Proceedings of the 26th Annual International Conference of the IEEE EMBS
San Francisco, CA, USA September 1-5, 2004
A Reconstruction Algorithm Based on Wavelet Network in Electrical Impedance
Tomography
Guizhi Xu, Qing Wu, Qingxin Yang, Ying Li, and Weili Yan
Department of electrical engineering, Hebei University of Technology, Tianjin, 300130, China
AbstractElectrical impedance tomography (EIT) is a
noninvasive technique to estimate the conductivity distribution
inside the object. In EIT, driving currents are injected through
the object and voltages are measured at the electrodes on the
surface. Algorithms to estimate the conductivity distribution
from the measured voltages are called reconstruction
algorithms. Image reconstruction is a nonlinear inverse
problem and typically ill-conditioned. Some algorithms for
image reconstruction including equal-potential back-projection,
regularization, and sensitivity matrix, modified Newton-
Raphson methods have been proposed in past, all of these
techniques, while giving reasonable solutions, require
simplifying hypotheses about the problem. This paper presents
a new reconstruction algorithm based on wavelet network,
which calculates conductivity distribution directly from finite
element simulations of forward problem. To indicate the
validity of this method, we have gotten the conductivity
distribution for the three-concentric-shell spherical head model
using this algorithm, the result show it is speed of image
reconstruction, conceptual and ease of implementation1.
KeywordsElectrical impedance tomography, wavelet
network, inverse problem, head model
I. INTRODUCTION
Electrical impedance tomography is an imaging
modality that produces images by computing electrical
conductivity within the subject. Sinusoidal current are
applied to the volume using electrodes and the resulting
voltages on the electrodes are measured. The internal
resistivity distribution is computed based on this boundary
data. Typically 16 electrodes are placed around the cross-
section to be imaged, current is applied to two drive
electrodes and measurements made of the resultant potential
difference across adjacent pairs of the other electrodes. The
drive electrodes are rotated to the next pair and the process
is repeated until a full data set is obtained. A resistivity
distribution map or image is then produced by means of a
suitable reconstruction algorithm [1].
Image reconstruction is also called as inverse problem
in EIT. Many linear algorithms for image reconstruction
have been proposed including equal-potential back-
projection, regularization methods, and sensitivity matrix
method [2]. The principal nonlinear algorithm used is the
modified Newton-Raphson method, which requires several
orders of magnitude more calculation time than linear
This work is supported by the Natural Science Foundation of Hebei
Province (E2004000054)
0-7803-8439-3/04/$20.002004 IEEE
techniques. All of these techniques, while giving reasonable
solutions, require simplifying hypotheses about the problem.
In EIT, finite element method (FEM) is mostly used to
describe the electric field of in homogeneous conductivity
distribution, by approximating it as a collection of
homogeneous elements. In this paper we present a new
reconstruction algorithm based on wavelet network for the
EIT inverse problem directly from finiteelement
simulations of forward problem.
II.THE MATHEMATICAL MODEL FOR EIT
The underlying relationships that govern the interaction
of electricity and magnetism are summarized by Maxwells
equations. For the case of EIT, several simplifying
assumptions can be applied to reduce the complexity of the
problem. Mathematically, the known quantities are the
voltages and currents at certain points on the body; the
unknown is the impedivity or resistivity within the body. At
low frequencies, these quantities are related by the Laplace's
equation
( ) = 0 (1)
Where is the potential in medium, is the electric
conductivity.
Expanding the divergence of this product gives
Poissons equation:
2 = (2)
Where the RHS term resembles a distribution of current
sources, equation (2) is non-linear.
A. Forward problem
The solution of Poissons equation for a given
conductivity distribution, is generally known as the forward
problem. The distribution of is dependent on the
conductivity distribution, the shape of the object and the
position of the electrodes. Analytical solution of (2) is only
possible for simple geometries and boundary conditions. For
most conductivity distributions, the forward problem must
be solved using numerical techniques such as the Finite
Element Method (FEM). In this case, the region is
partitioned into a number of discrete elements, each with
assumed uniform conductivity. The forward problem is then
solved for each element by ensuring continuity of voltage
and current flow between neighboring elements. Thus, FEM
resolution for (1) is converted into the algebraic problem:
1435
 K = Q (3) {
F = n ,m ( x) = a dn / 2 (a
n
x bm)n Z , m Z d }
Where K is symmetric element matrix, is node potential (5)
vector. a R and b R are respectively called dilation and
d

translation parameters (typically a = 2, b = 1 )[4].

B. Inverse problem:
If a single-scaling radial wavelet frame replaces the
Inverse problem is to reconstruct the conductivity radial basis functions in RBFN, and centers and radius
distribution from the given current sources and boundary parameter are replaced by the shift and scale parameters of
condition. Solution of the forward problem is an essential the wavelets respectively, we get a single-scaling radial
component of EIT image reconstruction. Iterative algorithms wavelet network:
h
repeatedly solve the forward problem as the conductivity
distribution is modified step by step until the computed f r (x) = wir (a ni x bm i ) , r = 1,2, L , s (6)
i =1
boundary measurements derived from the solution of the
forward problem, converge with the observed measurements. Where s , h and wi r are the dimension of output, the
A solution to the forward problem is also required for the number of hidden nodes and the connected weights of
computation of a sensitivity matrix, which is utilized in a network respectively. Fig.1 shows the calculating model of
number of reconstruction algorithms [3]. every wavelet neuron node.
N
(m) = B (m, i ) (i) (4)
i =1
d1
Where (m) is conductivity distribution, (i ) is measure
x1
voltage, and B(m, i ) is weights matrix. The value of dh
B(m, i ) matrix are calculated or trained by the back- x2 d1 s1
propagation learning rule using calculated electrode voltages -n
a y1
k
form forward solution as input vectors (i ) and . y2
conductivity values of elements as desired (known) output. . h1 hs s1
.
. () wir
xd
III.THE CONSTRUCTION WAVELET NETWORK FOR EIT ys
The main aim of constructing the wavelet network for bmk
EIT is to take advantage of the general mapping ability of
wavelet neural network to embed the regularization
h1
mechanism into the connected weights through extensive
learning from samples and to reduce the computational
complexity. Combining single-scaling wavelet frame and
radial basis function network (RBFN) model, a multi-
dimensional wavelet network (WN) is constructed. This
Fig .1 Calculation model of wavelet network
combination can hopefully remedy the weakness of each
other, resulting in networks with efficient constructive
method.
IV.THE INVERSE PROBLEM RESOLUTION WITH WAVELET
We know if a function f ( x) L ( R ) is radial, there
2 d
NETWORK

exists a function g : R R , such that x R , d

) We use wavelet network constructed above into solve

f ( x ) = g ( x ) hold, and its Fourier transform f () is
) )
radial too, that is f ( x) = g ( ) hold. A single-scaling
radial wavelet frame F of f ( x) L ( R ) can be built in
2 d
principle by scaling one mother radial wavelet function
L2 ( R d ) :
the EIT inverse problem of head, that is to find the
conductivity distribution within head from surface electrical
potential measured and the given current sources. A three-
shell concentric sphere head model with different
conductivity values respectively represents the brain and
cerebrospinal fluid (CSF), skull and scalp as shown in Fig.2,
and their parameters are showed in Table I.

1436
 N s
| ~
yir y ir |
i =1 r =1
TABLE I Aerror = (8)
sN
THREE-SHELL CONCENTRIC SPHERE HEAD MODEL
N s
| ~yir yir |2
PARAMETERS

Tissue CSF and Skull Scalp i =1 r =1

Brain Rerror = N S
100% (9)
Resistivity( m ) 1 1/ 0.0042 1 / 0.33
yir2
Resistivity( m ) i =1 r =1
0.91.1 200250 2.53.4
Min-max
Radius (cm) 8.7 9.2 10
Where s is the number of output vector N is the number of
training set or test sample, y ir is the real output value and
0.1
~
y ir is the output value computed by network. the calculation
0.08 results are listed in Table II, the training results and test
0.06 results of wavelet network for head model described above
0.04
are showed in Fig.3 and Fig.4 respectively.
0.02
TABLE II
0
RESULT OF WAVELET NETWORK IMPLEMENTATION
-0.02
Number of Number of Relative Relative
-0.04
training set test set training error test error
-0.06 950 50 3.56 9.09
-0.08
3 .4
-0.1
-0.1 -0.05 0 0.05 0.1
3 .3

3 .2
Fig.2. Three-shell concentric sphere head model
3 .1

The multi-channel data set x = [ x1 , x 2 , K x d ] (d is the 3

number of electrodes) measured at electrodes on object 2 .9

surface are taken as the network inputs and the 2 .8

corresponding parameters of conductivity distribution as the 2 .7

outputs. First, using available samples which can be

2 .6
obtained from the test examples or computations of forward
problem, WN can establish an approximation map from the 2 .5
0 5 10 15 20 25 30 35 40 45 50

measuring data space to the corresponding conductivity

distribution during the train process based on the assumption (a) Scalp resistivity distribution
that the map is smooth. Then, test the trained network with 250

measurement data without being used in training process. 245

The mother wavelet function which is called as 240

Mexican hat is chosen as following 235

x

( x ) = ( d - || x ||)e 2
(7)
230

225

where, x Rd , x = xTx . 220

We take 16 measurement electrodes on the head surface, 215

each shell resistivity value change from max to min with 10 210

steps, together with 10*10*10 kinds of conductivity 205

distribution which are divided into two parts, one is used as 200
0 5 10 15 20 25 30 35 40 45 50
samples to train the network, and other as test data for
reconstruct the real conductivity distribution. (b) Skull resistivity distribution
In order to evaluate the performance of the trained
wavelet network, we define the absolute error and relative
error as following:

1437
 1.1 1.1

1.08

1.06 1 .0 5

1.04

1.02 1

0.98 0 .9 5

0.96

0.94 0.9

0.92

0.9 0 .8 5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

(c) CSF and brain resistivity distribution (c) CSF and brain resistivity distribution
Fig.3 The training results of wavelet network for Three-shell concentric
sphere head model Fig.4 The test results of wavelet network for Three-shell concentric sphere
(notes :dot point shows the real value, circle o shows calculated head model
value; abscissa shows the number of training samples and ordinate shows (notes :dot point shows the real value, circle o shows calculated
the resistivity value ( m )) value; abscissa shows the number of test samples and ordinate shows the
resistivity value ( m ))
3 .8

3 .6 V. CONCLUSION
3 .4
In this paper we present a new reconstruction algorithm
3 .2
based on wavelet network for the EIT inverse problem
3 directly from finiteelement simulations of forward problem,
and have gotten the conductivity distribution for the three-
2 .8
concentric-shell spherical head model using this algorithm
2 .6 to indicate the validity of this method. The results show
2 .4
there are many advantages in its speed of image
reconstruction, conceptual and ease of implementation.
2 .2
0 5 10 15 20 25 30 35 40 45 50

(a) Scalp resistivity distribution REFERENCES

[1] KBoone, D.Barber, et al., Review Imaging with Electricity: Report
260
of the European Concerted Action on Impedance Tomography,
Journal of Medical Engineering & Technology, Vol.21, No.6, pp.201-
250 232, Nov./Dec., 1997
[2] F.Santosa, M.Vogelius, A back-projection algorithm for electrical
240 impedance imaging, SIAM J. Appl. math, pp216-24350 (1990)
[3] Beyhan KILIC, Mehmet KOUREK, a finite element method base
230
neural network technique for image reconstruction in electrical
impedance imaging, The 2nd International Biomedical Engineering
Days, pp100-102, 1998
220
[4] T.Kugarajah, Q.Zhang, multi-dimensional wavelet frame. IEEE
Trans. On Neural Networks, vol.6, pp1552-1556, Nov. 1995
210

200
0 5 10 15 20 25 30 35 40 45 50

(b) Skull resistivity distribution

1438