677
Power, S., Talbot, N., Kucharczyk, W., & Mandell, D. M. (2016). Allergic-like Reactions to the
MR Imaging Contrast Agent Gadobutrol: A Prospective Study of 32 991 Consecutive
Injections. Radiology, 281(1), 72-77. doi:10.1148/radiol.2016151066

108
Rempp, K. A., Brix, G., Wenz, F., Becker, C. R., Gückel, F., & Lorenz, W. J. (1994).
Quantification of regional cerebral blood flow and volume with dynamic susceptibility
contrast-enhanced MR imaging. Radiology, 193(3), 637-641.
doi:10.1148/radiology.193.3.7972800
Sengupta, S., Welch, E. B., Zhao, Y., Foxall, D., Starewicz, P., Anderson, A. W., . . . Avison, M.
J. (2011). Dynamic B0 shimming at 7 T. Magn Reson Imaging, 29(4), 483-496.
doi:10.1016/j.mri.2011.01.002
Setsompop, K., Wald, L. L., Alagappan, V., Gagoski, B. A., & Adalsteinsson, E. (2008).
Magnitude least squares optimization for parallel radio frequency excitation design
demonstrated at 7 Tesla with eight channels. Magn Reson Med, 59(4), 908-915.
doi:10.1002/mrm.21513
Shao, X., Ma, S. J., Casey, M., D'Orazio, L., Ringman, J. M., & Wang, D. J. J. (2019). Mapping
water exchange across the blood-brain barrier using 3D diffusion-prepared arterial spin
labeled perfusion MRI. Magn Reson Med, 81(5), 3065-3079. doi:10.1002/mrm.27632
Shao, X., Wang, Y., Moeller, S., & Wang, D. J. J. (2018). A constrained slice-dependent
background suppression scheme for simultaneous multislice pseudo-continuous arterial
spin labeling. Magn Reson Med, 79(1), 394-400. doi:10.1002/mrm.26643
Silver, M., Joseph, R., & Hoult, D. (1969). Highly selective π2 and π pulse generation. Journal
of Magnetic Resonance, 59(2), 347-351.
Spann, S. M., Shao, X., Wang, D. J., Aigner, C. S., Schloegl, M., Bredies, K., & Stollberger, R.
(2020). Robust single-shot acquisition of high resolution whole brain ASL images by
combining time-dependent 2D CAPIRINHA sampling with spatio-temporal TGV
reconstruction. Neuroimage, 206, 116337. doi:10.1016/j.neuroimage.2019.116337
Tannús, A., & Garwood, M. (1997). Adiabatic pulses. NMR Biomed, 10(8), 423-434.
doi:10.1002/(sici)1099-1492(199712)10:8<423::aid-nbm488>3.0.co;2-x
Teeuwisse, W. M., Webb, A. G., & van Osch, M. J. (2010). Arterial spin labeling at ultra‐high
field: all that glitters is not gold. International Journal of Imaging Systems and
Technology, 20(1), 62-70.
Tong, Y., Jezzard, P., Okell, T. W., & Clarke, W. T. (2020). Improving PCASL at ultra-high
field using a VERSE-guided parallel transmission strategy. Magn Reson Med.
doi:10.1002/mrm.28173
van den Bergen, B., Van den Berg, C. A., Bartels, L. W., & Lagendijk, J. J. (2007). 7 T body
MRI: B1 shimming with simultaneous SAR reduction. Phys Med Biol, 52(17), 5429-
5441. doi:10.1088/0031-9155/52/17/022
Wang, J., Zhang, Y., Wolf, R. L., Roc, A. C., Alsop, D. C., & Detre, J. A. (2005). Amplitude-
modulated continuous arterial spin-labeling 3.0-T perfusion MR imaging with a single
coil: feasibility study. Radiology, 235(1), 218-228.
109
Wang, K., Shao, X., Yan, L., Ma, S. J., Jin, J., & Wang, D. J. J. (2021). Optimization of adiabatic
pulses for pulsed arterial spin labeling at 7 tesla: Comparison with pseudo-continuous
arterial spin labeling. Magn Reson Med, 85(6), 3227-3240. doi:10.1002/mrm.28661
Wang, Y., Moeller, S., Li, X., Vu, A. T., Krasileva, K., Ugurbil, K., . . . Wang, D. J. (2015).
Simultaneous multi-slice Turbo-FLASH imaging with CAIPIRINHA for whole brain
distortion-free pseudo-continuous arterial spin labeling at 3 and 7 T. Neuroimage, 113,
279-288. doi:10.1016/j.neuroimage.2015.03.060
Werner, R., Norris, D. G., Alfke, K., Mehdorn, H. M., & Jansen, O. (2005). Improving the
amplitude-modulated control experiment for multislice continuous arterial spin labeling.
Magn Reson Med, 53(5), 1096-1102. doi:10.1002/mrm.20443
Williams, D. S., Detre, J. A., Leigh, J. S., & Koretsky, A. P. (1992). Magnetic resonance imaging
of perfusion using spin inversion of arterial water. Proc Natl Acad Sci U S A, 89(1), 212-
216. doi:10.1073/pnas.89.1.212
Wong, E. C. (2005). Quantifying CBF with pulsed ASL: technical and pulse sequence factors. J
Magn Reson Imaging, 22(6), 727-731. doi:10.1002/jmri.20459
Wong, E. C., Buxton, R. B., & Frank, L. R. (1998). Quantitative imaging of perfusion using a
single subtraction (QUIPSS and QUIPSS II). Magn Reson Med, 39(5), 702-708.
doi:10.1002/mrm.1910390506
Wright, P. J., Mougin, O. E., Totman, J. J., Peters, A. M., Brookes, M. J., Coxon, R., . . .
Gowland, P. A. (2008). Water proton T1 measurements in brain tissue at 7, 3, and 1.5 T
using IR-EPI, IR-TSE, and MPRAGE: results and optimization. MAGMA, 21(1-2), 121-
130. doi:10.1007/s10334-008-0104-8
Wu, W. C., Fernández-Seara, M., Detre, J. A., Wehrli, F. W., & Wang, J. (2007). A theoretical
and experimental investigation of the tagging efficiency of pseudocontinuous arterial spin
labeling. Magn Reson Med, 58(5), 1020-1027. doi:10.1002/mrm.21403
Wu, X., Auerbach, E. J., Vu, A. T., Moeller, S., Lenglet, C., Schmitter, S., . . . Uğurbil, K.
(2018). High-resolution whole-brain diffusion MRI at 7T using radiofrequency parallel
transmission. Magn Reson Med, 80(5), 1857-1870. doi:10.1002/mrm.27189
Wu, X., Auerbach, E. J., Vu, A. T., Moeller, S., Van de Moortele, P. F., Yacoub, E., & Uğurbil,
K. (2019). Human Connectome Project-style resting-state functional MRI at 7 Tesla
using radiofrequency parallel transmission. Neuroimage, 184, 396-408.
doi:10.1016/j.neuroimage.2018.09.038
Zhao, L., Vidorreta, M., Soman, S., Detre, J. A., & Alsop, D. C. (2017). Improving the
robustness of pseudo-continuous arterial spin labeling to off-resonance and pulsatile flow
velocity. Magn Reson Med, 78(4), 1342-1351. doi:10.1002/mrm.26513
Zhu, Y. (2004). Parallel excitation with an array of transmit coils. Magn Reson Med, 51(4), 775-
784. doi:10.1002/mrm.20011
110
Zimmer, F., O'Brien, K., Bollmann, S., Pfeuffer, J., Heberlein, K., & Barth, M. (2016). Pulsed
arterial spin labelling at ultra-high field with a B 1 (+) -optimised adiabatic labelling
pulse. MAGMA, 29(3), 463-473. doi:10.1007/s10334-016-0555-2
Zuo, Z., Wang, R., Zhuo, Y., Xue, R., St Lawrence, K. S., & Wang, D. J. (2013). Turbo-FLASH
based arterial spin labeled perfusion MRI at 7 T. PLoS One, 8(6), e66612.
doi:10.1371/journal.pone.0066612

111
APPENDIX A
Definition of the Adiabatic Pulses
HS pulse: The amplitude |𝐵
O
(𝑡)| is a hyperbolic secant function, and the frequency Δ𝜔(𝑡)
is a hyperbolic tangent function:
|𝐵
O
(𝑡)|=𝑆𝑒𝑐ℎ(𝛽𝑡)
Δ𝜔(𝑡)= −𝜇𝛽tanh(𝛽𝑡) [S1]

where 𝛽 defines the shape of the amplitude and 𝜇 adjusts the speed of the frequency sweep.
𝛽 and 𝜇 will be optimized to improve the pulse performance.
WURST pulse: The amplitude and frequency are defined as:

|𝐵
O
(𝑡)|=1−|sin(𝛽𝑡)|
f

Δ𝜔(𝑡)= 𝑘𝑡 [S2]

where −
Ø
n
<𝛽𝑡 <
Ø
n
, 𝑛 is the order of the modulation function and 𝑘 is the constant of the
frequency sweep rate. A larger 𝑛 yields a flatter waveform, and a higher 𝑘 yields a faster frequency
sweep and thus a larger bandwidth. 𝑘 and 𝑛 will be optimized.
FOCI pulse: Modulated from the HS pulse, the FOCI pulse uses an additional modulation
function with the aim of further sharpening the inversion profile. The same modulation function
was applied to the pulse amplitude, frequency sweep rate, and gradient amplitude to satisfy the
adiabatic condition. With a higher gradient, the FOCI pulse achieves a narrower transition band
compared with the original HS pulse. The most commonly used FOCI is constant-FOCI (C-FOCI)
for which the modulation function is designed such that the pulse waveform has a plateau when
exceeding a threshold (0.5 for this case):
112
𝐴(𝑡) =<
O
±c=(>Õ)
𝑤ℎ𝑒𝑛 𝑆𝑒𝑐ℎ(𝛽𝑡)>0.5
2, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒.

|𝐵
O
(𝑡)|=𝐴(𝑡)𝑆𝑒𝑐ℎ(𝛽𝑡)
Δ𝜔(𝑡) = −𝜇𝛽A(t)tanh(𝛽𝑡)
Gradient(t)=𝐺
ù±
𝐴(𝑡) [S3]

Same as the HS pulse, 𝛽 and 𝜇 will be optimized.
trFOCI pulse: The trFOCI pulse employs a piecewise modulation function that allows
more degrees-of-freedom to achieve an ideal inversion profile in the presence of 𝐵
N
and 𝐵
O
P

inhomogeneities. As shown in Fig 1 of (Hurley et al., 2010), there are three segments in the
modulation function, including two symmetrical linear segments and one curved segment, which
are defined as the following:
𝐴
O
(𝑡)=𝐴
½ÇÈ
@1−
𝑟
O
(𝑡+1)
𝑤
A
𝐴
n
(𝑡)=𝑏
O
𝑡
n
+𝑏
n
𝑡
Æ
+𝑏
É
𝑡
v
+𝑏
Æ
𝑡
1
+𝐴
½[f
[S4a]

where

𝑏
O
=
𝑟
É
(𝑐−𝐴
½[f
)
(𝑤−1)
n

𝑏
n
=
𝑟
Æ
(1−𝑟
É
)(𝑐−𝐴
½[f
)
(𝑤−1)
Æ

𝑏
É
=
𝑟
Ê
(1−𝑟
)Æ
(1−𝑟
É
)(𝑐−𝐴
½[f
)
(𝑤−1)
v

𝑏
Æ
=
(1−𝑟
Ê
)(1−𝑟
)Æ
(1−𝑟
É
)(𝑐−𝐴
½[f
)
(𝑤−1)
1

113
𝑐 =𝐴
½ÇÈ
(1−𝑟
O
) [S4b]

Given the time resampling function,
𝑇(𝑡) =

s
Õ
B
P
~
Õ
C
PÕ

s
P
~
PO
[S5]
the equations describing the new pulses become
|𝐵
O
(𝑡)|=𝐴(𝑡)𝑆𝑒𝑐ℎ(𝛽(𝑇(𝑡)))
Δ𝜔(𝑡)= −𝜇𝛽A(t)tanh(𝛽𝑇(𝑡))
Gradient(t)=𝐺
ù±
𝐴(𝑡) [S6]
Therefore, an 11-dimensional vector of parameters will be optimized:
[𝐴
½ÇÈ
,𝑤,𝑟
O
,𝑟
n
,𝑟
É
,𝑟
Æ
,𝑟
Ê
,𝜇,𝛽,𝜏
O
,𝜏
n
].

Abstract (if available)

Abstract Arterial spin labeling (ASL) is a perfusion magnetic resonance imaging (MRI) technique that utilizes magnetically labeled arterial blood water as an endogenous tracer to measure cerebral blood flow (CBF). The noninvasive nature and the ability to quantitatively measure tissue perfusion make ASL ideal for research and clinical studies. The main limitation of ASL technique is the low signal-to-noise (SNR) due to the intrinsically small fraction of the labeled arterial blood (~1%) and T1 relaxation of the label. ? Ultra-high field (UHF) benefits ASL with an increased intrinsic SNR of MRI signal (B?¹·??) due to the dielectric resonance of the magnetic field and a prolonged tracer half-life (blood T1). However, the implementation of ASL at UHF is not straightforward. Due to the inhomogeneity of the transmit B? (B??) field and the B? field, it is challenging to achieve high labeling efficiency with both pulsed ASL (PASL) and pseudo-continuous ASL (pCASL), which are the two commonly used ASL techniques. Meanwhile, parallel transmission (pTx) provides previously unavailable degree of freedom that allows full spatial and temporal control of the B?? field. With a transmit coil array that consists of several elements with spatially distinct B?? field pattern, the extra degrees of freedom can be exploited to overcome the effects of B?? inhomogeneities, which provides potential solutions to mitigate the issues in implementing ASL at UHF. ? The overall objective of this work is to develop a suite of reliable ASL technologies at UHF. First, the 7T PASL sequence was optimized by optimizing and evaluating the adiabatic inversion pulses with the B?? and B? inhomogeneities at 7T. The wide-band-uniform-rate-smooth-truncation (WURST) pulse achieved the lowest loss in simulation and achieved a superior performance compared with the other three pulses in the experiments. Second, the 7T pCASL parameters were optimized to achieve a high labeling efficiency with the B?? and B? inhomogeneities at 7T. The optimized pCASL sequence achieved robust labeling efficiency as well as good repeatability. Third, the optimized pCASL was incorporated with pTx B?? shimming aiming at further improving the labeling efficiency by increasing the B?? amplitude at the inflowing arteries. Both “indv-shim” (shimming weights calculated for each individual subject) and “univ-shim” (universal shimming weight calculated based on a group of subjects) successfully achieved increased perfusion signal intensity compared with the circular polarized (CP) mode. Fourth, an innovative continuous ASL (CASL) with Dynamic pTx pulse (noted as DCASL) was proposed. By utilizing channel-specific RF pulses for each transmit channel, the proposed DCASL showed intrinsic insensitivity to B? offset and relatively low Specific Absorption Rate (SAR) similar to conventional CASL sequence without violating the duty cycle limit of each RF transmit channel. ? In conclusion, reliable perfusion measurements were obtained with the PASL and pCASL sequences optimized for 7T, pTx B?? shimming methods further increased the LE of pCASL sequence, and the theoretical framework of an innovative DCASL implemented with dynamic pTx was proposed and evaluated.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Improving sensitivity and spatial coverage of myocardial arterial spin labeling

PDF

Improving the sensitivity and spatial coverage of cardiac arterial spin labeling for assessment of coronary artery disease

PDF

Improved myocardial arterial spin labeled perfusion imaging

PDF

Improving arterial spin labeling in clinical application with deep learning

PDF

Radiofrequency pulse performance for myocardial ASL

PDF

Assessment of myocardial blood flow in humans using arterial spin labeled MRI

PDF

Mapping water exchange rate across the blood-brain barrier

PDF

Characterization of lenticulostriate arteries using high-resolution black blood MRI as an early imaging biomarker for vascular cognitive impairment and dementia

PDF

Estimating liver iron non-invasively with high-field MRI

PDF

Developing optical instrumentation for detecting cutaneous skin perfusion level correcting transcutaneous ICG dilution curve cardiac output measurement

$Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease$

PDF

Quantitative MRI for the measurement of cerebral oxygen extraction fraction in sickle cell disease

PDF

Biophysical studies of nanosecond pulsed electric field induced cell membrane permeabilization

PDF

Development of high-frequency and high-field optically detected magnetic resonance

PDF

Optimization methods and algorithms for constrained magnetic resonance imaging

PDF

3D vessel mapping techniques for retina and brain as an early imaging biomarker for small vessel diseases

PDF

A virtual reality exergaming system to enhance brain health in older adults at risk for Alzheimer’s disease

PDF

Learning multi-annotator subjective label embeddings

PDF

Novel beamforming with dual-layer array transducers for 3-D ultrasound imaging

PDF

Computation of class groups and residue class rings of function fields over finite fields

PDF

Investigation of the electrode-tissue interface of retinal prostheses

Asset Metadata

Creator Wang, Kai (author)

Core Title Measurement of human brain perfusion with arterial spin labeling magnetic resonance imaging at ultra-high field

School Viterbi School of Engineering

Degree Doctor of Philosophy

Degree Program Biomedical Engineering

Degree Conferral Date 2021-08

Publication Date 07/16/2023

Defense Date 06/10/2021

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag arterial spin labeling,OAI-PMH Harvest,parallel transmission,perfusion,ultra-high field

Format application/pdf (imt)

Language English

Advisor Wang, Danny JJ (committee chair)

Creator Email kaiw1227@gmail.com,wang416@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-oUC15611965

Unique identifier UC15611965

Legacy Identifier etd-WangKai-9802

Document Type Dissertation

Format application/pdf (imt)

Rights Wang, Kai

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA

Repository Email cisadmin@lib.usc.edu