On the Weak and Numerical Solutions of the Forward Problem in EEG DIANA RUBIO1 and MARÍA INÉS TROPAREVSKY2 1 Instituto de Ciencias, Universidad Nacional de General Sarmiento José M. Gutiérrez entre José L. Suárez y Verdi 1613 - Los Polvorines - Prov. de Bs. As. 1 Escuela de Ciencia y Tecnología, Universidad Nacional de General San Martín Calle 91 Nº 3391, V. Ballester – Prov. de Bs. As. 2 Departamento de Matemática, Facultad de Ingeniería Universidad de Buenos Aires ARGENTINA Abstract: - The Forward Problem in EEG consists on finding the scalp potential due to current sources within the brain. Electrical activity in the brain is governed by the laws of electromagnetics. Thus, to solve the problem, Maxwell Equations must be solved in a volume representing the head. In this work we consider a simplified model of the human head and a static approximation of the dynamical model. We use BEM techniques to solve the resulting equations. We illustrate some properties of the solution by typical numerical examples. Key-Words: - Bioelectric Potentials, Boundary Element Method, Forward Model, Spherical Head Model 1 Introduction The potential distribution on the scalp produced by current sources within the brain can be measured by an Electroencephalograph (EEG). For several decades, neurologists have been interested in solving the problem of determining the location of these current sources from the measured potential on the scalp. In the case of epilepsy there are a few dipoles that give major contribution in the generation of the electric field and we are interested in finding their location based on the superficially measured potential values. In order to solve this problem, called the INVERSE PROBLEM in EEG, we must be able to calculate the superficial potential for any possible dipole configuration which is called the FORWARD PROBLEM (FP) in EEG. In this work we present some results concerning Numerical Solutions of the Forward Problem in EEG. Electrical activity in the brain is governed by the laws of electromagnetics. Thus, to solve the Forward Problem, Maxwell's equations must be solved in a volume representing the head. In order to solve numerically the problem of finding the location of the source of an epileptic seizure we need to propose a mathematical model, i.e., we need to choose the equations that will represent the process and the domain where these equations will be solved. At this stage of our work we simplified the model in two ways. The first one is about the differential system of equations arising from Maxwell's equations: we consider the static approximation to the equations. The static approximation of the dynamical model is standard in neurology and it is justified by the high speed of propagation of waves in the head (see [6], [9]). This approximation uncouples the magnetic and electric equations, in consequence the model for the electrical activity consists of the Poisson equation with boundary conditions (see [12]). The second simplification of the model is about the domain where we solve the equations: the human head is a complicated anisotropic media with tissues of different conductivity values. We simplify the geometry of the head representing it as three volumes with different conductivity values. Each volume is surrounded by the next one. They represent the brain, the skull and the scalp. In this case where the domain is described by concentric spheres it is possible to solve the differential boundary value problem exactly by a series of functions (see [3], [4], [11] among others). For general domains it is not possible to obtain the solution in this way. Therefore it is necessary to use numerical methods. As we are interested in finding a solution on a realistic head model, at a first step we propose an approximated solution although we are working with spherical models. We intend to adapt this scheme to the realistic head model. To solve numerically the differential problem we use the boundary element technique which allows us to reduce a problem in a volume to a problem over surfaces. In this way the numerical burden is alleviated. Other approaches with FEM techniques [15], the reciprocity theorem [13] and the cubic Hermite boundary element procedure [1] were also developed. The paper is organized as follows: in Section 2 we present the differential system and establish some properties of its weak solution. In Section 3 we present the integral formulation of the problem. Section 4 introduces the Discrete System. Numerical examples are shown in Section 5. Finally we present the conclusions in Section 6. 2 The Differential Equations 2.1 Maxwell Equations The electrical activity of the brain consists of currents generated by biochemical sources at the cellular level. The electric and magnetic fields that they produce can be measured and predicted because they obey physical laws: Maxwell's Equations [7] and the equation of conservation of the charge .J / t , where E is the electric field, B the magnetic field, the density of charge and J is the total current density. The density J can be decomposed in the current density due to the macroscopic electric field, and the current density caused by synaptic activity as follows J E J i (1) where is the conductivity of the media. If we accept the properties of the tissues involved (conductivity , magnetic permeability values and dielectric values to be constant, see [9]), the velocity of propagation of the electromagnetic waves caused by potential changes within the brain is about 10 5 m / s . This means that the effect of the potential changes may be detected simultaneously at any point in the brain or in the surrounding tissues. In other words, the currents caused by sources in the brain behave in a stationary way (see [6],[9]). The use of the static approximation of Maxwell's Equations to describe the process is then justified. We can simplify (1) to obtain: .( E ) 4 xE 0 4 J. c These equations together with the law of conservation of charge, that in this case can be written as (2) .J 0 give rise to (3) E u. Substituting (3) in (1), and taking the divergence of both sides we arrive to the Second Order Partial Differential Equation that describes the relationship between the measured potential u and the current density J i (the Poisson equation): .B 0 x( B ) .( ( x ).u( x )) .J i (4) with boundary condition u( x ) 0 , x G , n (5) where G denotes the volume representing the head and n represents the outward normal. This condition reflects the fact that the normal derivative to the head at the boundary must be zero since air is an insulating material that does not support current flow. As mentioned in the Introduction, we assume that G can be described as three concentric spheres where the radii and conductivity values are given (see [8]). We denote them from the inner one to the outer: G1 the brain, G2 the skull and G3 the scalp. The surfaces between them are denoted by: S1 , S 2 and S 3 (the scalp) respectively (Fig.1). Fig.1. The Three Sphere Model In neuromagetics a current dipole is widely used to approximate a primary current. In the case of epilepsy a dipole can be thought as a concentration of J i to a single point rq : J i1 M q ( r rq ), (6) where (.) is the Dirac delta distribution and M q is the dipole strength. In this work we considered that there is an unique dipole responsible for the seizure process. A current dipole is completely described by six parameters: three to determine its location and three to define its strength. There are physical considerations that must be taken into account. They can be described by the equations: u Si 0 u n Si 0 where . represents the difference between the values of the functions inside the brackets through the indicated surface. These equations represent the continuity of the potential and of its normal derivative across the different regions. The scalp potential u(x) is measured as a difference between the potential value at each point x S 3 and its value at a reference point x 0 S 3 , thus u( x0 ) 0 . Therefore, the differential equation to be solved is .( ( x ).u( x )) .J i with boundary condition u ( x ) 0, n x G and continuity condition for the normal derivatives at the transition surfaces u n and it must verify u( x0 ) 0 . u Si 0 Si 0, 2.2 Existence and Uniqueness of Solutions Equation (4) relates the impressed current J i with the electric potential u. This equation is a second order elliptic one and it can be written in the divergence form as Lu .( ( x ).u( x )) .J i . The function (x) that contains the conductivity of the different tissues at each point x, may be considered discontinuous and piecewise constant 1 , ( x ) 2 , , 3 x G1 x G2 . x G3 In order to assure the existence of a solutions of (4) with boundary condition (5) we need to work with a generalized solution [5]. It can be proved that there exist generalized solutions u of (4) if Ji verifies .J i values chosen for solving the equations are not accurate [17]. We state that the weak solutions of the equations are continuous with respect to . That is, if the values of two different are close, the corresponding scalp potential solutions will be close too. Suppose that u is the generalized solution obtained for a given , and that u is the generalized solution for . Considering the weak form of the equation (4) it can be proved that u u 2 C where C is a constant that depends on G, the electric field and the conductivities. We will illustrate this relationship with several examples in Section 5. In order to calculate a solution of (4), in the next Subsection we present Integral Equations for u. 3 Integral Equations for the Potential Integral equations can be derived from the differential boundary value problem using Green's integral theorems (see [6], [14]). The following formulas are obtained ( r )u( r ) 1v( r ) 3 r r' j j 1 u( r' ) dS' 3 4 j 1 r r ' S j for r S j , where v( r ) 0 and that these solutions differs only on a G constant. In the case we are considering, the above condition can be written as J i 0 and it is G automatically fulfilled because, under our hypothesis, J i has finite support inside G1 (dipole). The uniqueness of solution is also justified because at the reference point, x 0 , on the scalp the potential u( x0 ) 0 . Finally if u is a solution of the problem, it is also a generalized solution. 2.2.1 Continuity of the Solution with respect to the Conductivities When solving the equation (4) we need to know the conductivity of the brain, skull and scalp. There exist typical values for these constant [9]. However, it may differ from the real ones, thus it is important to estimate the error that appears if the conductivity 1 .J i r r' 41 G dr' . (7) Let r r j S j to obtain another equation for points on the surfaces of conductivity transition S j : k k 1 2 u( r ) 1v( r ) j j 1 r r' u( r' )n( r' ) dS' 3 4 j 1 r r ' S 3 (8) j for r S k . The above formulas provide the solution u at points on the surfaces of conductivity transition Sj. In the next Section we present a technique for the discretization of the integral equations (8) and describe the way we solved them. 4 The Discrete Problem In this section we propose a discretized version of the integral equations for the potential. 4.1 The BEM Technique The idea behind the Boundary Element Method (BEM) is to transform a differential equation given on a domain into an integral equation on the boundary of that domain. Thus, the dimension of the problem is reduced yielding to a reduced order numerical system (see [2]). In order to apply the BEM technique to our problem we must discretize the surface integrals (8), and approximate the volume integral over G for v(r), given by equation (7). at the nodal points, that is, at the vertices of the elements E j ,k of the grid. More precisely, if r is a nodal point, U(r,r') u(r')n(r') r r' H ( r , r' ) n( r' ) r r' 3 4.2.1 The Grid If we divide the surfaces into elements Ej,k such that verifies S j E j ,k then the grid coincides exactly k with each surface Sj j=1,2,3, and there is no error due to the surface discretization. We divide the surface into a set of spherical elements and choose the nodal points to be the vertices of these elements. We construct spherical rectangles as follows. We consider the spherical coordinates >0, 0,2 ) and 0, . On each sphere is fixed then the parameters are and . We divide [0,2) into n intervals and [0,] into n=n/2 intervals. This choice give rise to a grid of N=n(n nodal points and M=nn spherical elements. These elements are rectangles except for 0, h and - h , , v H1 where they become triangles as shown in Fig. 2. and , we approximate each u j ,k ,i E j ,k i 1 The discretization error depends on the grid and on the function and integral approximations. 3 Sj Nk The Discretization of the Equations r r' surface integral I S j ( r ) U ( r , r' )dS' by I S j ( r ) 4.2 r r' Nk H ( r ,r' )dS' , (9) E j ,k where uj,k,i are the value of the potential at vertex i of the element Ej,k and Nk is the number of vertices of the element Ej,k. As we are working on spheres, we can calculate the actual normal at each nodal point exactly. For simplicity we take the same amount of points, say N, on each surface that give rise to M elements per surface (see Section 4.2.1 for a description of the chosen grid). There exist other ways to discretize the potential. In [16] there are several comparisons of the results obtain for different approximations. 4.2.3 The Integral Approximation Finally, to calculate the integrals involved in (7) and (9) we need to select a numerical rule. The integral on (7) is easy to solve since we model the primary current Ji as a dipole (see equation (6)). The integral on (9), that is, r - r' H(r,r' )dr' dS , r - r' E j,k E j,k is the solid angle subtended by the element Ej,k at the point r of the grid. We apply the Simpson’s rule for r E j ,k and geometric properties of the solid angle in the improper case (see [10]), that is when r E j,k . 4.3 The Linear System Fig.2. The Spherical Grid 4.2.2 The Potential on the Grid The next step is to decide the way in which the function u will be approximated over the elements of the grid. We consider u to be an average of its values The implementation of all these approximation schemes lead to a Linear System of Equations (D -A) u=C (10) where D is the block diagonal matrix arising from the right hand side of (8), that is, D=diag(D1,D2,D3) where Di i i 1 2 for i=1,2,3. A is the matrix resulting from the discretization of the surface integrals ISj (9), C is the discretization of v(r) and u is the vector containing the values of the potential at the nodes uj,k,i. Once the grid is chosen, the matrices of the system, D and A, depends only on the conductivities of the brain, skull and scalp. Note that all the values of uj,k,i are involved in the calculation of the whole vector u. After solving (10) we can reconstruct the scalp potential interpolating the values obtained at the nodal points of the external surface (representing the scalp). 5 Numerical Results We solve the Forward Problem for a volume conductor G composed by three concentric spheres of radii .071, .078 and .085 mm. The conductivity values commonly used are (see [9]) ( 1 , 2 , 3 ) ( 0.33,0.0042,0.33 ) Am/m for the brain, skull and scalp respectively. We considered dipoles oriented in different ways. In Fig.3 the scalp potentials due to a transversal and radial dipoles are shown. Fig.4. Surface Potentials for Different Skull Conductivities. To quantify the differences between the potential distribution for different i i 1,2,3 we compare the 2-norm of the vector u containing the values of the potential over the three surfaces with the 2-norm of the vector . We present some results in the following table i j ui u j 4.50000e-003 2.20000e-003 2.30000e-003 4.03150e-003 2.75879e-003 1.27304e-003 2.23663e-004 9.95365e-008 5.86464e-008 4.16362e-008 4.58856e-009 3.09301e-009 1.50926e-009 1.46143e-010 Fig.3. Scalp Potential for a Tangential Dipole (on left) and Radial Dipole (on right). 5.1 Numerical Results Conductivity Values for 2,3 1, 3 1, 2 4, 5 5, 6 4, 6 4, 1 1 2 Brain Cond. 0.330000 0.330000 Skull Cond. 0.004200 0.006500 Scalp Cond. 0.330000 0.330000 3 0.330000 0.002000 0.330000 4 0.329900 0.004205 0.330200 5 0.327000 0.004150 0.333000 6 0.329000 0.004180 0.331100 Different As mentioned in Section 2 the solutions of the FP are continuous with respect to the values of . To illustrate this relationship we numerically solve the FP and found the potential distribution for different values of i i 1,2,3 around the references values mentioned above. We observe that for different values of i the patterns of the potential distribution are very similar, they only differ in their magnitude, not in their position. This fact can be observed in Fig.4 where the scalp potential for a dipole located at rq (0,0.0062,0.04) and with M q (1.2 10 10 ,0.6 10 10 ,0.6 10 10 ) , for different conductivities is shown. In the first column ( 0.33,0.0065,0.33 ) , in the second one ( 0.33,0.0042,0.33 ) and in the third one ( 0.33,0.0020,0.33 ) . It can be observed that as the order of i j decreases, the order of u i u j decreases too. 6 Conclusions In this work we established the continuity of the weak solution to the FP of EEG with respect to the vector parameter (conductivity) and shown how this fact is reflected in a numerical solution. From the examples presented in Section 5, it can be seen that the changes in the conductivity produces variations only in the magnitude of the potential as we pointed out. We proposed a spherical grid that exactly matches the surfaces of transition of the head model which reduces the discretization error. With respect to the potential, we discretized it at each grid element and used geometric properties to calculate the resulting integrals. These results were illustrated by some typical numerical examples. References: [1] Bradley, C. P., Harris, G. M. and Pullan, A. J., The Computational Performance of a High-Order Coupled FEM/BEM Procedure in Electropotential Problems, IEEE Trans. Biomedical Engineering, Vol.48, No.11, 2001, pp.1238-1250. [2] Brebbia, C.A., Telles, J.C.F.,and Wrobel, L.C., Boundary Element Techniques: Theory and Applications in Engineeering, Springer Verlag, N.Y. 1984. [3] de Munck, J.C., The Potential Distribution in a Layered Anisotropic Spheroidal Volume Conductor, J. Appl. Phys., Vol.64, No.2, 1988, pp. 464-470. [4] de Munck, J.C. and Peters, M.J., A Fast Method to Compute the Potential in the Multisphere Model, , IEEE Trans. Biomedical Engineering, Vol.40, No.11, 1993, pp. 1166-1174. [5] Evans, L.C., Partial Differential Equations, AMS Vol.19, 1999. [6] Hamalainen M., Hari R., Ilmoniemi R. J., Knuutila, J. and Lounasmaa O., Magnetoencephalography, Theory, Instrumentation and Applications to Noninvasive Studies of the Working Human Brain, Reviews of modern Physics, Vol.65, No.2, 1993, pp.414-487. [7] Jackson, J.D. , Classical Electrodynamics, J.Wiley, New York. [8] Lagerlund, T.D., EEG Source Localization (Model-Dependent and Model-Independent Methods), Electroencephalography: Basic Principles, Clinical Applications, and Related Fields, Chap. 46, 1999, pp 809-822. [9] Lopes da Silva, F. and Van Rotterdam, A., Biophysical Aspects of EEG and Magnetoencephalogram Generation, Electroencephalography: Basic Principles, Clinical Applications, and Related Fields, Chap.5, 1999, pp 93-109. [10] Meijs, J. W. H., Weier, O. W.,, Peters, M. J., and Van Oosterom, A., On the Numerical Accuracy of the Boundary Element Method, IEEE Trans. Biomedical Engineering, Vol.36, No.10, 1989, pp.1038-1049. [11] Mosher, J. C., Leahy, R. M., Ewis, P.S., EEG and MEG: Forward Solutions for Inverse Methods, IEEE Trans. Biomedical Engineering, Vol.46, No.3, 1999, pp. 245-259. [12] Plonsey, R., The Biomedical Engineering Handbook, Boca Raton, FL.CRC Press, 1995, pp 119-125. [13] Riera J., Fuentes M., Electric Lead Field for a Piecewise Homogeneous Conductor Model of the Head, IEEE Trans. Biomedical Engineering, Vol.45, No.6, 1998, pp.746-753. [14] Sarvas, J., Basic Mathematical and Electromagnetic Concepts of the Biomagnetic Inverse Problem, Phys. Med Biol., Vol.32, No.1 , 1987, pp.11-22. [15] Schimpf, P. H., Ramon, C., and Haneisen, J., Dipole Models for EEG and MEG, IEEE Trans. Biomedical Engineering, Vol.49, No.5, 2002, pp.409-418. [16] Schlitt, H., Heller, L., Aaron, R., Best, E.,Ranken, M., Evaluation of Boundary Element Methods for the EEG Forward Problem: Effect of Linear Interpolation, IEEE Trans. Biomedical Engineering, Vol.42, No.1, 1995, pp.52-58. [17] Tsai, Y, Will, J.A., ScottHubbard-Vsan Stelle, Cao, H., Tungjitkusolmun., Error Analysis of Tissue Resistivity Measurements, IEEE Trans. Biomedical Engineering, Vol.49, No.5, 2002, pp.484-494.