IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO.

7, JULY 1998 899

An Iterative Newton–Raphson Method to Solve

the Inverse Admittivity Problem
Peter M. Edic,* Member, IEEE, David Isaacson, Member, IEEE, Gary J. Saulnier, Senior Member, IEEE,
Hemant Jain, and Jonathan C. Newell, Member, IEEE

Abstract— By applying electrical currents to the exterior of denotes the body of the object, while denotes the
a body using electrodes and measuring the voltages developed surface of the object;
on these electrodes, it is possible to reconstruct the electrical complex potential distribution in and on ;
properties inside the body. This technique is known as electri-
cal impedance tomography. The problem is nonlinear and ill admittivity of . It is defined as , where
conditioned meaning that a large perturbation in the electrical is the electrical conductivity, is the angular
properties far away from the electrodes produces a small voltage frequency of the applied current waveform, and
change on the boundary of the body. is the electrical permittivity;
This paper describes an iterative reconstruction algorithm that current density applied on ;
yields approximate solutions of the inverse admittivity problem in
two dimensions. By performing multiple iterations, errors in the outward normal of defined on .
conductivity and permittivity reconstructions that result from a In the text that follows, is used to denote that is a
linearized solution to the problem are decreased. A finite-element complex quantity.
forward-solver, which predicts voltages on the boundary of the Equation (1) specifies that no source resides in the body
body given knowledge of the applied current on the boundary and
the electrical properties within the body, is required at each step and that charge is not allowed to accumulate in the body.
of the reconstruction algorithm. Reconstructions generated from Equation (2) constrains the current entering to be the integral
numerical data are presented that demonstrate the capabilities of the applied current density. Equation (3) is necessary for
of this algorithm. the existence of a unique solution to the system model.
Index Terms— Conductivity, impedance imaging, iterative re- With knowledge of all current densities applied on and
construction methods, least-squares reconstruction methods, per- the voltages measured on , can be determined for all
mittivity. positions in the body [2]. However, in practice, currents are
applied and voltages measured on at discrete regions using
I. INTRODUCTION electrodes. Since a finite number of voltage measurements are
recorded, it is possible to determine the admittivity distribution

E LECTRICAL properties of a patient or object may be

computed from electrical measurements made on the
surface of the body using a technology known as electrical
using a finite number of degrees of freedom. For a system with
electrodes, degrees of freedom can be
computed [1] for the admittivity distribution. The number of
impedance tomography (EIT). Ideally, the voltage on the degrees of freedom can be distributed as piecewise-constant
boundary of the object is measured while a current density is regions in the domain or multiplicative constants of basis
applied to its surface. The mathematical model that describes functions that exist over the entire region .
the system is [1] Several reconstruction techniques have been used to com-
pute the electrical properties within the region . Of par-
in (1) ticular interest are filtered backprojection and least-squares
techniques. In practice, the filtered backprojection technique
on (2)
projects a voltage difference signal, produced from measure-
ments on the boundary of the body made under two different
(3)
sets of conditions, along isopotential lines in the body [3]–[5].
This technique has been used mostly to generate dynamic
where images of the body and display the effects of a change in
the electrical properties of the body at two specific conditions.
Manuscript received August 27, 1996; revised February 2, 1998. Asterisk In the least-squares method, an objective function is se-
indicates corresponding author.
*P. M. Edic is with the General Electric Corporate Research and Develop- lected, and the residuals of this function are minimized [1],
ment, Schenectady, NY, 12309 USA (e-mail: edic@crd.ge.com). [6]–[11]. Iterative approaches used to minimize the objective
D. Isaacson is with the Department of Mathematical Sciences, Rensselaer function are presented in [8] and [9]; these papers consider
Polytechnic Institute, Troy, NY 12180 USA.
G. J. Saulnier is with the Department of Electrical, Computer, and Systems only the conductivity problem and do not discuss the admit-
Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180 USA. tivity problem.
H. Jain is with Epic Systems Corporation, Madison, WI 53711 USA. An important topic in EIT is the notion of distinguishability.
J. C. Newell is with the Department of Biomedical Engineering, Rensselaer
Polytechnic Institute, Troy, NY 12180 USA. This quantity is used to characterize the ability of a recon-
Publisher Item Identifier S 0018-9294(98)04445-0. struction algorithm in distinguishing two separate admittivity
0018–9294/98$10.00  1998 IEEE
900 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO. 7, JULY 1998

distributions. Although this topic is not discussed here, [12]

and [13] report on this topic for the inverse conductivity
problem.
This paper describes an iterative method to solve the ad-
mittivity problem. Better estimates of the true permittivity
and conductivity distributions are reconstructed when complex
voltage data on are used as input to the reconstruction
algorithm. Reconstructions of numerical data are presented
that demonstrate this effect.

II. DESCRIPTION OF THE ALGORITHM

An earlier reconstruction algorithm known as Newton’s
one-step error reconstructor (NOSER), developed in [1], uses
one step of a Newton–Raphson method to solve a nonlinear
system of equations that results from the minimization of an
error functional involving voltages measured on and those
predicted from a forward-solver. The forward-solver used in
that algorithm predicts the boundary voltages given knowledge
of either the conductivity or permittivity distribution in and
the current waveforms applied to .
The present complex Newton’s recursive-step error recon-
Fig. 1. The Joshua Tree Mesh used in the CNRSER algorithm.
structor (CNRSER) algorithm performs multiple steps of the
Newton–Raphson method in solving a nonlinear system of
equations, resulting from an error expression that is similar to The following error functional is used in the iterative
that used in the NOSER algorithm. However, the CNRSER reconstruction algorithm:
algorithm iteratively solves the full complex problem, over-
coming two limitations of NOSER which implements only one
step of the Newton–Raphson method to solve either the inverse
conductivity or permittivity problem separately. Yorkey et
al. demonstrated that the iterative Newton–Raphson method
is desirable since it converges to the smallest error between
measured and predicted voltages in the fewest iterations of (8)
the methods tested [10].

A. One Step of the CNRSER Algorithm

In the CNRSER algorithm, the admittivity distribution con-
where
sists of piecewise-constant regions and is defined as
number of the electrode on ;
number of the current pattern applied on to all
(4) electrodes simultaneously, ;
forward-solver voltage predicted on the th elec-
trode given that the th current pattern was applied
where is unity on the th mesh element and zero every-
on and that characterizes the electrical prop-
where else. Fig. 1 shows a picture of the mesh elements used
erties of ;
in the CNRSER algorithm, denoted as the Joshua Tree Mesh. It
experimental voltage measured on the th electrode
is possible to define another quantity which is the reciprocal
when the th current pattern was applied;
of , called the impedivity, and is defined as
error functional used in the NOSER algorithm;
gradient regularization term that forces the recon-
(5) struction to be spatially smooth by penalizing the
error functional by the squared-magnitude of the
gradient of the impedivity distribution;
where gain factor applied to the regularization term;
denotes the complex conjugate of .
(6)
In particular, denotes the complex, forward-solver
and voltage that results from the application of the th current
(7) pattern of the canonical set, shown in (9) at the bottom of the
next page.
EDIC et al.: ITERATIVE NEWTON–RAPHSON METHOD TO SOLVE INVERSE ADMITTIVITY PROBLEM 901

The error functional can be rewritten in terms of real to iteratively compute the solution to this nonlinear system of
quantities as equations. Here, is the Jacobian of

(16)

(10) Four separate expressions for the components of are given

in the Appendix. Although the second-order partial derivatives
of the potential terms in these expressions affect algorithm con-
where denotes the real part of and denotes vergence, they are difficult to compute. However, after a few
iterations, the difference between the predicted voltages and
the imaginary part of . the measured voltages, the factor multiplying these terms, will
The goal of the reconstruction algorithm is to find and be small. For this reason, these terms are not computed and
for that minimize this error functional. A simply neglected in the reconstruction algorithm. Likewise,
standard method [14], [15] for minimizing (10) is to let the second-order partial derivatives of the real and quadrature
components of the impedivity vanish with a suitable choice of
basis functions for these two quantities (discussed later).
As described in [1] and [10], Marquardt regularization is
a useful technique to improve the conditioning of the Jaco-
bian matrix used in solving the inverse admittivity problem.
However, after several steps of the CNRSER algorithm, the
computed impedivity distribution becomes very peaked using
this regularization method for certain impedivity distributions.
(11) The gradient regularization term, the expression in (8),
forces the reconstructed impedivity distribution to be spatially
smooth at the expense of increased error between measured
and predicted voltages on .
The orthonormal canonical current patterns specified in
(9) are denoted by . Neglecting
second-order partial derivatives, the expressions defined in the
Appendix can be written as
(12)

and solve the vector equation

(13) (17)

by a Newton–Raphson method. We define

(14)
(18)
and use this method
(19)
(15)

(9)
902 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO. 7, JULY 1998

(20)
Fig. 2. Basis functions used to develop an approximation to the piecewise
constant impedivity distribution.
where
with respect to and are also computed using these
expressions.
(21)
The gain factor in (8) is chosen so the root-mean-
squared (rms) error between measured and predicted voltages
and is defined as the vector inner product of vectors is comparable to the measurement precision of the tomograph.
and . The remaining terms in these equations can be inferred Hence, the value can be iteratively chosen to satisfy this
from the definition in (21). Using the complex equivalent of constraint.
a result derived in [1], one can show that
B. Multiple Steps of the CNRSER Algorithm
The quantities defined in the previous section must be
computed at each iteration of the CNRSER algorithm. For
(22) the first step of the CNRSER algorithm, the conductivity and
permittivity distributions are chosen to be constant. A guess is
made at these values by minimizing (10) with respect to two
constants: one representing the real part of the homogeneous
impedivity and one representing the imaginary part of the
(23) homogeneous impedivity.
To compute the quantities in (15), a forward-solver that
(24) uses the piecewise-constant admittivity distribution must be
employed at each step. A finite-element forward solution is
(25)
used for this purpose. Each element of the Joshua Tree Mesh,
where the dependence on has been omitted for simplicity. shown in Fig. 1, is segmented into several triangles. The
The integral is computed over the area enclosed by the triangulation of the Joshua Tree Mesh is used by the forward-
piecewise-constant mesh element, denoted as . solver to compute voltages at each node in and on . (A
Since the impedivity distribution is piecewise-constant, the node refers to a vertex of a triangle in the mesh.) These data are
gradient of this distribution is zero in the area enclosed by the then used to compute and at each iteration. The CNRSER
mesh elements and infinity at the edges of the mesh elements. algorithm computes a new estimate of , which is used by the
To approximate the gradient, the center-point of each mesh forward-solver to generate the predicted voltages. The process
element was computed and a basis function was constructed continues until the reconstruction converges to its final result.
that linearly interpolated in the radial and angular directions
between adjacent mesh element centers. Fig. 2 shows a typical III. METHODS
basis function; it has an amplitude of 1.0 at the center of The CNRSER algorithm was tested using numerical data.
the mesh element of interest and linearly interpolates to 0.0 The Joshua Tree Mesh was segmented into 9216 triangles for
at the boundaries of the basis function. The gradient of the the finite-element forward-solver algorithm.
impedivity distribution is then approximated by computing The electrode model used to generate the voltage data
the gradient of sets was the gap model [1]. In this model, the current in
each electrode is specified by (9) and the current density
(26) is assumed to be constant over the electrode region. The
current density in gap regions between electrodes is assumed
The gradients of and are to be zero. In [16], [17], and [18], more accurate models
for the electrodes are specified. Although these models will
(27) yield better reconstructions of experimental data, they are not
incorporated in our model since the reconstruction algorithm
presented here assumes the gap model for the electrodes. In
(28) this manner, electrode modeling errors will not corrupt the
results to be presented.
where , is a basis function typical of the To determine the precision of the forward-solver, a model of
one shown in Fig. 2. The partial derivatives of and a circular phantom consisting of a homogeneous conductivity
EDIC et al.: ITERATIVE NEWTON–RAPHSON METHOD TO SOLVE INVERSE ADMITTIVITY PROBLEM 903

Fig. 3. Admittivity distribution consisting of a homogeneous conductivity

distribution and a permittivity distribution with a step inhomogeneity. (Units
are millisiemens per meter.)

was developed. Both the finite-element forward-solver and an

analytical expression were used to compute boundary voltages
on when the current patterns specified in (9) were applied to
the electrodes. Using the voltages calculated with the analytical Fig. 4. Admittivity distribution simulating a cross section of a human chest
containing lungs and a heart. (Admittivity units are millisiemens per meter.)
model as the true values, the maximum error in voltages
predicted on by the finite-element forward-solver was 0.12% computed using complex voltage data to demonstrate the
of the maximum voltage amplitude that was computed. benefits of the CNRSER algorithm.
The real component of admittivity is conductivity while For all reconstructions, Marquardt regularization (gain equal
the imaginary component is permittivity scaled by the angular to 0.01) was applied to the Jacobian matrix. The gain of
frequency of the applied current waveform. However, in the the gradient regularization term in (8) was chosen to be zero.
text that follows, the imaginary part of the admittivity is simply
denoted as the permittivity. This nomenclature will make B. Numerical Voltage Data Generated
discussions more readable as well as allow direct comparison from a Simulated Human Chest
of magnitudes of both the real and imaginary components of Fig. 4 shows a cross section of the phantom used to simulate
the admittivity. a human chest. Two regions are included which simulate
human lungs as well as a region that simulates a human heart.
A. Numerical Voltage Data Generated Reasonable admittivity values for the objects in the phantom
from an Analytical Model were based on experimental data collected from living tissue
Given the current density on the boundary of the Joshua as reported in [19]. In particular, the admittivity of the lungs is
Tree Mesh, one can write the Fourier series for the current mS/m, the admittivity of the heart is mS/m,
density applied on since this function is periodic in angular and the admittivity of the remaining portion of the phantom is
position. For the analytical model, the conductivity distribution mS/m, the value typical of muscle tissue.
is circularly symmetric with a centered inhomogeneity, as A boundary-element forward-solver was used to generate
shown in Fig. 3. The admittivity distribution is homogeneous the boundary voltage data for this geometry. This algorithm
in conductivity (100 mS/m) and has a step inhomogeneity in used the gap model for electrode regions, which is compat-
permittivity (200 mS/m) at a radius equaling 0.5 of the total ible with the reconstruction algorithm. Suitable noise from a
radius of the object. uniform deviate was added to the data to simulate a voltmeter
The equation for the boundary voltage is developed by precision of 1 part in 4096 (12 bits). The voltage data corrupted
solving the boundary-value problem using Laplace’s Equation with noise were used as input to the CNRSER algorithm
in each homogeneous region in Fig. 3. This solution was used to generate a reconstruction of the phantom. Reconstructions
to generate complex voltage data that were used as input to of the real and quadrature voltage data were also computed
the CNRSER algorithm. A figure-of-merit for determining the separately to demonstrate the benefits of the CNRSER al-
benefit of the iterative solution will be the reduction in the gorithm and the reconstruction errors that occur if the real
squared error of the difference between measured voltages and and quadrature components of the voltages measured on
those predicted by the forward-solver. are considered separately. For all reconstructions, Marquardt
The real and quadrature voltage files were also used sepa- regularization (gain equal to 0.2) was applied to the Jacobian
rately as inputs to the reconstruction algorithm. Reconstruc- matrix; the gain of the gradient regularization term in (8)
tions of these data will be compared to the reconstructions was chosen to be zero.
904 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO. 7, JULY 1998

Fig. 5. Conductivity distributions as a function of the radial distance from Fig. 7. Conductivity distributions as a function of the radial distance from
the center of the Joshua Tree Mesh, calculated after each of four iterations the center of the Joshua Tree Mesh, calculated after each of four iterations
of the CNRSER algorithm using complex voltage data generated from the of the CNRSER algorithm using separate real and quadrature voltage data
geometry in Fig. 3. generated from the geometry in Fig. 3.

Fig. 6. Permittivity distributions as a function of the radial distance from Fig. 8. Permittivity distributions as a function of the radial distance from the
the center of the Joshua Tree Mesh, calculated after each of four iterations center of the Joshua Tree Mesh, calculated after each of four iterations of the
of the CNRSER algorithm using complex voltage data generated from the CNRSER algorithm using separate real and quadrature voltage data generated
geometry in Fig. 3. from the geometry in Fig. 3.

Reconstructions of the complex voltage data corrupted with used to display the output of the algorithm. Figs. 5 and 6 show
noise were computed using gradient regularization (gain equal the reconstructed conductivity and permittivity distributions,
to 1.0) to demonstrate the benefits of this technique. When respectively, after several iterations of the CNRSER algorithm
using experimental data with the CNRSER algorithm, should as a function of the radial distance from the center of the
be selected so that the rms value of the residual error in (8) is admittivity distribution shown in Fig. 3. The squared error
comparable to system noise in the tomograph. between measured and predicted voltages decreased by a factor
of 2.05e5 after four iterations.
IV. RESULTS Figs. 7 and 8 show cross sections of the reconstructed
conductivity and permittivity distributions, respectively, when
A. Output of the CNRSER Algorithm Using the real and quadrature voltage files were used separately
the Analytical Model as input to the reconstruction algorithm. For this case, the
squared-error between the real components of the measured
Since the admittivity distribution shown in Fig. 3 is cir- and predicted voltages decreased by a factor of 7.15e2 after
cularly symmetric, the output of the CNRSER algorithm is four iterations while the squared error between the imaginary
also circularly symmetric; hence, radial cross sections of the components of the measured and predicted voltages decreased
reconstructed conductivity and permittivity distributions are by a factor of 1.59e5.
EDIC et al.: ITERATIVE NEWTON–RAPHSON METHOD TO SOLVE INVERSE ADMITTIVITY PROBLEM 905

Fig. 9. Conductivity distribution as a function of angular position, calculated Fig. 11. Conductivity distribution as a function of angular position, cal-
after several iterations of the CNRSER algorithm using complex voltage culated after several iterations of the CNRSER algorithm using separate
data generated from the geometry in Fig. 4. Marquardt regularization of the real and quadrature voltage data generated from the geometry in Fig. 4.
Jacobian matrix was implemented for these reconstructions. Marquardt regularization of the Jacobian matrix was implemented for these
reconstructions.

Fig. 10. Permittivity distribution as a function of angular position, calculated

after several iterations of the CNRSER algorithm using complex voltage Fig. 12. Permittivity distribution as a function of angular position, calculated
data generated from the geometry in Fig. 4. Marquardt regularization of the after several iterations of the CNRSER algorithm using separate real and
Jacobian matrix was implemented for these reconstructions. quadrature voltage data generated from the geometry in Fig. 4. Marquardt
regularization of the Jacobian matrix was implemented for these reconstruc-
tions.
It is important to note that the reconstruction algorithm
could not exactly reconstruct the true distribution since the
algorithm. Marquardt regularization (gain equal to 0.2) was
boundary of the centered inhomogeneity was purposely chosen
used for each of these reconstructions. The squared error
so as not to coincide with a boundary of the Joshua Tree Mesh.
between the real components of the measured and predicted
voltages decreased by a factor of 4.47e3 after eight iterations
B. Output of the CNRSER Algorithm Using Voltage while the squared error between the imaginary components of
Data from a Simulated Human Chest the measured and predicted voltages decreased by a factor of
Figs. 9 and 10 show the conductivity and permittivity 2.85e2.
distribution, respectively, computed by the CNRSER algo- Figs. 13 and 14 show reconstructions of the conductivity
rithm, using the model shown in Fig. 4. The squared error and permittivity distributions, respectively, of the simulated
between measured and predicted voltages decreased by a chest phantom, computed using gradient regularization of
factor of 2.44e3 after eight iterations. Figs. 11 and 12 show the Jacobian matrix. The gain of gradient term [ in (8)]
the conductivity and permittivity distribution, respectively, was chosen to be 1.0. After four iterations of the CNRSER
computed by the CNRSER algorithm when the real and algorithm, the squared error between measured and predicted
quadrature voltage data were used as separate input to the voltages was reduced by a factor of 79.8.
906 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO. 7, JULY 1998

problem. The ill-conditioned nature of the problem is a result

of the characteristic that a large perturbation in the center
of the body can result in minor perturbations in the volt-
ages measured on the exterior of the body. As a result, the
reconstructions are highly susceptible to noise.
The reconstructions displayed in Figs. 5 and 6 show that
a linearized solution to the inverse problem results in an
approximate estimate of the true conductivity and permit-
tivity distributions. After the first iteration of the CNRSER
algorithm, both the conductivity and permittivity distributions
contain an inhomogeneity centered in the phantom. With more
iterations, the values of conductivity and permittivity become
more accurate, as seen in Figs. 5 and 6, respectively.
Accuracy of reconstructed data improves by computing
the conductivity and permittivity distributions simultaneously
using complex voltage data measured on as input. The
plots in Figs. 5 and 6, generated from reconstructions using
Fig. 13. Conductivity distribution as a function of angular position, calcu- complex voltages measured on , more closely resemble the
lated after four iterations of the CNRSER algorithm using complex voltage true distributions, after four iterations, than the plots in Figs. 7
data generated from the geometry in Fig. 4. Gradient regularization of the
Jacobian matrix was implemented for these reconstructions. and 8 which were generated from reconstructions which used
real and quadrature components of the measured voltage on
separately. In fact, Figs. 7 and 8 show that updates to the
conductivity and permittivity distributions actually progress
in a direction opposite of the actual distributions toward the
center of . This was true even though the squared error
between measured and predicted voltages for both the real
and imaginary components of the boundary voltages on
decreased after each of four iterations.
Reconstructions of voltage data generated from the sim-
ulated chest phantom better represent the true conductivity
and permittivity distributions in when complex voltage
is used as input to the algorithm. Figs. 9 and 10 show the
conductivity and permittivity distributions, respectively, after
several iterations of the CNRSER algorithm. By comparing
these reconstructions to reconstructions of the conductivity
(Fig. 11) and permittivity (Fig. 12) distributions computed
using the real and quadrature components of the voltage
on separately, one can see significant differences, par-
Fig. 14. Permittivity distribution as a function of angular position, calculated ticularly in the conductivity distribution. Fig. 11 shows that
after four iterations of the CNRSER algorithm using complex voltage data the conductivity of the heart region (H) is overestimated
generated from the geometry in Fig. 4. Gradient regularization of the Jacobian
matrix was implemented for these reconstructions. while the conductivity near the edges of the lungs (L1 &
L2) is underestimated after several iterations. These errors
V. DISCUSSION occur even though the squared error between measured and
predicted voltages for the real and imaginary components of
The results shown in the Section IV demonstrate two im- the boundary voltages on decreased after each of eight
portant characteristics of the CNRSER algorithm. iterations.
1) Reconstructions after multiple iterations of the CNRSER If the reconstructions in Figs. 9 and 10 are examined
algorithm more closely resemble the true conductivity closely, one notices a general trend in the reconstructions of the
and permittivity distributions of than the reconstruc- conductivity and permittivity distributions: the reconstructions
tions given by only one step of the algorithm. become less smooth as the CNRSER algorithm iterates. When
2) More accurate reconstructions of the conductivity and experimental data is used as input to the CNRSER algorithm
permittivity distributions result if these quantities are re- and Marquardt regularization is applied to the Jacobian matrix,
constructed simultaneously using complex voltage data this effect worsens. To force the reconstruction to be spatially
measured on . smooth, the gradient term was added to (8).
Although the inverse admittivity problem is ill conditioned, Figs. 13 and 14 show the conductivity and permittivity
it is possible to make better estimates of the reconstructed distributions calculated after four iterations of the CNRSER
conductivity and permittivity distributions by executing a algorithm using complex voltage data as input. The gain
multistep algorithm, instead of considering only the linearized of in (8) was selected to be 1.0. It is evident that
EDIC et al.: ITERATIVE NEWTON–RAPHSON METHOD TO SOLVE INVERSE ADMITTIVITY PROBLEM 907

use complex voltage data, instead of operating on real and

quadrature voltage data measured on separately. The errors
introduced in a linearized approach to the inverse problem are
(a) corrected as the number of iterations is increased.
APPENDIX
EXPRESSIONS FOR TERMS OF THE JACOBIAN MATRIX
There are four separate expressions for the components of
the Jacobian matrix defined in (16)

(b)

(c)

Fig. 15. (a) The ideal conductivity distribution and the reconstructed con-
ductivity distribution, calculated after (b) the first and (c) the fourth iterations (29)
of the CNRSER algorithm using complex voltage data generated from the
geometry in Fig. 4. Gradient regularization of the Jacobian matrix was
implemented for these reconstructions. The ideal reconstruction was generated
by computing the weighted-average of the conductivity values within a
mesh element by the percentage of area containing the values. (Units are
millisiemens per meter.)

the reconstructions in Figs. 13 and 14 are smoother when

compared to the reconstructions in Figs. 9 and 10, respectively.
Fig. 15 shows the static reconstructions of the conductivity
distribution after the first iteration [Fig. 15(b)] and after
the fourth iteration [Fig. 15(c)] of the CNRSER algorithm
using gradient regularization of the Jacobian matrix. From
this figure, one can see that the reconstructed range and the (30)
resolution of the conductivity image are not adversely effected
by this regularization method. Also included in Fig. 15 is an (31)
image of the ideal conductivity distribution [Fig. 15(a)]. This
image was generated by computing a weighted-average of
the conductivity values contained in a mesh element by the
percentage of area containing the values. This image shows
the best reconstruction that one would expect to compute
given the boundaries of the objects in Fig. 4.
VI. CONCLUSION
The CNRSER algorithm solves the inverse admittivity prob-
lem using an iterative Newton–Raphson method. This algo-
rithm is used in conjunction with a finite-element forward-
solver that computes boundary voltages on a body, given the
applied current vectors and the admittivity distribution within
the body. By performing only a few iterations of the CNRSER
algorithm, the estimates of the conductivity and permittivity
distributions approach the true distributions in the examples
presented. One is able to make better estimates of the true
(32)
distributions if reconstruction algorithms are devised which
908 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 45, NO. 7, JULY 1998

REFERENCES David Isaacson (M’86) received the Ph.D. degree in mathematics from New
York University’s Courant Institute of Mathematical Sciences, New York, in
[1] M. Cheney, D. Isaacson, J. C. Newell, S. Simske, and J. Goble, 1976.
“NOSER: An algorithm for solving the inverse conductivity problem,” He is a Professor of Mathematics at Rensselaer Polytechnic Institute, Troy,
Int. J. Imag. Syst., Technol., pp. 66–75, 1990. NY. He is currently working on problems arising in the use of electromagnetic
[2] J. Sylvester and G. Uhlmann, “A uniqueness theorem for an inverse fields for the diagnosis and treatment of disease.
boundary value problem in electrical prospection,” Commun. Pure, Appl.
Math., vol. 39, pp. 91–112, 1986.
[3] D. C. Barber, B. H. Brown, and N. J. Avis, “Image reconstruction
in electrical impedance tomography using filtered back projection,”
in Proc. Annu. Int. Conf. IEEE Engineering in Medicine and Biology Gary J. Saulnier (S’80–M’84–SM’95) was born in Fall River, MA. He
Society, 1992, pp. 1691–1692. received the B.S., M.E., and Ph.D. degrees in electrical engineering from
[4] B. H. Brown and A. D. Seagar, “The Sheffield data collection system,” Rensselaer Polytechnic Institute, Troy, NY, in 1980, 1982, and 1985, respec-
Clin. Phys., Physiolog. Meas., 8 suppl. A, pp. 91–97, 1987. tively.
[5] R. W. M. Smith, B. H. Brown, I. L. Freeston, F. J. McArdle, and D. In 1984, he joined General Electric Corporate Research and Development
Barber, “Real time electrical impedance imaging,” in Proc. Annu. Int. Center, Schenectady, NY, where he studied the design and implementation
Conf. IEEE Engineering in Medicine and Biology Society, 1990, pp. of bandwidth-efficient digital modulation techniques for fading channels.
104–105. Since 1986, he has been on the faculty of the Electrical, Computer and
[6] D. Isaacson and P. M. Edic, “An algorithm for impedance imaging,” Systems Engineering Department at Rensselaer Polytechnic Institute where
in Proc. Annu. Int. Conf. IEEE Engineering in Medicine and Biology he is currently an Associate Professor. His research interests include modems
Society, 1992, p. 1693. for mobile and mobile-satellite channels, antijam spread spectrum systems,
[7] P. M. Edic, G. J. Saulnier, J. C. Newell, and D. Isaacson, “A real-time and electronic instrumentation for biomedical applications.
electrical impedance tomograph,” IEEE Trans. Biomed. Eng., vol. 42,
no. 9, pp. 849–859, Sept. 1995.
[8] P. Hua, E. J. Woo, J. G. Webster, and W. J. Tompkins, “Iterative
reconstruction methods using regularization and optimal current patterns
in electrical impedance tomography,” IEEE Trans. Med. Imag., vol. 10, Hemant Jain was born in Meerut, India, on January 3, 1973. He received
no. 4, pp. 621–628, Dec. 1991. the B.Tech. degree in electrical engineering from Indian Institute of Tech-
[9] E. J. Woo, P. Hua, J. G. Webster, and W. J. Tompkins, “A robust image
nology, Kanpur, India, in 1993, and M.S. and Ph.D. degrees in biomedical
reconstruction algorithm and its parallel implementation in electrical
engineering from Rensselaer Polytechnic Institute, Troy, NY, in 1996 and
impedance tomography,” IEEE Trans. Med. Imag., vol. 12, no. 2, pp.
1997, respectively.
137–146, June 1993.
[10] T. J. Yorkey, J. G. Webster, and W. J. Tompkins, “Comparing recon- He is currently working at Epic Systems Corporation, Madison, WI. His
struction algorithms for electrical impedance tomography,” IEEE Trans. main research interests include noninvasive imaging techniques and high
Biomed. Eng., vol. BME-34, no. 11, pp. 843–852, Nov. 1987. performance computing for biomedical applications.
[11] M. R. Eggelston, R. J. Schwabe, L. F. Coffin, and D. Isaacson, “The
application of electric current computed tomography to defect imaging
in metals,” Rev. Quantitat. Nondestructive Evaluation, vol. 9A, pp.
455–462, 1990. Jonathan C. Newell (S’64–M’69) received the B.S.
[12] D. Isaacson, “Distinguishability of conductivities by electric current and M.S. degrees in electrical engineering from
computed tomography,” IEEE Trans. Med. Imag., vol. M1-5, pp. 91–95, Rensselaer Polytechnic Institute, Troy, NY, where
1986. he graduated in 1968, and the Ph.D. degree in
[13] M. Cheney and D. Isaacson, “Distinguishability in impedance imaging,”
physiology from Albany Medical College, Albany,
IEEE Trans. Biomed. Eng., vol. 39, no. 8, pp. 852–860, Aug. 1992.
[14] W. H. Press, Ed., Numerical Recipes in C: The Art of Scientific Com- NY, in 1974.
puting. New York: Cambridge Univ. Press, 1992. He is now Professor of Biomedical Engineering
[15] E. Isaacson and H. B. Keller, Eds., Analysis of Numerical Methods. at Rensselaer Polytechnic Institute and Professor of
New York: Wiley, 1966. Physiology and Surgery at Albany Medical College.
[16] K.-S. Cheng, D. Isaacson, J. C. Newell, and D. G. Gisser, “Electrode His research interests have included the regulation
models for electric current computed tomography,” IEEE Trans. Biomed. of the pulmonary circulation in hypoxia, and pul-
Eng., vol. 36, pp. 918–924, 1989. monary gas exchange in injured patients with acute respiratory failure. His
[17] E. Somersalo, M. Cheney, and D. Isaacson, “Existence and uniqueness recent work has been the development of an adaptive system for electrical
of electrode models for electric current computed tomography,” SIAM impedance imaging.
J. Appl. Math., vol. 52, no. 4, pp. 1023–1040, Aug. 1992.
[18] M. Pidcock, S. Ciulli, and S. Ispas, “Singularities of mixed boundary
value problems in electrical impedance tomography,” Physiolog. Meas.,
vol. 16, no. 3, suppl. A, pp. A213–218, Aug. 1995.
[19] H. P. Schwan and C. F. Kay, “The conductivity of living tissue,” Ann.
NYAS, vol. 65, pp. 1007–1013, 1957.

Peter M. Edic (S’93–M’95) was born in Utica,

NY, on December 25, 1959. He received the A.A.S.
degree in electrical engineering technology from
Mohawk Valley Community College, Utica, NY, in
1987, the B.S. degree in electrical engineering from
Syracuse University, Syracuse, NY, in 1990, and
the M.S. and Ph.D. degrees in electrical engineering
from Rensselaer Polytechnic Institute, Troy, NY, in
1991 and 1994, respectively.
He is currently working at the General Electric
Corporate Research and Development Center, Sch-
enectady, NY. His research interests include signal/image processing, medical
imaging, data visualization, and microprocessor-based medical instrumenta-
tion.