Professional Documents
Culture Documents
ch17nonpoint._cropped
ch17nonpoint._cropped
ch17nonpoint._cropped
Electrical Impedance
Tomography
17-1
S C
model of bulk tissue impedance based on this structure, which was proposed by Cole and Cole [1941], is
shown in Figure 17.1.
Clearly, this model as it stands is too simple, since an actual tissue sample would be better represented
as a large network of interconnected modules of this form. However, it has been shown that this model fits
experimental data if the values of the components, especially the capacitance, are made a power function
of the applied frequency ω. An equation which describes the behavior of tissue impedance as a function
of frequency reasonably well is
Z0 − Z∞
Z = Z∞ + (17.1)
1 + ( j(f /fc ))α
where Z0 and Z∞ are the (complex) limiting values of tissue impedance low and high frequency and fc is
a characteristic frequency. The value of α allows for the frequency dependency of the components of the
model and is tissue dependent. Numerical values for in vivo human tissues are not well established.
Making measurements of the real and imaginary components of tissue impedivity over a range of
frequencies will allow the components in this model to be extracted. Since it is known that tissue structure
alters in disease and that R, S, C are dependent on structure, it should be possible to use such measurements
to distinguish different types of tissue and different disease conditions. It is worth noting that although
maximum accuracy in the determination of the model components can be obtained if both real and
imaginary components are available, in principle, knowledge of the resistive component alone should
enable the values to be determined, provided an adequate range of frequencies is used. This can have
practical consequences for data collection, since accurate measurement of the capacitive component can
prove difficult.
where φ is the potential distribution within the object and σ is the distribution of conductivity (generally
admittivity) within the object. If the conductivity is uniform, this reduces to Laplace’s equation. Strictly
speaking, this equation is only correct for direct current, but for the frequencies of alternating current used
in EIT (up to 1 MHz) and the sizes of objects being imaged, it can be assumed that this equation continues
to describe the instantaneous distribution of potential within the conducting object. If this equation is
solved for a given conductivity distribution and current distribution through the surface of the object,
the potential distribution developed on the surface of the object may be determined. The distribution of
potential will depend on several things. It will depend on the pattern of current applied and the shape
of the object. It will also depend on the internal conductivity of the object, and it is this that needs to
be determined. In theory, the current may be applied in a continuous and nonuniform pattern at every
point across the surface. In practice, current is applied to an object through electrodes attached to the
surface of the object. Theoretically, potential may be measured at every point on the surface of the object.
Again, voltage on the surface of the object is measured in practice using electrodes (possibly different from
those used to apply current) attached to the surface of the object. There will be a relationship, the forward
solution, between an applied current pattern ji , the conductivity distribution σ , and the surface potential
distribution φi which can be formally represented as
φi = R(ji , σ ) (17.3)
If σ and ji are known, φi can be computed. For one current pattern ji , knowledge of φi is not in general
sufficient to uniquely determine σ . However, by applying a complete set of independent current patterns,
it becomes possible to obtain sufficient information to determine σ , at least in the isotropic case. This is
the inverse solution. In practice, measurements of surface potential or voltage can only be made at a finite
number of positions, corresponding to electrodes placed on the surface of the object. This also means that
only a finite number of independent current patterns can be applied. For N electrodes, N −1 independent
current patterns can be defined and N (N − 1)/2 independent measurements made. This latter number
determines the limit of image resolution achievable with N electrodes. In practice, it may not be possible
to collect all possible independent measurements. Since only a finite number of current patterns and
measurements is available, the set of equations represented by Equation 17.3 can be rewritten as
v = Ac c (17.4)
where v is now a concatenated vector of all voltage values for all current patterns, c is a vector of
conductivity values, representing the conductivity distribution divided into uniform image pixels, and
Ac a matrix representing the transformation of this conductivity vector into the voltage vector. Since Ac
depends on the conductivity distribution, this equation is nonlinear. Although formally the preceding
equation can be solved for c by inverting Ac , the nonlinear nature of this equation means that this cannot
be done in a single step. An iterative procedure will therefore be needed to obtain c.
Examination of the physics of current flow shows that current tends to take the easiest path possible in
its passage through the object. If the conductivity at some point is changed, the current path redistributes
in such a way that the effects of this change are minimized. The practical effect of this is that it is possible
to have fairly large changes in conductivity within the object which only produce relatively small changes
in voltage at the surface of the object. The converse of this is that when reconstructing the conductivity
distribution, small errors on the measured voltage data, both random and systematic, can translate into
large errors in the estimate of the conductivity distribution. This effect forms, and will continue to form, a
limit to the quality of reconstructed conductivity images in terms of resolution, accuracy, and sensitivity.
Any measurement of voltage must always be referred to a reference point. Usually this is one of the
electrodes, which is given the nominal value of 0 V. The voltage on all other electrodes is determined by
measuring the voltage difference between each electrode and the reference electrode. Alternatively, voltage
differences may be measured between pairs of electrodes. A common approach is to measure the voltage
between adjacent pairs of electrodes (Figure 17.2). Clearly, the measurement scheme affects the form
of Ac . Choice of the pattern of applied currents and the voltage measurement scheme used can affect the
accuracy with which images of conductivity can be reconstructed.
Electrical impedance tomography (EIT) is not a mature technology. However, it has been the subject
of intensive research over the past few years, and this work is still continuing. Nearly all the research effort
has been devoted to exploring the different possible ways of collecting data and producing images of tissue
resistivity, with the aim of optimizing image reconstruction in terms of image accuracy, spatial resolution,
and sensitivity.
1
16 2
3 Drive
15
14
4
13 5
6
12
7
11 Receive
10 8
9
FIGURE 17.2 Idealized electrode positions around a conducting object with typical drive and measurement electrode
pairs indicated.
Very few areas of medical application have been explored in any great depth, although in a number
of cases preliminary work has been carried out. Although most current interest is in the use of EIT for
medical imaging, there is also some interest in its use in geophysical measurements and some industrial
uses. A recent detailed review of the state of Electrical Impedance Tomography is given in Boone et al.
[1997].
Clearly, the magnitude of the voltage measured will depend on the magnitude of the current applied.
If a constant-current drive is used, this must be able to deliver a known current to a variety of input
impedances with a stability of better than 0.1%. This is technically demanding. The best approach to
this problem is to measure the current being applied, which can easily be done to this accuracy. These
measurements are then used to normalize the voltage data.
The current application and data-collection regime will depend on the reconstruction algorithm used.
Several EIT systems apply current in a distributed manner, with currents of various magnitudes being
applied to several or all of the electrodes. These optimal currents (see Image Reconstruction) must be
specified accurately, and again, it is technically difficult to ensure that the correct current is applied at each
electrode. Although there are significant theoretical advantages to using distributed current patterns, the
increased technical problems associated with this approach, and the higher noise levels associated with
the increase in electronic complexity, may outweigh these advantages.
Although most EIT at present is 2D in the sense given above, it is intrinsically a three-dimensional
(3D) imaging procedure, since current cannot be constrained to flow in a plane through a 3D object.
3D data collection does not pose any further problems apart from increased complexity due to the need
for more electrodes. Whereas most data-collection systems to date have been based on 16 or 32 elec-
trodes, 3D systems will require four times or more electrodes distributed over the surface of the object
if adequate resolution is to be maintained. Technically, this will require “belts” or “vests” of electrodes
that can be rapidly applied [McAdams et al., 1994]. Some of these are already available, and the applic-
ation of an adequate number of electrodes should not prove insuperable provided electrode-mounted
electronics are not required. Metherell et al. [1996] describe a three-dimensional data collection system
and reconstruction algorithm and note the improved accuracy of three-dimensional images compared to
two-dimensional images constructed using data collected from three-dimensional objects.
distribution c0 . This is usually taken to be uniform. Using Ac , the expected voltages v0 are calculated and
compared with the actual measured voltages vm . Unless c0 is correct (which it will not be initially), v0 and
vm will differ. It can be shown that an improved estimate of c is given by
c1 = c0 + c (17.6)
where Sc is the differential of Ac with respect to c, the sensitivity matrix and Sct is the transpose of Sc . The
improved value of c is then used in the next iteration to compute an improved estimate of vm , that is, v1 .
This iterative process is continued until some appropriate endpoint is reached. Although convergence is
not guaranteed, in practice, convergence to the correct c in the absence of noise can be expected, provided
a good starting value is chosen. Uniform conductivity seems to be a reasonable choice. In the presence
of noise on the measurements, iteration is stopped when the difference between v and vm is within the
margin of error set by the known noise on the data.
There are some practical difficulties associated with this approach. One is that large changes in c may
only produce small changes in v, and this will be reflected in the structure of Sc , making Sct Sc very difficult
to invert reliably. Various methods of regularization have been used, with varying degrees of success,
to achieve stable inversion of this matrix although the greater the regularization applied the poorer the
resolution that can be achieved. A more difficult practical problem is that for convergence to be possible
the computed voltages v must be equal to the measured voltages vm when the correct conductivity values
are used in the forward calculation. Although in a few idealized cases analytical solutions of the forward
problem are possible, in general, numerical solutions must be used. Techniques such as the finite-element
method (FEM) have been developed to solve problems of this type numerically. However, the accuracy of
these methods has to be carefully examined [Paulson et al., 1992] and, while they are adequate for many
applications, may not be adequate for the EIT reconstruction problem, especially in the case of 3D objects.
Accuracies of rather better than 1% appear to be required if image artifacts are to be minimized. Consider
a situation in which the actual distortion of conductivity is uniform. Then the initial v should be equal
to the vm to an accuracy less than the magnitude of the noise. If this is not the case, then the algorithm
will alter the conductivity distribution from uniform, which will clearly result in error. While the required
accuracies have been approached under ideal conditions, there is only a limited amount of evidence at
present to suggest that they can be achieved with data taken from human subjects.
17.3.2.2 Optimal Current Patterns
So far little has been said about the form of the current patterns applied to the object except that a
set of independent patterns is needed. The simplest current patterns to use are those given by passing
current into the object through one electrode and extracting current through a second electrode (a
bipolar pattern). This pattern has the virtue of simplicity and ease of application. However, other current
patterns are possible. Current can be passed simultaneously through many electrodes, with different
amounts passing through each electrode. Indeed, an infinite number of patterns are possible, the only
limiting condition being that the magnitude of the current flowing into the conducting object equals the
magnitude of the current flowing out of the object. Isaacson [1986] has shown that for any conducting
object there is a set of optimal current patterns and has provided an algorithm to compute them even
if the conductivity distribution is initially unknown. Isaacson showed that by using optimal patterns,
significant improvements in sensitivity could be obtained compared with simpler two-electrode current
patterns. However, the additional computation and hardware required to use optimal current patterns
compared with fixed, nonoptimal patterns is considerable.
Use of suboptimal patterns close to optimum also will produce significant gains. In general, the optimal
patterns are very different from the patterns produced in the simple two-electrode case. The optimal
patterns are often cosine-like patterns of current amplitude distributed around the object boundary
rather than being localized at a pair of points, as in the two-electrode case. Since the currents are passed
simultaneously through many electrodes, it is tempting to try and use the same electrodes for voltage
measurements. This produces two problems. As noted above, measurement of voltage on an electrode
through which an electric current is passing is compromised by the presence of electrode resistance, which
causes a generally unknown voltage drop across the electrode, whereas voltage can be accurately measured
on an electrode through which current is not flowing using a voltmeter of high input impedance. In
addition, it has proved difficult to model current flow around an electrode through which current is flowing
with sufficient accuracy to allow the reliable calculation of voltage on that electrode, which is needed for
accurate reconstruction. It seems that separate electrodes should be used for voltage measurements with
distributed current systems.
Theoretically, distributed (near-) optimal current systems have some advantages. As each of the optimal
current patterns is applied, it is possible to determine if the voltage patterns produced contain any useful
information or if they are simply noise. Since the patterns can be generated and applied in order of
decreasing significance, it is possible to terminate application of further current patterns when no further
information can be obtained. A consequence of this is that SNRs can be maximized for a given total data-
collection time. With bipolar current patterns this option is not available. All patterns must be applied.
Provided the SNR in the data is sufficiently good and only a limited number of electrodes are used, this may
not be too important, and the extra effort involved in generating the optimal or near-optimal patterns may
not be justified. However, as the number of electrodes is increased, the use of optimal patterns becomes
more significant. It also has been suggested that the distributed nature of the optimal patterns makes the
forward problem less sensitive to modeling errors. Although there is currently no firm evidence for this,
this seems a reasonable assertion.