1862 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO.

7, JULY 2022

System Matrix Based Reconstruction for Pulsed

Sequences in Magnetic Particle Imaging
Fabian Mohn , Tobias Knopp , Marija Boberg , Florian Thieben ,
Patryk Szwargulski , and Matthias Graeser

Abstract — Improving resolution and sensitivity will (-1.25, -1.25, 2.5) Tm-1 ) in x- and y-direction was achieved and
widen possible medical applications of magnetic particle a superior sensitivity for pulsed sequences was detected on
imaging. Pulsed excitation promises such benefits, at the the basis of reference phantoms.
cost of more complex hardware solutions and restrictions
on drive field amplitude and frequency. State-of-the-art Index Terms — Biomedical imaging, pulsed excitation,
systems utilize a sinusoidal excitation to drive superpara- high amplitudes, sequence design, MPI.
magnetic nanoparticles into the non-linear part of their
magnetization curve, which creates a spectrum with a I. I NTRODUCTION
clear separation of direct feed-through and higher har-
monics caused by the particles response. One challenge
for rectangular excitation is the discrimination of particle
I N MAGNETIC Particle Imaging (MPI) the spatial dis-
tribution of superparamagnetic iron oxide nanoparticles
(SPIONs) is determined by a superposition of a static gradient
and excitation signals, both broad-band. Another is the
drive-field sequence itself, as particles that are not placed field and one or several oscillating excitation fields [1]. The
at the same spatial position, may react simultaneously static gradient field, called selection field, generates a low-
and are not separable by their signal phase or shape. field-region (LFR) in its center, that includes a field-free-region
To overcome this potential loss of information in spatial (FFR), which could either be a field-free-point (FFP) or field-
encoding for high amplitudes, a superposition of shifting
fields and drive-field rotations is proposed in this work. free-line (FFL), depending on its shape. On the one hand, the
Upon close view, a system matrix approach is capable to oscillating excitation field drives the SPIONs through their
maintain resolution, independent of the sequence, if the magnetization curve, causing higher harmonics due to their
response to pulsed sequences still encodes information nonlinear characteristic. On the other hand, it drives the LFR
within the phase. Data from an Arbitrary Waveform Magnetic through the imaging volume and creates a specific trajectory.
Particle Spectrometer with offsets in two spatial dimensions
is measured and calibrated to guarantee device indepen- Most scanner topologies use narrow-band sinusoidal signal
dence. Multiple sequence types and waveforms are com- shapes to excite the tracer material (see Table II in [2] for an
pared, based on frequency space image reconstruction from overview). Due to the superposition of the oscillating fields
emulated signals, that are derived from measured particle and the selection field, these encoding schemes cause the
responses. A resolution of 1.0 mT (0.8 mm for a gradient of SPIONs in the LFR to respond at a specific point in time with
their maximal amplitude. In frequency domain this results in
Manuscript received August 9, 2021; revised January 26, 2022; a spectral fingerprint depending on space, which can be used
accepted February 2, 2022. Date of publication February 7, 2022;
date of current version June 30, 2022. This work was supported in to reconstruct the image by solving a linear system of equa-
part by the German Research Foundation (DFG) under Grant GR tions [1], [3]. Another advantage of narrow-band excitation is
5287/2-1 and Grant KN 1108/7-1, in part by the Forschungszentrum the discrimination of the frequency space in the two domains,
Medizintechnik Hamburg (FMTHH) under Grant 01fmthh2018, and in
part by the Fraunhofer Research Institution for Individualized and Cell- the narrow-band excitation band and the higher receive band,
Based Medical Engineering (IMTE). The IMTE is supported in part by which contains the harmonics caused by the particles non-
the European Regional Development Fund (EFRE) and in part by the linearity. Due to this discrimination, a separation between the
State Schleswig-Holstein, Germany, through Project IMTE under Grant
124 20 002 and Grant LPW-E1.1.1/1536. Publishing fees supported by strong feed-through of the excitation field and the by 10−6
the Funding Programme Open Access Publishing of Hamburg University to 10−10 lower particle signal can be achieved, using resonant
of Technology (TUHH). (Corresponding author: Fabian Mohn.) passive filtering [4]. Current developments show that such
Fabian Mohn, Tobias Knopp, Marija Boberg, Florian Thieben,
and Patryk Szwargulski are with the Institute for Biomedical Imag- encoding schemes can provide sub-millimeter resolution [5],
ing, Hamburg University of Technology, 21073 Hamburg, Germany, more than 46 volumes per second time resolution [6] and
and also with the Section for Biomedical Imaging, University Med- pico-gram sensitivity [7]. Developments in instrumentation
ical Center Hamburg-Eppendorf, 20251 Hamburg, Germany (e-mail:
[email protected]; [email protected]; [email protected]; f.thieben@ now reach for clinical scale [8]–[10] to address specific needs,
uke.de; [email protected]). which are currently only partly addressed by conventional
Matthias Graeser is with the Institute for Biomedical Imaging, Hamburg imaging systems. Possible medical applications reach from
University of Technology, 21073 Hamburg, Germany, also with the
Section for Biomedical Imaging, University Medical Center Hamburg- catheter imaging [11] in digital subtraction angiography, over
Eppendorf, 20251 Hamburg, Germany, and also with the Fraunhofer stent quantification [12], stroke imaging [13]–[15] and many
Research Institute for Individualized and Cell-Based Medicine and the more. Using multi-contrast image reconstruction [16], it is also
Institute for Medical Engineering, University of Lübeck, 23562 Lübeck,
Germany (e-mail: [email protected]). possible to distinguish between different particle systems [17]
Digital Object Identifier 10.1109/TMI.2022.3149583 or physical parameters in the vicinity of the particle system

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1863

like temperature [18], viscosity [19], or binding state [20], which is used to measure a hybrid system matrix with 1D
[21]. With sinusoidal excitation, large single-core particles excitation and 2D offset fields [27]. This raw dataset is later
show a strong relaxation behaviour, which broadens the processed to generate the sequence specific particle response.
point spread function (PSF) and reduces the signal response, b) Transfer function correction: To be able to compare the
therefore reducing sensitivity and resolution [22]. Recently, results with other imaging systems, the data needs to be
Tay et al. proposed a rectangular excitation and showed that handled in a comparable physical variable, in this case the
it has the potential to improve the achievable resolution by domain of the magnetic moment m ([m] = Am2 ), which is
using large particles in combination with a new reconstruction achieved by correcting the received signals with the transfer
approach [23]. They showed that the effect of broadening function (TF) of the system [28]. In this domain, the signal
the PSF can be significantly reduced under certain condi- is independent of the receiver and individual properties of a
tions, using rectangular excitation with small amplitudes [23]. given system, like receive coils or amplification [29]. This
In idealized rectangular excitation sequences, a LFR would allows cross-platform comparison of the SPIONs response to
jump between two resting points in space, causing all particles a given excitation.
in between to react simultaneously. Tay et al. proposed a c) Sequence noise: To avoid correlation of intrinsic mea-
reconstruction scheme, that integrates the receive signal, which surement noise by reusing a subset of the same dataset within
in turn encodes the magnetic moment m of the area between the emulated sequence, the signal needs to be overlaid with
the LFR locations before and after the pulse, as long as m dominant, digitally generated noise ũ snoise(t).
reaches a steady state within the resting time between pulses. d) Reconstruction noise: A model to accurately represent
Consequently, this approach requires low field amplitudes receive chain noise of a scaled MPI system, consisting of
(1 to 3 mT), such that the LFR remains within a single voxel resonant coils, low noise amplifier (LNA) and analog-to-
and no loss in spatial resolution in the excitation direction is digital converter (ADC), based on reference measurements
induced. Otherwise, the information of the particle distribution from a 40 mm receive coil, named ũ rnoise(t) [30]. Therefore,
between two LFR positions would be lost. Furthermore, a low calculated images are comparable to those of a small scanner
excitation frequency is required in order to let the particles system.
fully relax to ensure their steady state [23]. e) Phantoms: In spite of correcting data by a TF, the
In this paper, multiple sequences based on different exci- measured data from the AWMPS has a high signal-to-noise
tation waveforms and amplitudes are compared in system ratio (SNR), due to the close proximity of the receive coils to
matrix phantom reconstructions. A sequence for rectangular the tracer and the large sample compared to a voxel volume
excitation shapes is proposed, that uses shifts and rotations of a system matrix calibration. While the close proximity is
simultaneously for better spatial encoding. In this proposed corrected by the TF, a realistic model for scaling the measured
sequence, the rectangular excitation is rotated and shifted iron mass to the emulated voxel grid is needed.
orthogonal to the excitation. Similar to a radon sampling f) Independent system matrices: Two sets of system matrices
scheme, this allows to reconstruct measured data with large need to be acquired to avoid inverse crime, when the phantom
amplitudes and an isotropic sampling trajectory in 2D, even spectrum stems from the identical dataset as the system matrix
if the phase information between the resting points is lost. used for reconstruction. For each sequence that is investigated
If particles relax fast enough to follow the slew rate of the drive in this paper, two independent system matrices on different
field, their phases within the signals become distinguishable grids are calculated, which are based on separately measured
which allows a system matrix approach to reconstruct images datasets, to guarantee a realistic reconstruction.
at high resolution. For pulsed sequences, the results show g) Comparability across sequences: In a final condition, the
improved sensitivity while the proposed sequence has uniform total sequence measurement time is chosen as the common
and high resolution with short acquisition times. criterion to compare sequences of different design.
Following these criteria, the images reconstructed by this
II. M OTIVATION AND C ONCEPT method represent realistic images, achievable by a well
All signals and reconstructions are emulated, meaning they designed pulsed MPI system.
are based on measurements of an arbitrary waveform mag-
netic particle spectrometer (AWMPS) [24]–[26], which are III. M ETHODS
processed to resemble a specific particle response for a defined
sequence. By exciting SPIONs, superimposed by a range of Different sequences are generated on various sets of raw
different DC-offsets in two spatial dimensions, their response data, to compare and identify the effects that excitation wave-
is mapped on a 2D grid. This resembles the response to form, amplitude or sampling trajectory have on the image SNR
an overlaying 2D gradient field [27]. This approach has the and resolution. This includes pulsed and sinusoidal excitation,
benefit to deliver reliable results for a first evaluation, without as well as low and high excitation amplitudes. Individual
the time consuming process of building an actual system that sequences and their sampling trajectories are explained in
can produce the required fields on a large scale. Following Section III-C. Instead of depending on a narrow PSF for
conditions should be met to reproduce the data: high resolution, the proposed method in this work uses higher
a) Hardware: A device that is capable of producing arbitrary excitation field strengths above 10 mT, as this benefits not
waveform excitations, superimposed by two orthogonal spatial only the sensitivity, but also reduces acquisition time in future
DC-offsets. See Section III-A for details on the AWMPS, imaging systems due to the capability of using less averages.
1864 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 7, JULY 2022

Fig. 2. Pulsed excitation measurements with a 20 µL sample of

Perimag 17 µgFe µL−1 . To the left, three offsets in excitation direction
are plotted (14.88 kHz), the reference channel from the current monitor
is on the top, the TF corrected receive signal on the bottom. Plots on the
right show identical relaxation behavior of two overlaid signals at 2.5 kHz
and 14.88 kHz. The slew rate of both frequencies is identical, only the
hold time differs. Both signal curves are coinciding for Perimag.

Fig. 1. Sectional view of the AWMPS and flowchart of the entire transmit channels and records the receive signal, as well as
process chain from raw data to image. Using gradiometric receive the amplifiers reference current monitor for control purposes.
coils, the receive signal can be isolated by suppressing direct excita-
tion feed-through. Fine-tuning is achieved by turning the gear-wheel.
A Helmholtz-coil creates an additional DC-offset field in another spatial B. Measurements
dimension (y-direction). The general process of the methodology is
shown in the block diagram. All measurements in this work are based on a multi-
core tracer, Perimag (micromod Partikeltechnologie GmbH,
To highlight the advantages and limitations between the differ- Rostock, Germany) with an undiluted iron concentration cmeas
ent methods, the Cartesian sequence proposed in [23] is com- of 17 mgFe mL−1 (304.4 mmolL−1). The delta sample is
pared with the proposed method in this work. For a detailed filled with Vmeas = 20µL of undiluted tracer to guarantee
comparison of the benefits of pulsed excitation, the proposed high measurement SNR at low amplitudes. The excitation
shift-radial sequence is run with both, rectangular excitation frequency is chosen to be 14.88 kHz, to ensure steep flanks
as well as sinusoidal excitation. A radial sine sequence is also of the rectangle excitation and avoid slew rate artifacts on the
included, as proposed by Knopp et al. in [31], which does not edges while keeping a good reception for induction sensors.
use the shift orthogonal to the excitation direction. The system This differs from the choice of 2.5 kHz in [23], however the
matrix reconstruction approach is chosen for all sequences [3]. rise time tr of the pulsed excitation (square wave, tr = 3µs)
For an overview of the image reconstruction pipeline in this lies in a similar region (2 to 5 µs). On the right side in
study, the process is depicted in Fig. 1 (c). Fig. 2 these two frequencies are compared, to confirm the
quick and identical relaxation behaviour of the used tracer for
A. Arbitrary Waveform Magnetic Particle Spectrometer both frequencies. We note that Perimag is a tracer with low
A non-resonant AWMPS with two transmit coils was built relaxation times, in contrast to the long relaxations times of
to perform measurements that contain the one-dimensional the tracers used in [23]. Therefore, our frequency choice is
excitation, superimposed by DC-offsets in two orthogonal only valid for fast relaxing particles and not a general choice
spatial dimensions. The main transmit coil combines excitation for larger core SPIONs. Instead of averaging 25 times [23],
signal and DC-offset in x-direction within a single cylindrical datasets in this work are all recorded with 3 averages. Data
coil, whereas a second transmit coil in Helmholtz configuration is acquired sequentially, each y-offset is held constant during
is responsible for the orthogonal DC-offset in y-direction with which a sweep of all x-offsets is performed. Frequency and
up to ±50 mT. The AWMPS may measure any waveform excitation amplitude stay constant until a dataset acquisition is
up to 45 mT in amplitude and is limited for arbitrary pulse completed. Measurements are background corrected, TF cor-
shapes by a final slew-rate due to load and amplifier char- rected for device independence and arranged by time samples
acteristics, around 10 mTµs−1 . Fig. 1 shows a picture and and spatial offsets. The two resulting raw data sets for each
the cross-section of the design. The excitation and x-offset sequence type form the basis for the sequence generation
coil is connected to a 4-quadrant-amplifier (Dr. Hubert GmbH, process, yielding a low resolution system matrix ( ŜLR with
Bochum, Germany). To avoid direct feed-through, two receive 0.67 mT steps in x- and y-direction) and a high resolution
coils are arranged in opposite orientation for decoupling of system matrix ( ŜHR with 0.5 mT steps in x- and y-direction),
the receive path from the transmission line. The position of respectively.
one coil is adjustable by a gear wheel for fine-tuning of the
feed-through cancellation. The signal is then amplified by C. Sequence Generation
a custom-build LNA. The data acquisition card STEMLab Three sequence types are presented in this paragraph,
125-14, Red Pitaya is used for signal generation on two each based on two independent sets of measured raw data,
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1865

TABLE I
OVERVIEW OF S EQUENCE D ETAILS AND M EASUREMENT PARAMETERS OF THE AWMPS R AW D ATA

to compose ŜLR and ŜHR for reconstruction. The parameters Finally, the voltage signal is Fourier transformed (denoted
for excitation amplitude and waveform depend on the sequence by hat) and TF corrected, yielding the system function

â(k) T
−i2π kt
version, as detailed in Table I.
The sequence itself is built from the time series data ũ(r, t) ŝ(r, k) = ũ(T (r, t), t) + ũ snoise (t) e T dt (2)
in a way, that a subset of the dataset, like a virtual FOV, T 0
is continuously shifted over the raw dataset by the LFR where r ∈ 2 is the position and k ∈ 0 the frequency
sequence. If the sequence contains a rotation R(ϑ(t)), it is component. Negative frequency components are omitted, due
applied consecutively with the shifting b(t), as described by to the symmetry of Fourier coefficients of real signals. T =
the rigid transformations J tmax is the sequence time length, J the number of appended
⎧ periods and tmax the time length of a single period. Each voxel
⎨
⎨ R(ϑ(t)) (r + b(t)) seq. type a) contains a time signal of J periods.
T (r, t) = R(ϑ(t)) r seq. type b) To visualize the sequence generation process, a magnitude
⎨
⎩
r + b(t) seq. type c) plot is used in Fig. 3, 4 and 5. In this plot, the absolute value of
    the raw data ũ(r, t) is summed up over a whole period for each
cos ϑ(t) − sin ϑ(t) b (t)
for R(ϑ(t)) = , b(t) = x offset value. In other words, the signal intensity is integrated
sin ϑ(t) cos ϑ(t) b y (t) to quantify the system response for discrete, equally spaced
where ϑ(t), bx (t) and b y (t) are piecewise constant functions measurement values ũ i = ũ(r, ti ) as in
for each period. The measurement noise within the trans- i 1/ p
 max
formed measurement voltage ũ(T (r, t), t) from the AWMPS, (ũ i )i  p = |ũ i | p
. (3)
would appear identically numerous times within the sequence, i=0
which can create reconstruction artifacts. To avoid correlation
In this work, the 1 -normwith p = 1 is selected and referred
of measurement noise throughout the sequence, digitally gen-
to as the integrated signal intensity (ISI). Thus, the ISI is
erated noise is added with a higher noise level of
represented by bright yellow colors for positions of strong
ũ snoise(t) = 10 σmeas ũ n,1/f(t) (1) signal response and dark blue colors represent low signal
response. It is noted that the ISI-plot is for visualization
which is crucial to mask the measurement noise floor of the purposes only and the 1 -norm is not implemented in the
raw dataset. σmeas is the standard deviation of the measurement sequence calculation. Within the virtual FOV, the ISI-plot
background noise and ũ n,1/f is the digitally generated noise, visualizes a moving (and rotating) system response. The final
that has been fitted to the shape of the measured background sequence is generated by appending all time signals. Noise
noise. The overall shape resembles 1/f noise, which is typ- addition, fast Fourier transform and TF correction are done
ical for integrated circuits in the low frequency region. The equivalently for all sequences.
addition of the tenfold standard deviation is based on an a) Shift-radial sequence: Proposed in this work is a
analysis of the correlation coefficient versus the addition of sequence with the properties of combining a shift, orthogonal
n · σmeas with n ∈ + , such that the noise in the generated to its excitation direction, with a rotation of the excitation
sequence is dominated by uncorrelated noise. In general, ũ itself, referred to as shift-radial sequence. The excitation
is the unprocessed voltage signal whereas the TF corrected direction in the raw data is in x-direction and the shift is
signal u = a ∗ ũ is in the domain of the magnetic moment. in y-direction. The shift-radial sequence requires twice the
1866 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 7, JULY 2022

Fig. 3. Sequence generation diagram of the proposed shift-radial sequence. Raw data is recorded with twice the offset values in y-direction
than in x (excitation direction) with pulsed excitation. The absolute values of ũ(r , t) are summed up for each offset value to represent an integrated
signal intensity (ISI) plot. To generate a sequence, a virtual FOV is shifted from positive to negative y-offsets, mimicking an orthogonal shifting field.
Simultaneously, it is continuously rotated to perform a rotation of the fields or the object. To avoid inverse crime, due to reusing the same dataset
over and over, noise is added to mask repeating measurement background noise and prevent correlation from within the sequence. The sequenced
data is fast Fourier transformed and TF corrected to yield a system matrix, which is used to reconstruct images or create phantom spectra. Refer to
Table I for specific sequence parameter choices, excitation waveform and other measurement details.

Fig. 4. Sequence generation diagram of the radial sequence. Similar to the shift-radial sequence, the radial sequence is a simplified version,
without the FOV shifting component. A rectangular subset is solely rotated, as proposed by Knopp et al. in [31]. In contrast to Fig. 3 and Fig. 5,
the ISI-plot shows the response to a sinusoidal excitation waveform (in x-direction). Otherwise, the system matrix is generated likewise, by noise
addition, FFT and TF correction. Refer to Table I for specific sequence parameter choices, excitation waveform and other measurement details.

oversampling of raw data along the y-direction with ±40 mT Computed Tomography (CT) and restores the lost information
than in x-direction with ±20 mT, to produce the necessary caused by large amplitudes [32]. The trajectory is plotted in
oversampling for the shift (see Fig. 3). A high excitation ampli- Fig. 6 (a). For a better understanding of the influence of
tude of 15 mT is chosen, which corresponds to a wide PSF the excitation waveform, this sequence is generated for both,
spanning almost the entire FOV. An LFR sequence controls a sinusoidal and a pulsed rectangular excitation.
the virtual FOV shift over the raw data to produce a shifted b) Radial sequence: The radial sequence is an adaptation
dataset with a given number of pulses per shift JPpS . These of the shift-radial sequence, which solely utilizes a rotation
are rotated with a given number of shifts per rotation JSpR of the raw data, without the virtual FOV shift component.
with respect to their center. Subsequently, shifts and rotations Hence, a square-shaped subset with ±20 mT offsets in x- and
are interweaved to generate a continuous acquisition of both y-direction is rotated with a given number of pulses per
in the time domain. In this work, JSpR = 32 with JPpS = 31 rotation JPpR around its center, imitating a rotating excitation
are chosen, amounting to J = JSpR · (JPpS · 2 − 2) = 1920 as proposed in Knopp et al. [31]. The excitation amplitude is
total periods. For each pulse the virtual FOV is shifted one the same as the shift-radial sequence with 15 mT, but the
step in y-direction, starting from −35 mT until +35 mT is waveform depicted in Fig. 4 differs, as the radial sequence
covered, where the direction inverses without sampling the is only generated on the base of sinusoidal raw data. The
turning-point nor the start-end-point twice (15 mT amplitude sequence trajectory is shown in Fig 6 (b).
of shifts, virtual FOV of 40 × 40 mT2 ). This sequence is c) Meander sequence: A Cartesian sequence is implemented
similar to sequences known from MPI-FFL encoding and with a discretization of 31 steps in a resolution of 1 mT per
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1867

Fig. 5. Sequence generation diagram of the meander sequence proposed in [23]. Measured raw data of 3 mT amplitude, with offsets in two
spatial dimension is over-sampled in order to be able to move a virtual FOV, which results in a smaller shifted dataset. The sequence name stems
from the nature of this LFR movement and the excitation is in x-direction only. The system matrix is generated identically as for all other sequences,
by noise addition, FFT and TF correction. Refer to Table I for specific parameter choices of the LFR sequence and other measurement details.

step in x- and y-direction, and is referred to as meander

sequence due to its LFR movement. By over-scanning the
required square-shaped FOV of ±20 mT, with a larger offset
of ±40 mT, the initial measurement yields enough raw data
to move the virtual FOV, as illustrated in Fig. 5. The virtual
FOV is moved along the meandering sequence, resulting in
a shifted dataset. The measurement data inside the shifted
dataset has been moved along the trajectory and was sequen- Fig. 6. Sequence trajectory plot. The black line typifies the focus-field
trajectory of the LFR, as transformed by T (r , t). Orange and blue depict
tially appended to represent the particle response to the entire exemplary drive field excitations at individual time points, which then
sequence. A version of the sequence is generated for each of deflect the LFR from this trajectory. The length of these bars correlates
the two drive field amplitudes, 3 mT and 15 mT (x-direction). with the excitation amplitude and ticks are drawn in 15 mT increments.
One side length equals 40 mT. Type (b) is not exposed to any offsets.
The low excitation amplitude is chosen to be 3 mT, in a way Type (c) and (d) have identical FOV boundaries. Image sizes and scales
to trade off SNR and spatial resolution as proposed by Tay are consistent throughout this work. For demonstration purposes, the
et. al [23]. In order to keep the FOV boundaries identical density of the trajectories is reduced.
between versions, the step resolution for 15 mT amplitude
is reduced to 7/31 mT in x, as illustrated in Fig. 6 (d). (N voxel), to a realistic value, in order to match the total iron
Both sequences consist of 31 steps in x- and 31 steps in mass within one voxel. αs together with a noise addition from a
y-direction, tracing the same path backwards without sampling real life receiver system, results in realistic signal spectra. In all
the turning-point nor the start-end-point twice, amounting to undiluted tracer experiments, a targeted tracer concentration
J = 31 · 31 · 2 − 2 = 1920 periods. of ctarget = 5 mgFe mL−1 was used throughout this work. The
The sequence types differ in the shape of their FOV, concentration vector c, which is used to create the phantom
depending on the fact if a rotation was applied. Rotating a spectrum voltage ûP for reconstruction, is expressed by
rectangular dataset necessarily results in a circular sampled
c = αs cP
area (see Fig. 3 and Fig. 4, far right), whereas for the meander-
Vtarget ctarget
sequence (Fig. 5) the FOV stays rectangular. To create com- = cP (4)
mon ground between different sequence types, the same time Vmeas cmeas
length to compare sequence types is chosen, as this reflects the where Vmeas is the volume and cmeas the concentration of the
acquisition time. For a generated sequence length of 129 ms measured delta sample in the AWMPS. An assumed linear
at 14.88 kHz, this amounts to 1920 periods for each sequence gradient field G, with det G = 0 in T m−1 , and a voxel
in this work. grid vector g in mT with 3 dimensions, yield the target
concentration Vtarget in
D. Phantoms 3 
 
 −1 
The intensity of the signal response of SPIONs scales Vtarget =  G g  = 64 nL, for
i
linearly with the concentration of the tracer [1], [33]. Conse- i=1
⎛ ⎞ ⎛ ⎞
quently, the system matrix signal based on the AWMPS needs −1.25 0 0 0.5
to be scaled down. One option is to introduce a scaling factor G=⎝ 0 −1.25 0 ⎠, g = ⎝0.5⎠ (5)
αs that scales the phantom concentration vector cP ∈ [0, 1] N 0 0 2.5 1
1868 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 7, JULY 2022

where i denotes the i -th entry of the vector. This gradient field
results in an exemplary scaling factor of αs = 9.4 · 10−4 , for
a volume of 20 µL and a concentration of 17 mgFe mL−1 as
described in Section III-B. Further experiments are performed
with a lower phantom concentration cP to create a dilution
series for a sensitivity analysis of all sequences types. Due to
αs , the change in m contributed by a voxel with concentration Fig. 7. Phantoms used in this study. Reference for original shape. Left
to right: Gap phantom with 2.0 mT (1.6 mm) gap in x-direction, vessel
cP matches the real life magnetic moment change. phantom with stenosis in the left branch, large phantom to reveal borders
A second step involves a noise model to represent realistic of FOV and fine resolution phantom with 1.0 mT (0.8 mm) gap.
MPI receiver noise during reconstruction. Real background
noise measurements from a pre-clinical MPI scanner (Bruker,
Ettlingen, Germany), utilizing an optimized 42 mm coil [30], in practice leads to better spatial resolutions than the width
are used to select a noise level of nr = 70 · 10−15 mA2 for of the derivative of the particle magnetization curve would
the highest frequency. Using a 1/f function to model the noise predict [36].
curve results in a higher noise for every frequency compared to As this work contains emulated gradients only, distances
the measured reference in [14], guaranteeing a realistic SNR in phantoms are best defined in mT. However, to give the
scenario. The TF corrected noise level added during image reader a better classification of the results, the corresponding
reconstruction is dimensions for a gradient as in (5) are given in brackets e.g.
nr ζ 2.0 mT (1.6 mm) in x- and y-direction. This same gradient is
ûrnoise = T ∈ K (6) used for all emulations through-out this work. All phantoms
2π Tf,max
are oriented in x y-plane. Resulting images are reconstructed
with ζ ∼ CN (0, 1) a standard complex normal random with the low resolution system matrix ŜLR , yielding an image
variable with zero mean and variance of 1. T ∈ + K
contains resolution of 61 by 61 pixel, corresponding to a side length
the period time for each of the K frequency components and of 40.0 mT (24.4 mm). Assuming the gradient field from
Tf,max is the period length of the highest frequency in the (5), the fine system matrix ŜHR , that is used to constitute
receive band. ûrnoise will overlay the scaled phantom spectrum the phantom spectrum, accounts for an image resolution of
ûP , as visualized in Fig. 8, during the image reconstruction 0.5 mT/pixel · (1.25 T/m)−1 = 0.4 mm/pixel. System matrices
process. In Fig. 7, four phantoms are shown for original are denoised to reduce the impact of system matrix noise on
reference. the reconstructed image [37].

E. Image Reconstruction
F. Implementation
Images are reconstructed using the iterative Kaczmarz
method [34], which gives the calculated particle concentration Raw data measurements in the AWMPS are controlled by a
c∈ + N
for N voxel, solving the linear system of equations system software implemented in the open source programming
language Julia [38]. It controls the parameter sweep through
Ŝ c = ûSNR (7) the chosen x-offset range and gradually steps up the y-offset
for each excitation combination. The raw data is stored in an
where ûSNR ∈ K
is a filtered subset of the emulated
extended version of the Magnetic Particle Imaging data format
complex phantom voltage ûP with K frequency components
(MDF) [39] and the sequence is afterwards calculated within
and Ŝ ∈ K ×N is the system matrix in the frequency domain.
a custom simulation framework developed in Julia. System
For reconstruction the following least squares problem can be
matrix reconstructions are based on the MPI reconstruction
solved
 2 package developed in [40], accessible under [41].
 
cλreco = argmin  Ŝ c − ûSNR  + λ c22 (8)
N 2
c∈ + IV. R ESULTS
 2 Results are summarized in two main figures with phantom
 
where  Ŝ c − ûSNR  is the data discrepancy term and c22 image reconstructions, one focusing on sensitivity by means
2
is the penalization term used to dampen large oscillations of a dilution series (Fig. 9), the other on resolution and FOV
in the solution. cλreco is the solution of the concentration shapes (Fig. 10). Reconstruction parameters are individually
distribution, controlled by a relative regularization parameter determined to match the image noise across the different
λ ∈ + , that blurs the image noise at the cost of spatial reconstructions.
resolution. With increasing dilution, the SNR threshold is set Fig. 9 shows a dilution series of two phantoms to visually
higher and both, λ and the number of iterations, increase. compare the results of different excitation waveforms, ampli-
Additionally, the SNR selection PSNR is subjected to a minimal tudes and sequences, with an undiluted iron concentration of
frequency threshold, acting as a high-pass, to filter out any κ0 = 5 mgFe mL−1 . The left part of the figure depicts a
signal before the first harmonic of the fundamental frequency. simple two bar phantom (gap in x-direction), the right part
A reconstruction diagram is shown in Fig. 8, to visualize a more intricate vessel phantom with a stenosis in the left
necessary steps. Handling noise amplification in images is branch (see Fig. 7 for originals). For undiluted reconstructions,
described in detail in [35]. We note that the solution of (8) the shift-radial sequences and the meander sequence with
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1869

Fig. 8. Diagram of the implemented reconstruction process. A vessel phantom with a visible stenosis is scaled to represent a tracer distribution
c with a concentration of 5 mgFe mL−1 and a selectable dilution. The high resolution system matrix Ŝ HR multiplied by the phantom vector c results
in the phantom spectrum ûP , which is overlaid by a realistic measurement noise level û rnoise . After frequency selection (PSNR ), the low resolution
matrix Ŝ LR is used to reconstruct the image using the iterative Kaczmarz method with Tikhonov regularization. All images span 40 × 40 mT2 .

high excitation amplitude perform very similar, whereas the As this rectangle is rotated, only a circle with the diameter of
low amplitude meander sequence suffers from a low SNR. one side length is fully covered. As shown in Fig. 10, the area
Fine structures are resolved independent of the excitation in between the inner circle with diameter of a side length and
waveform. This effect becomes more visible with the intricacy a circle with a diameter of a diagonal length, is only occasion-
of the vessel phantom on the right hand side in Fig. 9. ally covered by the corner of the rectangular sampling area.
In undiluted reconstructions of the radial sequence, significant Thus, the resolution decreases as sampling becomes sparse.
blurring appears towards the edges of the FOV with the The radial sequence shows a very fine resolution in the center
exception of a small region in the center, where resolution of the FOV, which deteriorates continuously with increasing
is excellent. At a dilution of 78 µgFe mL−1 (1:64, κ2 ), the radius. As the excitation is rotated without shifting, the center
pulsed 3 mT meander sequence starts to lose the ability to of the FOV is sampled by every pulse, which in turn results
resolve any fine structures and for dilutions of 1.2 µgFe mL−1 in a fine local resolution. This comes with the drawback of
(1:4096, κ3 ) it stops to resemble even the rough shape of losing precision towards the edges of the FOV. A bar-phantom
the original phantom. However, the pulsed 15 mT meander with a 1.0 mT (0.8 mm) gap in y-direction is shown in the
sequence is able to display basic phantom features down to center row of Fig. 10, with the associated intensity profile
the highest dilution. The shift-radial sequence is able to depict over y in the bottom row. An average of three pixel left
phantom outlines of simple shapes at κ3 for both, pulsed and right of the center line is taken to calculate the intensity
and sinusoidal excitation. The pulsed excitation for extremely plot. The reconstruction based on the meander sequence is
high dilutions of 153 ngFe mL−1 (1:32768, κ4 ) seems to have not able to resolve the two bars in both cases, whereas
a slight advantage over the sinusoidal excitation, being still shift-radial and radial sequences are able to do so. This can
able to marginally reconstruct a simple phantom whereas the be explained by the limitation of Cartesian sequences to be
sinusoidal excitation is entirely dominated by system noise. an-isotropic in their nature, as the excitation is oriented solely
When comparing pulsed sequences with 15 mT amplitude, in one direction. In this work, x is chosen for excitation and
edges are traced more accurately by the pulsed shift-radial therefore a better resolution in x is obtained than in y, as also
sequence than the meander sequence at dilution κ3 for the left apparent in the an-isotropic FOV plot in the top row for
phantom (row 2 and 3). The sinusoidal radial sequence is able the meander sequence, independent of using small or large
to resolve most structures at κ2 , especially around its center, amplitude.
however it becomes heavily distorted beyond κ3 . In Fig. 11 the impact of changing a sequence parameter
In Fig. 10, the characteristic FOVs of all sequences are is shown by comparing shift-radial sequences with different
visualized using a phantom that extends over the full image trajectory densities. On the left, the identical system matrix
width (40 mT). For the meander sequence, the FOV spans to that was used throughout this work for sinusoidal shift-radial
the out-most LFR boundaries (36 × 30 mT2 ). Shift-radial is reconstructions with JSpR = 32 shifts per rotation was imple-
confined with full resolution to the radius of its shift
√ amplitude mented. On the right side, a sequence was generated with
(inner circle, 15 mT), although this extends with 2 to a sec- JSpR = 120 that has 7200 periods instead of 1920. All other
ond region of slightly degraded reconstruction ability (dashed, parameters and reconstructions are identical. Results indicate
white line). The shift in excitation, which can be described that for a good representation of edges and corners a high
as a pulse line, results in a rectangular shaped sampling area. density is advantageous.
1870 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 7, JULY 2022

Fig. 9. Dilution series emulation for two excitation waveforms and three different sequence types. Iron concentration κ0 of 5 mgFe mL−1 ,
which is diluted from left to right, for two different phantoms. On the left, the phantom consists of two rectangles with a gap of 2.0 mT (1.6 mm) in
x-direction, whereas the right side depicts a vessel phantom with a stenosis towards the left branch (originals in Fig. 7). Reconstruction parameters
are individually determined, to yield comparable noise in the image domain. System matrices are denoised and the intensity is normalized to 1 within
each image. Highlighted in orange are reconstructions of similar quality, produced by the identical sequence but differing in excitation waveform and
dilution. All images span 40 × 40 mT2 (24.4 × 24.4 mm2 ), sequence trajectories are shown to scale in Fig. 6.

In a final result in Fig. 12, the effect that phase information bear a slight advantage in sensitivity compared to sinusoidal
from within our measured data has on resolution, is isolated by excitation for the identical shift-radial sequence, if the two
creating an immobile sequence without any shifted or rotated images marked in orange in Fig. 9 are considered recon-
components in the trajectory. This can be done by setting the structions of the same quality. Note that a factor of 103
shift amplitude and step resolution to zero and the repetitions distinguishes column 3 from 4. The basis of this advantage
to 1920 periods, so the time length is identical to all other lies in the simultaneous response of many spatial positions,
sequences. The lack of any focus-field movement implies that resulting in a better SNR. The magnitude of this advantage
the LFR is solely moved by the excitation waveform along a becomes visible at higher dilutions only, and it depends on
single line in x, resulting in a 30 mT FOV line (dashed, white the phantom or on the medical application. The more homo-
line). Shown in Fig. 12 are reconstructions based on pulsed geneous the distribution of SPIONs (e.g. perfusion imaging),
and sinusoidal data, which are both capable of restoring a the higher seems the advantage of pulsed excitation in contrast
simple pattern in the FOV. The elliptic outline of this area is to the accumulation of tracer material in a bulk (e.g. stem cell
shaped similar to the outline of the ISI plots of raw data in tracking).
Fig. 3 and 4 for pulsed and sinusoidal data respectively. The drive field frequency was chosen to be 14.88 kHz as
this is the upper limitation of the amplifier to produce a
V. D ISCUSSION clean waveform with rise-times around 3 µs. The relaxation
By combining a shifting focus field with a rotating drive behaviour of the used tracer (Perimag, Fig. 2) allows the
field of large amplitude, the proposed shift-radial sequence use of 14.88 kHz in this work, due to its quick relaxation.
overcomes previous limitations in amplitude and acquisition Two points need to be clarified regarding this choice: First,
time for pulsed rectangular excitations. Shift-radial sequences tracers need to be asserted individually and measurement
provide an alternative to Cartesian sampling schemes, due to parameters such as frequency and rise time would need to
their ability to resolve 1.0 mT gaps (0.8 mm for a gradient be adapted for larger particles to match their relaxation.
of (−1.25, −1.25, 2.5) Tm−1 ) and their spatially isotropic Tay et al. proposed the pulsed excitation scheme for large
sampling trajectory. Furthermore, pulsed excitation seems to particles beyond a specific size of about 28 nm [23], [42]
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1871

Fig. 10. Resolution simulation and FOV comparison. The top row compares the reconstruction of a 5 by 5 squares phantom, separated by 2.0 mT
(1.6 mm) gaps to distinguish the boundaries of the respective FOVs (white lines). The phantom is shown in the rightmost column for reference.
The shape of the FOV depends on the sequence, resulting in a circular FOV for radial sequence types. In the center row, a phantom with a 1.0 mT
(0.8 mm) gap in y-direction is shown, with the corresponding intensity profile in the bottom row. The iron concentration is 5 mgFe mL−1 each and
the images span 40 × 40 mT2 (24.4 × 24.4 mm2 ).

Fig. 11. Trajectory density comparison. The image to the left stems
from the same system matrix for the sinusoidal shift-radial sequence
throughout this work, with the sequence parameter of JSpR = 32 shifts
per rotation. The image to the right has a denser trajectory with JSpR =
120, consequently the sequence is longer with 7200 periods instead of
1920. Image reconstruction parameters are identical. Fig. 12. Comparison of pulsed and sinusoidal data without a
sequence trajectory. In order to reveal the effect of information encoded
within the phase of the measured raw data for different excitation
waveforms, an immobile sequence is generated that does not use shifts
to circumvent limiting relaxation effects and improve reso- or rotations of any kind (also 1920 periods). Consequently, the LFR is
lution. The so called “relaxation wall” blurs the MPI image, only moved by the drive field (in x) and the resulting FOV is a single line
with 15 mT amplitude (white, dashed). The images span 40 × 40 mT2 .
as particles do not follow the Langevin model [22], and the Below, the intensity profile along this line in x is plotted.
optimum predicted resolution cannot be achieved [43], [44].
One precondition in order to reach optimum resolution for any
particle system is to ensure that the particle magnetization sigmoid-shaped excitation, which differs from an idealized
is sufficiently saturated to generate a strong signal, which rectangle, a small time-shift is observable in the raw data,
sets an upper limit to the excitation frequency. The second as shown in Fig. 2 on the left. Therefore, the phase information
point is that due to the quick relaxation of Perimag and the of the signal is distinguishable within the measured raw
1872 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 41, NO. 7, JULY 2022

data, despite our assumption that all phase information is For sinusoidal excitation this leads to an attenuation of the
lost by pulsed excitation. Due to this fact, the following signal at the FOV border, which is lower for pulsed sequences.
argument needs to be considered: Results in Fig. 9, 10 and This attenuation may help avoiding artifacts at the hard cut
11 show the combined effect of pulsed excitation including off at the end of the circular FOV for the presented radial-
phase and the effect of the sequence trajectory for system shift sequences. Sinusoidal drive fields performed equally in
matrix reconstruction. A separation of individual contributions almost all regards throughout this study, hence the effort for a
is not possible with our measurement data. Ultimately, system pulsed non-resonant scanner may only be reasonable for larger
matrix reconstruction of pulsed sequences with high amplitude particle systems. Further research on the application of the
combine the benefit in the SNR and profit from any phase proposed method to large single-core particles has to be done.
information within the particle response. This holds true for Improvements may be limited by peripheral nerve stimulation
meander as well as shift-radial sequences and explains why and the specific absorption rate, reducing the advantage of high
a meander sequence with 15 mT amplitude performs similar drive field amplitudes for in-vivo imaging in large animals or
compared to the pulsed shift-radial sequence (Fig. 9, rows 2 humans. Also, the optimum frequency-amplitude combination
and 3) for our measurement data and using system matrix for shift-radial and the choice of sequence parameters to trade
reconstruction. Shift-radial sequences profit from phase infor- off resolution and acquisition speed need to be investigated.
mation, but they would be capable to restore basic phantom Although rectangular excitation patterns did not outperform
features even without any phase information, similar to a the sinusoidal ones for medium-sized MPI particles in this
Radon transform [32]. For an ideal rectangular excitation (no work, they certainly have the potential to improve MPI not
phase discrimination) Cartesian meander sequences depend on only for large particles. By allowing to vary the excitation
a small excitation amplitude for achieving high resolution. For frequency as an adjustable hardware parameter, one can tailor
this reason a low amplitude (1 mT) was chosen in previous the excitation pattern to the currently used particle system,
works [23] to guarantee a small sampling kernel and sustain giving maximum flexibility. Potentially, this allows optimiz-
high resolution, with the draw back of a low SNR. ing the contrast in multi-contrast MPI applications, e.g. for
The sole benefit due to phase information for pulsed and discriminating different particle sizes [17].
sinusoidal raw data can be identified in Fig. 12. Surprisingly,
basic phantom features are resolved by the reconstruction R EFERENCES
based on pulsed raw data, which can only be attributed [1] B. Gleich and J. Weizenecker, “Tomographic imaging using the non-
to phase information as there is no additional information linear response of magnetic particles,” Nature, vol. 435, no. 7046,
provided by means of a sequence trajectory. Edges of the pp. 1214–1217, 2005.
[2] N. Panagiotopoulos et al., “Magnetic particle imaging: Current develop-
squares along the white FOV line are slightly better shaped ments and future directions,” Int. J. Nanomed., vol. 10, pp. 3097–3114,
for sinusoidal data, when the intensity profiles along the Apr. 2015.
x-center line are compared. However, the pulsed reconstruc- [3] T. Knopp et al., “Weighted iterative reconstruction for magnetic particle
imaging,” Phys. Med. Biol., vol. 55, no. 6, pp. 1577–1589, 2010.
tion clearly separates all 5 squares, relying only on the [4] M. Graeser, T. Knopp, M. Grüttner, T. F. Sattel, and T. M. Buzug,
small time-shift that is induced in the raw signal, as shown “Analog receive signal processing for magnetic particle imaging,” Med.
in Fig. 2 on the left. If this holds true for larger particle Phys., vol. 40, no. 4, p. 42303, 2013.
[5] P. Vogel et al., “Micro-traveling wave magnetic particle imaging—Sub-
systems remains to be proven. As a consequence, system millimeter resolution with optimized tracer LS-008,” IEEE Trans. Magn.,
matrix reconstruction may be able to sustain basic resolution vol. 55, no. 10, pp. 1–7, Oct. 2019.
even for large single-core particles, if the Nyquist criterion is [6] J. Weizenecker, B. Gleich, J. Rahmer, H. Dahnke, and J. Borgert, “Three-
dimensional real-time in vivo magnetic particle imaging,” Phys. Med.
not violated during waveform slopes (rise-time). Resolution Biol., vol. 54, no. 5, pp. L1–L10, 2009.
improvements of pulsed excitation compared to sinusoidal [7] M. Graeser et al., “Design of a head coil for high resolution mouse
excitation were neither observed nor expected for the given brain perfusion imaging using magnetic particle imaging,” Phys. Med.
Biol., vol. 65, no. 23, Dec. 2020, Art. no. 235007, doi: 10.1088/1361-
tracer and measurement parameters. However, the proposed 6560/abc09e.
shift-radial imaging sequence and reconstruction method does [8] M. Graeser et al., “Human-sized magnetic particle imaging for brain
not make any assumptions on the particle size and thus we applications,” Nature Commun., vol. 10, no. 1, Dec. 2019, Art. no. 1936.
[9] E. E. Mason, C. Z. Cooley, S. F. Cauley, M. A. Griswold, S. M. Conolly,
predict that it generalizes to different relaxations and larger and L. L. Wald, “Design analysis of an MPI human functional brain
particles, if individual particle dynamics are heeded and an scanner,” Int. J. Magn. Part. Imag., vol. 3, no. 1, 2017, Art. no. 1703008.
equilibrium state is reached. [10] J. Rahmer, C. Stehning, and B. Gleich, “Remote magnetic actuation
using a clinical scale system,” PLoS ONE, vol. 13, no. 3, Mar. 2018,
Concerning our implemented methods, the noise added Art. no. e0193546.
during sequence generation, as well as the noise added during [11] J. Haegele et al., “Toward cardiovascular interventions guided by mag-
image reconstruction, are overestimated and higher than in a netic particle imaging: First instrument characterization,” Magn. Reson.
Med., vol. 69, no. 6, pp. 1761–1767, Jun. 2013.
well designed scanner system. While rectangular excitation [12] F. Wegner et al., “Magnetic particle imaging: Artifact-free metallic stent
can be superior for large particles, the results in this work lumen imaging in a phantom study,” CardioVascular Intervent. Radiol.,
could not assert a clear advantage for standard MPI tracers vol. 43, no. 2, pp. 331–338, Oct. 2019.
[13] P. Szwargulski et al., “Monitoring intracranial cerebral hemorrhage using
with a size lower than the relaxation wall. In fact, sinusoidal multicontrast real-time magnetic particle imaging,” ACS Nano, vol. 14,
excitation seems superior at high concentrations, as it is no. 10, pp. 13913–13923, Oct. 2020, doi: 10.1021/acsnano.0c06326.
less prone to artifacts at the edges of the FOV. The signal [14] M. Graeser et al., “Design of a head coil for high resolution mouse
brain perfusion imaging using magnetic particle imaging,” Phys. Med.
intensity for each spatial position is linked to the derivative Biol., vol. 65, no. 23, Dec. 2020, Art. no. 235007, doi: 10.1088/1361-
of the excitation field at the time the LFR passes [45]. 6560/abc09e.
MOHN et al.: SYSTEM MATRIX BASED RECONSTRUCTION FOR PULSED SEQUENCES IN MAGNETIC PARTICLE IMAGING 1873

[15] P. Ludewig et al., “Magnetic particle imaging for real-time perfusion [30] M. Graeser et al., “Towards picogram detection of superparamagnetic
imaging in acute stroke,” ACS Nano, vol. 11, no. 10, pp. 10480–10488, iron-oxide particles using a gradiometric receive coil,” Sci. Rep., vol. 7,
Oct. 2017. no. 1, p. 6872, Jul. 2017.
[16] J. Rahmer, A. Halkola, B. Gleich, I. Schmale, and J. Borgert, “First [31] T. Knopp et al., “Trajectory analysis for magnetic particle imaging,”
experimental evidence of the feasibility of multi-color magnetic particle Phys. Med. Biol., vol. 54, no. 2, pp. 385–397, Dec. 2008.
imaging,” Phys. Med. Biol., vol. 60, no. 5, p. 1775, 2015. [32] T. Knopp, M. Erbe, T. F. Sattel, S. Biederer, and T. M. Buzug, “A Fourier
[17] C. Shasha, E. Teeman, K. M. Krishnan, P. Szwargulski, T. Knopp, and slice theorem for magnetic particle imaging using a field-free line,”
M. Möddel, “Discriminating nanoparticle core size using multi-contrast Inverse Problems, vol. 27, no. 9, p. 95004, 2011.
MPI,” Phys. Med. Biol., vol. 64, no. 7, p. 74001, 2019. [33] K. Lu, P. W. Goodwill, E. U. Saritas, B. Zheng, and S. M. Conolly, “Lin-
[18] C. Stehning, B. Gleich, and J. Rahmer, “Simultaneous magnetic particle earity and shift invariance for quantitative magnetic particle imaging,”
imaging (MPI) and temperature mapping using multi-color MPI,” Int. J. IEEE Trans. Med. Imag., vol. 32, no. 9, pp. 1565–1575, Sep. 2013.
Magn. Imag., vol. 2, no. 2, pp. 1–6, Dec. 2016. [34] S. Kaczmarz, “Angenäherte Auflösung von systemen linearer
[19] M. Möddel, C. Meins, J. Dieckhoff, and T. Knopp, “Viscosity quantifi- Gleichungen,” Bull. Int. Academie Polonaise Sci. et des Lettres,
cation using multi-contrast magnetic particle imaging,” New J. Phys., vol. 35, pp. 355–357, Oct. 1937.
vol. 20, no. 8, p. 83001, Aug. 2018. [35] M. Boberg, N. Gdaniec, P. Szwargulski, F. Werner, M. Möddel, and
[20] M. Möddel, F. Griese, T. Kluth, and T. Knopp, “Estimating orientation T. Knopp, “Simultaneous imaging of widely differing particle concentra-
using multi-contrast MPI,” Int. J. Magn. Part. Imag., vol. 6, no. 2, tions in MPI: Problem statement and algorithmic proposal for improve-
pp. 1–3, 2020. ment,” Phys. Med. Biol., vol. 66, no. 9, Apr. 2021, Art. no. 095004, doi:
[21] T. Viereck, C. Kuhlmann, S. Draack, M. Schilling, and 10.1088/1361-6560/abf202.
F. Ludwig, “Dual-frequency magnetic particle imaging of the [36] T. Knopp, S. Biederer, T. F. Sattel, M. Erbe, and T. M. Buzug,
Brownian particle contribution,” J. Magn. Magn. Mater, vol. 427, “Prediction of the spatial resolution of magnetic particle imaging using
pp. 156–161, Apr. 2017. [Online]. Available: http://www.sciencedirect. the modulation transfer function of the imaging process,” IEEE Trans.
com/science/article/pii/S0304885316329031 Med. Imag., vol. 30, no. 6, pp. 1284–1292, Jun. 2011.
[22] Z. W. Tay, D. W. Hensley, E. C. Vreeland, B. Zheng, and S. M. Conolly, [37] A. Weber, J. Weizenecker, U. Heinen, M. Heidenreich, and T. M. Buzug,
“The relaxation wall: Experimental limits to improving MPI spatial “Reconstruction enhancement by denoising the magnetic particle imag-
resolution by increasing nanoparticle core size,” Biomed. Phys. Eng. ing system matrix using frequency domain filter,” IEEE Trans. Magn.,
Exp., vol. 3, no. 3, May 2017, Art. no. 035003, doi: 10.1088/2057- vol. 51, no. 2, pp. 1–5, Feb. 2015.
1976/aa6ab6. [38] J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, “Julia: A fresh
[23] Z. W. Tay et al., “Pulsed excitation in magnetic particle imaging,” IEEE approach to numerical computing,” SIAM Rev., vol. 59, no. 1, pp. 65–98,
Trans. Med. Imag., vol. 38, no. 10, pp. 2389–2399, Oct. 2019. 2017.
[24] Z. W. Tay, P. W. Goodwill, D. W. Hensley, L. A. Taylor, B. Zheng, and [39] T. Knopp et al., “MDF: Magnetic particle imaging data format,” 2016,
S. M. Conolly, “A high-throughput, arbitrary-waveform, MPI spectrom- p. 9, arXiv:1602.06072.
eter and relaxometer for comprehensive magnetic particle optimization [40] T. Knopp, P. Szwargulski, F. Griese, M. Grosser, M. Boberg, and
and characterization,” Sci. Rep., vol. 6, no. 1, Sep. 2016, Art. no. 34180. M. Möddel, “MPIReco.jl: Julia package for image reconstruction in
[25] D. Pantke, N. Holle, A. Mogarkar, M. Straub, and V. Schulz, “Mul- MPI,” Int. J. Magn. Part. Imag., vol. 5, no. 1, 2019, Art. no. 1907001.
tifrequency magnetic particle imaging enabled by a combined passive [41] MPIReco.jl: Julia Package for Image Reconstruction in MPI.
and active drive field feed-through compensation approach,” Med. Phys., Accessed: Jan. 10, 2021. [Online]. Available: https://github.com/
vol. 46, no. 9, pp. 4077–4086, Jul. 2019, doi: 10.1002/mp.13650. MagneticParticleImaging/MPIReco.jl.git
[26] C. B. Top, “An arbitrary waveform magnetic nanoparticle relaxometer [42] C. Shasha, E. Teeman, and K. M. Krishnan, “Nanoparticle core size
with an asymmetrical three-section gradiometric receive coil,” TURK- optimization for magnetic particle imaging,” Biomed. Phys. Eng. Exp.,
ISH J. Electr. Eng. Comput. Sci., vol. 28, no. 3, pp. 1344–1354, vol. 5, no. 5, pp. 1548–1551, 2019.
May 2020. [43] R. M. Ferguson, K. R. Minard, and K. M. Krishnan, “Optimization
[27] A. von Gladiss, M. Graeser, P. Szwargulski, T. Knopp, and T. M. of nanoparticle core size for magnetic particle imaging,” J. Magn.
Buzug, “Hybrid system calibration for multidimensional magnetic Magn. Mater., vol. 321, no. 10, pp. 1548–1551, May 2009, doi:
particle imaging,” Phys. Med. Biol., vol. 62, no. 9, pp. 3392–3406, 10.1016/j.jmmm.2009.02.083.
Apr. 2017. [44] J. Weizenecker, B. Gleich, J. Rahmer, and J. Borgert, “Micro-magnetic
[28] A. von Gladiss, M. Graeser, A. Behrends, X. Chen, and T. M. Buzug, simulation study on the magnetic particle imaging performance of
“Efficient hybrid 3D system calibration for magnetic particle imaging anisotropic mono-domain particles,” Phys. Med. Biol., vol. 57, no. 22,
systems using a dedicated device,” Sci. Rep., vol. 10, no. 1, pp. 1–12, pp. 7317–7327, Nov. 2012.
Oct. 2020. [45] P. W. Goodwill and S. M. Conolly, “The X-space formulation of the
[29] T. Knopp et al., “Model-based reconstruction for magnetic parti- magnetic particle imaging process: 1-D signal, resolution, bandwidth,
cle imaging,” IEEE Trans. Med. Imag., vol. 29, no. 1, pp. 12–18, SNR, SAR, and magnetostimulation,” IEEE Trans. Med. Imag., vol. 29,
May 2010. no. 11, pp. 1851–1859, Nov. 2010.