1. Introduction
Electrical impedance tomography (EIT) is a non-invasive method to deliver cross-sectional images. The non-invasive character, available cost, portability, and safe use have allowed the technique to find wide application in diverse branches and disciplines of industry and science, such as biomedicine [
1,
2,
3], geophysics [
4,
5,
6,
7], industrial chemistry [
8,
9,
10,
11], and material engineering [
12,
13,
14,
15].
Principally, the approach relies on exciting the examined object by a harmonic current that passes through electrodes on its perimeter and, simultaneously, measuring the electric potential on the remaining electrodes according to the selected current pattern. The obtained values of the voltage or the phase difference between the current and the voltage (allowing admittivity reconstruction) enable us to reconstruct the specific conductivity distribution. In general terms, the presented reconstruction process can be described as a nonlinear and ill-posed inverse task. Considering the sensitivity of the process to changes in the individual parameters (the geometrical accuracy of the domain and positioning of the electrodes), the numerical finite element method (FEM)-based model of the domain must be set up in a precise manner. The various inaccuracies are advantageously reduced via optimizing the parameters of the forward solution model; such a step is beneficial because the inverse task, due to its character, does not enable us to estimate the parameters accurately. The solution is exactly defined by a physical and a numerical model, and its results should correlate with those of the real measurement. Precise domain modeling was previously proven to be a necessary precondition for high-quality image reconstruction. References [
16,
17], for example, showed that an inaccurate shape of the domain markedly affects the measured voltages, the impact being even greater than that exerted by the actual inhomogeneities. The deformed shape results in the blurring and incorrect localization or inferior recognizability of an inhomogeneity [
18]. Regarding the current patterns, the problem occurs more frequently in single-source than in trigonometric driving [
17]. The domain parameters were investigated in detail by, for instance, Hemant Jain et al. [
19], who examined the impact of non-circular boundaries on the reconstruction of comprehensive specific conductivity via the NOSER algorithm. During their experiments, the elliptically deformed circular domain caused the reconstructed image to exhibit substantial distortion, meaning that precise domain modeling has a fundamental influence on the reconstruction accuracy. Furthermore, the conductivity value error proved to be significantly lower in the axis than on the boundary of the investigated object.
Imprecise modeling and the distribution of the electrodes were discussed in reference [
17], with the outcome being that electrode misplacement may introduce in the reconstruction an error larger than the uncertainty of the entire measuring system. This type of inaccuracy results in blurring, artifacts, and incorrect localization; such spurious effects are more frequently encountered in single-source driving than in the trigonometric pattern and can be suppressed through accentuating regularization at the expense of losing the absolute value of conductivity. The same or associated problems, i.e., modeling inaccuracies, the positioning of the electrodes, and the measured voltage and current errors, were analyzed by Murphy et al. in [
20]. The authors examined the impact of the individual parameters on EIT reconstruction via the D-Bar approach, which allows one to solve the inverse task without iterative computation. The evaluation was carried out through simulation, exclusive of measurement, and concluded that incorrect modeling eventually generates artifacts.
The influences exerted by inappropriate electrode surface and contact impedance on the reconstruction process were focused on in references [
17,
18]. In addition to showing that electrode inaccuracies affect the reconstructed image to a markedly greater extent in multi-source than in single-source driving, the research also confirmed that the electrode size modeling error is indirectly proportional to the contact impedance deviation; thus, overestimating the electrode surfaces brings an effect similar to that of underestimating the contact impedance. From the perspective of measurement, the inaccuracy of the reconstructed image, given by the contact impedance, resembles in character an incorrectly set surface of the electrodes: while the issue will not manifest itself in neighboring patterns, in trigonometric driving, which comprises current-carrying electrodes, it substantially contributes to the formation of ring artifacts to impair the accuracy of the image.
In view of the higher signal-to-noise ratio (SNR) in the trigonometric pattern, a dedicated article presented the possibilities of compensating for contact impedance via triple measurement or a suitable structure with compound electrodes [
21]. By extension, other papers [
22,
23] centered on the effect of inaccurately known contact impedances, addressing the issue of contact impedance estimation based on the complete electrode model (CEM). For reliability purposes, the estimation was compared with measurement performed by means of an Oxford Brookes and a KIT4 tomograph; the distribution of the conductivity and contact impedance was reconstructed with real data. Finally, the results of the experiments showed that the contact impedance cannot be separated from the internal impedivity and estimated without measurement utilizing a uniform conducting medium.
In relation to clinical use, the problem of an unknown domain border and contact impedance was investigated in [
18]. The proposed novel approach challenged the inaccuracies numerically to successfully reduce the systematic errors in the reconstruction, with the original reconstructing procedure based on measurement; the deformation case, however, was verified solely via simulation.
The compensation of variable electrode contact in EIT was introduced in [
24], where the authors employed the CEM to demonstrate a hybrid nonlinear reconstruction algorithm that allows the substantial reduction of artifacts generated through poor contact. The compensation method was tested on a set of clinical data. Another compensation technique for modeling the unknown domain boundary error was presented in [
25]. This instrument exploits the approximation approach, and its efficiency was validated by experimental measurement on an EIT phantom. In [
26], fast and simultaneous statistical estimation of conductivity and electrode contact impedances for a 2D disc was outlined; the method fundamentally utilized the Toeplitz matrices to identify wrong contacts. The described approach proved operable and was verified via measurement.
The optimization of the electrode position in irregularly shaped domains was examined in [
27,
28], relying on an optimizing algorithm of the steepest descent type. Exploiting Fréchet derivatives of the CEM, the authors computed the electrode locations through simulation. In [
29], a procedure to compensate for imprecise electrode modeling and movement artifacts was described in detail, together with surface movement reconstruction options. The compensation was tested with the EIDORS tool, and the reconstruction exploited a real or a reproduced data set. The research characterized in [
30] targeted simultaneous reconstruction of time-varying images and contact impedance, yielding a significantly reduced electrode drift and motion artifacts. The method was evaluated on reconstructed clinical data, measured by using a GE GENESIS prototype.
A 2D-based procedure for optimizing the electrode position by means of deep learning was outlined in paper [
31]. The experiment involved simulating voltage data with added Gaussian noise upon 1%, 5%, and 10% standard deviation, and performing image reconstructions on non-circular domain shapes. The researchers concluded that optimized electrode positions can reduce the error in EIT reconstructions; as regards the experimental input, however, they employed only synthetic data, meaning that factors accompanying real measurement, such as interference, uncertainties of the instruments, and limited resolution, were not assumed. Further, the size and polynomial degree of the mesh elements gained attention in report [
32], which examined the relationship between the mesh density and the resulting error. The authors proposed that a higher density does not necessarily yield a lower reconstruction error, and they also specified the key variables causing unexpected behavior in the inverse task.
As pointed out above, our project comprised not only numerical modeling and simulation but also laboratory measurements. Utilizing the obtained data and knowledge, we set up an optimization procedure based on the Nelder–Mead algorithm; the input is the vector of measured voltages in a real-world tomograph, and the output consists of the approximate value of one or more required parameters (shape deformation of the domain, electrode placement, and initial conductivity). Principally, the actual procedure involves verifying the physical model of the tomograph via computing an FEM-based numerical model, in such a manner that we obtain the best possible match between the resulting vectors of simulated and measured voltages. The optimization was implemented with the EIDORS library and the relevant Matlab toolbox, the aim being to correlate the numerical model and the real-world measurements. The entire set of tasks has been designed to improve the image reconstruction accuracy and to reduce the artifacts. In more general terms, this article contributes to the state of the art by introducing a novel approach to optimize the parameters of the mathematical model, such as the initial conductivity, electrode positions, and domain shape deformations; importantly, the entire procedure is supported by the measurement of the physical model. The research comprises, above all, laboratory experiments and their analysis to improve the applicability of the tomography in the real-world monitoring of large-sized objects. We assume broader and deeper cooperation with the Department of Water Structures, Faculty of Civil Engineering, Brno University of Technology; in this context, the optimized mathematical model will be employed in monitoring the functionality of dams and predicting their material composition.
2. Materials and Methods
This section outlines the measurement patterns and the Precise, Low Impedance (PLI) EIT prototype device for data acquisition. In this context, the forward and inverse tasks are derived and described; we also show the Nelder–Mead algorithm with an implementation in computable unknown EIT parameters and demonstrate image reconstruction to develop relevant comprehensive discussion within the last chapter.
2.1. Measurement and Device
The principle of acquiring EIT image reconstruction data exploits voltage measurement and signal phase shift with respect to the input current on the electrodes of the tomograph. Thus, multiple stimulation patterns have been created over the decades, allowing the application of alternating current to predefined electrodes of the system; simultaneously, suitable combinations of electrodes to facilitate the measurement are specified. The most popular single source stimulation patterns include Adjacent, Opposite, and Skip-X driving, all of which are shown in
Figure 1 [
33,
34,
35].
However, a multi-source pattern category is also available, with the best known element being the trigonometric current pattern. Each driving option exhibits certain advantages and drawbacks. The benefits of adjacent driving consist of good edge sensitivity and contact impedance insensitivity; conversely, the disadvantages include poor sensitivity at the center; sensitivity to the boundary shape and electrode position; high measurement error; and electrode noise interference. Similarly, the opposite pattern is characterized by good overall performance and uniform current distribution, but also poor edge sensitivity and low current injection. Compared to single-source driving, the trigonometric pattern has strengths such as sufficient center sensitivity, high current injection, and improved SNR; the weaknesses are then embodied in the more independent current drivers and considerable susceptibility to unknown contact impedance [
17,
36].
Utilizing our knowledge of driving patterns, we designed a single-source precise, low impedance EIT system for tomographic data acquisition; the entire setup is presented in greater detail within [
37]. The images below display a tomographic data processing diagram and a version 1 EIT switching card, whose components are represented in the drawing (
Figure 2).
The designed diagram consists of a function generator and a voltage-controlled current source (VCCS) to deliver a constant AC current amplitude for the stimulation. The feeding part is connected to a shunt resistance Z
B, which operates as an AC current detector. The shunt resistance is joined to the electrode system of the tomograph through a PLI EIT switch (
Figure 2b). The electrode system and shunt resistance are connected with differential amplifiers and digital multimeters to perform the voltage measurement. The amplified voltage drops in the current sensor, and on the tomograph electrodes are evaluated with an oscilloscope to determine the phase shifts. The oscilloscope, multimeters, and function generator are driven by a computer in LabView. The measurement is fully automatic, and the combination of the electrodes is programmable as desired [
37].
Figure 2b displays a prototype of the switch, independently extendable and designed with respect to a low on-state impedance (max 0.7 Ω in the range from 10 Hz to 100 kHz). The PLI EIT system also contains the VCCS with a range of ±15 V; two differential amplifiers (AD 620) with a shunt resistance for current sensing; and a battery power supply for the amplifiers. The innovated variant of the PLI EIT system will contain an analog-to-digital converter (ADC) to replace the digital multimeters, and also hysteresis comparators with a counter to substitute for the oscilloscope; currently, this version is being developed [
37].
In the actual optimization, we employed only the module of voltages, expecting the voltage values to be equal to the impedivity module. The impact of the signal phase shift will be explored within the future research.
2.2. Forward Task
The physical model of the forward task is based on a description of the domain Ω in an R
N, N = 1, 2, 3 sized space. Let us assume that the domain has smooth, continuous boundaries, to which are connected
L electrodes equidistantly distributed on the domain’s surface. An alternating current
Il is supplied via the electrodes to the actual domain; when passing through, the current then induces corresponding voltage drops, and these are detected by the electrodes. The forward task, current supply, and measurement of the voltage drops can all be most accurately defined by the complete electrode model (CEM), describable as follows:
where
x is a coordinate belonging to the domain Ω,
σ(
x) denotes the specific conductivity of the investigated medium,
u(
x) represents the electric potential inside the domain Ω,
Ul and
Il stand for the voltage and current passing through an electrode
l,
zl is the contact impedance between an electrode and unknown conductivity, and
n characterizes the vector that is normal to the surface of the domain ∂Ω. The last two of the equations describe Kirchhoff’s laws, or, by extension, the law of the conservation of energy, in a physical expression of the forward task [
38,
39].
The numerical solution of the above-defined physical model exploits discretization. The partial differential equations (PDEs) are then approximated via the finite element method (FEM). Using the FEM to discretize the domain is regulated by the following formula:
where
u(
x) denotes the voltage at any point in the system,
ui,j are the voltages at the nodal points of the mesh, and
Ni,j represents the linear basis approximation function relating to the created mesh.
With the finite element method, the equation system of the physical model (CEM) (1)–(4) can be assembled into the form:
where the individual elements of the matrix are expressed by the following relationships:
where
N denotes the linear approximation function,
σ represents the conductivity,
zl is the contact impedance, and
El defines the electrode surface area. The derivation of the system of equations is described in detail within [
29,
39].
2.3. Inverse Task
In general terms, the inverse task is characterized as a difficult-to-solve mathematical function. Within EIT, the operation consists of estimating computationally the distribution of specific conductivity in a domain Ω; the actual computation is nonlinear, ill-posed, and problematic to execute. There are two essential approaches to the EIT inverse solution: difference and absolute imaging. The former procedure reconstructs two diverse measurements, either in time (time-difference) or depending on the frequency (frequency-difference); these measurements are then utilized for computing the relative change of the distribution of conductivity inside the domain. By comparison, the latter option, or absolute imaging, relies on reconstructing the conductivity via a single measurement [
3].
Principally, the inverse task rests in searching for such a conductivity matrix that will satisfy the conditions of Equation (1) while preserving the values of the voltage and current within the domain Ω. In EIT, we have the equation:
where Ψ(σ) is the conductivity change vector,
A denotes the Jacobian, and
b represents the voltage error [
39].
By applying the Moore–Penrose inverse, we yield the equation:
which resolves with the least squares method formula:
In the ill-posed matrix
A, the least squares procedure may fail to perform properly while the inverse task is being solved, rendering the problem unresolvable. Thus, the method was complemented with regularization, exploiting options such as the Tikhonov approach, where the L2 term is applied to penalize (suppress) sharp changes to a greater extent than possible with the L1 term. The least squares method including the Tikhonov regularization is characterized as:
where α denotes the regularization parameter [
29].
By applying the above functional directly to the inverse task, we yield the objective function formula:
where
UM is the vector of the voltages measured at the border of the domain Ω;
UFEM(σ) represents the vector of the voltages on the electrodes, obtained via the forward solution;
α stands for the regularization parameter; and
R denotes the regularization matrix.
The algorithm to iteratively compute the conductivity by utilizing the Newton–Raphson method is obtained through the original estimation, which follows from the Taylor development, assuming that the conditions of the Euler–Lagrange equation have been satisfied to allow Equation (18) to be zero. The iterative conductivity calculation can read:
where
σi+1 is the new approximation of the desired conductivity,
σi represents the approximation in the previous iteration, and
ε can be written as:
where
J is the Jacobian, expressing the sensitivity of the discretized electrical potential over the FEM domain model [
39,
40,
41,
42,
43].
2.4. Optimization
The optimization of the parameters (namely, the geometry of the model; regularity, and electrode placement) of the domain fundamentally influences the image reconstruction, as already proposed within the introductory section of the paper. The problem remains very topical because specifying and determining the parameters are critical steps in the effort to achieve more accurate results, reduce the image artifacts, and limit the computational intensity; all of the goals are reached by decreasing the number of the degrees of freedom to be determined through the inverse task. In this context, it is then more advantageous to define as many parameters as possible via pre-calculation before initiating the inverse task.
The actual procedure involves verifying the physical model of the tomograph via computing an FEM-based numerical model to calculate the best possible match between the vectors obtained from forward task-simulated and measured voltages. To finally and finely optimize the above-listed parameters, exploiting the measured values, we employed the Nelder–Mead simplex algorithm (
Figure 3). Included in the Matlab optimization toolbox, this heuristically based method finds frequent use in solving nonlinear optimization problems where the function derivative can be unknown; in our case, the focus is on the relationship between the measured quantities and the properties of the domain.
When operating, the algorithm utilizes the geometrical structure referred to as the simplex. The structure is a geometrical shape in an R
n space with n + 1 points; of these, one point defines the origin of the simplex, while the others determine the direction of the R
n space vector. By setting the point of origin, we can generate
n further points through geometrical transformations (reflection, expansion, and outside/inside contraction). Depending on the transformation applied, the simplex moves, expands, or contracts; moreover, after each transformation, the worst vertex is replaced with a better one. The main advantage of the algorithm over alternative optimization methods rests in its easy implementability, ensuring a highly effective, rapid search for the local optimum [
44,
45].
The algorithm is initiated by selecting the first point, which then enables the formation of an n‑dimensional topology. Subsequently, the individual vertices, x
1 to x
n + 1, are arranged according to the optimization function gradient, f(x
1) to f(x
n + 1), with f(x
1) and f(x
n + 1) being the best and worst points of the function, respectively. The result of an iteration consists of a new point to substitute for the worst point of the function in the following iteration. If shrinking occurs, the substitution proceeds with
n new points, which, together with x
1, are the input simplex points of the following iteration [
44,
45,
46].
To optimize by using the Nelder–Mead algorithm, we defined the minimization function as:
where
f(
p) is the minimization term of the least squares method,
UM represents the vector of the voltages measured on the laboratory tomograph, and
UFEM(
p) denotes the voltage on the electrodes evaluated via the forward solution, which is parameterized by the variable
p.
The implementation stages of the algorithm are set out in the diagram below (
Figure 4).
The optimization input data consist of one or more selected parameters, including:
Parametric deformation of the domain (in our case, circular/elliptical deformation);
Calculation of the initial conductivity;
Misplacement of the electrodes on the border of the domain.
In addition to the selected optimization parameter, the algorithm also comprises a vector of measured voltages, this being a measured sequence for a particular combination of non-excited electrodes. Based on the preset parameters and the values obtained from the tomograph, the optimization facilitates the generation of a parametric FEM model via Netgen; the model is then solved via the forward task by utilizing EIDORS library [
47]. The solution procedure yields a vector of the voltages that were obtained through the simulation, and this vector is evaluated by means of the optimization. If the sum of squares of the differences is greater than the convergence criterion, the algorithm will begin generating a new FEM model and solving the forward task. If the convergence criterion has been satisfied, the optimization will stop, and the return value of the function will contain the nearest possible value of the parameter preset by the user within the selected tolerance.
Considering the large number of parameters, gradual optimization appears to be convenient. Generally, it holds true that the more unknowns at the input of the optimization, the longer the computational time; by extension, we can also point out that if the optimization parameters are set simultaneously, the algorithm may not find a correct solution. To verify the functionality of the algorithm, we used a laboratory model containing a homogeneous medium; this model allowed us to confirm the results physically. This approach then facilitates the determination of factors such as shape deformation, area size, and irregularity in the placement of the electrodes, thus simplifying the subsequent solution of the remaining parameters (initial conductivity and, possibly, contact impedance). The actual impact of the parameters is demonstrated on the results of the reconstruction involving inhomogeneities.
6. Conclusions
This article examined and discussed several factors that affect conductivity imaging, including but not limited to initial conductivity, shape domain deformation, and electrode misplacement. In the introductory chapter, a brief overview of the prior research is outlined. Based on the previous work and the problems we encountered when correlating the laboratory tomograph with the numerical model, we prepared optimization via the Nelder–Mead algorithm. The real dataset was created by means of measurements exploiting single-source driving, or, more concretely, the adjacent and the opposite patterns on the 8-electrode tomograph. For this reason, we employed a precise, low impedance EIT system. The optimization algorithm processed the forward and the inverse tasks by utilizing the EIDORS library. The three-dimensional parametric FEM models with approximately 15,000 elements were generated via Netgen (
Figure 5). The results of the optimization were evaluated by the conductivity error rate (Equation (22)) and the inhomogeneity area ratio (Equation (23)) in
Table 2 and
Table 3.
As regards the actual impact on the consumers, optimizing the properties of the mathematical model significantly improves the image reconstruction accuracy. In this context, we propose a novel approach to integrating the properties of the mathematical model with those of the physical one. This paper also outlines a viable procedure for optimizing a model of an experimental impedance tomograph by utilizing a simple measurement based on Matlab Optimization Toolbox and the EIDORS platform. By extension,
Figure 10 shows that opposite feeding and sensing render the conductivity distribution more sensitive to undesired variations of the domain shape, and it also demonstrates that the method is not usable in the optimization of electrode positions.
At present, the optimizing process solves only one parameter during a run. Another limitation rests in an increasing number of degrees of freedom (by including more parameters to optimize simultaneously), an effect that markedly affects the computational intensity. Although the paper does not discuss experiments involving the simultaneous optimization of two or more parameters of the mathematical model, currently available results indicate that such an option features insufficient stability and may not yield the desired outcome. Yet another problem is inadequate measurement accuracy, where the optimization error can be caused by not only the voltage and phase shift uncertainties but also high noise levels. The future efforts will be directed towards the following aspects or procedures: improving the accuracy of the results achievable with the tomograph, using better parameterizable models and a different mesh generator; reducing the computational intensity and cost; and including the contact impedance, which is important in the measurement of current-carrying electrodes. We also intend to focus on producing a novel and more precise version of the low impedance EIT data acquisition unit, designed to feature a programmable and easily extendable switch with low on-state impedance.