Machine Learning for Microwave Imaging

M. Ambrosanio, S. Franceschini, F. Baselice, V. Pascazio,

Department of Engineering, University of Naples Parthenope, Naples, Italy.
e-m ails: {michele.ambrosanio, stefano.franceschini, fabio.baselice, vito.pascazio}@uniparthenope.it

Abstract—This paper proposes a fully-connected artificial

neural network (ANN) approach for addressing the full-wave
inverse scattering problem in a quantitative fashion. The pro- Γ
posed scheme processes the scattered field samples collected at
receivers locations and provides as output an estimate of the
unknown complex permittivity in strongly non-linear scenarios.
The proposed approach requires a proper training step, which is
also addressed via an automatic randomly-shaped complex 𝜖𝑠 1𝒓
profile generator inspired by the statistical distribution of breast
biolog- ical tissues, and is almost real-time in the recovery step.
Several representative numerical tests were carried out to 𝜖𝑠 2𝒓
evaluate the performance of the proposed method and to
validate the use of ANN for quantitative imaging purposes in Ω
biological-inspired scenarios.
Index Terms—Microwave imaging, inverse scattering, artificial
neural network, MIMO systems, biomedical imaging.

I. I NTRODUCTION
Fig. 1. Multiview-multistatic imaging system. Only one transmitter per time
Inverse scattering (IS) techniques have a wide range of is active, while the others act as receivers to collect the field coming from the
applications in various fields [1]– [2], since they are investigation domain (i.e., imaging domain).
concerned with non-invasive and non-destructing imaging
modalities for retrieving either physical or geometrical
properties of unknown scatterers located inside a inaccessible II. M ATHEMATICAL B ACKGROUND AND THE FORWARD
domain from some measures of the scattered field. For these PROBLEM
reasons, this imaging modality seems suitable for biomedical
applications [3]–[5].
The drawbacks of this approach mainly lies in the intrinsic In order to simplify the mathematical formulation, in the
ill-posedness and non-linearity of the IS problem at hand. following a scalar two-dimensional (2D) scenario will be
Generally, iterative optimisation schemes with regularisation considered. The reference geometry is depicted in Fig. 1.
are employed to solve the problem and find an estimate of The background medium is assumed to be homogeneous in
the unknown profiles. One of the main issues related to the space and includes the imaging domain (ID) Ω 2, which
⊂ antennas,
use of these iterative strategies concerns the fact that they hosts the targets. Both transmitting and receiving
are time-consuming and thus not suitable for real-time appli- modelled as filamentary z-directed electric currents (e.g.,
cations. Even though some linear/non-linear approximations transverse-magnetic polarization), are located on a curve Γ
can be proposed to hasten the procedure, they are still not which surrounds the ID in a multiview-multistatic fashion. All
accurate in several scenarios, especially when strong scatterers the scatterers, as well as the background medium, are assumed
are involved. In this framework, methodologies based on invariant along the z-direction and characterised by a
artificial neural network (ANN) can play an important role magnetic
for quantitative imaging purposes [6]. Recently, deep learn- permeability equal to µ0 = 4π · 10−7 H/m.
ing has attracted attention with interesting results for image Due to the previous assumptions, it is possible to focus only
classification and segmentation. Nevertheless, neural network on the electric field. The two fundamental equations which
with regression features have also provided impressive results govern the scattering phenomena for the considered simplified
on inverse problems, such as denoising, deconvolution and geometry can be expressed in an integral form (electric-field
interpolation. integral equation) [7]:
In this paper, we propose a neural network approach for
the quantitative imaging of strongly non-linear profiles via a Et(r, rT , ω) − Ei(r, rT , ω) =
direct inversion scheme. A fully-connected neural network is 2
/ G(r, r/ , ω)χ(r/ , ω)E (r /, r , /
=
employed and a proper data set is generated for the training = t T
Ω
phase. The output of the network consists in an estimate of
the complex permittivity profile. = Ai[χEt], r, r/ ∈ Ω, rT ∈ Γ, (1)

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2020 at 20:11:50 UTC from IEEE Xplore. Restrictions apply.
 Es(rR, rT , ω) = between data and unknowns can be obtained, accelerating the
= k / G(rR , r /, ω)χ(r/, ω)Et(r/, rT , ω)dr/
2 inversion considerably.
b Ω
= A. Data-set generation
= Ae[χEt], r/ ∈ Ω, rT , rR ∈ Γ, (2)
As previously stated, ANN and DNN can represent an
in which kb is the wavenumber in the background medium,
interesting, fast alternative to classical analytical inversions
ω is the working frequency, Et, Ei, Es are the total, in-
for imaging purposes. Nevertheless, the main drawback of
cident and scattered fields respectively, and G(r, r /, ω)
these approaches is related to the need of exploiting large data
=
j sets
− H(2) (kb|r − r/|) is the Green’s function for the
4 0
case of which are not always available to the research community.
homogeneous background. Therefore, the data set selection still represents an important
The function “χ (r, ω)” relates the electric properties of the step for the learning procedure. Aware of this limitation, and
objects inside the imaging domain to those of the homoge- inspired by the relatively simple geometry of breast profile,
neous hosting medium as: we propose an automatic procedure for the random generation
Es(r, ω) of internal breast tissue profiles.
χ(r, ω) = 1, (3)
E (ω) − Since the biological tissues of the human female breast
b
in which E (r, ω) = E/ (r) − j σ s( r )
and E (ω) = E/ − j can be mainly grouped into three classes, i.e. fibro-glandular,
σb s s ω b b ω transitional and adipose tissues [9], it is possible to exploit a
are the complex permittivities of the unknown targets and random-shape generator to model the fibro-glandular tissues
background, respectively. as scatterers located into an adipose tissue, which is treated as
It is worth to note that both the radiating operators i and a background medium, while the transitional one is spatially-
A
e can be evaluated in a fast and efficient way by means of fast constrained to surround the fibro-glandular tissue. In this first
A Fourier transform (FFT) codes, due to the fact that the simplified 2D test-case, the skin ring which embeds all the
integrals in (1)-(2) represent a convolution [8]. These other tissues has not been considered.
equations need to be treated and solved. Therefore, after a In order to model the spatial variability of the fibro-
proper discretization of the domain Ω and after choosing glandular tissue, a random-shape generation based on a univer-
properly the measurement points for the scattered field Es, a sal multi-fractal random field generator is proposed [10]. It is
more handy expression of worth to note that the class of the generated breast (according
these formulas can be obtained [8]. to the percentage of fibro-glandular tissue) can be selected by
As well known, the problem of retrieving the contrast properly setting the combination of three parameters which
function χ(r, ω) from the measurement of the scattered field govern the multi-fractal behvoiur (Levy parameter, the level
samples Es(rR, rT , ω) is a non-linear, ill-posed problem [7]. of sparsity of the field and the factional integration, i.e.
The ill-posedness is due to the loss of information related to the ruggedness or smoothness of the profile). Regarding the
the intrinsic properties of the operator relating the unknown complex permittivity employed for the considered objects
objects to the scattered field, which entails that the solution mimicking the biological tissues, the statistical distributions
can be sought only in a finite dimensional space, therefore reported in [9] have been exploited.
efficient and reliable regularization strategies are required in
order to obtain stable solutions. Furthermore, the non-linearity B. Fully-connected framework
of the problem at hand also contributes to make the solution
harder to be sought and computationally expensive. The use of ANN and DNN for imaging purposes goes
In order to face these drawbacks and provide an efficient, back till to the nineties for simple imaging scenarios [11],
almost real-time imaging strategy also in complicated, and recently has also been exploited by other groups in
strongly non-linear scenarios, the framework of machine more complicated scenarios as a complementary strategy in
learning can be exploited. the inversion procedure for regularisation and super-resolution
issues [12]– [13]. In this manuscript, the authors want to
III. M ACHINE L EARNING FOR QUANTITATIVE propose a preliminary network architecture in order to test
MICROWAVE IMAGING the imaging potentialities of the ANN framework which aims
The solution of IS problems for imaging purposes has a at providing quantitative maps of the electric properties of the
long history in the research community and it is mainly based targets under test starting from measures of the scattered field
on use of analytical and iterative methods. An alternative, samples.
more recent and less computationally-expensive approach A DNN can be described as a multilayer stack of simple
exploits the theory of artificial neural networks (ANN) and modules organised in series, i.e. the output of each module
deep learning (DNN) for solving the inverse problem at hand represents the input of the next one. Each module, known
via minimising a proper non-linear functional with the aim of as layer, is composed of multiple units or neurons named
finding an estimate of the inverse non-linear operator. More nodes. Per each pair of consecutive layers, all the nodes are
in details, if ANN are involved, the weights of the network connected one another and the result of the weighted sum of
are learnt by means of large data sets with pairs of inputs and the input units is followed by a nonlinear activation function.
outputs. As a result of the learning procedure, a direct
mapping

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2020 at 20:11:50 UTC from IEEE Xplore. Restrictions apply.
 𝜖′
Biases
( b1 )

Fully- ReLu
Hidden layer 2 Hidden
connected + layer 3
( W1 ) Ω

Esct
Hidden layer 1

𝜖′′
Ω

Fig. 2. Architecture of the proposed network. This direct inversion scheme has the samples of the scattered field (Esct) as input and provides an estimate
of unknown complex permittivity (El, Ell) profile maps as output.

The weights between each pair of units are learnt via a certain Regarding the training of the proposed network, the default
training procedure. values of the initial weights belong to a Gaussian distribution
Inspired by the universal approximation theorem [14], with zero-mean and standard deviation equal to 0.01, and
which states that a fully-connected neural network with a default initial bias equal to zero. The method employed to
large number of neurons in its hidden layers can represent minimise the loss function is the stochastic gradient descent
any function to be learnt under some assumptions on the algorithm with momentum which updates the network param-
activation function, we explored a three-layer network whose eters by taking small steps in the direction of the negative gra-
architecture is illustrated in Fig. 2. Each of the three hidden dient of the loss. Moreover, in the updating rule, a momentum
layers (forty nodes per each) in the network is composed term is added to reduce the oscillation due to the contribution
of a fully-connected layer which multiplies the input by a of the previous gradient step at each iteration. Finally, a
weighting matrix and then adds a bias. As suggested by the regularisation term for the weights of the loss function is
name, all neurons in a fully-connected layer connect to all added to reduce the overfitting.
the neurons in the previous layer. Each layer combines all of In order to perform the training, a constant learning rate
the features (local information) learned by the previous layers equal to 10−5 was employed for a maximum of 100 epochs,
across the image to identify the largest patterns. After each and a constant L2-norm regularisation parameter equal to 10−4
fully-connected layer, a non-linear activation ReLu function is was used. The whole training procedure lasts approximately
added which performs a threshold operation to each element 50 minutes on a Dell Alienware Area-51 workstation with 64
of the input, where any value less than zero is set to zero. GB RAM memory DDR4 XMP 2933 MHz and a graphics
Finally, a regression layer is included in the last hidden layer, board NVIDIA GeForce GTX 1080. The accuracy of the
which is a common strategy in this kind of architectures. retrieved profiles was measured quantitatively via introducing
a normalised mean square error:
IV. N UMERICAL RESULTS lEx (r) − Eref 2,
err =
l (4)
In order to test the performance of the proposed neural
lEref l 2
network, a data set of N = 14000 profiles has been whose values for the considered numerical examples are
employed. The profiles were generated according to the reported in Fig. 3.
information reported in Subsection III-A and the V. C ONCLUSION
corresponding scattering matrices were evaluated via
In this manuscript, a fully-connected neural network ar-
running a method of moments (MoM) solver. The whole
chitecture for full-wave quantitative microwave imaging in
population was split into three main classes, i.e. 80% of
strongly non-linear scenarios is presented. In order to train
this data set was used for the training, 15% for the
the network, a database of randomly-shaped profiles has been
validation and 5% for the testing.
generated with electric parameters randomly selected in three
The imaging domain in which the electric profiles are lo-
cated is λb ×λb m2, with λb the wavelength in the different ranges according to proper statistical distributions, as
background medium, E/r,b = 4 and σb = 5 10−3 S/m reported in [9]. These profiles were selected to simulate the
being the electric features · of the background. The internal tissues of human female breast for biomedical
operating frequency is fixed at imaging purposes, which represents a strongly non-linear
1 GHz and the circle on which the receivers and transmitters scenario.
are located has radius equal to √2λb. Regarding the number The neural network architecture consists of three fully-
of antennas, thirty elements acting as transmitters/receivers in connected layers and proved to have good retrieving perfor-
a multiview-multistatic fashion have been employed. mance in comparison with distorted Born iterative method, a

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2020 at 20:11:50 UTC from IEEE Xplore. Restrictions apply.
 ((a)) ((b)) ((c))

((d)) ((e)) ((f))

Fig. 3. Retrieved profiles. (a),(d): real and imaginary parts of the reference complex permittivity, respectively; (b),(e): retrieved profiles via the proposed
neural network architecture, err = 0.23; (c),(f): classical non-linear approach (distorted-Born iterative method), err = 0.30.

classical well-known inversion scheme, especially in [9] M. Lazebnik, D. Popovic, L. McCartney, C. B. Watkins, M.J. Lindstrom,
retrieving the imaginary part of the complex permittivity, J. Harter, S. Sewall, T. Ogilvie, A. Magliocco, T. M. Breslin, et al., “A
which is an issue for several classical imaging schemes. large-scale study of the ultrawideband microwave dielectric properties
of normal, benign and malignant breast tissues obtained from cancer
Future work will focus on studying a proper training proce- surgeries,” Physics in Medicine & Biology, vol. 52, no. 20, pp. 6093,
dure to effectively train the network and on finding the 2007.
optimal design of the network. [10] D. Schertzer and S. Lovejoy, “Nonlinear variability in geophysics:
Multifractal simulations and analysis,” in Fractals’ Physical Origin and
Properties, pp. 49–79. Springer, 1989.
REFERENCES [11] S. Caorsi and P. Gamba, “Electromagnetic detection of dielectric cylin-
ders by a neural network approach,” IEEE transactions on geoscience
[1] M. Pastorino, Microwave imaging, vol. 208, John Wiley & Sons, 2010. and remote sensing, vol. 37, no. 2, pp. 820–827, 1999.
[2] M. Ambrosanio, M. T. Bevacqua, T. Isernia, and V. Pascazio, “The [12] P. Shah and M. Moghaddam, “Super resolution for microwave imaging:
tomographic approach to ground-penetrating radar for underground A deep learning approach,” in 2017 IEEE International Symposium
exploration and monitoring: A more user-friendly and unconventional on Antennas and Propagation & USNC/URSI National Radio Science
method for subsurface investigation,” IEEE Signal Processing Meeting. IEEE, 2017, pp. 849–850.
Magazine, vol. 36, no. 4, pp. 62–73, 2019. [13] A. Ashtari, S. Noghanian, A. Sabouni, J. Aronsson, G. Thomas, and
[3] M. Ambrosanio, P. Kosmas, and V. Pascazio, “A multithreshold iterative S. Pistorius, “Using a-priori information for regularization in breast
dbim-based algorithm for the imaging of heterogeneous breast tissues,” microwave image reconstruction,” IEEE Transactions on Biomedical
IEEE Transactions on Biomedical Engineering, vol. 66, no. 2, pp. 509– Engineering, vol. 57, no. 9, pp. 2197–2208, 2010.
520, 2018. [14] K. Hornik, M. Stinchcombe, and H. White, “Universal approximation of
[4] M.T. Bevacqua and R. Scapaticci, “A compressive sensing approach an unknown mapping and its derivatives using multilayer feedforward
for 3d breast cancer microwave imaging with magnetic nanoparticles as networks,” Neural networks, vol. 3, no. 5, pp. 551–560, 1990.
contrast agent,” IEEE transactions on medical imaging, vol. 35, no. 2,
pp. 665–673, 2015.
[5] M Ambrosanio, R Scapaticci, and L Crocco, “A simple quantitative
inversion approach for microwave imaging in embedded systems,”
International Journal of Antennas and Propagation, vol. 2015, 2015.
[6] A. Lucas, M. Iliadis, R. Molina, and A. K. Katsaggelos, “Using deep
neural networks for inverse problems in imaging: beyond analytical
methods,” IEEE Signal Processing Magazine, vol. 35, no. 1, pp. 20–36,
2018.
[7] D. Colton and R. Kress, Inverse acoustic and electromagnetic scattering
theory, vol. 93, Springer Science & Business Media, 2012.
[8] T. Isernia, V. Pascazio, and R. Pierri, “A nonlinear estimation method in
tomographic imaging,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 35, no. 4, pp. 910–923, 1997.

Authorized licensed use limited to: Carleton University. Downloaded on August 25,2020 at 20:11:50 UTC from IEEE Xplore. Restrictions apply.