Comparing two electrical impedance tomography algorithms: Gauss-Newton and Topology Optimization

Motta-Mello L. Aa and Lima C. R.a Aya J. C.b Pai C. N.b Moura F. S.b Silva E. C. N.a Gonzalez-Lima R.b
a b

Mechatronics Dept. Univ. of S ao Paulo, Av. Prof. Mello Morais 2231, S ao Paulo, Brazil; Mechanical Eng. Dept. Univ. of S ao Paulo, Av. Prof. Mello Morais 2231, S ao Paulo, Brazil
ABSTRACT

Electrical Impedance Tomography (EIT) seeks to recover the impedance distribution within a body using boundary data. More specically, given the measured potentials, the model of the body - an elliptic partial dierential equation - and the boundary conditions, this technique solves a non-linear inverse problem for the unknown impedance. In this work, an algorithm called Topology Optimization Method (TOM) is applied to EIT and compared to the Gauss-Newton Method. The Topology Optimization has solved some non-linear inverse problems and some of its procedures were not investigated for EIT, for instance, the use of Sequential Linear Programming. Assuming a pure resistive medium, the static resistivity distribution of a phantom was estimated using a 2-D nite element model, by two dierent iterative algorithms: one based on Newton-Raphson scheme and the other on Topology Optimization Method. While the rst method essentially solves several algebraic systems, the second solves several linear programming problems. Results using experimental data are shown and the quality of the images obtained, time and memory used are compared for both algorithms. We intend to use these methods, in future works, for the visualization of a human lung subjected to mechanical ventilation. Keywords: 2-D electrical impedance imaging, topology optimization, Gauss-Newton method, nite element method, absolute resistivity

1. INTRODUCTION
Electrical Impedance Tomography (EIT) allows us to obtain internal images of a body (domain). Electrodes are xed on the boundary of the body and some low intensity electric currents, following a sinusoidal pattern, are applied through them. The resulting electric potentials are measured and, considering this data, the unknown electric impedance (resistivity and permittivity) distribution within the body is obtained by the EIT algorithm. This impedance distribution is represented by an image and constitutes the solution of a non-linear and ill-posed inverse problem.1, 2 A current pattern denes the location of the current-carrying electrodes and the current intensity on each current-carrying electrode. Several current patterns are applied and therefore many electric potential values are available for image reconstruction. The Topological Optimization Method has successfully solved some non-linear inverse problems. Some procedures of TOM have not been investigated in Electrical Impedance Tomography, namely, the use of sequential linear programming and the change of variables that in the TOM literature is often called material model. The objective of this work is to evaluate whether this procedures have advantages for EIT algorithms. In this work we compared two EIT algorithms. The rst is based on Gauss-Newton Method (GNM),3 a simplication of Newton-Raphson Method, and the second is based on the Topology Optimization Method4 (TOM). The Finite Element Method (FEM) is used to model the domain, which is a cylindrical vessel with a glass object immersed on saline. The complete electrode model2 is also considered and thus the impedances arising from the contact between electrodes and saline are taken into account. Both methods minimize a performance index that contains dierence between electric potential measurements and computations. TOM minimizes the performance index using Sequential Linear Programming5, 6 (LP) problems. On the other hand, GNM
Further author information: (Send correspondence to L. A. Motta-Mello) L. A. Motta-Mello: E-mail: [email protected], Telephone: 55 11 30919851
Medical Imaging 2006: Physiology, Function, and Structure from Medical Images, edited by Armando Manduca, Amir A. Amini, Proc. of SPIE Vol. 6143, 614339, (2006) 1605-7422/06/$15 doi: 10.1117/12.654059 Proc. of SPIE Vol. 6143 614339-1

minimizes the performance index using a generalized Gauss-Newton iterative update rule, which requires some matrix inversions. Both methods look for static images. Image quality, convergence rate, amount of memory used and computing time are compared.

2. FINITE ELEMENT MODEL

In EIT, the Maxwells equations must be used to describe the current density ow within the body. However, these equations can be simplied to the following elliptic partial dierential equation: ( V ) = 0, in (1)

where is the electric conductivity, V is the scalar electric potential and is the conductive body. The implemented algorithms depend on the solution of eq. 1 for the electric potentials, also known as forward problem. This solution, or the forward solution, can be uniquely determined with the knowledge of Newmann and Dirichlet boundary conditions on ; however, it cannot be obtained analytically for arbitrary geometry, conductivity distributions or boundary data. Thus, FEM was employed. Following this method, the body is conveniently divided into nite elements in this work, three node triangular elements and the electric potential is approximated within each element by a known function which depends on the electric potential values on nodes of the element, the new unknowns. The electric potential is described by a nite dimensional space and the problem of nding the nodal electric potentials, V , turns into an algebraic problem or the following linear system of equations: [K ]{V } = {I } (2)

where {V } is the vector of nodal electric potentials, {I } is vector of nodal electric currents and [K ] is the conductivity matrix of the system. A detailed reference in FEM can be found in the literature.7 When the electrodes are added to the model, the eect of the contact impedances must be taken into account. In this work, we considered the complete electrode model2 to represent these eects. This model gives rise to FEM matrices [Kel ] and vectors {Iel } corresponding to each electrode attached to the body, which are assembled on [K ] and {I }, respectively. Each matrix is given by: 1/2 0 0 1/2 ael el 1 0 1 0 [Kel ] = (3) 0 1/2 1/2 tel 0 1/2 1 1/2 2 and each vector by: 0 0 {Iel } = 0 I

(4)

where ael is the width and tel is the thickness (both assumed constant) of each quadrilateral nite element which comprises the model, el is the conductivity of the interface and I is the electric current applied to the electrode (if it is the case). el /tel is called electrode parameter, an unknown value in practice. The fourth node in the vector corresponds to the electrode and the others must be connected to nodes of the body. Assembling the matrices and vectors on the linear system (eq.2), we obtain the following system: [KT ]{VT } = {IT } (5)

Once the electric potential of a reference node is chosen, the system can be solved for the vector {VT }, a non-linear function of the resistivity distribution.

Proc. of SPIE Vol. 6143 614339-2

3. INVERSE PROBLEM
The inverse problem is the problem solved for the electric conductivities or resistivities, given the electric currents, electric potentials on the boundary, the structure of the model, and the geometry of the domain. Mathematically, the problem can be stated as: ne np 1 F = (ij ij 0 )2 (6) 2 i=1 j =1 where ne is the number of current load cases, np is the number of measurement points, ij 0 is the relative potential obtained from the i-th pair of electrodes in the j-th current load case and ij is the corresponding relative potential obtained from {VT }, e.g., through the solution of the FEM linear system, where the relative potential is the dierence between the potential of one electrode and the potential of the neighboring electrode, skipping one electrode (see Fig. 1). The minimization of this equation gives the best estimate for the conductivities and therefore for the image.

77
Figure 1. The measurement of relative potential.

This problem is ill-posed, which means that the best estimate for the image, based on eq. 6, may not be the best image. Therefore, regularization techniques must be applied. Each method studied, TOM and GNM, corresponds to a dierent regularization approach. We discuss these regularization approaches in the next subsection, in conjunction with characteristics of each method.

3.1. Topology Optimization Method

From the point of view of Bendse and Sigmund,4 TOM distributes material properties inside a xed domain with the purpose of optimizing the allowed performance of the corresponding system represented (the allowed performance), in this case, by eq. 6 and some constraints or regularizations. In this work, the material properties in each point vary, continuously, from the properties of a material A to a material B based on SIMP, or Simple Isotropic Material with Penalization. The mathematical form of this statement is given by: = p A + (1 p )B , in (7)

where A and B are the maximum and minimum allowed conductivities, respectively, and is the normalized design variable, which denes and varies from 0 to 1. Besides, when p is bigger than 1, penalization is introduced into the model, which means that intermediate values of can be reduced on the nal result. We used p=1 to match the GNM algorithm. We assumed in this work that the design variables are constant in each nite element. Thus, the conductivities were given by: (8) m = m A + (1 m )B where the subscript m indicates a nite element. Finally, the regularized problem solved for the conductivity distribution was: minimize w.r.t m subjected to F =
1 2 ne i=1 np j =1 (ij

ij 0 )2 (9)

0m 1

Proc. of SPIE Vol. 6143 614339-3

where M is the number of design variables. This proposed non-linear problem was solved by a Sequential Linear Programming (SLP) algorithm,6 which solves a sequence of linear problems through Linear Programming5, 6 (LP). In each LP problem (iteration), the non-linear problem above stated is linearized and the design variables are constrained between additional adjustable lateral constraints, or moving limits, which validates the linearization. A simplex based method solved each LP problem.6 Besides the mentioned constraints (box constraints), we considered another regularization scheme: the spatial lter8 (SF), which averages the moving limits based on limits of neighboring elements, weighted by the distances from the considered element (distance between centroids). The lter radius specied denes how many elements are taken into account in the averaging process (see Fig. 2). The lter equation is given by: mlm = and
q

mlq Wmq
q

Wmq

(10)

F R distmq Vq (11) FR where mlm is the m-th moving limit, F R is the specied lter radius, distmq is the distance between m-th element and q-th element and Vq is the volume of the q-th element. The SF is applied at each iteration and, mathematically, drives the optimization into regions in the space of feasible solutions where the distribution of the design variables is smooth. Wmq =

mc s h

Figure 2. A piece of a two-dimensional FEM mesh showing some elements within the lter radius FR, which are therefore considered in the ltering process.

Finally, the derivatives of F with respect to the design variables m which are used in the linearization of the problem were obtained through the adjoint method.5

3.2. Gauss-Newton Method

The second algorithm addressed in the present work is the Gauss-Newton Algorithm in the generalized form.3 The update of the vector of unknowns, , follows eq. 12
t t k+1 = k + (Jk Jk + W )1 (Jk W (z h(k )) W k )

(12)

where denotes the vector of the unknown physical property, for instance, resistivity or conductivity, is the step size, is a regularization parameter, Jk is the Jacobian matrix and each element of this matrix is Jk,i,j = (zi hi ()/j , z is the vector of measured electrical potentials, W is a regularization matrix and h(k ) is the observation model derived from a nite elements model.

Proc. of SPIE Vol. 6143 614339-4

The regularization matrix W = Lt L, where L is a high pass lter, penalizes the high frequency content on the vector. The high pass lter is computed using a low pass gaussian lter,9 G, according to L = I G, where I is the identity matrix. Since the Topology Optimization Method uses a change of variable according to eq. 7 with much success, in this work, two dierent implementations of the Gauss-Newton Algorithm are performed. One of them uses conductivity as unknowns and consequently p = 1 on eq. 13. The other implementation uses resistivity as unknowns and consequently p = 1 on eq. 13. p (13) i = i where p belongs to the set of real numbers, i is conductivity of the i-th element ans i is the i-th element of the vector of unknowns.

4. RESULTS
In this section, images and other results are presented. Examples using experimental data are shown. These data were obtained from a cylindrical container with 30 bar electrodes attached to its boundary (see the sketch in Fig. 3) with 35 mm high and 10mm wide (the thickness of each electrode is not necessary for the algorithm). They are equally spaced along the container, which was lled up to 35 mm with a 0.3 g/L saline solution (NaCl). Its resistivity is approximately 17 m. The inner diameter of the container is 300 mm. This diameter limits the conductive body. The small circle shown in Fig. 3 represents a glass object with diameter and resistivity10 equal to 32 mm and approximately 106 m, respectively, which was immersed into the container. The minimum distance between the center of the object and the boundary is 30 mm.

g1ass object

electrode

Figure 3. Sketch of the cylindrical container.

To obtain the electrical potentials in the electrodes, a pair of them was electrically excited following the load pattern seen in Fig. 4 (this gure corresponds to one current load case). Then, the relative potentials were measured, except for those which share current carrying electrodes, due to hardware limitations. The pair of excited electrodes was successively changed until a satisfactory number of observations (or until enough information as potential values) under dierent angles was obtained, thus providing the necessary data for a high quality image. 30 current load cases were applied following this pattern.

electric current
4.

Figure 4. Load or excitation pattern considered.

Proc. of SPIE Vol. 6143 614339-5

The image reconstruction was carried out in a mesh with 2946 elements (including the elements of the electrode model) or 2886 unknowns, and 1564 nodes (see Fig. 5). The initial guess was 35 m. Electrode parameters, which were previously obtained, are equal to 0.02 m2 . All computations are carried out on a 3.6GHz PC with 1024MB of RAM.

Y Z X

Figure 5. Mesh used for image estimation.

The Fig. 6 is the resistivity estimate obtained by TOM algorithm. The lter radius was chosen so that only the neighboring elements were considered in the averaging process. The position and size of the glass object were recovered. However the amplitudes and spatial variation are small, which is attributed to the SF. The resisitivity of the saline is about 200 m, which seems ten times higher than expected taking into account the amount of NaCl dissolved in water. One hundred iterations where required and the elapsed time was 1023 s.
Topology Optimization

1000 900 800 700 600 500 400 300 200 0.15 0.1 0.05 0 0.05 0.1 0.1 0 0.05 0.05 0.1 0.15

Figure 6. Image obtained by the Topology Optimization Algorithm TOM.

Proc. of SPIE Vol. 6143 614339-6

The Fig. 7 is the resistivity estimate obtained by Gauss-Newton algorithm when the unknowns are conductivities, p = 1. The position of the glass object was recovered properly. The peak resistivity of the glass object, = 112.04 m, is much smaller than expected. The diameter of the object also resulted smaller than expected. The resisitivity of the saline is about = 20 m, which is on the order of the expected taking into account the amount of NaCl dissolved in water. The standard deviation of the gaussian lter is s = 0.02 and the regularization parameter = 0.75. Smaller values of the regularization parameter lead to meaningless images. Ten iterations where required and the elapsed time was 821 s.
Newton Raphson

110 100 90 80 70 60 50 40 30 0.15 0.1 0.05 0 0.05 0.1 0.1 0 0.05 0.05 0.1 0.15

Figure 7. Image obtained by the Gauss-Newton Algorithm using conductivity as unknowns, p = 1

When the Gauss-Newton Algorithm was implemented using resistivity as unknowns, p = 1 the position of the object was recovered properly and the diameter of the object resulted as expected, but the resistivity of the object was obtained much smaller than expected = 62.0 m, as shown on Fig. 8.The resisitivity of the saline is about = 20 m, which is on the order of the expected taking into account the amount of NaCl dissolved in water. The standard deviation of the gaussian lter is s = 0.04 and the regularization parameter = 1.0e 7. Smaller values lead to meaningless images. Four iterations where required and the elapsed time was 341 s.
Newton Raphson

60 55 50 45 40 35 30 25 20 0.15 0.1 0.05 0 0.05 0.1 0.1 0 0.05 0.05 0.1 0.15

Figure 8. Image obtained by the Gauss-Newton Algorithm using resistivity as unknowns, p = 1

Proc. of SPIE Vol. 6143 614339-7

The Fig. 9 is the resistivity estimate obtained by Total Variation11 (TV), which is dened in eq. 14, for a given continuous and dierentiable function of one variable x: T V ( (x)) = | (x)|dx (14)

Newton Raphson

250

200

150

100

50 0.15 0.1 0.05 0 0.05 0.1 0.1 0 0.05 0.05 0.1 0.15

Figure 9. Image obtained by the Gauss-Newton Algorithm using resistivity as unknowns, p = 1, using Total Variation

The GNM converged much faster than TOM algorithm. We can see that 4 iterations were required for the, approximately. Memory needs and approximate computing times are seen in Tab. 1. GNM algorithm was faster than TOM. However, it required a lot of storage, which is attributed to the large dimensional matrices necessary for computing the update of the resistivity (or conductivity) vector.
Table 1. Approximate memory needs and computing times.

GNM (resistivity) Memory (MB) Time (s) 547 341

GNM (conductivity) 547 821

TOM 195 1023

5. FINAL COMMENTS
From these preliminary results some comments can be made. In general, the resisitivity of the glass object was underestimated, to achieve better values the electrodes parameters should also be estimated. The image obtained by the Topology Optimization Method presented the highest resistivity, with a satisfactory diameter for the glass object and good position of the object. However, the algorithm overestimated the resistivity value of the saline. The resistivity of the saline is expected to be near 15 m, judging from the NaCl concentration. The smaller resistivity obtained using the Gauss-Newton method suggests that the method requires a regularization of larger magnitude than TOM. This is probably related to numerical error propagation of the Sequential Linear Programming compared to the Parallel Direct Sparse Solver of the MKL Intel R . Further investigations are required to verify this statement. The overestimation of the resisitivity of the saline by TOM can be explained by a restriction on the spatial derivative of the resisistivity. In spite of the overestimation of the resistivity of the

Proc. of SPIE Vol. 6143 614339-8

saline and larger computing time, the use of the Sequential Linear Programming on TOM seems to be benecial for the quality of the image and for the memory used. Besides, Some sharp restrictions, for instance, the nonnegativeness of the conductivity, are implemented using Linear Programming without increase of numerical error propagation and has regularizing eects.

ACKNOWLEDGMENTS
The authors are grateful for the nancial support from The State of S ao Paulo Research Foundation, through a masters schollarship, FAPESP number 02/10574-0, and the nancial support under FAPESP number 01/05303-4 .

REFERENCES
1. A. L. Hyaric and M. K. Pidcock, An image reconstruction algorithm for three-dimensional electrical impedance tomography, IEEE Trans. Biomed. Eng. 48, pp. 230235, 2001. 2. P. Hua and et al., Finite element modeling of electrode-skin contact impedance in electrical impedance tomography, IEEE Transactions on Biomedical Engineering 40, pp. 335343, 1993. 3. P. J. Vauhkonen, Image Reconstruction in Three-Dimensional Electrical Impedance Tomography. PhD thesis, University of Kuopio, Kuopio, Finland, 2004. 4. M. P. Bensoe and O. Sigmund, Topology Optimization: Theory, Methods and Application, Springer Verlag, Berlin, Heidelberg, New York, 2003. 5. R. T. Haftka and Z. Gurdal, Elements of Structural Optimization, Kluwer Academic Publishers, Netherlands, 1999. 6. W. H. Press and et al., Numerical Recipes in C The Art of Scientic Computing, Cambridge University Press, Cambridge, 1999. 7. K. J. Bathe, Finite Element Procedures, Prentice-Hall, New Jersey, 1996. 8. E. L. Cardoso, Controle de complexidade na otimiza c ao topol ogica de estruturas cont nuas, Masters thesis, Universidade Federal do Rio Grande do Sul, Porto Alegre, Brazil, 2000. 9. R. C. Gonzalez and R. E. Woods, Digital Image Processing, Prentice-Hall, New Jersey, 2002. 10. C. J. Phillips, Glass: Its Industrial Applications, Reinhold Publishing Corporation, New York, 1960. 11. V. Kolehmainen and et al., A bayesian approach and total variation priors in 3d electrical impedance tomography, in Proceedings of the 20th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 20, pp. 10281031, 1998.

Proc. of SPIE Vol. 6143 614339-9