CORE Metadata, citation and similar papers at core.ac.

uk
Provided by Glamorgan Dspace

IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY

Meas. Sci. Technol. 19 (2008) 045501 (11pp) doi:10.1088/0957-0233/19/4/045501

A magnetic induction tomography system

for samples with conductivities below
10 S m−1
S Watson1, R J Williams1, W Gough2 and H Griffiths3
1
Faculty of Advanced Technology, University of Glamorgan, Pontypridd CF37 1DL, UK
2
Department of Physics and Astronomy, Cardiff University, Cardiff CF24 3AA, UK
3
Department of Medical Physics & Clinical Engineering, Singleton Hospital, Swansea SA2 8QA, UK

Received 23 August 2007, in final form 12 December 2007

Published 7 February 2008
Online at stacks.iop.org/MST/19/045501

Abstract
A 16-channel magnetic induction tomography (MIT) system has been constructed for imaging
samples with low conductivities (<10 S m−1) such as biological tissues or ionized water in
pipelines. The system has a fixed operating frequency of 10 MHz and employs heterodyne
downconversion of the received signals, to 10 kHz, to reduce phase instabilities during signal
distribution and processing. The real and imaginary components of the received signal,
relative to a synchronous reference, are measured using a digital lock-in amplifier. Images are
reconstructed using a linearized reconstruction method based on inversion of a sensitivity
matrix with Tikhonov regularization. System performance measurements and images of a
pipeline phantom and a human leg in vivo are presented. The average phase precision of the
MIT system is 17 millidegrees.

Keywords: bioimpedance, tomography, magnetic induction, MIT, EMT, sub-sea oil production

1. Introduction for example in the brain, through the skull, and into pipes or
vessels constructed from non-conducting materials.
Magnetic induction tomography (MIT, also known as Initial progress in MIT was most rapid in imaging metals
electromagnetic induction tomography (EMT)) is a technique because their high conductivities produce large eddy-current
for imaging the passive electromagnetic properties of an signals and the metals industry is very familiar with inductive
object. MIT applies a magnetic field from a current-carrying sensors; from this a range of industrial applications is now
coil to induce eddy currents in the object which are then emerging [2–4]. The development of MIT for biomedical
sensed by an array of other coils. Compared with the related use has been much more difficult because the electrical
technique, electrical impedance tomography (EIT), MIT has
conductivities of biological tissues (0.02–2 S m−1) are many
the advantage that it does not require direct contact with the
orders of magnitude lower than those of metals and give
object but operates through an air gap. Another technique,
much weaker signals. Nevertheless, our work and that of
electrical capacitance tomography (ECT), can operate through
an air gap but is unsuitable for conductive materials because others has shown that the technological difficulties are being
the air layer dominates the response. The contactless overcome, that the signals are measurable and that images
nature of MIT means that errors due to electrode contact can be reconstructed [2–10]. Proposed applications include
impedances are avoided completely and that the positions of imaging cerebral stroke, body composition, wound healing
the coils are known and remain fixed, an advantage for image and bioelectrical spectroscopy. Industrial applications in
reconstruction. In principle, more coils could be added to an imaging conductivities similar to those of biological tissues
MIT array to increase spatial resolution without any loss of have also been suggested, such as the imaging of ionized
practicality. Alternatively, the array can be translated along water in pipelines and separators and imaging foodstuffs
an object to provide cross-sectional images at different levels [2–4]. All these applications, biomedical and industrial, where
[1]. Magnetic coupling allows eddy currents easily to be the conductivity is less than about 10 S m−1, can conveniently
induced in samples surrounded by high-resistivity barriers, be termed ‘low-conductivity’ applications and pose research

0957-0233/08/045501+11$30.00 1 © 2008 IOP Publishing Ltd Printed in the UK

Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

questions of a generic nature especially with regard to signal square of the frequency. A non-unity relative permeability also
capture and image reconstruction. gives rise to a real component, but with a value independent
In this paper, we describe the design and performance of frequency. For the experiments described in this paper, it
of a 10 MHz, 16-channel, MIT system and show preliminary will be assumed that µr = 1.
imaging experiments on a human leg in vivo.
In addition, we address what is potentially a new
3. System design
industrial application of MIT in off-shore oil production,
where knowledge of the seawater fraction (‘water cut’) in 3.1. Overview
multiphase flow in pipelines and separators is important for
controlling productivity. Various commercial sensors are The MIT system comprises four subsystems:
available employing capacitance, microwaves, ultrasound or
• the front end, consisting of the chassis/screen, the 16
γ -rays, but their accuracy has been limited by factors such
coil modules and the transmitter (excitation) and receiver
as scaling of the pipes, a limited range and the gas content
(detection) circuitry;
of the mixture. Eddy-current techniques could provide
• the signal-distribution and control system;
a complementary method because they would be sensitive
only to the conductive component of the mixture, i.e. the • the measurement system;
process water, which has a typical conductivity of 5 S m−1. • the image-reconstruction system comprising the data stor-
Furthermore, the use of several coils and possibly multiple age, forward-modelling software, image-reconstruction
frequencies would enable the conductivity of the process algorithm and image-visualization software.
water to be determined. Albrechtsen et al [11] demonstrated The operating principle is as follows. One of the 16
an experimental system using a single excitation coil with channels is selected as the active transmitter channel, and the
phantoms simulating different flow regimes. With non- oscillator and power amplifier on the transmitter circuit board
cylindrically-symmetrical flow (e.g. stratified), the derived are enabled. The transmitter circuits of all other 15 channels
water fraction depended strongly on the position of the coil are disabled. The receiver circuits of each of these 15 channels
relative to the water/oil interface; the authors concluded that are then sequentially enabled.
to overcome this problem, a tomographic system should be The signals are de-multiplexed to the measurement system
developed. More recent work by another Norwegian research which measures each signal as a complex number, i.e. with
group using a single-channel inductive system has also in-phase and in-quadrature (real and imaginary) components
concluded that the work should be extended to tomography relative to a synchronous reference waveform. One frame
[12]. of data consists of 240 such measurements covering all
In this paper, we report what we believe to be the first the transmitter/receiver coil combinations (16 × 15). For
such measurements and describe MIT imaging of a phantom imaging, two frames are measured, a data frame with a sample
simulating multiphase flow in an oil pipeline. placed within the detector space and the other, a reference
frame for an empty detector space or some other reference
2. The MIT signal condition. An image of the conductivity or permittivity
distribution of the sample is then reconstructed.
There are two contributions to the signal detected by the Details of the subsystems are given in the following
sensing coil in MIT. The first is directly induced by the field sections.
from the excitation coil (the primary signal, V). The second
is from the eddy currents induced in the material which in
turn produce their own magnetic field which induces a further 3.2. Chassis/screen
emf in the detector coil (the secondary signal, V). For a The MIT system front end consists of a cylindrical
sinusoidally-time-varying excitation at angular frequency ω, electromagnetic screen of aluminium (350 mm in diameter,
the skin depth of the electromagnetic field in the material 250 mm in height) within which are attached 16 coil modules
(strictly for a plane wave) is given by δ = (2/ωµ0µrσ )1/2, (see figure 1). Sixteen corresponding transmitter and receiver
where σ and µr are the electrical conductivity and relative circuit modules are housed in metal enclosures attached to the
permeability of the sample and µ0 is the permeability of free outside of the screen. In concept, this construction is very
space. If δ is large compared with the thickness of the sample, similar to that of the Moscow MIT system [5] but the detailed
which will often be so for a segment of the human body, design is original to our research group.
V The electromagnetic screen (i) provides a rigid chassis
= Qωµ0 [ωε0 (εr − 1) − iσ ] + R(µr − 1), (1)
V to support the coils, (ii) reduces interference from external
where ε r is the relative permeability of the sample, ε0 is electromagnetic fields, (iii) confines the excitation field within
the permittivity of free space, and Q and R are geometrical the imaging volume, thereby removing the potential for
constants [13]. So, the conduction currents induced in the interference by external conductive or magnetic objects and
sample give rise to a component of V which is proportional (iv) acts as a ground plane that helps to reduce undesirable
to frequency and conductivity and is imaginary and negative, electric-field (capacitive) coupling between the excitation and
meaning that it lags the primary signal by 90◦ . Displacement detection coils. In a recent paper, we have shown that the
currents cause a real (in-phase) component proportional to the capacitive coupling in this system is very small indeed [13].

2
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

+5V

10
+1
Tx
10 coil
-1
1/2 AD8056
10 MHz
oscillator OPA2682
module -5V

Figure 3. Transmitter module design.

2
Rx
coil
OPA682
Figure 1. The MIT front end showing the coils fixed inside the OPA2682
electromagnetic screen.

50 mm
3 4 5

10 mm
Rx coil
TUF1-H
× ×2 ×2 ×4
Output
Tx coil AD8056 AD8056
LO Low-frequency amplifier
and filter
30 mm Perspex former
Figure 4. Receiver module design. The numbers, 1–5, shown
against the components relate to measurements presented later in
section 5.2. LO = input from local oscillator.

Part of electromagnetic screen

an excitation current of approximately 100 mA r.m.s. The
Figure 2. Detail of individual coil module. Rx = receiver, Tx = output amplifiers (OPA2682) have a disable function which
transmitter. places their outputs into a high-impedance state. This function
was used to minimize the loading of the transmitter coil when
3.3. Coils not in use and hence minimize its interaction with any of the
The constructional details of the coil modules are shown other coils by inductive coupling. The specifications of the
in figure 2. The transmitter coils consist of two turns of OPA2682, for the disabled state, quote an impedance from
0.6 mm diameter, PVC-insulated, solid, copper wire; this its output to ground of 800  in parallel with 4 pF for the
number of turns was found to provide the best combination of inverting configuration and 1 M in parallel with 4 pF for
magnetic field strength (∝ current × number of turns) and load the non-inverting. Including also the two 10  resistors in
impedance at 10 MHz. The receiver coils also had two turns series with the amplifier outputs (figure 3), this means that
which set the self-resonance frequency above 50 MHz. This when not in use, the coil will be shunted by an impedance of
reduced the potential for phase instability which was found to (805–4133i) .
increase when operating near the resonant frequency. It also
avoided resonance at harmonics of the operating frequency; 3.5. Detection circuitry
with four turns, for instance, the coil exhibited resonances
at approximately 30 MHz and 50 MHz, causing significant The receiver module circuitry (figure 4) consists of a wideband
distortion of the 10 MHz signals. differential amplifier with unity gain, formed from two devices
(OPA2682 and OPA682). This allows conversion of the
received signal from balanced to unbalanced. The balanced
3.4. Excitation circuitry
input stage helps to provide further rejection of unwanted
The design of the transmitter module is shown in figure 3. interference due to capacitive coupling with the transmitter coil
The excitation signal source is a 10 MHz, temperature- [4]. The signal is passed to a mixer (Minicircuits TUF1-H)
compensated, crystal-oscillator module which is enabled when where it is mixed with a 9.99 MHz waveform from a local
the channel is to be used as the transmitter. When the oscillator oscillator, thus downconverting it to 10 kHz. The 9.99 MHz
is disabled, its power supply is also disconnected by means of a waveform is derived from a single, low-phase-noise generator
relay. The output of the oscillator is passed to a buffer and then (Marconi 2022C) and is distributed to each receiver channel
to a coil-driver amplifier with a balanced output which provides via a power amplifier (AD815) and a 1:16 resistive power

3
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

splitter. The mixer output then undergoes low-pass filtering where σ is the unknown conductivity distribution and b is
(single pole ft = 30 kHz) to remove the sum frequency, and the measurement vector, the conductivity distribution, σ λ,
three stages of amplification in a circuit which exhibited a low reconstructed using Tikhonov regularization, is given by
phase drift and phase skew [14].

σ λ = min Sσ − b22 + λ2 L(σ − σ 0 )22 , (4)
σ
where λ is the regularization parameter, L is the regularization
3.6. Signal distribution and measurement
matrix and σ 0 is an a priori estimate of σ . For all the
The downconverted signals from the receivers are distributed reconstructions presented in this study, L was set to the identity
to two analogue multiplexers (ADG406), one of which selects matrix and no a priori information was incorporated, i.e. σ 0 =
the received signal, and the other the reference signal (2.1 V 0. Images were reconstructed using a MATLAB regularization
r.m.s.) derived from the receiver coil of the channel active toolbox [17]. λ was selected by a subjective assessment of
as the transmitter. It is assumed here that the signal derived the smallest value, consequently giving the highest spatial
from the receiver coil on the same coil former as the active resolution, which did not introduce significant noise artefacts.
transmitter coil can be employed as a reference since, due to
its proximity to the transmitter coil, it will be sensitive almost 5. Measurements
entirely to the primary field and will be hardly affected by a
sample placed within the array. 5.1. Noise and drift
The measurement system is a digital lock-in amplifier
To determine the noise of the system, a reference frame
(Stanford Research Systems 830) controlled by a PC via a
was collected. A second ‘data’ frame was then collected
GPIB interface. The lock-in amplifier operates as a vector
after an interval of 10 s; both frames were collected with
voltmeter, providing the real and imaginary components of the
no sample present in the array. Im(V/V) was computed
received signal relative to the phase of its internal reference
for each transmitter/receiver combination, as described in
signal.
section 3.6. This was repeated ten times and the average noise
Once the reference frame (empty detector space) and for each channel was calculated as the standard deviation of the
data frame (sample in place) have been acquired as sets ten measurements. The system’s average noise was calculated
of complex numbers, the differences between the data and by averaging over all 240 channel combinations.
reference measurements, V, are divided by the reference To determine the system’s drift, 30 frames were collected
measurements, V, to form the quantities Re(V/V) and spaced at 10 s intervals, giving a total measurement time of
Im(V/V), compatible with equation (1). Since these 48 min. The drift was calculated by taking the difference
quantities contain ratios of signals, the influence of any between the maximum and minimum values of Im(V/V)
variations in the receiver gains between channels and over over the 30 measurements for each transmitter/receiver
time is reduced. The total time taken to acquire one frame of combination. The system’s average drift was then defined as
data (240 complex numbers) is 90 s. the average over all channel combinations. The maximum drift
was defined as the figure obtained for the channel combination
4. Forward modelling and image reconstruction showing the highest drift figure.

Images of conductivity were reconstructed from the measured 5.2. Temperature sensitivity of receiver components
values of Im(V/V). First, the sensitivity matrix, S, was
computed using a finite-difference forward model of the As temperature changes are a common cause of drift
system [15]. The volume was divided into cuboidal voxels in electronic systems, an investigation was performed to
and S was computed using two different methods. For determine the sensitivity of the phase of the signal to changes
perturbations in conductivity relative to empty space, S was in temperature of the active devices in the receiver circuit.
computed by setting the conductivity of each voxel in turn Heat-sink compound (HTCP, Electrolube, Berkshire, UK) was
applied to the tops of five components of the receiver circuit—
to 1 S m−1. For perturbations relative to a conductive
the first and second stages of the RF input amplifier, the mixer,
background, S was computed using the reciprocity principle
the low-frequency amplifier and the filter and amplifier section
[16]. For each transmitter/receiver coil combination, the
(labelled 1–5 in figure 4).
sensitivity Si of voxel i, with conductivity σ i, is given by
Measurements of the receiver output were made at the rate
JTi · JRi of 2 s−1 for a period of 10 min during which time a broad-tipped
Si = , (2)
σi soldering iron was held against each device in turn for 5 s, via
where JTi is the eddy current density induced within voxel i by the heat-sink compound. The sequence of measurements was
the transmitter coil, and JRi is the eddy current density which repeated twice. A further three sequences were acquired, but
would be induced by the receiver coil if it were employed as the this time with the soldering iron applied to the corresponding
transmitter. S then has as many rows as there are voxels in the components in the receiver circuit of the reference channel.
volume and as many columns as there are coil combinations. A second experiment was then performed in which
separate circuit boards for the individual devices were
Expressing the forward problem as
constructed, allowing each to be placed in a temperature-
Sσ = b, (3) controlled oven. The change in phase of the signal in response

4
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

Figure 5. The MIT system imaging a simulated oil/gas/water

mixture in a cylindrical pipe phantom.

Table 1. Details of the four subjects imaged.

Height Weight Mean thigh
Subject (m) (kg) diameter (mm) Figure 6. Upper graph: average noise in Im(V/V) for the 240
1 1.80 88 143 different channel combinations. Lower graph: the primary signal, V,
2 1.66 67 146 in Volts r.m.s. at the receiver output (see figure 4). The channel
3 1.82 85 152 separation is the separation of the receiver from the transmitter in a
4 1.88 94 156 clockwise sense (in units of 22.5◦ ).

were obtained for different values of the conductivity of the

to temperatures in the range 30–60 ◦ C was recorded. The solution.
measurements were carried out four times for each device.
5.5. Simulated femur
5.3. Imaging-simulated multiphase flow
As a comparison with the thigh measurements, a plastic rod
A cylindrical plastic tube (diameter 200 mm, length 500 mm) (diameter 36 mm) was immersed in the solution to simulate
was positioned coaxially through the coil array. It was then the femur. The rod was positioned parallel to the axis of the
filled to different levels with saline, 5 S m−1 in conductivity, cylinder, with its own axis displaced from that of the cylinder
to simulate stratified flow, and a data set was acquired for each by half of the radius of the cylinder, roughly to represent the
level. It was assumed that the air in the pipe represented the anatomical position of the femur in the leg. Further MIT
oil and gas in a real pipeline as oil is non-conductive and has a images were obtained.
low relative permittivity (εr < 5) compared with that of water.
The tube was supported by a horizontal platform along which
6. Results
it could be slid out of the array for the reference frame of data
to be acquired (see figure 5). 6.1. Noise and drift

5.4. In vivo imaging The average noise measured for the different channel
combinations is shown in figure 6 (see section 5.1). Here,
In a second imaging experiment, the purpose was to obtain the value of the standard deviation of Im(V/V) has been
in vivo images of the human thigh. Four volunteers from multiplied by 100 so that the noise is expressed as a percentage
university staff were recruited (table 1). The MIT array was of the primary signal, V. The average noise for each channel
positioned again with its axis horizontal to enable the leg of a combination varied from 0.005% to 0.08%, the larger values
supine subject to be imaged. First, a reference set of data was tending to be when the distance between the excitation and
acquired with the array empty. The leg was then inserted detection coil was greatest, since the primary signal was least
through the array such that the coils were approximately in these positions. The percentage noise averaged across all
midway between the knee and the groin. The mean diameter channel combinations was 0.03%, corresponding to a phase
of the thigh at this level was measured and also its vertical noise figure of 17 millidegrees. The large scatter in the
position within the coil array. An MIT data set was then results for a given channel separation is likely to be due to the
acquired. fact that some drifting was occurring during the measurement
As a calibration, a plastic cylinder of saline solution sequence, which lasted 190 s (two frames with a 10 s gap).
(diameter 150 mm, length 200 mm) was positioned at the same The average drift in the received signal, measured as
vertical position in which the leg had been. MIT measurements described in section 5.1, was 0.17% and 0.23% for the real

5
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

6 H True cross- S free space S uniform

section λ = 2×10-5 conductivity
-7
λ= 1×10
4 1
Ref
Rx
2
Im(∆V /V ) %

0.9
0
0 400 800 1200
sample no.
-2
0.8

-4

-6 0.7
1 2 3 4 5

Figure 7. Change in Im(V/V) over a 10 min period during which

the active devices in the receiver module were heated in turn, at
times indicated by the numbered markers (see also figure 4). Two
0.6
curves are shown, one of measurements when the components of the
receiver module of the receiver channel were heated (Rx) and the
other when the corresponding components in the receiver module of
the reference channel were heated (Ref). Each curve is the mean of
three measurements; representative error bars are shown, equal to 0.5
±1 standard error.

Table 2. Temperature coefficients of phase change (mean ±1

standard error) measured by heating the devices in an oven.
0.4
Temperature coefficient of phase
Component change (millidegrees ◦ C−1)
OPA2682 11 ± 0.5
OPA682 11 ± 0.5 0.3
TUF-1H 94 ± 5
AD8056 7±1

0.2
and imaginary components, respectively, while the maximum
drift was 1.5% and 1.3% (860 and 745 millidegrees). These
measurements demonstrate that the drift was very large
compared with the random noise and a significant source of
0.1
error in this MIT system.

6.2. Temperature sensitivity of receiver components

The results of heating of the individual active components Figure 8. MIT images of a simulated oil/gas/seawater mixture in a
in the receiver circuit are shown in figure 7. A change in pipeline for different fractional depths, H, of saline solution. The
Im(V/V) was observed when each of the five devices in second column indicates the true cross-section of the saline in the
the receiver circuit of the receiver channel was heated with the pipe (white area). The third and fourth columns contain images
reconstructed with the two different sensitivity matrices; in each
soldering iron. By far the most sensitive device was component case, the regularization parameter, λ, is given. The white circle
3, the TUF1-H mixer, which showed a change of −4.2% indicates the position of the pipe wall.
(2400 millidegrees in phase). When the corresponding
components in the receiver circuit of the reference channel
6.3. Imaging-simulated multiphase flow
were heated, very similar changes were observed but, as
expected, with the opposite sign. Figure 8 shows the images obtained for different volume
For the second experiment, in which the devices were fractions of the saline solution in the cylindrical phantom
placed separately in an oven, the results are shown in table 2. simulating stratified flow of an oil/gas/seawater mixture in
The observation from the first experiment was confirmed: that a pipeline. The images were reconstructed from the measured
the TUF1-H mixer was by far the most temperature-sensitive values of Im(V/V), using a sensitivity matrix computed
component, having a temperature coefficient more than eight by dividing the cylinder into 20 × 20 × 50 cubic voxels,
times greater than that of any of the other devices. of side-length 10 mm. The first set of images (column 3)

6
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

1.0 Imag Real

NSUM = 0.111f 3 + 0.444f 2 + 0.447f - 0.0102 (a) Absolute images
Thigh Subject 1
0.8

0.6
NSUM

(b) Absolute images

Saline cylinder
0.4 1.0 S m-1

0.2 S freespace
S uniform conductivity (c ) Absolute images
Saline cylinder
0.0 Rod in place
0.0 0.2 0.4 0.6 0.8 1.0
water fraction f

Figure 9. Sum of the image values, NSUM, normalized to the figure (d ) Difference
for a full pipe, versus water fraction, for the two series of images in images
Saline cylinder
figure 8. A third-order polynomial is fitted to the points for the
Rod in place
images reconstructed with a free-space sensitivity matrix (equation
given).
Figure 10. Images reconstructed from the real and imaginary parts
were reconstructed using a sensitivity matrix computed for of the data for (a) human subject 1 and (b)–(d ) the saline phantom.
free space as described in section 4. The second set of The inner white circle is the region of interest applied for summing
the image values. The regularization factor, λ, was 2 × 10−6 for
images (column 4) were reconstructed using a sensitivity
rows (a)–(c) and 1 × 10−7 for row (d ). The sign of the image was
matrix computed for a uniform conductivity within the pipe, changed for images (a)–(c) in the left-hand column, so that in all
using the reciprocity principle (equation (2)). In both cases, cases, the main image feature appeared as a bright area. All images
the reference frame of data was measured with the cylinder were normalized to the maximum image value (even if located in an
removed from the MIT array (i.e. for free space). artefact).
The main image feature resulting from the conductive
saline solution (bright area) is negative. Any positive image sufficient spatial resolution to detect any internal anatomical
values occurring near the edge of the images were set to zero. structure.
The images reconstructed with the sensitivity matrix A cylindrical saline phantom of varying conductivity
relating to free space give a better impression of the cross- was used as a calibration, allowing the equivalent uniform
section of the simulated seawater in the pipe. The images conductivity and relative permittivity of the thigh to be
reconstructed for the conductive background show the main calculated. Figure 10(b) shows the images obtained for one
area of conductance compressed inwards from the wall and a value of the conductivity of the phantom (1.0 S m−1). A
void at the centre; they also show artefacts at the pipe wall. circular region of interest was applied to the images to include
In order to generate a ‘response curve’, the image values the image of the phantom but to exclude any noise and edge
were summed over the whole image and normalized to the artefacts in the rest of the image. The sum of the image
figure obtained for a full pipe (H = 1). This sum is plotted values within the region of interest was computed: this is
in figure 9 against the water fraction, f, calculated from the termed SUMReal and SUMImag for the images reconstructed
value of H, which can readily be shown to be related to f by from the real and imaginary parts of the data, Re(V/V) and
the formula Im(V/V), respectively. The two quantities are plotted against
1  saline conductivity in figure 11.
f = [cos−1 (1 − 2H ) − 2(1 − 2H ) H − H 2 ]. (5) SUMImag is negative and proportional to conductivity, in
π
Although the two sets of images are different in appearance, the close agreement with the theoretical equation (equation (1)).
normalized responses are very similar and closely represented The straight line fitted to the values of SUMImag was now used
by a third-order polynomial as shown. as a calibration line (see figure 11). On the images of the
thigh, a similar region of interest was applied (figure 10(a)).
From the value of SUMImag measured, the equivalent uniform
6.4. In vivo imaging
conductivity of the thigh was calculated from the equation of
Figure 10(a) shows typical images obtained from the thigh the line and is given in table 3. For the four subjects, the
of one of the four subjects. These are absolute images, values lie between 0.27 S m−1 and 0.50 S m−1 and are within
reconstructed from the real and the imaginary parts of the the published range for fat and muscle at 10 MHz, the main
measurements, referenced to free space, and using a sensitivity constituents of the thigh (fat 0.03 S m−1, muscle 0.67 S m−1
matrix computed for free space. From equation (1), it is [18]).
assumed that the image obtained from the imaginary parts For calculating the equivalent uniform relative
represents the conductivity of the thigh and the image from permittivity of the thigh, equation (6) was used; this is
the real parts, its permittivity. Clearly, the images do not have an adaptation of the equation derived from equation (1) by

7
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

Table 3. The sums of the image values from the images of the thighs of the four human subjects. Estimates of the average conductivity and
relative permittivity of the thigh are shown. BMI is the body mass index, = weight/(height)2, calculated from the anthropometric data given
in table 1.
BMI Equivalent uniform Equivalent uniform
Subject (kg m−2) SUMImag SUMReal conductivity of thigh (S m−1) relative permittivity of thigh
1 27.2 −692 55.8 0.27 40
2 24.3 −837 77.7 0.33 55
3 25.7 −1140 94.8 0.44 67
4 26.5 −1290 122 0.50 86

although Im(V/V) closely maintained its proportionality to

0
0 0.2 0.4 0.6 0.8 1
conductivity (as does SUMImag in the present study), Re(V/V)
departed from equation (1) even when the skin depth was
greater than the dimensions of the phantom used. This means
-1000 that the value of εr obtained using equation (6) will be accurate
SUM Imag

only at very low conductivities and will be underestimated at

higher conductivities.
-2000 For the lowest conductivity of the saline phantom
(0.016 S m−1), the value of SUMReal was 102 (see figure 11)
SUM Imag = -2569.8 σ + 5.3753 and equation (6) leads to a value, εr = 72. This is comparable
R2 = 0.9999 with the known value for water (80 at 20 ◦ C) within the
-3000
measurement noise. At a conductivity of 1.0 S m−1, SUMReal
100
had fallen to 52, giving εr = 37, even though the diameter of
SUM Real

the phantom (150 mm) was still only equal to 0.94δ.

The extent of the error in the measurements of the relative
0
0 0.2 0.4 0.6 0.8 1
permittivity for the thigh, due to the skin effect, can be judged
conductivity σ (S m-1) from the measurements of SUMReal for the phantom, given in
figure 11. For the range of conductivities estimated for the
Figure 11. Plots of SUMImag and SUMReal against the conductivity thigh (0.27–0.50 S m−1, see table 3), the value of SUMReal
of the saline cylinder. The straight line fitted to SUMImag is drawn
for the phantom has dropped by 5–10% relative to its value at
and its equation and coefficient of determination, R2, given. The
standard deviation of the measurements was 10. very low conductivities. This implies that the derived values
of the relative permittivity of the thigh should be increased by
Griffiths et al [13], but now using the values of SUMReal 5–10%. With this adjustment, they still lie within the known
and SUMImag obtained from the images, instead of the real range for fat and muscle.
and imaginary parts of the measurements themselves. The
derivative in the denominator is simply the gradient of the 6.5. Simulated femur
calibration line used above (−2570 S−1 m). The derived values
Imaging the saline phantom containing a plastic rod
of the equivalent uniform relative permittivity of the thighs of
(section 5.5) was performed in order to investigate the fact
the four subjects are given in table 3 and lie between 40 and
that even a high-contrast structure such as the femur could not
86. Again, they are within the range for fat and muscle at 10
be distinguished on the images of the thigh (figure 10(a)).
MHz (fat 15, muscle 135–153 [18]). Neither the equivalent
Figure 10(c) shows the absolute images of the phantom
uniform conductivity nor the relative permittivity appears
obtained. They do not reveal the rod.
to be correlated with the body mass index of the subjects
Figure 10(d ) shows difference images reconstructed using
(table 3).
a frame of data measured for the uniform saline cylinder as a
−SU MReal reference rather than a frame obtained for empty space. The
εr = ∂
+ 1. (6)
ωε0 ∂σ (SU MImag ) rod is now clearly visible, demonstrating that the information
Equation (1) predicts that Re(V/V), and hence SUMReal, on this internal structure is indeed contained within the
should be proportional to the relative permittivity of the target. measurements, well above the noise level. It appears,
The relative permittivity of saline solution falls slightly with therefore, that the inability of the absolute images to show
increasing salinity; Hilland [19] reported a drop of 1–3% as the plastic rod (figure 10(c)) or the femur (figure 10(a)) is due
the conductivity increased from 0 S m−1 to 1 S m−1. In the to the inadequacy of the image reconstruction algorithm and
results presented in figure 11, SUMReal fell by a much larger not the performance of the MIT data-collection system.
amount, by almost 50% over the same conductivity range; this
cannot be explained by a change in the relative permittivity 7. Conclusions
of the saline solution. Griffiths et al [13] observed a similar
behaviour in measurements made with a single channel and A 10 MHz experimental MIT system has been developed as
demonstrated that it was due to the skin effect. They found that a first stage towards a non-contacting biomedical imaging

8
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

system and an industrial imaging system for targets with (ii) the settling time for the filters and (iii) the time taken
conductivities of less than about 10 S m−1 that need to be to make the measurement and transfer it to the PC via the
viewed through an air gap or some other insulating barrier. GPIB. The total time taken by these processes is 375 ms. An
In a new application of MIT, a phantom simulating improved MIT system design would be fully synchronous,
stratified flow of an oil/gas/seawater mixture in a pipeline with signals derived from a single master oscillator distributed
was used to demonstrate imaging of the water component. The to all transmitter and receiver modules so that phase locking
images allowed the cross-section of the water to be visualized. would be needed only at the start of the measurement cycle.
It was found that the response curve of the normalized Higher-order filters with a shorter settling time and a more
integrated image value plotted against water fraction was efficient data transfer could also shorten the measurement time.
insensitive to the sensitivity matrix used for the image A very large improvement, by a factor of 15, could be achieved
reconstruction. Further development of the reconstruction by using parallel data acquisition rather than the present serial
algorithm must now take place and the system must be tested method.
with other flow regimes, e.g. bubble and annular (see [11]). In the reconstruction of images, the choice of the
The reference was obtained by acquiring a frame of data with regularization parameter, λ, was subjective, whereas an
the simulated pipeline removed from the MIT array. This objective method would have been preferable, e.g. the
would not be feasible in an industrial context as the MIT array L-curve or generalized cross-validation (GCV) methods.
would need to be permanently attached around the pipeline. We and others have previously used the L-curve method
Other methods of referencing the image must therefore be successfully for reconstructing differential images of low-
developed, e.g. the use of multiple frequencies. The present contrast perturbations relative to a conductive background
imaging system is very slow (one frame per 90 s) and it is likely [20]. For the absolute images reconstructed in this study, the
that much higher frame rates will be required for industrial method was found to fail because the curves did not form well-
use. However, the high conductivity of the process water defined ‘L’ shapes, instead showing multiple, weak maxima
(≈5 S m−1) relative to targets of interest in other applications of curvature. Similarly, the GCV method did not produce a
(e.g. biological tissues), should mean that a good SNR can be well-defined minimum and both methods resulted in under-
achieved with a much shorter integrating time and a consequent regularized images. Other workers have reported similar
increase in speed. difficulties and cited ‘geometry noise’ as an explanation, i.e.
Imaging of the thigh, in vivo, in a group of volunteers when the model used for computing the sensitivity matrix is too
was demonstrated. No internal anatomical structure was poor a fit to the true measurement system and systematic errors
discernable on the absolute images (referenced to empty dominate over random measurement noise [21]. An additional
space), but a method for estimating the equivalent uniform difficulty is that, even with accurate geometry, a sensitivity
conductivity and relative permittivity of the limb segment was matrix suitable for single-step reconstruction of absolute
demonstrated. A phantom experiment, using a saline cylinder images is unlikely to exist unless sufficient information is
representing the thigh with an insulating rod simulating the known about the conductivity distribution a priori. As far
femur, similarly showed that the rod could not be distinguished as we are aware, no one has yet demonstrated a successful,
on absolute images. However, difference images were able general, single-step method for reconstructing absolute images
clearly to show the rod. This indicates that the spatial of extended conductivity distributions (although it can be done
resolution and contrast was not limited by the noise in for well-separated small objects). It is likely that for absolute
the measurement system but by the image reconstruction imaging, iterative, nonlinear reconstruction methods will be
algorithm. The spatial resolution could be increased by necessary and this is the subject of current work in the MIT
using a smaller coil array so that the target occupied a larger community.
proportion of the imaging space, but improvement of the The short-term random noise level in the data-collection
image reconstruction algorithm is probably the main area system was measured as 17 millidegrees but the phase drift in
that needs development. The ability to perform difference the system was much larger, typically 800 millidegrees, and
imaging in vivo must also be explored and may allow was the main limitation to the accuracy of the measurements.
better distinction of the internal structures. Difference It is likely that the drift was due to temperature variations in
imaging was not possible in this in-vivo experiment (as it the components of the receiver. Two experiments in which
was not possible to remove the femur and then replace it!) the active components in the receiver circuit were heated,
but frequency-difference imaging in future might provide directly revealed that the mixer was by far the most sensitive
a suitable method. Time-difference imaging may also be component and should now be the focus of attention for the
appropriate for some biomedical applications (e.g. imaging next design.
regional lung ventilation), and some industrial ones (e.g. In order to reduce the short-term noise level, consideration
monitoring contamination in food products on a production must also be given to ‘backing off’ the primary signal, V, at
line) but would need a faster imaging system than the present the receiver coil. Watson et al [22], using a single channel,
one. achieved a primary-field cancellation factor of 300 over the
The speed of the system is not limited by the integration frequency range 1–10 MHz. They showed further that the
time but by a number of other processes within the lock-in noise level fell by a factor of over 40 when the primary
amplifier: (i) the time it takes to ‘acquire lock’, i.e. to measure field was backed off, suggesting that a significant contribution
the signal frequency and synthesize its own internal reference, to the noise was from short-term phase fluctuations between

9
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

the primary and reference signals, most probably from phase frequencies below 1 MHz are normally chosen as this is
noise in the reference. Rosell-Ferrer et al [10] have achieved where the β-dispersions of tissues mainly occur and there
primary-field cancellation using planar gradiometers and have is best contrast between different tissues. For performing
successfully constructed a multi-frequency MIT partial array tissue characterization by MIT, the method of Rosell-Ferrer
consisting of one excitation coil and 14 receivers. Additional et al [10] using gradiometers and multiple frequencies, may
projections were acquired by rotating the target relative to prove to be the most effective. For imaging haemorrhagic
the array, allowing MIT images to be reconstructed. In an cerebral stroke, it is interesting to note that even at 10 MHz,
accompanying paper, the group demonstrated retrieval of the the conductivity of blood is still at least three times higher than
conductivity spectrum over frequency (100–500 kHz) from that of brain tissue [18] so these higher frequencies, with their
the images of a potato immersed in saline solution [23]. resulting larger signals, may be suitable. For the new industrial
Thermal drift was identified as the main source of error in application discussed here, imaging the seawater fraction in an
the measurements. oil pipeline, a frequency of at least 10 MHz with single coils
An indication that phase noise in the reference signal may be optimal, possibly combined with an alternative, faster,
was significant in our MIT system is seen from figure 6. method of demodulation such as direct phase measurement [5,
The amplitude of the primary signal ranged over nearly three 14]. A focus by system designers on specific biomedical and
decades between the receivers closest to the transmitter and industrial applications is now necessary.
those furthest away (figure 6, lower curve). The noise in
the receiver amplifiers is expected to have a constant level
so when expressed as a percentage of the primary signal, V, it Acknowledgements
will vary by the same factor as V itself. The observed variation
was very much less than this and by only a factor of 4 over This work was supported by the Engineering and Physical
the range of channel separations (from approximately 1–4%, Sciences Research Council (grants EP/E009832/1 and
taking the average for each channel separation: figure 6, upper EP/E009697/1). The authors thank Mr C Craven for
curve). Another source of noise is the lock-in amplifier which mechanical engineering assistance.
continually synthesizes its internal reference signal from the
analogue signal applied to its reference channel. This synthetic References
reference will have a small and varying phase error relative to
the actual reference signal [24]. This will contribute noise to [1] Peyton A, Mackin R, Goss D, Wan-Daud W, Crescenzo E,
the measurement of V that will be proportional to V. Thus, Saunders N and Tapp H 2003 Addressing the difficulties in
this component of the noise in V/V will be independent of V using inductive methods to evaluating human body
composition Biomét. Hum. Anthropol. 21 63–71
and is the most likely explanation of the much smaller range [2] Griffiths H 2001 Magnetic induction tomography Meas. Sci.
of values seen in the upper curve of figure 6 than in the lower. Technol. 12 1126–31
A possible source of systematic error is crosstalk in the [3] Tapp H S and Peyton A J 2003 A state of the art review of
multiplexers as they have to handle the complete range of electromagnetic tomography Proc. 3rd World Congress on
amplitudes of the 10 kHz signals (see section 3.6). The inter- Industrial Process Tomography, Banff, Canada pp 340–6
[4] Griffiths H 2005 Magnetic Induction Tomography Electrical
channel isolation was measured to be at least 70 dB between Impedance Tomography: Methods, History and
adjacent channels of the multiplexer and 98 dB for channels Applications ed D S Holder (Bristol: Institute of Physics
with the greatest separation. In the least favourable case, the Publishing) pp 213–38
primary signal received in the channel diametrically opposite [5] Korjenevsky A, Cherepenin V and Sapetsky S 2000 Magnetic
the excitor will have an amplitude of 1.4 mV (figure 6). The induction tomography: experimental realization Physiol.
Meas. 21 89–94
reference will contribute a crosstalk of about 0.03 mV (i.e. [6] Korzhenevsky A and Sapetsky S 2001 Visualisation of the
2.1 V reduced by 98 dB), and this is equal to about 2% of internal structure of extended conducting objects by
the received primary signal. The secondary signals of interest magnetoinduction tomography Bull. Russian Acad. Sci.:
(V) will typically be a few per cent of V or less, so the Phys. 65 1945–9
crosstalk will be significant in comparison. However, if the [7] Watson S, Morris A, Williams R J, Gough W and Griffiths H
2002 The Cardiff MIT system Proc. IFMBE:EMBEC02,
cross-talk does not vary with time, it will effectively become Vienna, Austria (Dec 4–8) vol 3 part 1, pp 116–7
part of the measured primary signals and will have very [8] Karbeyaz B U and Gencer N G 2003 Electrical conductivity
little effect on the measured values of V/V. Nevertheless, imaging via contactless measurements: an experimental
it is an undesirable contamination of the true signals so should study IEEE Trans. Med. Imag. 22 627–35
be borne in mind in future system design. [9] Igney C H, Watson S, Williams R J, Griffiths H and Dösse1 O
2005 Design and performance of a planar-array MIT system
It is likely that the design of the next generation of MIT with normal sensor alignment Physiol. Meas. 26 S263–78
systems will depend on the particular applications targeted, [10] Rosell-Ferrer J, Merwa R, Brunner P and Scharfetter H 2006
even within the ‘low-conductivity’ range (<10 S m−1). These A multifrequency magnetic induction tomography system
differences could be, for example, in the type of coils using planar gradiometers: data collection and calibration
used (single coil or gradiometer), the choice of operating Physiol. Meas. 27 S271–80
[11] Albrechtsen R A, Yu Z Z and Peyton A J 1995 Preliminary
frequency and the method of demodulation (vector voltmeter, experiments on the investigation of the inductive technique
direct phase measurement or direct digitization of the high- for measuring water content in multiphase flow Proc.
frequency signals). For biological tissue characterization, ECAPT, Bergen pp 205–13

10
Meas. Sci. Technol. 19 (2008) 045501 S Watson et al

[12] Hammer E A and Fossdal G 2002 A new water-in-oil monitor [19] Hilland J 1997 Simple sensor system for measuring the
based on high frequency magnetic field excitation Proc. 2nd dielectric properties of saline solutions Meas. Sci.
Int. Symp. on Process Tomography, Wroclaw, Poland (Sept. Technol. 8 901–10
11–12) pp 9–16 [20] Watson S, Igney C H, Dösse1 O, Williams R J and
[13] Griffiths H, Watson S, Williams R J and Gough W 2007 Griffiths H 2005 A comparison of sensors for
Residual capacitive coupling and the measurement of minimising the primary signal in planar-array
permittivity in magnetic induction tomography Physiol. magnetic induction tomography Physiol. Meas.
Meas. 28 S301–11 26 S319–31
[14] Watson S, Williams R J, Griffiths H, Gough W and Morris A [21] Johnson P R and Gulrajani R M 2000 Selecting the corner in
2003 Magnetic induction tomography: phase vs. vector the L-curve approach to Tikhonov regularisation IEEE
voltmeter measurement techniques Physiol. Trans. Biomed. Eng. 47 1293–6
Meas. 24 555–64 [22] Watson S, Williams R J, Griffiths H and Gough W 2004 A
[15] Morris A, Griffiths H and Gough W 2001 A numerical model primary field compensation scheme for planar array
for magnetic induction tomographic measurements in magnetic induction tomography Physiol. Meas.
biological tissues Physiol. Meas. 22 113–9 25 271–9
[16] Geselowitz D 1971 An application of electrocardiographic [23] Brunner P, Merwa R, Missner A, Rosell J, Hollaus K and
lead theory to impedance plethysmography IEEE Trans. Scharfetter H 2006 Reconstruction of the shape of the
Biomed. Eng. 18 38–41 conductivity spectra using differential multi-frequency
[17] Hansen P C 1994 Regularization tools: a Matlab package for magnetic induction tomography Physiol. Meas.
analysis and solution of discrete ill-posed problems Numer. 27 S237–48
Algorithms 6 1–35 [24] Meade M L 1983 Lock-in Amplifiers: Principles and
[18] Gabriel C, Gabriel S and Corthout E 1996 The dielectric Applications, IEE Electrical Measurement Series 1
properties of biological tissues: I. Literature survey Phys. (London: P. Peregrinus) (Chapter 4 ‘Lock-in amplifier
Med. Biol. 41 2231–49 specifications’, pp 70–1)

11