1 Introduction
Electrical Impedance Tomography (EIT) is an emerging technology for obtaining the internal conductivity of a physical body from boundary measurements of currents or voltages on the surface of the body Holder2005 ; Yorkey1990 ; Calderon1980 . It is an illposed problem due to the fact that boundary measurements show little sensitive to (even large) changes of interior conductivity distribution Ammari2008 . Intensive research exists on this topic Uhlmann2009 ; Brown2009 ; many regularization methods have been proposed to overcome the illposedness and to improve the imaging quality Han1999 ; Hsiao2001 ; Chung2005 ; Knudsen2009 .
More recently it has been suggested to augment the measurement setup in EIT with an ultrasonic device thus yielding the hybrid imaging method known as AcoustoElectric tomography (AET) Ammaribook2008 . The resulting modality has been investigated theoretically and numerically, and AET seems to have the potential to dramatically increase the contrast, resolution, and stability of the conductivity reconstruction Bal2012 ; Kuchment2012 .
The idea of AET it is to conduct a usual EIT experiment while a known focused ultrasonic wave travels through the object. The high intensity of the acoustic pressure will create a small local deformation in the physical body and thus of the electrical conductivity due to the acoustoelectric effect Geselovitz1971 .
A physical body imaged by EIT is modeled as a bounded Lipschitz domain for . The changes caused by the acoustic wave can be recorded by EIT measurements on the boundary Jossinet1999 ; Jossinet2005 . The power density in is then
(1) 
Here, is the conductivity and is the electrical potential inside produced by applying an electric field on the boundary . We call the power density operator. With the assumption that measurements are carried out in low temporal frequency and that contains no interior sources or sinks of charges the governing equation is the generalized Laplace equation
(2) 
subject to suitable boundary conditions. Given noisy measurements of the true power densities the problem is to reconstruct We approach the problem by optimization
(3) 
In AET the power density can be computed from the boundary measurements Ammari2008 ; Kuchment2012 ; Jensen2019 . The nonlinear relationship between and renders the problem nonlinear Song2017 . Some methods have been developed in the literature for reconstructing from Ammari2008 ; Kuchment2012 ; Song2016 ; Song2017 ; Ammaribook2008 ; Bal2012 ; Bal2013 , but all these methods are based on the continuum model with Dirichlet or Neumann boundary conditions. See also Adesokan2018 ; Jensen2018 ; roy2018a ; gupta2019a for optimization approaches to the problem different from (3). The limited data problem was considered in Hubmer2018 .
The Complete Electrode Model (CEM) is a practical model for EIT, and it can simulate the electrical potential with a very good accuracy Somersalo1992 . In the model, electrodes are attached on boundary . A known total current injected through the th electrode is given as
(4) 
Here,
is the outward unit normal vector to the boundary
. is the th electrode, and indicates the derivative of in the direction of the outward unit normal vector . Since there is no current flowing out through boundary regions without electrodes, one has(5) 
On the electrode the electric potential is assumed to be constant. This boundary potential consist of a part due to the interior potential and a part due to the electrode contact, and it is comprised in the model
(6) 
where denotes the socalled contact impedance assumed to be constant on the th electrode. The partial differential equation (2) with boundary conditions (4)(6) gives the CEM. To ensure existence and uniqueness of the solution, this model also needs to include the law of charge conservation
(7) 
and to determine the potential’s grounding by
(8) 
The CEM problem (2)(6) has a unique solution for any Hyvonen2017 , however the mixed boundary conditions (5)(6) allow singularities near the edges of the electrodes. In order to increase the regularity of the electrode conductance is introduced as a function
(9a)  
(9b)  
(9c) 
Assuming that for some then Hyvonen2017 . In particular for , and . We call (9) the smoothed CEM (SCEM).
The main aim of this paper is to develop an iterative method for reconstructing from Inspired by Bal2012 the method will be based on the LevenbergMarquardt method for solving the least squares problem (3). The main novelty here being the use the SCEM to accurately and stably model electrodes and the electric current in the forward problem. In addition, a reconstruction strategy based on a combined use of both CEM and Continuum model with Dirichlet boundary condition (DCM) is proposed to increase the computational efficiency.
The outline of the paper is as follows. In Section 2, the LevenbergMarquardt method is briefly introduced. The nonlinear problem is linearized, and the adjoint problem is setup. In Section 3, the iterative reconstruction method is developed based on CEM and LevenbergMarquardt iteration. A linear system is built to calculate the updating step for each iteration. The algorithm for increasing efficiency by exploiting DCM is also introduced in this section. These algorithms are implemented and applied to reconstruct the conductivity distribution of several phantoms in Section 4, numerical performances of different algorithms are discussed in detail. The conclusion of the presented work is given in Section 5.
2 Reconstruction algorithm
In this section we will first recap the LevenbergMarquardt Algorithm for solving (nonlinear) optimization problems. Then we will for the particular problem in AET with CEM derive the necessary ingredients, that is the Fréchet derivative of and it’s adjoint.
2.1 The LevenbergMarquardt Algorithm
Let be a (possibly nonlinear) operator between Hilbert spaces and For some with the problem is to solve at least approximately the equation and often the minimization problem
(10) 
is considered. If is (Fréchet) differentiable, the linear approximation
yields the iterative scheme
solved by the Newton algorithm
This is applicable only when is leftinvertible. In general one can instead minimize the following Tikhonov functional Kaltenbacher2008
(11) 
where is the regularization parameter; this minimization problem is solved by the LevenbergMarquardt Algorithm (LMA)
(12) 
Here, is the adjoint of and is identity operator. When the operator satisfies a certain nonlinarity condition, a proper choice of the parameter and the initial guess sufficiently close to the desired solution, the LMA converges to a solution of (10).
The LMA can be thought of as a combination of steepest descent and GaussNewton method. When the current solution is far from the correct one, a large value is assigned to , and LMA behaves like a steepestdescent method which converges slowly. When the current solution is close to the correct solution, a small is used, and LMA behaves like a GaussNewton method, which has faster convergence.
2.2 The Fréchet derivative and its adjoint
In the present work we consider the operator in (1). In order to apply (12) we therefore need to calculate the Fréchet derivative of the power density operator at and its adjoint
We start with the Fréchet derivative of i.e. the operator The Fréchet derivative can be calculated (see the general approach in Lechleiter2008a ) in the following way:For a given with compact support inside the difference is approximated by a function that is linear in Indeed, define as the solution to the modified CEM problem
(13a)  
(13b)  
(13c) 
With the grounding (13) has a unique weak solution which obviously is linear with respect to . Moreover,
showing that is indeed the Fréchet derivative of at
The Fréchet derivative of in the direction is obtained as in Bal2013 by
(14) 
We now compute the adjoint first as an operator in Consider for some
(15) 
We focus on the second part of the integral (the first part is selfadjoint). Introduce the pair and defined by the weak PDE form (for all )
(16) 
The strong form reads
Then insert the choice in (16) to calculate the latter term in the RHS of (15)
where the last equality follows from the weak form of (13). Thus we find the form
that is
(17) 
To get to the adjoint in higher order Sobolev spaces for the chosen (e.g. in 2D) we lift the operator to higher order spaces, i.e. we solve for
This is a fourth order PDE problem when With an embedding operator , can be written for as
(18)  
(19) 
where is the adjoint of .
We now restrict ourselves to i.e. and consider the presence of noise. Since noise can not be assumed to be differentiable . Since is a bounded operator, its Hilbert adjoint is bounded as well Kreyszig1978 . With an embedding operator , can be written as
(20)  
(21) 
where is the adjoint of . .
Instead of directly calculating in , the form of in is firstly found, which is given as
(22)  
(23) 
Here, with be a linear operator defined by
(24a)  
(24b)  
(24c)  
(24d) 
3 Iterative reconstruction algorithm based on LMA
According to the LevenbergMarquardt iteration given in (12), the formulation for calculating the th updating step for the presented problem is explicitly given as
(25) 
with
(26) 
Here, is easily obtained with (14), (21), and (23). After computing from (25), the conductivity map obtained from the th iteration is updated by for a new iteration. All PDEs are coupled and collected into the PDE system
(27a)  
(27b)  
(27c)  
(27d)  
(27e)  
(27f)  
(27g)  
(27h)  
(27i)  
(27j)  
(27k)  
(27l)  
(27m)  
(27n)  
(27o)  
(27p)  
(27q) 
with , , , , , and . With the number of measurements , the system is formulated with and . Since equations (27g)(27o) need to be solved for each measurement, one additional measurement will need 3 additional partial differential equations.
It may be possible to simplify the above PDE system if the interior potential near the measurement boundary and the measured potential both turn out to converge significantly faster (with the number of iterations of the solution procedure) than the conductivity estimate
. In this case, we would expect and to hold early in the iteration. From (27h), we would then get that the change in the current at any resulting from a change in is approximately zero after only a few iterations, and conditions (27h) and (27i) might justifiably be substituted with the much simpler Dirichlet boundary condition on .The iterative reconstruction method based on the above system is here named as LMSCEM which is demonstrated in Algorithm 1 for a single measurement. The measured data of is simulated with SCEM in this paper. The relative error is given by where and denotes the true and reconstructed conductivities. The parameter should theoretically be updated according to the value of . If leads to a reduction of the relative error in , is decreased and is accepted. Otherwise, is discarded and is increased. Since can not be determined in practice, a relatively large value is asisgned to in this paper, and is slowly decreased to ensure convergence. The iteration is stopped when the norm of is smaller than a given value or the maximum number of iteration is achieved.
3.1 The LMDCM method and the mixed reconstruction algorithm
If DCM is considered instead of SCEM, the system for reconstructing can be built by a similar calculation Bal2013 . The resulted system for computing remains the same as the one for LMSCEM except that the boundary conditions (27h) and (27i) need to be replaced with on . The computation of in LMCEM is actually expensive because of the additional unknowns in . To obtain a good reconstruction accuracy of LMSCEM, multiple measurements are usually considered, but the computational efficiency will decrease quickly with the increasing number of measurements. Our investigation on the convergence of boundary potential and the conductivity shows that the boundary potential converges much faster, examples are given in numerical experiments. This is mainly because EIT measurement is not very sensitive to the internal change of the conductivity distribution Ammari2008 . This property renders EIT an illposed problem, but it will be taken as the foundation here to build a faster reconstruction approach by mixing LMSCEM and LMDCM, which is abbreviated by LMSCEMDCM.
This mixed reconstruction method is illustrated in Algorithm 2 for a single measurement. Here, , which is a vector composed of the voltages on the electrodes. The true values can be measured, which is produced when simulating the power density with CEM, therefore no additional computation is required. The information of is here exploited to define a stopping criteria for reconstructing the boundary potential with LMSCEM. With a relative error given by , the value of is checked in each iteration, and LMSCEM is terminated when an expected relative error is achieved. Since the regularity of potentials computed from CEM is not so good on , a smaller region is defined with a small which can “smooth” out the possible irregularity of close to . The boundary potential is here defined on and used for the reconstruction with LMDCM. The power density in is firstly reconstructed from . This can be achieved with the method introduced in Ammari2008 but will be simulated with DCM here. Because there are no additional unknowns in LMDCM, the computation will be more efficient, especially for the computation with multiple measurements. Meanwhile, the conductivity map produced by LMSCEM is used as the initial guess for LMDCM. This good initial guess will also help LMDCM converge faster. Therefore, the mixed reconstruction method can provide a practical and efficient computational model for AET.
4 Numerical investigation
4.1 Phantom preparation and numerical setup
The variational forms of the linear systems defined by equations (27) in Section 3 can easily be obtained through integration by parts. They are solved with a mixed finite element method Boffi2013 which is implemented using FEniCS AlnaesBlechta2015a . To illustrate the stability and accuracy of the presented approaches we the focus on numerical examples in a 2dimensional (2D) problem. Two phantoms are considered.
The first example is a heartlung model Zlochiver2003 , see Figure 0(a). The considered three tissues are heart (red, ), lung (cyan, ), and softtissues (blue, ). The model is placed into a circular region with a background material (white, ) and a radius .
). All electrodes are uniformly distributed with the same corresponding central section angle. It is assumed that the electrical conductivity
in the region close to boundary (between solid and dashed black lines) is known.The second example is the human brain model shown in Figure 0(b). The considered tissues in this model include scalp (green, ), skull (blue, ), cerebrospinal fluid (red, ), gray matter (yellow, ) and white matter (cyan, ). Refer to Andreuccetti1997 for conductivities of different tissues. The shape of this model is close to an ellipse whose semimajor and semiminor axes are and . The model is placed in an ellipse region with a background material (white, ). The semiminor and semimajor axes of the region are and , respectively. The conductivity maps for the phantoms are piecewise constant functions which will be mollified with
(28) 
to produce with for 2D problem. The constant is selected so that . The value of is and for heartlung model and humanbrain model, respectively. The true smoothed distributions of for the two models are shown in Figure 2.
To work with CEM and LMSCEM, 16 electrodes (red rectangles shown in Figure 1) are uniformly attached on the boundary (solid black lines). The section occupied by each electrode has the same central angle in both the heartlung and humanbrain models. The following computations assume that the conductivity in a small region close to boundary (between solid and dashed black lines) is known. The distance between the solid and dashed lines is given by . This known region helps to improve the convergence of the algorithm. Three current patterns based on Fourier basis functions are used in the computations, which are for , and . The regularization parameter is chosen to decrease exponentially, and with . In what follows, a relatively large value is given to , and a value close to 1 is given to for a slow decreasing of to ensure the convergence of the iterations.
4.2 Performance of LMSCEM
Numerical experiments on heartlung model are carried out here to investigate the performance of LMSCEM. A mesh of the circular domain with 77101 triangles is used for the reconstruction. The power density for each current pattern is simulated with CEM, and Gaussian white noise is added to avoid an inverse crime. Here, the noise level is measure with signal to noise ratio (SNR)
, where is a Gaussian white noise distribution. The LevenbergMarquardt iteration is stopped when or total number of iterations greater than 15. These values were chosen to balance the quality in the reconstructions versus the computational speed. The reconstruction with different current patterns and different level of noise are considered to check the convergence of and the stability of LMSCEM. The parameters for the reconstruction are given as , , and . These parameters are chosen to ensure that the reconstruction with current pattern converges. Other measurements are taken into the reconstruction without changing the parameters.To simulate the power density , SCEM is used to yield better regularity of the electrical potential. The electrode conductance of is chosen as
(29) 
Here, is the length of the electrode, which is proportional to the central angle corresponding to the electrode. is then scaled to have the required maximum value. The distribution of for the calculation in this paper is given in Figure 2(a) with a maximum value 1. The power density simulated with SCEM is given in Figure 2(b) for a region near one electrode. Here, the current pattern is . A smooth distribution of near the edge of the electrode is observed. For the same region, the power density simulated from CEM is also given in Figure 2(c) with for . Two singularity points are observed at the edges of the electrode. These singularities will cause instability in the reconstruction, and therefore SCEM is used in LMSCEM.
The LMSCEM algorithm is used to reconstruct the distribution of . The initial guess is given as which is the value of the background tissue. The conductivity in the region for is supposed to be known. The values of are truncated to only update for with . The discontinuity caused by this truncation can be removed by applying (28) properly (either by mollification or simply by replacing the discontinuous values), but it does not cause any numerical problems since is small, so no special treatment was done in the following computation. The values of for the first 15 iterations are shown in Figure 3(a). The relative error is also given in Figure 3(b). With noise, the reconstruction with uniformly converges to with 15 iterations. The reconstructed is shown in Figure 4(a). To achieve a level of , it takes more than 40 iterations. A reconstruction with and is then carried out, but a similar speed of convergence and relative error is observed, result is in Figure 4(b). When the current pattern is further considered into the reconstruction, an obvious improvement of convergence is seen, and a relative error level is achieved with 14 iterations, the conductivity map is shown in Figure 4(c). Therefore, the convergence of LMSCEM depends not only on the regularization parameter and the scaling parameter , but also on the current patterns for the measurements. Since indicates 0.1% of Gaussian noise in the simulated , the reconstructed result with is already good, and more iterations will not improve the result. To verify, a reconstruction with , and is carried out with (1% noise). Relative errors are shown in Figure 3(b), the reconstruction converges to within 7 iterations, and more iterations did not bring any improvements. The reconstructed result is given in Figure 4(d).
4.3 Performance of the mixed reconstruction approach
Though LMSCEM performs well for reconstructing the conductivity map, its efficiency of the computation decreases quickly with the increase of measurements. Since EIT is not very sensitive to the change of interior conductivity, the electrical potential should converge faster than the convergence of the conductivity. The humanbrain model is used here for numerical experiments with LMSCEM. The elliptic domain is characterized with major and minor axes, and the domain is meshed with 36893 triangular elements. The parameters are given as , and . The current patterns and are used in the reconstruction. With noise, the iteration is terminated when or the maximum number of iterations equals 30. The reconstruction based on LMSCEM is shown in Figure 6(a). of each current pattern is computed with SCEM and for th iteration, easily follows then. The variation of the relative error and are shown in Figure 5(a) and Figure 5(b), respectively. As can be seen, slowly converges to . This error is much larger than the input noise level, this is mainly caused by the complexity of the phantom and the high contrast of among different tissues. But the potentials on the boundary for both and converge fast to a level of in few iterations. Therefore, the boundary potential converges much faster. This property is exploited here to accelerate the computation by mixing LMSCEM and LMDCM, as demonstrated in Algorithm 2. In this computation, the LMSCEM is stoped when for all current patterns are smaller than . The LMDCM is performed in the region with . The potential on is computed with SCEM and the reconstructed from LMSCEM. The power density in can be reconstructed with the method introduced by Ammari et al Ammari2008 . However, it requires the knowledge on the deformation caused by the ultrasonic waves, therefore, we compute it with DCM instead. Noise with SNR is added, and LMDCM is used for the reconstruction. The relative error is given in Figure 5(a). The conductivity map is reconstructed with in 30 iterations, and the result is given in Figure 6(b). Here, the time required for 30 LMDCM iterations is about 20 minutes which is approximately the time needed for one LMSCEM iteration. So the reconstruction efficiency is greatly improved, and better results are obtained. A similar computation with noise is further considered here. As seen in Figure 5(b), increasing noise does not influence much the convergence of the boundary potential, therefore, this mixed approach can be a good way to remove noise from the measured power density. With 40dB noise in the reconstructed power density in , the distribution of obtained with LMDCM is shown in Figure 6(d). Comparing it to the results obtained with LMSCEM, as shown in Figure 6(c), a better noise tolerance is observed in LMDCM.
5 Conclusion
The work first developed a computational approach for AET by incorporating CEM into the LevenbergMarquardt algorithm. Since the regularity of the power density obtained with traditional CEM is limited because of the Robintype discontinues boundary conditions, a recently proposed smoothed CEM is used in this paper. Numerical investigation shows that this iterative method can stably reconstruct the conductivity map of complex phantoms, and a good accuracy can be achieved even with a certain level of noise though this also depends on the current patterns used in the measurements.
However, the reconstruction algorithm becomes inefficient quickly when the number of measurements is increased. Since EIT is not very sensitive to the changes of the internal conductivity, the boundary potential converges to its true value much faster than the convergence of the conductivity. This fact is then used to build a mixed computational approach. LMSCEM is used to reconstruct the boundary potentials in a few iterations, and LMDCM is applied to reconstruct the conductivity distribution based on the obtained boundary potential. It is observed in the given example that the time for one LMSCEM iteration is enough for the whole calculation of LMDCM, so the reconstruction efficiency is greatly improved with the mixed strategy.
Here, reconstruction with LMDCM requires a first step to compute the power density from the obtained boundary potential from LMSCEM. This additional step can be considered as a step to smooth the electrical potential and to remove the noise in the measured power density, in addition to the better noise tolerance of LMDCM compared to LMSCEM, a better reconstruction can be obtained with the mixed algorithm.
Several numerical experiments are carried out with a heartlung phantom and a complex human brain model. The performance of the presented reconstruction approaches are well demonstrated with different number of measurements. The proposed method applies also to 3dimensional acoustoelectric tomography and anisotropic conductivy distributions; we leave the implementation in these scenarious for future work.
Acknowledgement
The majority of the work was done while Changyou Li was postdoc at the Department of Applied Mathematics and Computer Science, Technical University of Denmark in 2018. We thank the Danish Council for Independent Research — Natural Sciences (grant 400200123) for financial support.
References
 (1) D. S. Holder, Electrical impedance tomography :, IOP Press,, 2005 (2005).
 (2) T. J. Yorkey, Electrical impedance tomography with piecewise polynomial conductivities, Journal of Computational Physics 91 (1990) 344 – 360 (1990).
 (3) A.P. Calderón, On an inverse boundary value problem, in: Seminar on Numerical Analysis and its Applications to Continuum Physics (Rio de Janeiro, 1980), Soc. Brasil. Mat., Rio de Janeiro, 1980, pp. 65–73 (1980).
 (4) H. Ammari, E. Bonnetier, Y. Capdeboscq, M. Tanter, M. Fink, Electrical impedance tomography by elastic deformation, SIAM Journal on Applied Mathematics 68 (2008) 1557–1573 (2008).
 (5) G. Uhlmann, Electrical impedance tomography and Calderón’s problem, Inverse Problems 25 (2009) 123011 (2009).
 (6) B. Brown, Electrical impedance tomography (EIT): a review, Journal of Medical Engineering and Technology 27 (2009) 97–108 (2009).
 (7) D. K. Han, A. Prosperetti, A shape decomposition technique in electrical impedance tomography, Journal of Computational Physics 155 (1999) 75 – 95 (1999).
 (8) C.T. Hsiao, G. Chahine, N. Gumerov, Application of a hybrid genetic/powell algorithm and a boundary element method to electrical impedance tomography, Journal of Computational Physics 173 (2001) 433 – 454 (2001).
 (9) E. T. Chung, T. F. Chan, X.C. Tai, Electrical impedance tomography using level set representation and total variational regularization, Journal of Computational Physics 205 (2005) 357 – 372 (2005).
 (10) K. Knudsen, M. Lassas, J. Mueller, S. Siltanen, Regularized dbar method for the inverse conductivity problem, Inverse Problems and Imaging 3 (2009) 599–624 (2009).
 (11) H. Ammari, An Introduction to Mathematics of Emerging Biomedical Imaging, 1st Edition, Mathématiques et Applications, SpringerVerlag Berlin Heidelberg, 2008 (2008).
 (12) G. Bal, Hybrid inverse problems and internal functionals, in: Inverse Problems and Applications: Inside Out II, MSRI Publications, 2012, pp. 325–368 (2012).
 (13) P. Kuchment, Mathematics of hybrid imaging: A brief review, in: The Mathematical Legacy of Leon Ehrenpreis, Springer Milan, Milano, 2012, pp. 183–208 (2012).
 (14) D. B. Geselowitz, An application of electrocardiographic lead theory to impedance plethysmography, IEEE Transactions on Biomedical Engineering BME18 (1971) 38–41 (1971).
 (15) J. Jossinet, B. Lavandier, D. Cathignol, Impedance modulation by pulsed ultrasound, Ann. N. Y. Acad. Sci. 873 (1999) 396–407 (1999).
 (16) J. Jossinet, C. Trillaud, S. Chesnais, Impedance changes in liver tissue exposed in vitro to highenergy ultrasound, Physiological Measurement 26 (2005) S49–S58 (2005).
 (17) B. C. S. Jensen, A. Kirkeby, K. Knudsen, Feasibility of acoustoelectric tomography – a numerical study, in preperation (2019).
 (18) X. Song, Y. Xu, F. Dong, Linearized image reconstruction method for ultrasound modulated electrical impedance tomography based on power density distribution, Measurement Science and Technology 28 (4) (2017) 045404 (2017).
 (19) X. Song, Y. Xu, F. Dong, Sensitivity matrix for ultrasound modulated electrical impedance tomography, in: 2016 IEEE International Instrumentation and Measurement Technology Conference Proceedings, 2016, pp. 1–6 (May 2016).
 (20) G. Bal, W. Naetar, O. Scherzer, J. Schotland, The LevenbergMarquardt iteration for numerical inversion of the power density operator, Journal of Inverse and Illposed Problems 21 (2013) 265–280 (2013).

(21)
B. J. Adesokan, K. Knudsen, V. P. Krishnan, S. Roy,
A fully
nonlinear optimization approach to acoustoelectric tomography, Inverse
Problems 34 (10) (2018) 104004, 16 (2018).
doi:10.1088/13616420/aad6b1.
URL https://doiorg.proxy.findit.dtu.dk/10.1088/13616420/aad6b1  (22) B. C. S. Jensen, K. Knudsen, B. Jin, B. J. Adesokan, Acoustoelectric tomography with total variation regularization, Inverse Problems (2018). doi:10.1088/13616420/aaece5.
 (23) S. Roy, A. Borzí, A new optimization approach to sparse reconstruction of logconductivity in acoustoelectric tomography, Siam Journal on Imaging Sciences 11 (2) (2018) 1759–1784 (2018). doi:10.1137/17m1148451.
 (24) M. Gupta, R. K. Mishra, S. Roy, Sparse reconstruction of logconductivity in current density impedance tomography, Journal of Mathematical Imaging and Vision (2019) 1–17 (2019). doi:10.1007/s10851019009295.
 (25) S. Hubmer, K. Knudsen, C. Li, E. Sherina, Limitedangle acoustoelectrical tomography, Inverse Problems in Science and Engineering (2018) 1–20 (2018). doi:10.1080/17415977.2018.1512983.
 (26) E. Somersalo, M. Cheney, D. Isaacson, Existence and uniqueness for electrode models for electric current computed tomography, SIAM Journal on Applied Mathematics 52 (1992) 1023–1040 (1992).
 (27) N. Hyvönen, L. Mustonen, Smoothened complete electrode model, ArXiv eprints (Mar. 2017). arXiv:1703.08022.
 (28) B. Kaltenbacher, A. Neubauer, O. Scherzer, Iterative Regularization Methods for Nonlinear IllPosed Problems, Berlin, Boston: De Gruyter, 2008 (2008).
 (29) A. Lechleiter, A. Rieder, Newton regularizations for impedance tomography: Convergence by local injectivity, Inverse Problems 24 (6) (2008) 065009 (2008). doi:10.1088/02665611/24/6/065009.
 (30) E. Kreyszig, Introductory Functional Analysis with Applications, John Wiley & Sons, New York, 1978 (1978).
 (31) D. Boffi, F. Brezzi, M. Fortin, Mixed Finite Element Methods and Applications, Springer Berlin Heidelberg, 2013 (2013).
 (32) M. S. Alnæs, J. Blechta, J. Hake, A. Johansson, B. Kehlet, A. Logg, C. Richardson, J. Ring, M. E. Rognes, G. N. Wells, The fenics project version 1.5, Archive of Numerical Software 3 (100) (2015).
 (33) S. Zlochiver, M. Rosenfeld, S. Abboud, Inducedcurrent electrical impedance tomography: a 2D theoretical simulation, IEEE Transactions on Medical Imaging 22 (2003) 1550–1560 (2003).

(34)
D. Andreuccetti, R. Fossi, C. Petrucci,
An internet resource for the
calculation of the dielectric properties of body tissues in the frequency
range 10 Hz  100 GHz, based on data published by C.Gabriel et al. in
1996. (1997).
URL http://niremf.ifac.cnr.it/tissprop/