91
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. MI5, NO. 2, JUNE 1986
Distinguishability of Conductivities by Electric
Current Computed Tomography
DAVID ISAACSON
AbstractWe give criteria for the distinguishability of two different
conductivity distributions inside a body by electric current computed
tomography (ECCT) systems with a specified precision.
It is shown in a special case how these criteria can be used to determine the measurement precision needed to distinguish between two different conductivity distributions.
It is also shown how to select the patterns of current to apply to the
body in order to best distinguish given conductivity distributions with
an ECCT system of finite precision.
urVU.
In the forward problem we assume that a and the normal component of the current density vector on the surface S are given.
In other words, on S we are given
J n
aU =j
a
an
where n denotes the unit outward normal vector on S.
If we specify U at one point in B, and we choose a j for
T HE problem of reconstructing the conductivity inside which
a body from lowfrequency measurements on the
body's surface has been discussed in [1][10], [13][20].
jdA = 0
Descriptions of impedance cameras designed for medical
s
imaging are given in [11], [16].
The first purpose of this paper is to give precise defi then there is a unique solution U(p) to the preceding equanitions of what it means for two different conductivity dis tions. We denote its restriction to the boundary S by
tributions to be distinguishable by finite precision meaV = V(p)  U(p)
surements. The second purpose is to give a careful discussion of how to select the patterns of current to apply for p on s.
Thus, the forward problem is to find the voltage V on
to the surface of a body to best distinguish between two
conductivity distributions by an ECCT system of finite the surface, given the conductivity a and currentj. When
precision. We also give simple examples that illustrate the we want to emphasize that the voltage V is a linear funcdefinitions and the discussion of current pattern selection. tional of j and a nonlinear functional of a, we write
Since the conductivities of blood, muscle, and lung difV = V(p; a, j).
fer considerably, an approximate reconstruction of the
This forward problem can be solved numerically by ficonductivity throughout the thorax would contain morphological information as well as electrophysiological in nite difference or finite element methods.
The inverse problem is to find the conductivity a in B
formation needed to make more accurate solutions in the
from
a knowledge of
forward and inverse problems of electrocardiography [11],
[21].
Vk  V(p; Or, ik)
II. DESCRIPTION OF FORWARD AND INVERSE PROBLEMS for a sequence of currentsik, k = 1, 2, 3,
.
of
Descriptions attempts to solve this inverse problem
Throughout this paper we let B denote a body and S its
surface. We assume that B is a linear conductor with sca are in [1][10], [13][20].
lar conductivity a = a(p) where p is a point in the body.
III. DISTINGUISHABILITY
The voltage or potential U = U(p) is assumed to satWe say that two conductivities a, and 92 are distinisfy
guishable
(in the mean square sense) by measurements of
V caVU = 0
precision e iff there is a current j for which jil = 1 and
in B. The current density vector is denoted by J = J(p)
1 V( ;Ol, j)  K ; O2, j)11 > E
and is given by
Here if f = f(p) and g(p) are any functions on S
Manuscript received November 27, 1985; revised January 27, 1986.
I. INTRODUCTION
The author is with the Department of Mathematical Sciences and the
Department of Computer Science, Rensselaer Polytechnic Institute, Troy,
NY 12181.
IEEE Log Number 8607881.
(f, g _
02780062/86/06000091$01 .00 1986 IEEE
f(p) g(p) dA
92
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. MI5, NO. 2, JUNE 1986
U(R , 0) = U(R +, 6)
and
Ilf!!
and
flif(p)2 dA 1
au
au
We also say that the conductivities a1 and U2 are not
distinguishable by measurements of precision E iff for all
j with lj 11 = 1 we have that
IIV( ; O01, j)  V( ; 02, j)1 <
We illustrate these two definitions with a simple twodimensional example [14], [15].
Let B be the disk x2 + y2 < l and S the circle x2 +
y2= 1. Using the polar coordinates x = r cos 0 and y =
r sin 0, we define
{a
if 0 c r .< R
ai = U1(r, 6) _
1
if R < r c I1
and
for O c r 1.
O'2 = a2(r, 6) 1
We next ask which inhomogeneous distributions a, are
distinguishable from the homogeneous distribution U2 by
measurements of precision E?
For U2
1, we have that
V . VU = 0.
If
00
E Cn cos n6 + Sn sin nO
j(6)= n=I
(R, 0) = 1(R +, ).
ar
ar
It is
V(6; ,1j) =
cOD
and
00
U(r, 0) =  E= (r'ln) [Cn cos nO + Sn sin nO]
n 1
and
00
V = V(6; c2, j) =  E
n=
n1[C, cos nO
+ Sn sin nO].
Sn sin n6}
compute
IIV( ; (02, j)  V( ; Oi, j)II

oo
1/2
n~~~~~~~~itR22n))2
Since R2nln(1 + /lR2n) decreases as n increases, and since
= Vi 2Ii L E (R2nln(I +
lij 11
00
(C2 + S2)
n
(C2
n
S2)
n
we have that
IV(*; 2,1j)  V(; aOl,J)I
< 21/L R21(1 + /LR2).
Thus, if
+
AR2)
' E
then a, and u2 are not distinguishable.
We note also that the currents j(0) = (cos 6)IV or (sin
0)/vG7 will distinguish between o1 and 02 if
21/1l R21(1 + AR2) > 6.
Thus, the curve in the a  R plane given by
LR2)
= C
divides the plane into regions in which a1 and 92 are distinguishable by measurements of precision E from those
regions in which they are not distinguishable.
It follows that the disk of radius R and conductivity a
is indistinguishable from the homogeneous distribution
with conductivity 1 if a and R lie in the region
U {(U,R)Iam(R) . a < acm(R), and 0 c R 1}
where
The solution to the inhomogeneous problem is again
found by separation of variables and by imposing the con om(R)
ditions
and
rM(R)
_(a 1)I(a + 1).
Note that if a > 0, then I,u < 1.
To determine the distinguishability of ar2 from a,, we
21/lI R21(1
U(r = 0, 0) = 0,
then by separation of variables
2n
where
21/I R21(1
ir
n 1 
{C, cos nO
where
Cn = =1 2r2 j(6)o cos nO dO
Sn = 1 2r2 j(O, sin nO dO
I 1 ,t,R In
ni
max
{O, (R2  cI(2 + E))I(R2 + cI(2 + E))}
1 + (2E/(2  6))I(R2 E/(2  c)),
for 2
2
+00,
forR2
6E
<
R2
6e
<
93
ISAACSON: DISTINGUISHABILITY OF CONDUCTIVITIES
e
0ll,
>
X2
>
0. In other words,
* *
Dik
Xk Ik
D2ik
Xkjk
and
R
Fig. 1.
It also follows that disks which lie in the subset of {0
< a; 0 < R c 1 } that is outside of U are distinguishable
from the homogeneous distribution by measurements of
precision e. See Fig. 1 for a diagram of these regions.
This simple calculation tells us that we cannot distinguish any conductivity variations at all for disks of radius
Rm or less where
Rm = [61(2 + 6)] 1/2.
fork= 1,2,3,
It follows from the minimax principal [12] that the best
current (or currents) to distinguish between ul and a2 are
the eigenfunctions of D2 having the largest eigenvalue.
If, for example, X2 is nondegenerate, thenjl would be the
best current. It also follows that
IDjiII =1 XI
max 11A(02)  A(ol)] j 1
111 11 1I
Thus, a, is distinguishable from u2 by measurements of
precision e (using j1) if XXI > c, and they are not distinguishable if 1 I < e.
The next best current we can apply to S to distinguish
a1 from U2 isj2 where
= max IlDIll = iX21
(Jl) =0
Thus, if we want to see conductivity variations in disks (assuming X's are nondegenerate).
of radius R, we need a precision e smaller than
This next eigenfunction will also distinguish between
a1
and 2 to precision e iff X21 > 6. By continuing this
2R2/(1  R2).
10D1211
process we see from the minimax principal [12] that the
eigenfunctions jI, j2, * * * , 1k of D for which Xk > e and
Xk +11 e are in the least squares sense the best currents
to apply to S in order to distinguish a1 from a2 by measurements of precision 6. It is pointless to use the eigenfunctions j, with 1 > k because their use would result in
measurements of voltages whose differences would be
IV. SELECTION OF CURRENTS TO APPLY
smaller than our given precision.
We illustrate these ideas with a simple example. Let B,
In a certain sense we give an answer to the question:
=
S,
which currents ik, k 1, 2, * * * , should we apply to S
a,, and (r2 be the same as before. Then the best currents
to best distinguish between two conductivities a, and a2? to distinguish between a1 and ar2 are the eigenfunctions of
The current ji with iijl  = 1 is said to be the current D2 corresponding to its largest eigenvalue X2, where
that best distinguishes between a1 and a2 iff
D2j = [A(al)  A(a2)]2j
If R = 0.1, this expression says e must be smaller than
0.025.
We suspect that to distinguish conductivity variations
in regions of the same size but nearer the boundary, the
same estimates will yield sufficient precision.
V(; ar2, j,)  V( ; (,1, il) 1
= max i V( ; a2,j) 
00
V(K; al, 1)1
Let A(a) denote the linear operator defined by
A(a) j(p) _ V(p; a, j).
The best current j is then the current j that maximizes
I IA(a2)
Z [(2ltR2,)/n(I
E
n I
A(or ) j 112
and has lij I = 1. Since A(r) is selfadjoint with respect
to the inner product previously defined, we have that j
[A(2)
A((J1)]2 j >I<j, j >
Sn sin nO).
00
The eigenvalues
by
>
n =I
Cn cos nO
forn = 1, 2, 3,
Sn sin nO.
are doubly degenerate and are given
A /1\2
K*
[A(a2)  A(o1)]. In general D2 is a compact self to Xn are
adjoint nonnegative operator with a complete set of orthonormal eigenfunctions jl, ]2, j3, * * * , and eigenvalues
Let D
AR2,)]2
and
maximizes the expression
(],
(Cn cos nO
2ItI
R2n A
R
)L +iR2ns
..The eigenfunctions corresponding
iCn
cos
nOi
IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. MI5, NO. 2, JUNE 1986
94
and
insin nOIV7.
The largest eigenvalue is X2 = [21ipl R2/( 1 + tt R 2)]2
and there is a twodimensional space of best eigenfunctions spanned by cos OiV7 and sin o/1.
If our measurements are made with precision c we only
need the eigenfunctions with Xnl > E, i.e., those with
1 211 R2n
n 1 + R2n
Since, for small R, this expression goes to zero rapidly
as n increases, only a few functions will be effective. For
example if R = 0.1, a = 2, and e = 103 only cos 0/
v&r and sin oiV will yield data of significance.
If B is the unit disk, a1 = ul(r) and a2 = a2(r), i.e.,
they are both independent of 0, then the eigenfunctions of
D are again cos nO/IV 7 and sin nOi'Jr; for n = 1, 2, * .
For arbitrary functions, vi = ai (r, 0), i = 1, 2, the best
currents are, in general, not trigonometric functions.
As a general rule, low (spatial) frequency currents yield
voltages that are the most sensitive to changes in conductivity far from the boundary, while highfrequency currents yield voltages sensitive mostly to changes near the
boundary.
We next point out that it is possible to use a set of currents j + and j  that span the same space as the best currents j] and 12 but whose use results in meaningless measurements.
Suppose Dj, > c and 10D1211 < c. Choose an a > 0
but small and define
j+
(ail
=
[((aXI)2
+
((1 ))X2)2)/(a2
(1
aY)2)]1/2
ACKNOWLEDGMENT
The author would like to thank R. Kohn, M. Vogelius,
J. Sylvester, G. Uhlmann, T. J. Yorkey, J. G. Webster,
and W. J. Tompkins for informing him of their work and
making some of their work available before publication.
He would like to thank the referees for informing him
of the references [14][17] and for pointing out the similarity between the definition of distinguishability given
here and the notion of visibility in [14], [15].
He would especially like to thank D. Angwin and J.
Newell for their collaboration in designing and carrying
out the experimental studies in conductivity imaging that
inspired this work.
REFERENCES
[1] R. E. Langer, "An inverse problem in differential equations," Amer.
Math. Soc. Bull. Ser 2, vol. 29, pp. 814820, 1933.
[2] R. J. Lytle and K. A. Dines, "An impedance camera: A system for
determining the spatial variation of electrical conductivity," Lawrence Livermore Laboratory, Livermore, CA, Rep. UCRL52413,
(1 () j2)I(a 2 + (1  a)2)1/2
Clearly, j + span the same space as Ii and i2 but
tIDjII
is distinguishable from each member of E by measurements of precision e.
We point out that it is impossible to identify a conductivity distribution a in E, much less image it, by an ECCT
system if this system cannot distinguish a from each
member of E. Hence, a systematic manner in which one
may go about designing an ECCT system is to first select
the resolution that one would like by choosing the class E
that is to be identifiable. Second, determine the precision
of the instrument that is needed by first estimating the
smallest size and conductivity variations possible in a
member of E. Then use estimates similar to those given
in Section III to determine the precision needed. Finally,
use the best patterns of current described in Section IV.
<
1978.
[3] R. H. Bates, G. C. McKinnon, and A. D. Seager, "A limitation on
systems for imaging electrical conductivity distributions," IEEE
Trans. Biomed. Eng., vol. BME27, pp. 418420, July 1980.
[4] R. L. Parker, "The inverse problem of resistivity sounding," Geophys., vol. 42, no. 12, pp. 21432158, Dec. 1984.
[5] H. Schomberg, "Nonlinear image reconstruction from projections of
ultrasonic travel times and electric current densities," in Mathematical Aspects of Computerized Tomography: Proceedings, G. T. Herman and F. Natterer, Eds. New York: SpringerVerlag, 1981, pp.
270291.
[6] A. P. Calderon, "On an inverse boundary problem," in Seminar on
Numerical Analysis and its Applications, W. H. Meyer and M. A.
Raupp, Eds. Rio de Janeiro, Brazil: Brazilian Math Society, 1980.
[7] R. V. Kohn and M. Vogelius, "Determining conductivity by boundary measurements," Commun. Pure Appl. Math., vol. 37, pp. 289298, 1984.
, "Determining conductivity by boundary measurements II. In[81
terior results," Univ. Maryland, Tech. Note BN 1028, 1984.
[9] J. Sylvester and G. Uhlmann, "A uniqueness theorem for an inverse
boundary value problem in electrical prospection," preprint 1985.
[10] T. Muari and Y. Kagawa, "Electrical impedance computed tomography based on a finite element model," IEEE Trans. Biomed Eng.,
if a is chosen small enough.
Voltage measurements V(p; a,, j ) and V(p; a2, j +)
differ from each other by less than the precision E and are
therefore meaningless while the use of jI alone would have
yielded a meaningful measurement in attempting to distinguish between conductivity distributions.
In practice, one applies currents to K electrodes. Thus,
there are K  1 independent currents. If the measurements were done with perfect precision (e = 0) any K 1 independent currents would have the same ability to distinguish between conductivities. This example shows that
when the precision is finite (e > 0) different K  1 independent currents have different abilities to distinguish
between conductivities. Thus, the use of the best currents
is necessary in order to decide whether the conductivity
inside the body is distinguishable from a conjectured con [11]
ductivity by a system of finite precision.
vol. BME32, no. 3, pp. 177184, Mar. 1985.
R. P. Henderson and J. G. Webster, "An impedance camera for spatially specific measurements of the thorax," IEEE Trans. Biomed.
Eng., vol. BME25, no. 3, pp. 250254, May 1978.
V. SUMMARY
[12] R. Courant and D. Hilbert, Methods of Mathematical Physics. Vol.
I. New York:
1953, pp. 405406.
A conductivity distribution a is identifiable in a given [13] R. H. T. Bates,Wiley,
"Fullwave computed tomography, Part 1: Fundamental theory," IEE Proc., vol. 131, A, no. 8, pp. 610615, 1984.
class of functions 2 by measurements of precision e iff it
95
ISAACSON: DISTINGUISHABILITY OF CONDUCTIVITIES
[141 A. D. Seagar, T. S. Yeo, R. H. T. Bates, "Fullwave computed tomography, Part 2: Resolution limits," IEE Proc., vol. 131, A, no.
8, pp. 616622, 1984.
[15] A. D. Seagar and R. H. T. Bates, "Fullwave computed tomography,
Part 4: Low frequency electric current CT," IEE Proc. vol. 132, A,
no. 7, pp. 455466, 1985.
,
[16]
[17]
D. C. Barber and B. H.
Brown, "Applied potential tomography,"
Phys. E: Sci. Instrum., vol. 17,
Y.
pp.
723733, 1984.
J.
Kim, J. G. Webster, and W. J. Tompkins, "Electrical impedance
imaging of the thorax," J. Microwave Power, vol. 18,
245257, 1983.
no.
3,
pp.
[18] T. J. Yorkey, J. G. Webster, and W. J. Tompkins, "Errors caused
by contact impedance in impedance imaging," in Proc. IEEE 7th
Conf. Eng. Medicine and Biol., 1985, pp. 632637.
[19] K. A. Dines and R. J. Lytle, "Analysis of electrical conductivity
imaging," in Geophys., vol. 46, no. 7, pp. 10251036, 1981.
[201 T. J. Yorkey, "A quantitative comparison of impedance tomographic
systems for a new instrument," Ph.D. proposal, Univ. Wisconsin,
1985.
[21] T. C. Pilkington and R. Polnsey, Engineering Contributions to Biophysical Electrocardiography. New York: IEEE Press, 1982.