IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO.

5, MAY 2010 1193

A Fast Forward Problem Solver for the Reconstruction of Biological Maps

in Magnetic Induction Tomography
Nuno B. Brás, Raúl C. Martins, A. C. Serra, and A. Lopes Ribeiro
Instituto de Telecomunicações, Lisboa 1049-001, Portugal

The magnetic induction tomography (MIT) image reconstruction is a distributed parameter estimation problem normally solved it-
eratively, where each iteration engages several 3-D harmonic eddy-current problems. These calculations are the main time-consuming
process in the MIT reconstruction. In this paper, an approach to accelerate the resolution of a set of similar eddy-current problems
to be used in the reconstruction process is presented. Each eddy-current problem is described by a reduced magnetic vector potential
eddy-current formulation using the finite integration technique (FIT) framework defined over an octree based subgridding scheme with
several performance precautions. The solver accuracy is compared with analytical solutions of appropriate geometrical scenarios. A
performance assessment is presented for a full set of eddy-current problems that constitutes a forward problem. The new method re-
duced the computing time approximately to 11% of a similar problem without performance concerns maintaining a relative mean error
inferior to 1.5% relatively to analytical simulations.
Index Terms—Eddy current, FIT, magnetic induction tomography, RMVP, subgridding.

I. INTRODUCTION methods are very useful when the number of measurements is

low. They require calculating a solution for each pair of drive
AGNETIC INDUCTION TOMOGRAPHY (MIT) has and measurement coils in each iteration of the inverse problem
M been studied as a candidate method for tomographical
imaging of biological tissues [1], [2], industrial [3] and geophys-
resolution. Nevertheless, even for a simple acquisition setup,
there are a high number of eddy-current solutions to be found.
ical applications [4]. The MIT objective is to reconstruct the This performance issue is shown to be of critical importance in
passive electrical properties of a body by measuring the mag- [7] where it is reported that each inverse problem iteration time
netic field generated in its interior when exposed to an external was of 53 min in a 1.7 GHz single processor personal computer.
variable magnetic field. Measurements are usually performed by Increasing the number of measures will turn this process even
coils in a passive open loop setup. The magnetic field is gener- more critical. The presented forward method is suitable for this
ated by a coil fed by a source current. inversion strategy.
This imaging method received a special attention in recent Another type of methods has been adopted to solve general
experimental and numerical studies in biomedical applications distributed parameter estimation problems that use a truncated
due to its low cost, contactless sensors attaining a good penetra- version of the Gauss-Newton descent direction in each recon-
tion in biological bodies with bone like properties, contrarily to struction iteration. When the number of measurements is large,
what happens with electrical impedance tomography. In the case like in a moving sensing coil system [5], this kind of methods
of biological tissues, the major electrical property to be charac- has the nice property that the number of forward problems to
terized is the complex conductivity. Bodies to be analyzed have solve does not increase with the number of measurements. The
some tens of centimeters, described by small changes on the presented forward problem solver is also suitable for this kind
conductivity between adjacent tissues, typically 0.5 to 2 S/m. of inversion strategy, where it is necessary to know explicitly a
Frequencies ranging from tens of kilohertz to some megahertz measurement description.
are used in the excitation magnetic field [5]. This context de-
fines the forward problem as a set of diffusion like eddy-current A. Discretization Methods for the Eddy-Current Problem
problems where the conductivity maps are defined by approxi-
The edge finite-element method (eFEM) has been widely
mately constant patches with abrupt discontinuities.
studied to solve eddy-current problems [8], [9] since it is
The MIT inverse problem is an ill-posed nonlinear opti-
configurable to any geometric shape.
mization problem that uses several magnetic field measures
Other class of methods to solve the problem is based on the
around the body for several positions of the magnetic source.
eddy-current integral formulation. In this case, the problem
The number of acquisitions should be high in order to reduce
is defined just over the body, which reduces the number of
the ill-posedness of the problem. However, this increases the
points and simplifies the edge/interface conditions. However,
number of eddy-current problems to solve during the inverse
the problem is formulated by a full global matrix [10] which
problem resolution. Methods based on the reciprocity theorem
inherently generates difficulties in a fast resolution process.
and resulting adjoint fields have been used to decrease the
Here it is explored the idea that the refined space should
number of calculations per iteration, namely in the calculation
change with the solution knowledge during the reconstruction
of the so-called sensitivity matrices of the problem [6]. These
process. An orthogonal meshing with multiple resolution vol-
umes (or subgrids) has a very quick global stiffness matrix build,
Manuscript received August 03, 2009; revised October 16, 2009; accepted making it an interesting candidate to be used as the eddy-current
November 28, 2009. Current version published April 21, 2010. Corresponding
author: N. Brás (e-mail: nuno.bras@ix.it.pt). discretization method in this context where the resulting matrix
Digital Object Identifier 10.1109/TMAG.2009.2039486 has to be built along iterations. There are several methods that
0018-9464/$26.00 © 2010 IEEE
1194 IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO. 5, MAY 2010

discretized consistently the Maxwell equations in their integral

form over simple orthogonal or nonorthogonal meshes. They are
generically called Yee like staggered schemes [11]. Examples
are the classic Yee discretization of the space/time, the mimetic
finite differences, proposed in [12] or the finite integration tech-
nique (FIT) [13].
In this paper, an eddy-current description using the reduced
magnetic vector potential and explicitly imposing the electric
potential is used. This method is adapted to be used over the
FIT framework using a fast subgridding scheme to solve each
eddy-current electromagnetic problem that composes a forward Fig. 1. Conceptual figure of the eddy-current problem.
problem of magnetic induction tomography. The resulting for-
ward solver was then optimized, allowing a reduction of 89%
in the computation time compared with a naive approach to Here, is the eddy-current matricial operator that maps fields
the problem while maintaining a relative mean error inferior to into their sources is the measure matrix that trans-
1.5% relatively to analytical simulations. forms the fields in measurements, is the measured data array,
Special care is taken in order to accelerate the implemented is a regularization term with a priori information about
method: 1) a multiproblem approach based on the choice of an the conductivity map and is a regularization parameter
optimum incomplete LU factorization threshold parameter to that adjusts the regularization term weight in the reconstruc-
act as a preconditioner of the set of eddy-current problems is tion. A typical redefinition of this problem is to substitute the
used; 2) discretization model simplifications like reducing the constrain condition in the cost function, resulting in the uncon-
grid size over the sensing coils and defining a fast subgrid inter- strained problem:
polation are done in order to reduce the problem size yielding
however, much of its initial accuracy.
The paper is organized as follows. In Section II and III, the (2)
electromagnetic problem and the discretization method is pre-
sented. In Section IV the multisolver technique is presented. In
Each iterating step of the resulting nonlinear problem embraces
Section V, accuracy comparison with analytical known solu-
solving several forward problems. The complete forward func-
tions and performance assess of the proposed methodology are
tion is given by , where and is the ma-
quantified.
trix composed by the column vectors . The matrix , cor-
responding to the eddy-current problem, is defined in the next
II. MODEL DEFINITION sections.

The electromagnetic model is described by an isotropic pa- B. The Eddy-Current Mathematical Model
rameter model. Although not real in the biological context [15],
isotropy of the complex conductivity strongly reduces the ill A brief comprehensive description of the used differential
condition of the problem, turning the reconstruction much more eddy-current equations is here presented in order to establish a
feasible. All models in literature are based on this simplification, connection to the used discrete version. Consider a typical eddy-
(see, e.g., [7] or [4]). It is also considered a constant magnetic current problem of a biological MIT application, depicted in
permeability , as is generally accepted when studying mag- Fig. 1. Air region is delimited by and the interface boundary
netic properties of biological tissues excepting specific patholo- between Air and a Body is called . In the Air region the con-
gies ([16], [17]). This simplification takes the forward problem ductivity is null and there are current sources and sensing coils.
to a much easier context and since conductivity is the main re- The followed approach uses a steady-state problem approx-
constructed map, it covers a real scenario. imation for a harmonic description of the Maxwell equations.
The governing differential equations are given by
A. The Inverse Problem
(3)
The MIT parameter estimation problem could be stated as
a minimization problem between the measured data array for (4)
each of the source field positions and the simulated forward (5)
problems that incorporate the conductivity map information (6)
:
where . Constitutive equations relate the fields with
the correspondent flux densities. For a constant permeability and
an isotropic conductivity one gets

(1) (7)
BRÁS et al.: FAST FORWARD PROBLEM SOLVER FOR THE RECONSTRUCTION OF BIOLOGICAL MAPS 1195

Substituting (3) in (4), the electric description of the general calculated by the Biot-Savart law using a numerical approx-
eddy-current law for constant permeability is given by imation of an elliptical integral defined in [18]. The relation
is fulfilled and (11) and (12) take the form
(8)
(13)
where . In this equation the current divergence free
condition is directly imposed because the divergence of the first (14)
term in the left-hand side is null then ( ,
In the interface , since there are no surface currents, the
even over the source coil, which is true for an inductive source).
continuity of the normal component of and the continuity of
The electric scalar potential and the magnetic vector po-
tangent component of is directly fulfilled. In this interface the
tential are often used to solve this problem because they
relation between and is established by forcing the current
are continuous functions when no skin charges and skin cur-
density normal component to be null since there is no conduc-
rents are considered. The electric scalar potential is often
tivity material outside this boundary:
substituted by a modified electric scalar potential , related
by . This modified potential is used in order to (15)
make the final matrix symmetric in terms of the used potentials.
The decomposition , with This Newmann condition needs not to be explicitly applied
without a gauging condition for do not uniquely define the since (14) defined at the interface fulfills it automatically. A
potential. Although several ungauged formulations have been homogenous Dirichlet boundary condition is imposed at
used to solve eddy-current problems [25], here it was defined for the normal component of which in practice is done by
that a nonsingular stiffness matrix should be attained in order forcing .
to be handled in the inverse problem formulation. The gauging
process allows defining a uniqueness condition for potentials III. THE DISCRETIZATION METHOD
and, hence, a unique solution for the eddy-current problem so- Several discretization methods were recently developed or re-
lution. Using the defined potentials and the so-called Coulomb defined based on topological principles. These methods define
gauge given by , (8) can be written as the Maxwell equations as metric free equations, glued by the
metric dependent constitutive laws in accordance with the un-
derlying cell complex topology. The discrete formulation based
on chains and cochains defined by Tonti [19] is the foundation
(9)
of a consistent approach to electromagnetic discrete description,
based on algebraic topological concepts.
where the added term does not change the equation.
Here, a brief introduction is given in order to show that the
Since is taken as constant, using the relation
problem can be reduced to a simple discretization model. Gener-
ically, the space domain D is completely mapped by disjoint ad-
(10) jacent 3-cells that define a main grid .
A dual grid to , is defined as well, using an orthogonal
the following differential equation is obtained: one-to-one relation between edges from and faces from . In
the case of nonorthogonal grids, these could be obtained by the
(11) so-called Voronoi tessellation obtained from a Delaunay main
grid. In the case of orthogonal grids they are defined just as two
The referred gauge is imposed implicitly in (11) but it should cubic meshes where the dual vertices coincide with each main
be added to a boundary condition that satisfies in grid cube center (see Fig. 2). Given an inner orientation in
. However, after applying this gauge the current continuity and an outer orientation in it is possible to define a way to
condition is not forced anymore by (11). This equation should sum p-cells (points, lines, areas or volumes corresponding to
then be added to the system of equations, establishing a relation ). Such a definition is a p-chain. When a global
between potentials: quantity is mapped into a chain, e.g., by integration, a scalar is
obtained that is directly connected to the correspondent p-chain.
(12) This map is called a p-cochain and, intuitively, it is a way to
integrate a field over a certain p-chain.
This is the typical formulation for a constant , The resultant scalar values are attached to specific geometric
described, e.g., in [27]. forms and there defined. Vector fields are replaced by discrete
In order to avoid discretizing the source coil over the grid, forms defined in and in along their facets, edges or nodes.
the reduced magnetic vector potential (RMVP) is used in the As an example, the current density is the resulting scalar from
so-called formulation [8]. The magnetic vector a 2-cochain map defined over a 2-chain (surface) into a scalar
potential is divided into allowing imposing cur- value that corresponds to a mean current density value through
rent sources by defining the imposed potential . This descrip- this surface. The following nomenclature is followed: a quantity
tion also avoids numerical instabilities when conductivity is low defined over faces is written as ; a quantity defined over edges
which generates large differences between and is is written as ; quantities integrated over the cell volume and
1196 IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO. 5, MAY 2010

2-form quantities and 1-form quantities and averaged material

properties. For example

(20)

The integration forces the definition of a metric system in the

specification of the constitutive relations.

B. The Eddy-Current Discrete Problem Formulation

Fig. 2. Geometrical representation of the discretized quantities in use. The eddy-current problem based on potentials could be
derived in the same way as in the differential approach for
constant in all space. Using the Helmholtz decomposition
TABLE I
DISCRETE TOPOLOGICAL OPERATORS discretization

(21)

Equation (9) is written as

(22)

where again . This is possible because both dis-

crete constitutive quantities are defined in the same topological
structure. Since is considered to be homogenous and the new
incorporates , a discrete Laplacian operator can be discretized,
avoiding the stiffness matrix singularity. The Coulomb gauge
cannot be directly written since divergence is not defined for
1-form cochains. The following discrete Laplacian operator over
defined as a mean value at the center of the cell (the dual corner the dual grid is defined:
in the orthogonal case) are defined as . An array of values of
each type is represented in bold. In Table I each electromagnetic
quantity is topologically defined and identified over the grids. (23)
A. The Discrete Operators where and are the discrete Hodge operators defined
The discrete topological operators rot, grad, div are defined for constant conductivity and permeability, correspondingly.
for each cell. Furthermore, they can be defined as different enti- The operator resembles the 7–point finite difference stencil
ties for each grid (e.g., rot and ). Note that the tilde relates an from the discrete vector Laplacian operator. Since the discrete
operator to the dual grid. It is known that due to Coulomb gauge is not feasible in FIT, a similar discrete gauging
the topological relation between the dual grids. The existence of can be defined using to relate 1-forms with the divergence
these operators just needs the definition of an orientation. They definition defined by [25]. A new formulation
result immediately from the notion of p-chain boundary [19] of the eddy–current equations is obtained, similar to the differ-
and correspond to the classic differential operators, in the topo- ential scenario, with the finite difference Laplacian operator:
logical sense. For instance, rot transforms 1-forms in 2-forms
and the div transforms 2-forms in 3-forms. The Maxwell equa- (24)
tions written in terms of the presented discrete tools are known (25)
in the FIT nomenclature as Maxwell grid equations [13]:
where is a scalar value. The current continuity equation closes
(16) the system of equations, establishing a relation between poten-
tials, as in the differential case. Final equations are obtained
(17)
from the application of the relation (corresponding
(18) to (13) and (14)):
(19)
(26)
The constitutive relations need a metric to be defined and so (27)
it materializes the first numerical approximation of the method.
As in [20], the constitutive relations (discrete Hodge operators) The resultant system is nonsingular and the used gauging en-
are defined by diagonal matrices that link quantities defined in sures a robust regularization of the system whichever is the con-
the both grids. The constitutive relations present a ratio between ductive boundary, multiply connected or single connected [25].
BRÁS et al.: FAST FORWARD PROBLEM SOLVER FOR THE RECONSTRUCTION OF BIOLOGICAL MAPS 1197

Note that the use of in the entire space takes in consider-

ation nonexistent nonzero conductivity over the air-space. So,
the choice of value should in one sense be able to minimize
the impact of nonzero current flux through the interface and
in other sense it cannot be too small which could affect drasti-
cally the condition number of the matrix problem. In Fig. 2, the
geometrical definitions of the relevant fields for this problem are
depicted.

C. Subgridding and Stiffness Matrix Assembly Fig. 3. Low to high resolution virtual neighbors. The 1-form to which the vir-
tual neighbor is related is represented at the center of a gray face and the cor-
A performance-oriented octree-based subgridding technique responding virtual neighbor “a” is represented by a gray arrow and its value is
given by the expression represented below each schema.
is applied to improve accuracy and refine the conductivity map
to be analyzed (see [4] for details in the octree). As a compo-
nent of the inverse problem, one can see the subgridding refine-
ment as a way to reduce the number of degrees of freedom of a
problem, an essential issue in this case. Subgridding in harmonic
problems is less demanding than in transient ones since there
is no stability repercussions in the time coordinate. Though,
whichever is the problem type, final accuracy is mainly depen-
dent on how interpolations are done in the interface region.
A subgridding scheme for (26) and (27) discrete equations
needs only to concern with 1-forms and 0-forms in the grid
( and ) and constitutive relations. The presented scheme is
very quick to perform since it uses just the direct neighbors to
interpolate the virtual neighbor information.
In any implementation of the orthogonal subgridding scheme
the fundamental information to be gathered is the neighbor in-
formation, here defined in a set of matrices . It maps the
positioning between p-forms in each direction. Different ma-
trices exist for each direction ( index), for each p-form type (
index) and for each p-form orientation ( index). The directions
could be and . The used p-forms types in
this context are lines (1-form) and points (0-form) and there are
three orientations for the lines. An example of the application
of the Neighbors set of matrices is given: Fig. 4. High to low resolution virtual neighbors. The same logic as in Fig. 3 is
used.
(28)

This equation can be read as follows: is a 4. Note that the represented planes are the dual grid facets
neighbor matrix for 1-forms oriented in that applied to a that are crossed by the 1-forms in analysis and “ ” in the pre-
specific array of this type, , generates , an array of sented weight equations is the approximation of the grayed vir-
values that are the neighbors of the values. tual neighbor:
Each matrix is a matrix where is the number of In the case of high to low interface neighbors the following
cells that define the main grid. It is formulated by identifying schemes and correspondent virtual neighbor weights are
the neighbor cell of the cell , in each line. If the neighbor ex- applied.
ists in the same subgrid the matrix entry is 1. If there is not a The expressions define the nearest point of the plane or line
direct neighbor, as in the case of subgridding interface cells, a defined by the used points to describe the virtual neighbor. For
set of weighted neighbors define the virtual neighbor point such the 0-forms, the same kind of analysis is done in Fig. 5.
that all points have its neighbor relation satisfied, based on the The set of matrices allows building easily all the opera-
metrical position of the nearest neighbors. The interpolation and tors and constitutive relations. For instance, the gradient for the
correspondent coefficients are explicitly represented in each ma- oriented 0-forms can be built using
trix. The used virtual neighbors location and weighting resem-
bles the method proposed in [21], here adapted to this particular (29)
problem with 0-forms and 1-forms and simplified in its approx-
imation order to have a quicker building and a simple stiffness In the definition of the interpolation coefficients was again
matrix. Without loss of generality for the other coordinates, the necessary to use a metrical system, which is the second numer-
1-form neighbors in the interface are presented in Figs. 3 and ical approximation of the method.
1198 IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO. 5, MAY 2010

Moreover, the number of points that define each spire is de-

fined in such a way that between two points there is a distance
of relating it with the space between points.
Another important issue in terms of performance and accu-
racy is to understand the importance of the resolution grid level
that should be adopted over the sensing coils, and hence, the
order of interpolation that a measurement matrix should
have if used. For example, for the test problem defined in
Section IV-C with a typical MIT problem, the reduction in the
number of space points while using a high resolution level over
Fig. 5. Low to high (left) and high to low (right) resolution virtual neighbors. the object and sensors or just over the body is about 70%.
The same logic as in Fig. 3 is used, except that the scalar points to which the
virtual neighbor is related is marked in bold.
F. The Measurement Matrix
The measurement matrix is a linear transformation ap-
D. The Large and Sparse Matrix Linear Problems plied to the magnetic vector potential that generates for each coil
the corresponding electromotive force given by
As stated previously, the forward problem is composed of
a set of eddy-current problems for several angular placements
of the object in respect to the source field. The full forward (30)
problem is given by with and the mea-
surement matrix. Each matrix has the size
and where is the coil number, is the number of points that define
where is the number of angular positions of the object in re- a sensing coil, and is the coordinate interpolated over the
spect to the source field, is the number of cells of the grid coil in the position. For inverse problem methods based in
and is the number of non null conductivity cells. The resul- the reciprocity theorem there is no need to define explicitly this
tant matrix of the eddy-current problems is a large and sparse matrix. However, for other kind of methods that need to know
complex nonsymmetric matrix. To solve each problem an in- explicitly the measured quantities this matrix have to be defined,
complete LU factorization is calculated and used for example in the case of truncated or inexact Newton methods.
as the problem preconditioner. A biconjugate gradient stabi- is stated as
lized (bicgstab) method [22] is then applied to solve the linear
problem. The ILU threshold is an important parameter in terms .. .. ..
. . . (31)
of performance. The smallest the threshold, the better the ap-
proximation of LU matrix to and so the preconditioning.
However, the ILU calculating time can increase substantially as
the threshold goes to zero. This value has then an optimum value where is given by
that could be balanced for the defined number of problems
to be determined in order to increase the overall performance. (32)

E. Numerical Considerations Each is an array with the tangent weight for each co-
ordinate in each point of the sensor coil . is the
Several numerical simplifications are presented that allow normalized weight of each of the neighbors for the interpola-
accelerating the solver without substantially degrading its tion order of each of the points that compose the coil. Note
performance. that the matrix is calculated only once for the same set of
The source field calculation can become a heavy process de- sensing coils, turning its calculation with higher order a rela-
pending on the number of points where it is determined and on tively small task in the overall process.
the number of points that define the source coil. In each cell,
the potential vector should be determined in three distinct posi-
tions corresponding to each facet in each cell ( and ). IV. RESULTS
The process time is reduced to approximately 1/3 if the calculus
A. Typical Experimental Scenario
is approximated by the center of each cell and calculated over
each face using a linear approximation with the help of the men- The geometrical dimensions of the space and the used param-
tioned neighbor matrices. eters are based on a typical experimental setup [5]. The space
On the other hand, windings of the source coil could be de- where the object is normally placed is defined as a cylinder with
termined over the same position if the space between them, 20 cm diameter and an equal height. Source coil radius is of
, is smaller than the minimum cell side where the field 2.5 cm with 10 windings and sensing coils have 1.5 cm radius
is calculated, . Given this, as a consequence of the numer- also with 10 windings. Sensing coils and source coils have a
ical tests, a set of windings separated by are calculated winding spacing of 0.5 mm. The current source amplitude value
by the same centered winding. The same is implemented for the is 1 A with a fixed frequency of 700 kHz. The center of the space
sensing coils. is the zero coordinate point.
BRÁS et al.: FAST FORWARD PROBLEM SOLVER FOR THE RECONSTRUCTION OF BIOLOGICAL MAPS 1199

Fig. 7. [ Abs A( ] ) in the y = 0 plane. The circles represent the simulated

results and the lines correspond to analytical results.

Fig. 6. Profile of the simulated geometrical space and corresponding resolution

levels over the sphere. Used plane for analytical validation.

B. Analytical Validations
The adopted test scenario is based on a harmonic eddy-cur-
rent problem of a conducting sphere model whose analytical
solution is proposed in [23], for the case of a sphere with con-
stant conductivity. This test allows understanding how the ob-
ject meshing due to orthogonal gridding affects the solution and
mimics a typical spherical phantom, used in [10] to test a MIT
scenario.
The simulated space is a cube with 72 cm side, 6 levels of res-
olution, with a maximum resolution cell side of 2.5 mm. Con-
sequently the largest cells have 8 cm side. The studied object
has 5 cm diameter and a real conductivity value of 1 S/m. The
used was able to solve the problem without compromising its
convergence time and avoiding significant changes in the results
due to nonexistent currents through interface. An insightful Fig. 8. Abs A
( ) Relative error in the y = 0 plane.
study about the choice of will be published in an upcoming
work.
The first test compares the analytical solution of the with
the results obtained by the numerical method. A plane segment
with and aligned with the source coil axis is chosen as
the space to be tested without loss of generality for any plane
that includes the z axis since the conductivity space is axisym-
metric. In Fig. 6, the simulated object in the used resolution pro-
file along with the comparison plane is depicted.
In this test scenario, all the numerical simplifications de- Fig. 9. Profiles of simulated geometrical spaces to calculate axisymmetrical
scribed in III-E are applied and just the object is covered with EMF values in sensing spires. In (a) a third resolution level is used over the
the higher resolution subgrid. This scenario is referred from sensing spires and in (b) the first resolution level is used over the sensing spires.
now on as the performance oriented simulation scenario.
Several profiles defined over the depicted plane, are
compared in Fig. 7. plane is 1.5% and the maximum error is of %. The error is
The error in the defined plane is given in the next figure again larger near the z-axis because the field is also near zero which
for the absolute value: enlarges the relative error.
Defining the mean relative error as The same geometrical setup was used to compare the elec-
tromotive force in the sensing coils, placed coaxially with the
source coil, in two distinct scenarios. The first one is the per-
formance oriented simulation scenario defined in the previous
test. The sensing coils are placed over the third resolution level.
(33) In the second scenario, called standard simulation, the solution
that relates quantities in the numeric model solution is found without simplifications and the higher mesh resolution
with the same quantities in the analytical so- also covers the sensing coils. Both geometries are represented
lution , the mean relative error of the analyzed in Fig. 9.
1200 IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO. 5, MAY 2010

Fig. 10. EMF absolute values in sensing coils placed in a vertical axis over the
sphere, for the performance oriented simulation.
Fig. 12. Performance comparison of a unique eddy-current forward problem
for three scenarios: a performance scenario with performance improvements
and with the highest resolution just over the object; a standard scenario with the
highest resolution just over the object, and a standard scenario with the highest
resolution over the object and sensors.

Fig. 11. Relative error of the EMF values in sensing coils placed in a vertical
axis over the sphere.

Fig. 13. Typical MIT scenario with two positions of the conducting sphere: left
side: centered; right side: 1 cm deviated to the right. Gray coils are source coils
The resulting electromotive force (EMF) for a single spire and black coils are sensing coils. Eight coils of each type are represented.
sensing coil placed in 50 different positions along a vertical axis
is presented in Fig. 10 for the performance oriented simulation.
Corresponding relative errors are presented in the next figure setup, with 32 sensing coils and 16 source coils. As in the pre-
for both simulation profiles: vious section, the performance oriented simulation and the stan-
All results have a small offset error that could be explained by dard simulation are tested. The performance oriented simulation
the approximation of the sphere. Error oscillations are created is constituted by 25 k cells and the standard by 127.5 k cells. The
by the linear approximation of the method to the real values, left schema in Fig. 13 illustrate the geometric setup.
which generates oscillations with (approximately) the grid size The simulations run in a dual core with 2.0 GHz PC and the
period. The EMF from spires placed over the higher resolution necessary RAM memory was of 1.5 GB.
zone in the performance mesh, approximately between 3 cm and A performance plot is presented in Fig. 12 for a unique eddy-
4 cm, are good approximations of the same values for the stan- current simulation (one source coil only). It allows comparing
dard mesh. Over the second resolution zone for the simple first separately each task that composes an eddy-current simulation.
order approximation the error has spikes over the interfaces and Nine tasks are reported as significant during the process. One of
the error level is around 8%. When a second order interpolation the most relevant in terms of time consuming is the analytical
is used, the error on the interfaces does not have spikes. More- calculation step which is, in the case of the performance ori-
over, the error over these zones is reduced to the same order of ented simulation, reduced from 195 to 23 seconds. Also, ILU
magnitude of the higher mesh (compare gray line with black preconditioning, and the linear system solving are heavy pro-
line) and inferior to 5%. The interpolation order that reduces cesses. The last one is clearly the most demanding component,
more effectively the relative error was of second order in each taking approximately 55% of the processing time. The perfor-
direction, that is, a set of 27 points is used for each interpolation. mance oriented simulation was able to reduce this time up to
The mean relative error is approximately 1.5%. 18% of the standard simulation.
A full problem is now considered, with 16 source coils. Since
C. Performance Results of the Forward Problem Solver some of the time is spent in tasks that are common for all source
In order to test the performance of the method, a full forward coil problems the improvement is even better, as can be seen in
problem is simulated. A sphere is placed inside a classic MIT Table II.
BRÁS et al.: FAST FORWARD PROBLEM SOLVER FOR THE RECONSTRUCTION OF BIOLOGICAL MAPS 1201

TABLE II position of the active source coil. For the deviated sphere setup,
COMPARISON BETWEEN THE STANDARD AND THE PERFORMANCE ORIENTED stronger values are found over the nearest sensing coil to the
SIMULATIONS IN TERMS OF ERROR AND PERFORMANCE
body (24th coil) when the source soil is at the same position
(12th coil), as expected from the system sensitivity [24].

V. CONCLUSION
A new approach to the discretization of the MIT forward
problem has been proposed with focus on the performance in
order to solve the MIT inverse problem.
The formulation avoid simulating the coil
current source and also avoid numerical instabilities typically
present in the formulation in the case of low conduc-
tivity eddy-current problems due to large differences between
and . A simplified formulation of the eddy-current
problem using a hybrid formulation using the finite integration
technique and a discrete finite difference Laplacian stencil was
implemented that allowed to formulate a simple nonsingular
matrix problem equivalent to the differential continuous case.
Although the weak gauging property of the iterative Krylov
subspace methods could be used to solve the ungauged singular
problem [25], a gauged version was chosen since the nonsin-
gularity of the stiffness matrix can influence the numerical
robustness of the inversion strategy.
A simple yet efficient discretization of the problem with a fast
subgridding method allowed solving 16 eddy-currents problems
(each for each source coil) in approximately 11% of the time
due to a reduction in the number of points and other described
numerical improvements.
Some of the calculation steps during a full forward problem
are done only once for all the inversion process, namely the
source field and the measurement matrix calculation. So, al-
though its weight in the 3-D forward problem is considerable,
in the overall inverse problem it only affects performance once.
In terms of accuracy, the method has an error over a bounded
Fig. 14. Complete set of absolute values of EMF measurements of a forward
simulation with 32 sensing coils and 16 source coils. plane that crosses the sensing coils positions always inferior to
5% and with a total relative error of 1.5%. For the EMF calcu-
lation the same errors were found for the maximum value and
The overall time is reduced to 11% approximately, which the mean relative error.
gives a mean resolution time for each eddy-current simulation The method does not easily mimic conductive boundaries
of 19 s. since it is limited to orthogonal cells. This constitutes the main
drawback of the method, as it is implemented.
D. Full Forward Problem Solution New approaches for solving large measurement based in-
verse problems will now be considered using this forward
Two distinct positions of the sphere are used to show a final solver and the specific use of the measure matrix in the problem
absolute values for the EMF signals for the described setup. formulation.
Two situations, with a sphere centered in the space in analysis
or slightly deviated (1 cm) to the right, are tested as depicted in ACKNOWLEDGMENT
Fig. 13. The “o” symbolizes the space center.
The authors would like to thank the contribution of Prof. M.
Using both geometries a complete set of absolute values of
Clemens for the helpful discussion on the FIT subject and valu-
EMF measurements of a forward simulation, again with 32
able suggestions on the topic. The work of N. Brás was sup-
sensing coils and 16 source coils, was calculated. In Fig. 14 the
ported by the Portuguese Ph.D. 309 scholarship program from
results are presented in a form of a 2-D image where the rows
FCT (SFRH/BD/17405/2004). This work was sponsored by the
are the sensing coils and columns are the source coils. The
FCT research project PTDC/EEA-ELC/105333/2008.
order by which sensors and source coils are used is represented
in Fig. 13 by its numbering, although just eight coils of each REFERENCES
type were depicted to simplify the representation.
[1] H. Griffiths, W. R. Stewart, and W. Gough, “Magnetic induction to-
In the case where the sphere is absolutely centered (left) it mography, a measuring system for biological tissues,” Ann. NY Acad.
is possible to see a pattern which changes accordingly to the Sci., no. 873, pp. 335–345, 1999.
1202 IEEE TRANSACTIONS ON MAGNETICS, VOL. 46, NO. 5, MAY 2010

[2] A. Peyton et al., “Development of electromagnetic tomography (EMT) [16] R. Casanas, H. Scharfetter, A. Altes, A. Remancha, P. Sarda, J. Sierra,
for industrial applications: Sensor design and instrumentation,” in Proc. R. Merwa, K. Hollaus, and J. Rosell, “Measurement of liver iron over-
1st World Congr. Industrial Process Tomography, 1999, pp. 306–312. load by magnetic induction using a planar gradiometer: Preliminary
[3] M. Hamsch, C. H. Igney, and M. Vauhkonen, “A magnetic induction human results,” Phys. Meas., vol. 25, pp. 315–323, 2004.
tomography system for stroke classification and diagnosis,” in 52nd Int. [17] D. Simunic, “Calculation of energy absorption in a human body model
Wissenschaftliches Kolloquium Technische Universität Ilmenau, 2007. in a homogeneous pulsed high-frequency field,” Bioelectrochem.
[4] E. Haber and S. Heldmann, “An octree multigrid method for quasi- Bioenerg., vol. 47, no. 2, pp. 217–230, 1998.
static Maxwell’s equations with highly discontinuous coefficients,” J. [18] M. Plonus, Applied Electromagnetics, 2nd ed. New York: McGraw-
Comput. Phys., vol. 223, no. 2, pp. 783–796, 2007. Hill, 1978, ch. 8.
[5] N. B. Brás, A. Pestana, R. C. Martins, and A. C. Serra, “Dimensioning [19] E. Tonti, , G. Ferrarese, Ed., “On the Geometrical Structure of Electro-
and characterization of a new magnetic induction tomography experi- magnetism,” in Gravitation, Electromagnetism and Geometrical Struc-
mental setup,” I2MTC Proc., pp. 2131–2136, 2008. ture, for the 80th Birthday of A. Lichnerowicz. Italy: Pitagora Editrice
[6] R. Merwa, K. Hollaus, P. Brunner, and H. Scharfetter, “Solution of Bologna, 1995.
the inverse problem of magnetic induction tomography (MIT),” Phys. [20] M. Clemens and T. Weiland, “Magnetic field simulation using con-
Meas., no. 26, pp. 241–250, 2005. formal FIT formulations,” IEEE Trans. Magn., vol. 38, no. 2, pp.
[7] M. Soleimani and W. R. B. Lionheart, “Absolute conductivity recon- 389–392, Mar. 2002.
struction in magnetic induction tomography using a nonlinear method,” [21] D. Brown and L. Lowe, “Multigrid elliptic equation solver with adap-
IEEE Trans. Med. Imag., vol. 25, no. 12, pp. 1521–1530, Dec. 2006. tive mesh refinement,” J. Comput. Phys., vol. 209, no. 2, pp. 582–598.
[8] O. Bíró, “Edge element formulations of eddy current problems,” Comp. [22] R. Barrett et al., Templates for the Solution of Linear Systems: Building
Meth. Appl. Mech. Eng., vol. 169, pp. 391–405, 1999. Blocks for Iterative Methods. Philadelphia: SIAM, 1994.
[9] A. Bossavit, ““Stiff” problems in eddy-current theory and the regular- [23] T. P. Theodoulidis and E. Kriezis, “Coil impedance due to a sphere
ization of Maxwell’s equations,” IEEE Trans. Magn., vol. 37, no. 5, pp. of arbitrary radial conductivity and permeability profiles,” IEEE Trans.
3542–3525, Sep. 2001. Magn., vol. 38, no. 3, pp. 1452–1460, May 2002.
[10] M. H. Pham and A. J. Peyton, “A model for the forward problem in [24] J. Rosell, R. Casanas, and H. Scharfetter, “Sensitivity maps and system
magnetic induction tomography using boundary integral equations,” requirements for magnetic induction tomography using a planar gra-
IEEE Trans. Magn., vol. 44, no. 10, pp. 2262–2267, Oct. 2008. diometer,” Physiol. Meas., vol. 22, pp. 121–130, 2001.
[11] A. Bossavit and L. Kettunen, “Yee-like schemes on staggered cellular [25] M. Clemens and T. Weiland, “Numerical algorithms for the FDiTD
grids: A synthesis between FIT and FEM approaches,” IEEE Trans. and FDFD simulation of slowly varying electromagnetic fields,” Int. J.
Magn., vol. 36, no. 4, pp. 861–867, Jul. 2000. Numer. Model., vol. 12, pp. 3–22, 1999.
[12] M. Hyman and J. M. Shashkov, “Mimetic discretizations for Maxwell’s [26] M. Clemens and T. Weiland, “Regularization of eddy-current formu-
equations,” J. Comput. Phys., vol. 151, no. 2, pp. 881–909, 1999. lations using discrete grad-div operators,” IEEE Trans. Magn., vol. 38,
[13] M. Clemens and T. Weiland, “Discrete electromagnetism with the finite no. 2, pp. 569–572, Mar. 2002.
integration technique,” Progr. Electromagn. Res., vol. 32, pp. 65–87, [27] O. Biró and K. Preis, “On the use of the magnetic vector potential in
2001. the finite element analysis of three-dimensional eddy currents,” IEEE
[14] A. Bossavit, “Generalized finite differences in computational electro- Trans. Magn., vol. MAG-25, no. 4, pp. 3145–3159, Jul. 1989.
magnetics,” Progr. Electromagn. Res., vol. PIER 32, pp. 45–64, 2001.
[15] O. Grimmes, Bioimpedance and Bioelectricity Basics. New York:
Academic, 2000.