IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO.

11, NOVEMBER 2010 1851

The X-Space Formulation of the Magnetic Particle

Imaging Process: 1-D Signal, Resolution, Bandwidth,
SNR, SAR, and Magnetostimulation
Patrick W. Goodwill* and Steven M. Conolly

Abstract—The magnetic particle imaging (MPI) imaging process

is a new method of medical imaging with great promise. In this
paper we derive the 1-D MPI signal, resolution, bandwidth require-
ments, signal-to-noise ratio (SNR), specific absorption rate, and
slew rate limitations. We conclude with experimental data mea-
suring the point spread function for commercially available SPIO
nanoparticles and a demonstration of the principles behind 1-D
imaging using a static offset field.
Despite arising from the nonlinear temporal response of a
magnetic nanoparticle to a changing magnetic field, the imaging
process is linear in the magnetization distribution and can be de-
scribed as a convolution. Reconstruction in one dimension is exact
and has a well-behaved quasi-Lorentzian point spread function.
Fig. 1. (Left) The magnetization of the system when the FFP is at location x
The spatial resolution improves cubically with increasing diameter and SPIO nanoparticles positioned at the small circles. The magnetization is
of the SPIO domain, inverse to absolute temperature, linearly shown for one particle and with two particles. (Right) Signal produced by the
with saturation magnetization, and inversely with gradient. The magnetization when rapidly scanning the FFP back and forth with trajectory
bandwidth requirements approach a megahertz for reasonable x (t). The signal is shown graphed against the position of the FFP. The signal
imaging parameters and millimeter scale resolutions, and the SNR changes in sign when the FFP is scanning in the opposite direction.
increases with the scanning rate. The limit to SNR as we scale MPI
to human sizes will be patient heating. SAR and magnetostimula-
tion limits give us surprising relations between optimal scanning In tandem with progress in building new hardware, there have
speeds and scanning frequency for different types of scanners. been efforts to develop a working theory on the MPI process and
Index Terms—Biomedical imaging, magnetic particle imaging, how to rapidly reconstruct images in a linear, repeatable manner.
signal detection, x-space. Excellent progress towards this goal is shown in Rahmer et al. [4],
where it is shown that the frequency-space signal contribution of a
I. BACKGROUND magnetic particle when using a linear or Lissajous excitation pat-
tern can be described using Chebychev polynomials of the second
kind, which form an excellent basis set for reconstruction. This
AGNETIC particle imaging is a new medical imaging
M modality that holds significant promise for high-sensi-
tivity, high-resolution imaging in small animals and humans.
shows that MPI excitation gives a readily invertible signal that can
be converted to an image through algebraic reconstruction tech-
niques or a Chebychev basis set. Simulation based approaches are
Recent papers show real time imaging of tracer at physiologic discussed in [2] and [5]. However, there is not yet a unified theory
concentrations being passed between the chambers of a mouse describing the MPI signal, resolution, bandwidth, signal-to-noise
heart [1], methods for single-sided magnetic particle imaging ratio (SNR), SAR, and magnetostimulation.
(MPI) that do not require surrounding the patient or animal with In this paper, we will show that MPI can be understood as a
a gradient magnet [2], and novel methods for improving receiver x-domain scanning process and, as such, reconstructed quickly
coil matching with inter-modulation excitation [3]. and simply. We begin by approaching MPI as a 1-D system
Manuscript received March 23, 2010; revised May 29, 2010; accepted May and solve for the point spread function (PSF). The PSF gives us
31, 2010. Date of publication June 07, 2010; date of current version November bandwidth requirements, which we use to derive the signal-to-
03, 2010. This work was supported in part by the California Institute for Regen-
erative Medicine (CIRM) graduate fellowship under Training Grant T1-00007,
noise ratio. We then separately discuss SAR and magnetostim-
in part by CIRM Tools and Technology Grant RT1-01055-1, in part by a Univer- ulation, which give surprising results regarding excitation slew
sity of California Berkeley Bioengineering graduate fellowship, and in part by rates and excitation frequencies. We conclude with a brief test
the National Institutes of Health training grant. The contents of this publication
are solely the responsibility of the authors and do not necessarily represent the
of reconstruction and system linearity with a real instrument.
official views of CIRM or any other agency of the State of California. Asterisk
indicates corresponding author.
*P. W. Goodwill is with the Department of Bioengineering, University of II. FUNDAMENTAL RELATIONS OF MPI IN ONE DIMENSION
California, Berkeley, CA 94720 USA (e-mail: goodwill@berkeley.edu).
S. M. Conolly is with the Department of Bioengineering, University of Cali- Let us consider a 1-D MPI system with a convenient linear
fornia, Berkeley, CA 94720 USA (e-mail: sconolly@berkeley.edu). gradient and an time-changing offset field
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. such as what would be created by a Helmholtz
Digital Object Identifier 10.1109/TMI.2010.2052284 coil (Fig. 2). If the gradient is zero at the origin, we can find the
0278-0062/$26.00 © 2010 IEEE
1852 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO. 11, NOVEMBER 2010

H t
Fig. 2. Addition of a time varying homogeneous offset field ( ) to a gradient
field 0Gx x t
causes a shifting of the field free point ( ). For a gradient 0G ,
x t G H t
the location of the FFP can be solved for as ( ) = ( ).

magnetic field at position

(1)
Fig. 3. Simulated MPI 1-D image of a complex phantom (solid line) and the
Fundamentally, MPI relies on the rapid movement of a field free source distribution (dotted line).
point (FFP) across the sample to elicit a signal. We can describe
the location of the field free point by solving the above equation
for . This yields where the , , and axes correspond to the physical , , and
axes. This tells us the total magnetization in the system is a
convolution of the magnetic particle density with a Langevin
function kernel. This assumes that the magnetic nanoparticles
Substituting FFP position into (1), the magnetic field at
do not change the H field significantly. This is generally true
position can be rewritten
since the average of magnetite at physiologically well toler-
ated concentrations [1] is small [6].
However, the point spread function (PSF) of a 1-D MPI
The magnetization of magnetic nanoparticles in response to system is not the Langevin function. We typically receive the
an applied magnetic field in one dimension can be described by MPI signal using an inductive detector that sees changing mag-
the Langevin equation (see Section III) netic fields with sensitivity [7]. Visually described
in Fig. 1, we obtain
MPI 1-D signal equation in volts

where is the magnetic moment of the magnetic

nanoparticle, is a property of the magnetic nanopar-
ticle, is the applied magnetic field, and is
the nanoparticle density. By using the Langevin function, we
make a number of physical assumptions. Here we assume that
the system is superparamagnetic, magnetic domains align in-
stantaneously with the applied magnetic field (i.e., quasi-static
with no particle relaxation), and the magnetic nanoparticle
concentration is dilute enough that the particles are non-inter-
acting.
Now consider an available magnetization density that is a
continuous distribution of magnetic nanoparticles, .
If we assume that the nanoparticle distribution is distributed
along the axis and zero elsewhere, we define
. Then, the magnetization at a single point along
this continuous distribution when the FFP is at position is
We can convert this 1-D magnetization into a flux, . Assuming
the magnetization density only changes along results in a con-
volution
(2)
For an image, what we desire is a magnetic particle distri-
bution convolved with a PSF. We assign the derivative of the
Langevin curve to be our system PSF and divide out the extra
terms, and this results in
MPI 1-D image equation
(3)
We see a simulated 1-D image that is reconstructed using this
method in Fig. 3. This implies MPI in one dimension while
using a super-paramagnetic nanoparticle is a linear, space in-
variant (LSI) system, a hallmark of most imaging modalities.
III. LANGEVIN EQUATION
The Langevin function (Fig. 4) accurately describes the mag-
netization of an ensemble of magnetic nanoparticles in response
to an applied magnetic field in the case where relaxation is
not dominant [8]. For [particles/ ] magnetic nanoparticles,
each with magnetic moment , the magnetization as a function
GOODWILL AND CONOLLY: HE X-SPACE FORMULATION OF THE MAGNETIC PARTICLE IMAGING PROCESS 1853

Fig. 4. (Top) The Langevin function, [kH ]. (Bottom) The derivative of the
Langevin function _ [kH ]. The derivative of the Langevin function is analyti- Fig. 5. Intrinsic MPI resolution for various SPIO nanoparticle diameters when
cally FWHM  4:16. imaged using different gradient field strengths.

TABLE I
VALUES OF k FOR VARIOUS DIAMETERS are separated by the full-width at half-maximum (FWHM) [10].
OF SPHERICAL SPIO NANOPARTICLES We use this classical criterion in this paper. We also note that
super-resolution methods exist for increasing image resolution
beyond the classical limit at the expense of SNR [11].
From the image (3), we see that the point spread function
of the MPI process is the derivative of the Langevin function.
Solving for the FWHM of this well behaved function, we can
estimate the intrinsic resolution of the MPI process analytically
as . This gives [4]
of applied field is

where
and is Boltzmann’s constant, is the temperature, and is
the vacuum permeability. We note that gives
the magnetic moment of a spherical, single domain magnetic
nanoparticle [4]. For magnetite SPIO nanoparticles,
. nets. This would even enable adjusting the intrinsic resolution
The derivative of the Langevin function is a well behaved
even function
The operand of the Langevin function, , is a dimensionless
value. Scalar has units of [m/A] and example values are shown
in Table I.
IV. SPATIAL RESOLUTION
Spatial resolution of the MPI system can be thought of as the
ability to accurately depict two distinct points of equal intensity
in space [9]. There are many classical criteria for resolution in-
cluding Rayleigh, Schuster, Houston, and Buxton. Houston pro-
posed a criterion where two points are just resolved if two points
which shows that system resolution can be improved by in-
creasing the diameter of the magnetic nanoparticles , or by in-
creasing gradient strength . We can then calculate the intrinsic
resolution of the 1-D MPI process as a function of particle size
and gradient strength (Fig. 5 and Table II). Because spatial res-
olution improves with the cube of magnetic nanoparticle diam-
eter, it may be possible to use a large particle with resistive mag-
of the system in real time by simply changing the strength of
the gradient. Note that operating at reduced temperatures, such
as 77 K or 5 K, could dramatically increase and the resulting
resolution for preserved histological samples, provided we do
not go below the blocking temperature which would causes the
SPIO particles to behave ferromagnetically.
V. BANDWIDTH
To design a MPI system, we must know the bandwidth re-
quirements to achieve the desired resolution. We can approxi-
mate the bandwidth required to represent the Langevin function
derivative through Fourier analysis.
We found earlier that the derivative of the Langevin function
is our PSF and defines the intrinsic resolution of the 1-D MPI
process. The derivative of the Langevin function does not have
1854 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO. 11, NOVEMBER 2010

TABLE II
INTRINSIC RESOLUTION FOR GIVEN SPIO NANOPARTICLE SIZES AND
GRADIENT STRENGTHS. LARGER CORE DIAMETERS ENABLE SIGNIFICANTLY
IMPROVED RESOLUTION. A 5:1 T=m= MAGNETIC FIELD GRADIENT IS
REASONABLE TO BUILD FOR SMALL ANIMALS

a simple Fourier transform, and so we will approximate our re-

sults with a Lorentzian. To reasonable accuracy of peak
error, we can approximate the derivative of the Langevin func-
tion as a Lorentzian function with
Fig. 6. Bandwidth requirements of the MPI imaging process. The energy con-
tent of the frequency spectrum decays exponentially with increasing frequency.

We can model the rate of change of a triangular, sinusoidal,

TABLE III
or arbitrarily changing FFP position as a linearly changing FFP INTRINSIC RESOLUTION AND BANDWIDTH SCALING WITH RESPECT TO
position. A linearly ramping field with ramp rate gives THE PARTICLE SIZE. RATIOS ARE NORMALIZED TO A 30 NM
us a time varying position . This corresponds to a PARTICLE. SMALLER PARTICLES HAVE LOWER BANDWIDTH
REQUIREMENTS AND POORER RESOLUTION
time varying magnetic field of

This lets us introduce what we term the magnetic field slew rate,
[A/m/s], which is the product of the scanning rate and the
gradient size . We will also find that governs a number of
important parameters including bandwidth requirement, SAR, , ) results in a . We
and magnetostimulation. can see the relationship between the particle size, the intrinsic
For particles located at the origin, resolution, and the bandwidth in Table III. The particle
, the 1-D signal (2) and substituting a Lorentzian size strongly controls the resolution and required bandwidth.
for the derivative of the Langevin curve gives
VI. RESOLUTION-BANDWIDTH TRADEOFFS
The finite receiver bandwidth of a real system reduces the
bandwidth from the intrinsic resolution to what we term the
Taking the Fourier transform of yields achievable resolution. As a result, we can trade off receive band-
width, achievable resolution, and SNR. To see this, let us assume
a receive bandwidth, , with a brick-wall filter. Then, the re-
ceived signal in Fourier and real space is

which describes the frequency content of the MPI experiment as

seen in Fig. 6. We note that since the MPI signal occurs
at baseband and does not have a carrier frequency.
For a simpler representation of required bandwidth, we can
solve for the bandwidth in Hertz This implies that limiting the bandwidth in frequency space with
a lowpass filter will cause widening of the PSF, reducing our
achievable resolution. However, we cannot increase our band-
width indefinitely, as this will reduce our SNR (see Section VII).
For example, choosing a bandwidth whose FWHM is the same
as the intrinsic resolution of the system will almost halve the
and we see that MPI’s bandwidth requirements increase lin- achievable resolution of the system [Fig. 7(b)]. From the graph,
early with nanoparticle properties , gradient strength , and we see that we asymptotically approach the intrinsic resolution.
scanning rate [see Fig. 7(a)]. Evaluating the FWHM with re- We do not reach 150% of the intrinsic resolution until the re-
alistic parameters for a 3 cm field of view ( , ceiver bandwidth is , and 110% of the in-
, , , trinsic resolution until .
GOODWILL AND CONOLLY: HE X-SPACE FORMULATION OF THE MAGNETIC PARTICLE IMAGING PROCESS 1855

Fig. 7. MPI F bandwidth and maximum resolution as a function of multiples of the F bandwidth (a) Linear relationship between F bandwidth for
various particle sizes and magnetic field slew rate. Note that the magnetic field slew rates are much faster than those seen in MRI and are given in T=ms= (b)
Relationship between F bandwidth and resolution assuming a brick-wall receive filter. The achievable resolution of the MPI process is 150% of the intrinsic
4 
resolution when the receiver bandwidth is f 2:2F . We do not reach 110% of the intrinsic resolution until f 4  3:8F .

VII. SIGNAL-TO-NOISE RATIO SNR, which we define as the peak signal divided by the stan-
dard deviation of the noise, is
We can calculate the signal and noise from first principles.
We note that our primary assumption is that the system acts
quasi-statically since the SNR may be significantly less if we ac-
count for particle relaxation. From the MPI Signal (2), the peak
signal for a linearly ramping magnetic field for a point source
of particles with total magnetization
is
which assumes the nanoparticles are at the coil temperature,
. This describes the SNR for a single pass across the sample,
and does not take into account time averaging or acquisition
We can consider as the maximum magnetization the SPIO time.
nanoparticles at the point source can produce. This is useful for
understanding the SNR as a function of the available magneti- A. 3-D SNR With Averaging
zation, which is constant for a given weight of magnetic tracer. The signal to noise ratio should be normalized to scan time
The received noise occurs from three sources, the noise figure and take averaging into account. If we assume a regular sam-
of the preamplifier, NF, the noise from the receive coil, , pling of the field of view with readout in the direction and
and body noise, . Since the MPI received signal is typ- acquisition of and lines in the and directions, re-
ically at a higher frequency than noise present in Si-BJT spectively, during a total sampling time of , we can estimate
and Si-JFET preamplifiers [12], we can model the body and coil the number of averages as
noise as spread across a noise bandwidth, . Then, assuming
Boltzmann noise, the standard deviation of the resistive voltage
noise after the preamplifier is

Then, the 3-D SNR can be estimated

which we simplify by assuming coil noise dominance. Body (4)

noise dominance is difficult to achieve at the kilohertz frequen-
cies utilized in MPI [3]. We must decide what the desired resolu- and as a function of particle diameter
tion of the system is since the resolution drops as we limit band-
width Fig. 7(b). Suppose we desire a bandwidth that worsens
the FWHM of the imager by approximately half, then we can
express the bandwidth requirements as a function of the 3 dB (5)
bandwidth: . As a result, our
1856
TABLE IV
THEORETICAL F , SAR AND SNR, AND DETECTION LIMIT FOR REAL-TIME (RT) AND HIGH RESOLUTION (HR) SCANS FOR d
SAR LIMIT OF 4 W/KG WITH AVERAGING TIME IN SECONDS. NOTE THAT WE REQUIRE BW
WE HAVE CHOSEN THE AVERAGING TIMES TO REFLECT TYPICAL TIMESCALES OF INTEREST TO THE REAL TIME OR HIGH-RESOLUTION SCANS
where arises from the increased SNR due to the con-
tribution of receiver coils in multiple axes. This is perhaps the
only imaging modality where our signal increases linearly with
the scanning rate as faster scanning not only increases the raw
signal, but also increases the number of averages possible. As
expected, the SNR increases with greater time averaging, ,
and decreases with increasing field of view size. This does not
take into account the increased resistance of the body as we
move to high frequencies [13] since , however at
the low frequencies employed in MPI we typically operate in a
coil-noise dominated regime.
Optimizing a tracer to the MPI process has many benefits.
Increasing the particle diameter increases the signal dramati-
cally as . This also reduces the gradient magni-
tude required for the same resolution. Increasing the saturation
magnetization, , of the magnetic nanoparticles will also en-
hance SNR. However, the saturation magnetization is difficult to
change since this is an intrinsic property to the magnetic mate-
rials used.
System modifications can also increase the signal to noise
ratio. SNR increases with gradient strength, the FFP speed, re-
ducing coil noise, and improvement of the preamplifier noise
figure. MPI’s SNR increases as we increase the scanning speed,
even though we will continue to increase our bandwidth require-
ments. Reducing the gradient strength, , also increases SNR
for a fixed sample size because the scanning rate can be in-
creased while maintaining the same magnetic field slew rate.
From (4) we estimate the SNR for various scenarios in
Table IV. The calculations are made for specific applications
with well defined fields of view and resolutions at the SAR
limit assuming a single receive coil and sensitivity and noise
values from Gleich et al. [1]. For a small animal, we assume
a coil sensitivity value of and a noise of
. For a human scanner and
. It is clear that, theoretically, MPI should detect
a single nanogram of magnetic material with reasonable SNR.
These sensitivity numbers are similar to those calculated in
an illuminating simulation study on the optimal SPIO core
diameter for MPI imaging [14].
VIII. SINUSOIDAL EXCITATION
Ideally, a MPI system would excite the sample with a trian-
gular excitation waveform. However, as the excitation magnetic
field is typically orders of magnitude larger than the received
signals generated by the SPIO nanoparticles, it is not a simple
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO. 11, NOVEMBER 2010
= 30 nm
PARTICLES AT THE
=22
: F TO ACHIEVE 150% OF THE INTRINSIC RESOLUTION.
task to magnetically isolate the excitation field. Since a trian-
gular excitation field would be composed of a sum of odd har-
monics, we instead excite with a simple sinusoid that has limited
frequency content. We then isolate geometrically and with pas-
sive electrical filters.
Geometric isolation [1], [15] reduces coupling between the
transmit and receive coils by putting the receive coil in a gra-
diometer configuration. This also reduces coupled noise from
outside the system bore. Unfortunately, even perfect geometric
isolation between the transmit and receive coils is difficult—if
not impossible—to achieve because of eddy current coupling. If
we place a conductive sample in the bore, eddy currents in the
sample induce voltage in the receive coil.
Because we cannot perfectly geometrically isolate the two
coils, existing systems isolate with passive notch and high-pass
filters [1], [3], [15]. The combination can achieve million-fold
reduction in the excitation frequency from the received signal.
Unfortunately, this comes with drawbacks. With a sinusoidal
excitation, the SNR drops near the edges of the scan volume
because of the reduced magnetic field slew rate. As a result, we
do not obtain optimal SNR throughout. In this paper, we do not
discuss filtering, but we do excite with a sinusoidal signal.
IX. SPECIFIC ABSORPTION RATE
For a sinusoidally oscillating magnetic field, SAR can be an-
alytically estimated for a cylinder as [16]
where is the excitation frequency, is the excitation mag-
nitude, is the diameter of the sample, is the tissue resistivity
at the excitation frequency, and is the specific gravity of the
sample. If we assume sinusoidal excitation, FOV, gradient size
, and scanning rate , we can estimate the excitation magni-
tude and excitation frequency as
and
which implies that SAR in terms of magnetic field slew rate
is
GOODWILL AND CONOLLY: HE X-SPACE FORMULATION OF THE MAGNETIC PARTICLE IMAGING PROCESS
Fig. 8. (Top) Maximum magnetic field slew rate for a 4 W/kg specific absorp-
tion rate. (Bottom) Optimal scanning frequency at the 4 W/kg RG limit.
Surprisingly, the SAR presented in this manner does not depend
on the field of view.
At present, the United States Food and Drug Administra-
tion (FDA) limits SAR deposition in the body to 4 W/kg. We
can then solve for the maximum magnetic field slew rate for
a given sample size. Since SNR increases with the magnetic
field slew rate, this will give us our optimal scanning rate. We
see the maximum sample diameter as a function of magnetic
slew rate in Fig. 8. Note that the chest diameter is estimated at
, which is significantly larger than the chest field of
view . XI. EXPERIMENTAL
X. SLEW RATE AND MAGNETO-STIMULATION
A rapidly changing magnetic field in conductive tissue in-
duces an electric field that can stimulate peripheral and cardiac
nerves. The fundamental law of magneto-stimulation states [17]
where is the peak amplitude for magneto-stimulation
as a function of excitation time constant, , is the min-
imal peak-to-peak excitation amplitude for frequencies going to
infinity, and is the chronaxie time constant. This model has
been studied extensively [18], [19], and is similar to both FDA
and IEC regulations for MRI scanners [20].
1857
Because current MPI scanners are constructed using a reso-
nant excitation coil, we can estimate the minimum sinusoidal
excitation frequency to prevent magneto-stimulation of the pe-
ripheral nerves. For a sinusoidal excitation, we can approximate
[21] the excitation rise time as
and so the fundamental law of magneto-stimulation with a si-
nusoidal excitation signal with respect to the desired magnetic
field slew rate is
We can then solve for the optimal frequency of excitation for a
magnetic field slew rate product while maximizing SNR
Ideally, in a system we will operate with the maximum pos-
sible slew rate that does not cause magneto-stimulation. Be-
cause is large compared to the slew rates used in MRI gra-
dients, is essentially the magnetic stimulation threshold.
That is, . For the peripheral nerves, we use the
estimate Schaefer et al. [20] found for the Z gradient in an MRI
with for a peak-to-peak excitation magnitude of
and a chronaxie time constant of .
We can see the optimal excitation frequencies for various
sample sizes in Fig. 8. This formulation becomes more impor-
tant at lower excitation frequencies as the ratio
. At the maximum magnetostimulation excitation fre-
quencies to achieve the SAR limited magnetic field slew rate,
imaging a human heart can have up to a 20-kHz line scan rate,
which has already been shown in practice in mice [1]. There is
no benefit to a higher fundamental frequency as the only impor-
tant criteria is the FFP movement rate. In fact, lower frequency
fundamentals are simpler to filter out because they take a smaller
percentage of the available bandwidth. Unfortunately, imaging
at the magnetostimulation limit requires increasing the scan-
ning frequency to prevent stimulation of the peripheral nerves
in order to reach the desired magnetic field slew rates.
To test the principles described in this paper, we built a zero-
dimensional MPI spectrometer as shown in Fig. 9. The system
is constructed with a excitation electromagnet and a biasing
magnet. The outer electromagnet generates the bias field,
which simulates placing a point source sample at location
. Varying the bias field enables us to simulate moving
the sample in a gradient field. The virtual FFP is scanned using
a resonant transmit coil, , and the signal produced is re-
ceived with a gradiometer receive coil. Using the system, we
measured the MPI signal from an undiluted SPIO nanoparticle
sample (Chemicell 50 nm fluidMAG-D, Berlin, Germany). The
point spread functions were identical when diluted.
The bias coil is driven by a dc coupled audio amplifier
(Crown M-600, Elkhart, IN) at fields up to while
1858 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 29, NO. 11, NOVEMBER 2010

Fig. 9. MPI spectrometer for testing point spread function and bandwidth of
the MPI process. The excitation magnet generates a 160 mT peak-to-peak os-
cillating magnetic field at 6.23 kHz. The bias coil supplies a dc magnetic field
6
of up to 80 mT. The signal received from the gradiometric receive coil is dig-
itized at 1.25 MSPS without filtering.

dissipating . The resonant excitation coil generates

160 mT peak-to-peak at 6.23 kHz and is driven by an audio
amplifier (AE Techron LVC5050, Elkhart, IN) with
of instantaneous power at a pulsed 1.5% duty cycle. The signal
from the receive coil is digitized by a 16-bit data acquisition
system with a 1.25 MSPS sampling rate (National Instruments
USB-6259, Austin, TX) controlled by custom software written
in MATLAB (Mathworks MATLAB, Natick, MA).

XII. RESULTS
In Fig. 10(a), we see the received signal plotted as a function
of virtual FFP position. This is equivalent to the explanatory
figure shown in Fig. 1. The measured data fits remarkably well
to theoretical predictions. Normalizing to the relative speed of
the FFP position shows that the point spread function does not
change across the FOV.
The magnitude of the magnetic fields used here are signif-
icantly beyond the magnetostimulation limit. Large excitation
fields are necessitated by the small core diameters of the SPIO
nanoparticles used in the study.
We fit the measured point spread function to what would be
Fig. 10. Experimental data from zero-dimensional MPI imager (a) [Top]
measured if we considered a distribution of nanoparticle diam- Parametric plot of FFP position and received signal with multiple offset field
eters. Assuming a log-normal distribution, we estimate that the strengths. The peaks correspond to the offset field generated by the bias coil.
mean magnetic core diameters is 12.3 nm with a standard devia- [Bottom] Received signal divided by instantaneous FFP velocity. Only half the
data is shown due to alignment issues which we believe are due to hysteresis.
tion of 3.2 nm. This corresponds well to particles from the same This is equivalent to the 1-D image [see (II.3)] (b) [Top] Measured PSF
manufacturer measured with a TEM [22] as well as Resovist compared with simulated PSF of a particle distribution, also shown with the
particles measured using a different type of MPI-based instru- point spread function of a hypothetical 30 nm particle. [Bottom] Log-normal
particle size distribution used to generate simulated data.
ment [23].

XIII. CONCLUSION
Magnetic Particle Imaging is a new imaging modality with
great promise to detect nanogram quantities of iron oxide tracers
deep in the body of an animal or a human. To develop the tech-
nology into human capable systems, we need to understand the
basis for the imaging method. In this paper we have derived the
1-D x-space formulation of the MPI signal. We use this founda-
tion to discuss resolution, bandwidth requirements, SNR, spe-
cific absorption rate, and slew rate limitations. We concluded
by briefly discussing the construction of a system to test our
theoretical predictions, which we tested by measuring the point
spread function and estimating the core diameter of a commer-
cially available SPIO nanoparticle.
XIV. ACKNOWLEDGEMENTS
The authors would like to thank A. Tamrazian and R. Pida-
parthi for their excellent assistance in acquiring data.
REFERENCES
[1] J. Weizenecker, B. Gleich, J. Rahmer, H. Dahnke, and J. Borgert,
“Three-dimensional real-time magnetic particle imaging,” Phys. Med.
Biol., vol. 54, pp. L1–L10, Feb. 2009.
[2] T. F. Sattel, T. Knopp, S. Biederer, B. Gleich, J. Weizenecker, J.
Borgert, and T. M. Buzug, “Single-sided device for magnetic particle
imaging,” J. Phys. D: Appl. Phys., vol. 42, p. 022001, Dec. 2008.
[3] P. Goodwill, G. Scott, P. Stang, and S. Conolly, “Narrowband mag-
netic particle imaging,” IEEE Trans. Med. Imag., vol. 28, no. 8, pp.
1231–1237, Aug. 2009.
GOODWILL AND CONOLLY: HE X-SPACE FORMULATION OF THE MAGNETIC PARTICLE IMAGING PROCESS
[4] J. Rahmer, J. Weizenecker, B. Gleich, and J. Borgert, “Signal encoding
in magnetic particle imaging: properties of the system function,” BMC
Med. Imag., vol. 9, no. 1, p. 4, 2009.
[5] T. Knopp, S. Biederer, T. Sattel, and T. Buzug, “Singular value analysis
for magnetic particle imaging,” in IEEE Nucl. Sci. Symp. Conf. Rec.
(NSS’08), Jan. 2008, pp. 4525–4529.
[6] J. F. Schenck, “The role of magnetic susceptibility in magnetic res-
onance imaging: MRI magnetic compatibility of the first and second
kinds,” Med. Phys., vol. 23, pp. 815–850, Jun. 1996.
[7] D. Hoult and R. Richards, “The signal-to-noise ratio of the nuclear
magnetic resonance experiment,” J. Magn. Reson., vol. 24, pp. 71–85,
Jan. 1976.
[8] T. Yoshida and K. Enpuku, “Simulation and quantitative clarification
of ac susceptibility of magnetic fluid in nonlinear Brownian relaxation
region,” Jpn. J. Appl. Phys., vol. 48, no. 12, 2009.
[9] A. J. den Dekker and A. van den Bos, “Resolution: A survey,” J. Opt.
Soc. Am. A, vol. 14, no. 3, pp. 547–557, 1997.
[10] W. Houston, “A compound interferometer for fine structure work,”
Phys. Rev., vol. 29, pp. 478–485, Jan. 1927.
[11] M. Shahram and P. Milanfar, “Imaging below the diffraction limit: A
statistical analysis,” IEEE Trans. Image Process., vol. 13, no. 5, pp.
677–689, May 2004.
[12] V. Radika, “Low-noise techniques in detectors,” Ann. Rev. Parr. Sci.,
vol. 38, no. 1, pp. 217–277, 1988.
[13] P. Röschmann, “Radiofrequency penetration and absorption in the
human body: Limitations to high-field whole-body nuclear magnetic
resonance imaging,” Med. Phys., vol. 14, no. Jan., pp. 922–31, 1987.
[14] R. M. Ferguson, K. R. Minard, and K. M. Krishnan, “Optimization of
nanoparticle core size for magnetic particle imaging,” J. Magn. Magn.
Mater., vol. 321, pp. 1548–1551, Jan. 2009.
1859
[15] B. Gleich and J. Weizenecker, “Tomographic imaging using the
nonlinear response of magnetic particles,” Nature, vol. 435, pp.
1214–1217, Jun. 2005.
[16] P. A. Bottomley and E. R. Andrew, “RF magnetic field penetration,
phase shift and power dissipation in biological tissue: Implications for
NMR imaging,” Phys. Med. Biol., vol. 23, pp. 630–643, Jan. 1978.
[17] W. Irnich and F. Schmitt, “Magnetostimulation in MRI,” Magn. Reson.
Med., vol. 33, pp. 619–623, May 1995.
[18] B. A. Chronik and B. K. Rutt, “Simple linear formulation for magne-
tostimulation specific to MRI gradient coils,” Magn. Reson. Med., vol.
45, pp. 916–919, May 2001.
[19] B. J. Recoskie, T. J. Scholl, and B. A. Chronik, “The discrepancy be-
tween human peripheral nerve chronaxie times as measured using mag-
netic and electric field stimuli: The relevance to MRI gradient coil
safety,” Phys. Med. Biol., vol. 54, pp. 5965–5979, Oct. 2009.
[20] D. Schaefer, J. Bourland, and J. Nyenhuis, “Review of patient safety
in time-varying gradient fields,” J. Magn. Reson. Imag., vol. 12, pp.
20–29, Jan. 2000.
[21] J. P. Reilly, “Maximum pulsed electromagnetic field limits based on
peripheral nerve stimulation: Application to IEEE/ANSI C95.1 elec-
tromagnetic field standards,” IEEE Trans. Biomed. Eng., vol. 45, no. 1,
pp. 137–141, Jan. 1998.
[22] V. Schaller, G. Wahnström, A. Sanz-Velasco, P. Enoksson, and C. Jo-
hansson, “Monte carlo simulation of magnetic multi-core nanoparti-
cles,” J. Magn. Magn. Mater., vol. 321, no. 10, pp. 1400–1403, 2009.
[23] S. Biederer, T. Knopp, T. Sattel, K. Lüdtke-Buzug, B. Gleich, J.
Weizenecker, J. Borgert, and T. Buzug, “Magnetization response
spectroscopy of superparamagnetic nanoparticles for magnetic
particle imaging,” J. Phys. D: Appl. Phys., vol. 42, p. 205007,
2009.