Content uploaded by Panagiotis Koulountzios

Author content

All content in this area was uploaded by Panagiotis Koulountzios on Apr 04, 2022

Content may be subject to copyright.

Available via license: CC BY 4.0

Content may be subject to copyright.

20896 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

A Triple-Modality Ultrasound Computed

Tomography Based on Full-Waveform

Data for Industrial Processes

Panagiotis Koulountzios, Tomasz Rymarczyk ,

Member, IEEE

, and Manuchehr Soleimani

Abstract

—Ultrasound computed tomography (USCT) is

gaining interests in many application areas in industrial

processes. The recent scientiﬁc research focuses on the

possible uses of USCT for varied ﬁelds of industry such

as ﬂow monitoring in pipes, non-destructive inspection, and

monitoring of stirred tanks chemical processes. Until now,

most transmission tomography (UTT) and reﬂection tomog-

raphy (URT) have been demonstrated individually for these

applications. A full waveform USCT contain large amount

of information on process under evaluation. The developed

approach in this paper is focusing on demonstration of a

triple modality USCT. First, an optimised transmission image

is formed by fusion of time-of-ﬂight (TOF) and acoustic atten-

uation (AA) images. Secondly,a reﬂectionimage is being opti-

mised by using the information from the transmission image.

This triple modality method enables integration of a shape-

based approach obtained by URT mode with the quantitative

image-based approach UTT mode. A delicate combination of

the different information provided by various features of the

full-wave signal offers optimal and increased spatial resolution and provides complementary information. Veriﬁcation

tests have been implemented using experimental phantoms of different combinations, sizes, and shapes, to investigate

the qualitative imaging features. Moreover, experiments with different concentrations solutions further validate the

quantitative traits to beneﬁt from both reﬂection and transmission modes. This work displays the potential of the full-

waveform USCT for industrial applications.

Index Terms

—Ultrasound computed tomography (USCT), ultrasound process tomography (UPT), industrial processes,

multi-modality ultrasound tomography, TOF imaging, AA imaging, reﬂection imaging, full-waveform rich tomography.

I. INTRODUCTION

ULTRASOUND computed tomography (USCT) has been

studied lately on a broad spectrum of industrial appli-

cations with signiﬁcant success [1]–[10]. Its usage has drawn

a special attention notably relating to the imaging of bipha-

sic medium and liquid mixtures in pipe ﬂows and stirred

reactors environments [11]–[15]. USCT works by analysing

Manuscript received June 24, 2021; revised July 22, 2021; accepted

July 23, 2021. Date of publication July 27, 2021; date of current ver-

sion September 15, 2021. This work was supported by the European

Union’s Horizon 2020 Research and Innovation Program through Marie

Skłodowska-Curie Grant 764902. The associate editor coordinating the

review of this article and approving it for publication was Prof. Yongqiang

Zhao.

(Corresponding author: Manuchehr Soleimani.)

Panagiotis Koulountzios and Manuchehr Soleimani are with the

Electronic and Electrical Engineering Department, University of

Bath, Bath BA2 7AY, U.K. (e-mail: p.koulountzios@bath.ac.uk;

m.soleimani@bath.ac.uk).

Tomasz Rymarczyk is with Research and Development Centre, Netrix

S.A., 20-704 Lublin, Poland (e-mail: tomasz@rymarczyk.com).

Digital Object Identiﬁer 10.1109/JSEN.2021.3100391

the acoustic wave propagation, via sound velocity or pulse

amplitude decays of different materials. It aims to the mapping

of medium’s acoustical properties. It is non-invasive and non-

destructive, compatible with high dynamical processes, like

oil and gas ﬂow. A better understanding of the measurement

process and fast reconstructions algorithms are imperative for

the use of USCT within the industry. Moreover, complex

stirred tank processes, require sophisticated algorithms, which

can provide accurate results. Due to the complex physical

behaviour of acoustics propagation, there are multiple recon-

struction modes, which use different waveform’s properties.

The transmission and reﬂection modes have been traditionally

used in ultrasound tomography reconstructions, accounting for

transmitted diffracted and reﬂected waves. Functional features

of these reconstruction modes can be complementary. For

instance, reﬂection tomography offers good resolution at the

boundary of different media, while the transmission method

has better resolution in distinguishing discontinuities along the

signals’ propagation path. Transmission mode properties, such

1558-1748 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.

See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on November 23,2021 at 10:30:19 UTC from IEEE Xplore. Restrictions apply.

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20897

as acoustic attenuation (AA) or the time-of-ﬂight (TOF), can

be used to determine the amplitude and sound-speed proﬁles of

the region of interest (ROI), offering quantitative information.

All the AA, TOF, and reﬂection modes have drawbacks

with artefacts under certain biphasic medium distributions.

Thus, the performance of single-modality Ultrasonic Process

Tomography (UPT) is limited. However, these shortcomings

may be compensated with a multi-modality reconstruction.

For instance, in a liquid-liquid mixture, the reﬂected signals

will be signiﬁcantly low, leading to a possible reﬂection

reconstruction failure. However, transmission image should

be more meaningful in such scenarios. On the other hand,

in a liquid-solid or liquid-gas medium, reﬂected waves may

be exploited. For instance, in many stirred tanks chemi-

cal processes like fermentation and crystallization, localised

super-saturated suspensions may be formed due to process

malfunction e.g., stirrer malfunction. In this case, a drastic

structural phase difference may occur. Thereafter, reﬂection

mode might be used as a malfunction detection by detecting

the localised high concentrated suspensions. It would be also

possible to detect complex dynamic phenomena such as a gas-

ﬂow or vortexes coming from a high stirring effect. Although,

the use of reﬂection mode in that direction needs to be further

investigated.

The multiple mechanisms (transmission, diffraction, reﬂec-

tion) during ultrasound propagation and the rich information

(attenuation, time-delay, distortion) contained in an ultrasonic

full signal, establishes the need for a multi-modality method.

Therefore, a multi-modality approach, that can facilitate mul-

tiple reconstruction methods, is expected to result in more

accurate imaging, as it can process measurements coming

from different signal’s features. Several studies show that dual-

modality UPT (transmission/reﬂection) is superior to single-

modality, conﬁrming that multi-modality offers tremendous

beneﬁts [16], [17]. As a result, a novel triple-modality image

reconstruction method combining AA and TOF transmission

as well as TOF reﬂection is proposed.

In a small circular setup ﬁlled with a non-homogeneous

medium, signiﬁcant back-scattering and reﬂections are

expected to happen. Thus, the “noise” levels are higher. Most

common issues in ultrasound tomography revolve around the

estimation of TOF and AA from the full-waveforms, especially

in instruments that are not calibrated or that are characterised

by high “noise” levels [18]–[21]. A novel reﬂection TOF

picking method was built to tackle this problem. It exploits

a forward reﬂection solver based on ray acoustics, to opti-

mise the recorded reﬂected TOF values and subsequently the

reﬂected image. The proposed method ﬁts well in the robust

triple-modality proposed approach. The amplitude of the trans-

mitted pulses and time-of-ﬂight of both the transmitted and

reﬂected pulses have been used to produce three different

reconstructions (TOF, AA, reﬂection). Finally, a method of

image fusion, using the results of TOF-UTT, AA-UTT and

URT methods, was developed and used to generate the ﬁnal

image.

The paper is organised as follows. Section II presents

the main functionality of the tomographic system and

details its speciﬁcations. Moreover, the methods undertaken

for transmitted and reﬂected TOF and AA picking are

Fig. 1. Triple-modality ultrasonic tomography (transmission/reﬂection).

Design of transmitted and reﬂected signals’ paths.

further characterised. Section III describes the reconstruction

formulas for transmission and reﬂection tomography, while

section IV presents the proposed algorithm for reﬂected TOF

picking, which helps optimising the recordings. Section V

presents the developed fusion method. Finally, in section VI,

the experimental results are presented and evaluated, and in

section VII, the conclusions and discussion occur.

II. MEASURING SYSTEM

Figure 1 depicts the design of such a triple-modality ultra-

sound tomographic concept. For a 16-channel transducer

system each sensor acts as both transmitter and receiver

and complete tomographic data is collected by exciting each

sensor in turn. When Tx1 is an excitation transducer Rx1,

Rx2,…,Rx15 represent the receivers. The transmission mode

uses AA and TOF data for Rx3, Rx4, Rx5, ..., Rx13 and

for reﬂection mode Rx1, Rx2, Rx14, Rx15 data are used.

The actuated receivers in transmission mode are those that

are included in the fan beam of 120-degree angle where the

best quality of transmission data is possible. Those, located

in the neighbourhood of the transmitter are excluded from

transmission mode data but used in reﬂection mode.

The multi-modality USCT approach utilizes TOF and

amplitude information from transmission and reﬂection waves.

These waves come from the interaction of different phase

structure within the medium. Acoustical properties are depen-

dent on changes in the structural phases (i.e., acoustic

impedance, velocity). Sound transmission and reﬂection

would, therefore, change within these phases. Acoustic

impedance Z affects the propagation and nature of pulse

excitation and is dependent on the material’s structural phase.

It indicates the intensity of a medium’s regions to block the

vibrations of the particles in the acoustic ﬁeld [22]. The

ratio of the reﬂected pulse’s amplitude, Pr,to the incident

wave’s amplitude, Po, is called the acoustic pressure reﬂection

coefﬁcient R [23], and it is deﬁned as:

R=Pr

Po

=Z1cosθ0−Z2cosθt

Z1cosθ0+Z2cosθt

(1)

By the same way, the acoustic pressure transmission coef-

ﬁcient, T, is deﬁned as:

T=Pr

P0

=2Z1cosθ0

Z1cosθ0+Z2cosθt

(2)

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on November 23,2021 at 10:30:19 UTC from IEEE Xplore. Restrictions apply.

20898 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

where Pr,Ptdeﬁne the reﬂection and transmitted wave’s

acoustic pressure; θ0,θtrepresent the wave’s angle of inci-

dence and angle of transmission, respectively; P0is the

acoustic pressure of the incident wave. Z1and Z2are the

acoustic impedances of medium 1 and medium 2.

When a sound wave propagates through a medium, its

intensity decreases with the distance travelled, as expressed

in eq. (3).

A=A0e−μh(3)

where A0is the amplitude of the propagating wave at a given

location. A is the reduced amplitude at another location. In this

case, his the distance travelled between the two locations,

and μis the attenuation coefﬁcient in Neper (Np)/length. The

two primary mechanisms that cause the attenuation of sound

energy are absorption and scattering. Industrial processes

usually consist of multiple phase media with a drastic dif-

ference in structural phase. Such conditions are favourable

for a multi-modality approach in ultrasonic reconstructions,

exploiting attenuation, sound-speed and acoustic impedance

change within the medium.

A. Tomographic Device

The ultrasonic tomograph has 32 independently working

channels that can perform measurements in transmission and

reﬂection modes. In this paper, we use 16 channel sensors in a

single plan 2D USCT mode. The bottom layer was used for the

data collection. At the same time, a sensor sends an ultrasonic

signal of 5 cycles (tone burst), while remaining sensors record.

Respectively, receivers measure the full-waveform signal. The

sequence repeats until every sensor produces an excitation

signal. The system inside the reinforced suitcase consists of

eight four-channel measurement cards connected via a FD

CAN bus to the measurement module. The measurement mod-

ule is a bridge between a microprocessor measuring system

and a touch panel or external control application (Figure 2).

Each device channel has its own analogue signal processing

module and its own 12-bit ADC 4MSPS converter. In TOF

and amplitude measurement mode, the signal is normalised to

voltages from 0-3.3V, according to the transducer’s reference

voltage. The sampling frequency is 0.25 samples per micro-

second, which results directly from the ADC converter speed.

A Built-in envelope converter was used for converting an

analogue acoustic signal to the envelope with the possibility

of switching its conﬁguration to a frequency of 400 kHz.

In addition, the measurement module can monitor the mea-

suring sequence, store user-speciﬁc parameters, control the

high voltage inverter, and switch the USB HS bus between the

front panel’s socket and the touch panel. The touch panel was

made using a RaspberryPi 4B 2GB RAM board and a 7-inch

capacitive touch screen. The most crucial data buses have been

led to the front panel of the device. Each sensor has its own

signal conditioning for both transmission and receiving mode

as shown in green box in ﬁgure 2a.

B. TOF/ AA Picking Method for Transmission

Tomography

Transmission signals directly travel from the transmitter to

the receivers without any reﬂection. These signals undergo

Fig. 2. (a) Measurement system: Ultrasonic tomograph block diagram.

(b) ultrasound tomographic system. (c) tank with sensors.

either diffraction or direct transmission, without signiﬁcant

change of direction. A transmission pulse usually travels faster

and transmits at a larger amplitude. Figure 3(a) shows a full-

waveform signal and its envelope, recorded by Rx6 upon

Tx1 excitation. Moreover, it illustrates signal speciﬁc TOF

and AA picked points, calculated by the applied method.

Figure 3(b),(c) present background and full calculated TOF

and amplitude data. The picking method of transmitted TOF

values is described below.

First, the analytic envelope of the signal processed using

Hilbert transform. Using the envelope, signal oscillation can

be removed, facilitating more accurate peak detection. The

advantage of using the envelope is to migrate the effect of arbi-

trariness and to eliminate the effect of phase changes. Then,

the enveloped is processed to detect the transmitted pulse and

record its x-value, deﬁning the pulse’s travel-time, and y-value,

determining the recorded pressure. The method is described in

eq. (4-8), where v(i)represents an enveloped signal, in our

experiments the receiving signal contains 1665 samples or

time steps. A minimum threshold of 10%, th,is used to

cut down the minor pulses caused by back-scattering or

equipment-related noise, eq. (4). TOF value is determined

by the projection of the ﬁrst signal’s point after threshold, th,

to the x-axis. According to the sampling frequency, the TOF

must be multiplied by 0.25, to be converted in μsec. Eq. (5-7)

describe this in a linearised fashion. Comparatively, the biggest

y-value within a 20% signal’s window in the transmitted pulse

“region” indicates the recorded pulse’s amplitude, eq. (8).

th =0.1max(v(i)) ,where i ={1,2,...,n},q=1665

(4)

v(i)=v(i)for v(i)>th

0for v(i)<th (5)

kj=i|v(i)= 0,j={1,2,...,m}where m ≤n(6)

TOF =0.25k1(7)

AA =max(v(l)),

where{l=k|l={i,i+1,...,i+0.2q}(8)

Authorized licensed use limited to: UNIVERSITY OF BATH. Downloaded on November 23,2021 at 10:30:19 UTC from IEEE Xplore. Restrictions apply.

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20899

Fig. 3. (a) Recorded full-waveform signals from Tx1-Rx6 pair, with its envelope. Each timestep is 1/4 µsec. (b) TOF data computed from the

enveloped signals. (c) AA data calculated from the enveloped signals.

All reconstructions are generated by using difference imag-

ing, collecting background data (reference data), TOFback ,and

full data (data collected by scanning a non-uniform medium),

TOFfull. TOF measurement data, TOFtr, originates from the

subtraction of full data from the background data and deﬁne

the travel-time delays (μs), eq. (9).

TOF tr =TOF back −TOFfull where

TOF tr 0forTOF

tr <0

TOF tr for TOFtr >0(9)

AA measurement data are computed by eq. (10) [24].

AAtr =1

fc

ln AAback

AAfull (10)

where AAback is the signal’s amplitude at each receiver when

there is only water (reference data) in the ﬁeld of view (FOV)

and AAfull is the amplitude of the full data. fcis the centre

frequency of the excitation pulse.

In both TOF and amplitude data, the “Deleting Outliers”

statistical ﬁltering method was used to handle this noise for all

the datasets [25]. Speciﬁcally, “outlier” TOF values usually are

generated from back-scattering or reﬂected signals. Iterative

implementation of the Grubbs Test was used to identify the

outlier signals. In any given iteration, the tested value is either

the highest or lowest value, represented by the furthest value

from the sample mean.

C. TOF Picking Method for Reﬂection Tomography

The “traditional” picking method of reﬂected pulses is

described in this section. This method makes full use of all

the four transducers positioned in pairs, on each side of the

emitter. Contrasting full and background measurements, can-

cels tank-related back-scattering and reﬂection, as expressed

in eq. (11). As both background and full data are assumed

to present similar tank speciﬁc back-scattering and reﬂection.

The recorded maximum pulse peak is assumed as the observed

reﬂected TOF, TOF

rfl, as shown in eq. (12).

pi=|

v(i)back −v(i)full|,

where i ={1,2,...,1665}(11)

TOF

obs

rf =max (pi),where i ={1,2,...,900}(12)

Figure 4(a) illustrates the experimental setting. Figure 4(b)

represents the plots of full (background and inclusions) and

background waveforms and their respective envelopes. The

ﬁrst peak comes from the transmission pulse and exists in

both signals. The two second peaks come from the reﬂected

pulse within the inclusion surface and exist only in Full

data. Lastly, the third peak represents tank-related back-

scattered signal, existing in both signals as well. Absolute

subtraction optimises reﬂected pulses measurements, reducing

overall noise and back-scattered effects. As the reﬂected pulse

remain unchanged, it is easily trackable. Figure 4(c) depicts

the absolute subtracted signal.

III. METHODS

A. Transmission Mode Reconstruction

Transmission can be measured either via a travel-time or

an acoustic attenuation technique. Transmission reconstruction

would take place via either travel-time or the amplitude decay

of the ﬁrst-arrival pulse [26]. The most used approximation for

20900 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

Fig. 4. (a) Schematic of the setup. (b) Background and Full measurements in full waveforms. (c) Difference data in full waveforms. Each timestep

is 1/4 µsec.

transmission USCT is the ray-based method. It is fundamental

in most tomographic schemes, as the line integral deﬁnes

the path of a high frequency propagating pulse between an

emitter and a receiver. It is a simpliﬁed approach, which

does not account for the diffraction effect caused by non-

homogeneous medium. Therefore, a computational model

based on diffraction on the 1st Fresnel zone [27], was used.

Fresnel volume or ‘fat ray’ tomography is an appealing

compromise between the efﬁcient ray theory tomography and

the computationally intensive full waveform tomography [28].

Using a ﬁnite frequency approximation to the wave equation

leads to a sensitivity kernel where the sensitivity of the travel-

time delay also appears in a zone around the fastest ray path.

The delay time is given as:

t(x)=t(s,x)+t(x,r)−t0(s,r)(13)

Here t(s, x) and t(x, r) are the travel-time from

the source (s) to x and from x to the receiver (r) and t0(s,r)

is the travel-time along the ray path from the source to

receiver. The times of travelling can be evaluated using the ray-

tracing method. A point x is always within the ﬁrst Fresnel

zone if the corresponding travel-time satisﬁes the eq. (14),

in which T deﬁnes the emitted wave’s period:

|t(x)|<T

4(14)

The following function deﬁnes the sensitivity of a Frechet

kernel based on the ﬁrst Fresnel zone:

S(x)=KV(s,x)V(x,r)cos 2πt(χ)

T

×exp ⎛

⎝−at(x)

T

42⎞

⎠(15)

where S(x)is the sensitivity at x,V

(s,x)and V (x,r)are

the amplitude values of the waveﬁeld at xpropagating from

sto xand from xto r, respectively, andk is a constant

value for normalisation purpose. The cosine factor models

the alternating sensitivity being positive in the odd Fresnel

zones and negative in the even Fresnel zones. The a,inthe

Gaussian factor controls the degree of cancellation in Fresnel

zones. The geometrical spreading approximates the amplitude

factors in a homogeneous medium. The normalisation of the

kernels is achieved by ensuring that the integrated sensitivity

over the whole medium is equal to the length of the reference

ray path [29]. SIPPI MATLAB software has been used to

generate these sensitivity kernels [30]. These kernels represent

the acoustic distribution of the medium of each sensor’s

excitation, forming the sensitivity matrix. A Normalisation

method based on the geometric wave path was applied to

the generated kernels to ensure an accurate time-of-ﬂight

(TOF) and acoustic-attenuation (AA) mapping, as described

in eq. (16).

Ai,j=Hi,j

m

i1=1j1=jHi1,j1

(16)

where Hi,jis the sensitivity matrix based on the Frechet

method and Ai,jis the normalized matrix, used for recon-

structions, and for m measured data and n number pixels,

i=[1,...m]and j =[1,...,n].

In a tomographic approach, the transmission sensitivity

matrix simulate the propagation of the measured energy from

a excitation sensor. The measurement data for UTT includes

TOF and AA data. The so-called forward problem is formed

by the multiplication of the sensitivity matrix with the mea-

surement data. Below the notation deﬁnes Mstands for

both TOF and AA data for TOF and AA reconstructions,

respectively. A generalised tomographic forward problem can

be expressed as:

M=AS+e(17)

where Sis the reconstructed distribution based on acoustic

features, A is the modelling operator which expresses the sen-

sitivity distribution in the FOV, Mis the sensor’s recorded

data, and eis the noise in the measurements. A simpliﬁed

inversion can be done using back projection.

S≈ATM(18)

Total Variation regularisation (TV) [31], [32] was used,

which has more signiﬁcant potential in solving the regularised

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20901

Fig. 5. (a) Functionality of the reﬂection tomography and geometrical

computation of reﬂection points

C

1and

C

2.(b) Reﬂection image by

superimposing all the ellipses.

inverse problem in a stabilised fashion. The TV problem is

deﬁned as an optimisation problem, minimizing

||AS−M||2+a||∇S||1(19)

where a, the regularisation parameter, ∇is the gradient and

||.||1is the l1−norm. Then the problem to be solved is the

constrained optimisation problem, as shown in eq. (20).

xa=argminS(α||∇S||1)such that

||AS−M||2<p,(20)

where p is determined based on our knowledge of measure-

ment noise. The problem is solved by the Split Bregman

based TV algorithm [33], [34]. Then, carefully choosing the

regularisation parameters, we optimise the image by deleting

undesired artefacts.

B. Reﬂection Reconstruction

A time-of-ﬂight reﬂection method has been applied to

reconstruct the captured reﬂected pulse travel-times. The

objective of this method is to locate the reﬂected pulses, which

lie between the interaction of the object’s boundaries with

the medium. Figure 5 (a) presents a geometric representation

of the sensors and directly transmitted and reﬂected waves.

Tx1 emits a tone burst pulse while Rx1, Rx2, Rx14, and

Rx15 record the reﬂection signals. In this case, a relevant

algorithm is developed to connect every Tx with its four Rx

points. For instance, in Tx1-Rx1, the algorithm connects the

two points, ﬁnds the mid-point P of the line, then connects P

to the centre of the circle (centre of the circular object); the

intersection point C is the estimated reﬂection point. Using

the coordinates of C, one can compute the travelling distance

of the pulse and subsequently the reﬂected TOF data. This

method comprises the reﬂection forward problem and can be

used to calculate simulated reﬂection TOF data.

To reconstruct the acoustic proﬁle of the medium using

captured reﬂected TOF data, a reﬂection reconstruction

model based on an ellipse intersecting algorithm was used

[35], [36], [37], [38]. If transmitter and receiver are different,

the back-projection is an ellipsoidal locus with the ellipse’s

foci at transducer positions. The image is reconstructed by

drawing arcs of an ellipse along the reﬂection path. Input TOF

values are translated to the pulse’s travelled distance by using

the prior information of the sensors’ coordinates, as shown in

eq. (21).

d=s0TOF

ref l

where d =dT×1−C+dR×1−C(21)

where dR×1−Cand dT×1−Cdenote the axial distance between

the reﬂection point and the receiver and between the reﬂection

point and the transmitter, respectively. TOF

ref l represents

the time of ﬂight and s0is the sound speed in the water.

Superimposing the arcs of ellipses generate an image where

the intersection of these ellipses highlights the boundary of

the circular object. The eq. (16) is used to produce all these

ellipses that can deﬁne the edges of the medium that allow

reﬂection:

AC +CB =2a=d(22)

A and B are two foci of the ellipse, and C is a point located

in the ellipse curve, astands for the long axis length of

the ellipse. A and B represent the transmitter and receiver;

respectively, C stands for a particular point of the target

surface, reﬂecting the ultrasound wave. The value of acan be

easily calculated, using also the ellipse equation of eq. (23).

a2=b2+c2(23)

The value of b and c can also be easily obtained where b is

the short axis length of the ellipse and c is the distance between

focus and the ellipse centre. The distance can be calculated by

the equation:

c=1

2(xr−xR)2+(yr−yR)2(24)

where xr,xR,yr,yRare the transducer coordinates, their

subscripts indicate the transducer mode. When all ellipse

parameters are obtained, a particular ellipse can be drawn in a

determined position and dimension. At last, the target image

can be found by many ellipses that are mutually intersected.

Figure 5 (b) presents ellipses generated by the developed

reﬂection reconstruction program.

IV. NOVEL REFLECTION PULSE PICKING APPROACH

AIDED BY TRANSMISSION RECONSTRUCTION

Dual modality ultrasound imaging, fusing transmission and

reﬂection reconstructions, have been recently researched as

an optimised ultrasound tomography method [9], [16], [39].

Indeed, combining the two different modalities, which use

different full-waveform’s features, can aid reconstructions by

combining complementary information. Despite the good per-

formance of transmission imaging, reﬂection imaging can aid

more towards the improvement of outcomes, especially in

well-characterizing the domain boundaries. Therefore, a robust

algorithm for reﬂection reconstruction needs to exploit the

medium’s boundaries. In practice however, picking algorithms

struggle to locate the correct reﬂection pulse many times,

and noise is added to the measurements. This is a common

issue of all ultrasonic tomographic instruments, caused by

the back-scattering effect [16]. Therefore, a reﬂection TOF

picking method guided by transmission image was developed.

20902 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

Fig. 6. (a),(d) True images. (b),(e) Fused transmission images. (c),(f) Segmented Images used as domain in forward reﬂection solver.

(g),(h) Reﬂection simulated data.

The developed method is based on the reﬂection for-

ward solver to produce better-reﬂected TOF values than

those coming from picking the reﬂected pulses, described

in section IV.

A. Image Segmentation & Reﬂection Forward Solver

An acoustic proﬁle domain of the ROI is created by trans-

mission image and used from the reﬂection forward solver

to produce the simulated reﬂected data, TOF

sim

rf .Inthat

case, the fused transmission image is used as described below

in Section V. A segmentation approach has been developed

to deﬁne the acoustic distribution on the fused transmission

image. First, the global Otsu’s thresholding method was

used [40]. Then, a labelling method calculates the number

of noncontinuous detected regions [42]. For every region,

the centre of mass and its boundaries’ shape are calculated.

The solver uses the regions’ information to ﬁnd the closest

region to the corresponding pair of transmitter-receiver, to

avoid multiple reﬂection signals.

This aims to improve accuracy, notably considering the

impact of more complex distributions on the reﬂective forward

solver. Then, as shown in Figure 5(a), a line is computed from

the middle point of sensors to the region’s centre of mass. The

intersection of this line with the region’s boundaries forms

the intersection point C. Using eq. (22), given that A and

B represent the two sensors, calculates the wave’s travelled

distance d. Finally, the simulated reﬂected data, TOF

sim

rf ,are

generated by eq. (25).

TOF

sim

rf =s0

d(25)

Figure 6(a-f) present the true, transmission reconstructed

and segmentation images of two single and double inclusions

cases. Figure 6(g-h) present the recorded, TOF

obs

rf ,andthe

simulated, TOF

sim

rf , reﬂection data, using the reﬂection for-

ward solver. In the reﬂected data, there are lower-values region

for every positioned object, coming from the time-delays that

they introduce. In Figure 6(g), a single low-values region

can be noticed, while Figure 6(h) two of them. Furthermore,

a clear resemblance between simulated and observed data

can be noticed, showing the good performance of the reﬂec-

tion forward solver. The number of reﬂection points reduces

signiﬁcantly in multiple inclusions cases, which is clearly a

disadvantage of the method. However, a ring setup with more

sensors would increase the spatial resolution and the accuracy

of the method.

B. Optimised Reﬂection TOF Picking Method

This method incorporates an optimised travel-time picking

method. It picks a signal’s value, using as an a priori infor-

mation the simulated reﬂection data, TOF

sim

rf , to optimise the

observed data, TOF

obs

rf . At ﬁrst, an appropriate threshold was

set, and all the potential reﬂected peaks above that threshold

were stored, (Pm,n).

In this way, no peak is being excluded. Then, the algorithm

calculates the “closest” peak to the corresponding simulated

TOF value, TOF

sim

rf . The calculated point represents the

predicted travel-time value, TOFpred

rf , eq. (26).

TOFpred

rf =P|min

nεN(P−TOF

sim

rf )(26)

where Pis a m=[1,...,M]by n=[1,...,N]matrix,

containing the peaks of the full-waveforms; Mis the number

of the measurements and Nthe number of peaks for each

measurement. Usually, the generated TOFpred

rf data were opti-

mised compared to the observed ones, TOF

obs

rf . Nevertheless,

the data were occasionally highly affected by the simulated

data, leading to overﬁtting cases. Therefore, a polynomial least

square ﬁtting model was used between the observed reﬂection

data and the “optimal” data, to smooth the data. A conver-

gence criterion of the average percentage of similarity of the

observed data, TOF

obs

rf ,and the predicted data, TOFpred

rf ,was

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20903

Fig. 7. (a) Subtracted full-waveform signal with Peaks, observed, simulated, optimised data depicted. The corresponding pulse is represented

zoomed. (b) The plot of experimental, simulated, and “optimal” data of all waveforms.

also used. Eq. (27) describes this criterion

C=TOF

obs

rf −TOFpr ed

rf

TOF

obs

rf

with C <0.1 (27)

If the convergence criterion is true, then the optimisation is

applied to the data. Such a technique is widespread in concepts

of full-waveform inversion [43], [44], [45] The simulated data

are used to optimise the already captured data, solving the cost

function in eq. (27).

ε=1

2

M

m=1TOF

obs

rf (m)−TOFpred

rf (m)2

(28)

ξopt =arg min||ε(ξ)|| (29)

where ξrepresents the acoustical property distribution to be

recovered, and ε(ξ) is the error functional. This ﬁnal step

produces the last reﬂection data TOF

opt

rf , which is optimal.

Figure 7(a) shows the difference data signal, computed by

eq. (12), and a zoomed window of the reﬂection pulse. The

black dot represents the travel-time of the reﬂected pulse com-

ing from the straightforward method of section II.B (TOF

obs

rf ).

The green dot is the simulated travel-time, TOF

sim

rf .

The red dots represent all the captured peaks of the

waveform above the threshold value. Finally, the blue circle

represents the “optimal” reﬂected travel-time value that results

from the “minimal distances method”, TOFopt

rf . The effect of

the method is noticeable in Figure 7(b). The blue function

represents the optimised data, there red function the simulated

data and the black function the recorded data. The black

function missed calculating a correct TOF value in few cases,

where zeros are observed. However, the zeros have been

replaced with estimated values, in the optimised data, with

the aid of simulated data, concluding in the well-behaviour

of the optimization. These travel-time values are assumed to

be reﬁned when compared to the observed data. This method

ensures a signiﬁcantly richer dataset than the straightforward

way of picking the reﬂected pulses. This method is named as

“minimum distances method”.

Figure 8 presents the reﬂection data optimisation process

for two single and double inclusions cases. Speciﬁcally,

Figure 8 (a),(d), show the true images, while Figures 8 (b)-(e)

present reﬂection reconstructions using the straightforward

“traditional” way for reﬂection TOF picking data from

and “proposed” reﬂection TOF picking methods. Finally,

ﬁgure 8 (f) presents the reﬂection TOF of observed, simulated

20904 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

Fig. 8. (a) True images of two tested conﬁgurations. (b),(d) Reﬂection images generated by the “traditional” method. (c),(e) Reﬂection images

generated by the “proposed” method. (f) Plots of observed, simulated and optimal reﬂection data.

and optimal data. The observed data come from the “tradi-

tional” method, the simulated data come from the reﬂection

forward solver, and the optimal one come from the “proposed”

method. Both two stages of “minimal distances” and “least-

square ﬁtting” were plotted for optimal data. In the ﬁrst case,

the two functions coming from optimal data are the same,

which means that the convergence criterion is not valid. On the

contrary, in the second case, these functions differ. In both

cases, the effect of the simulated data for the computation

of optimal data is pronounced. Optimal data seem to be a

processed function dragged by the optimised ones.

The novel reﬂection data picking algorithm consists of all

the previously described methods and aims to provide optimal

reﬂection TOF data. The proposed algorithm ﬁts the optimal

data to the captured ones with respect to a priori information

of the simulated data, as shown in Figure 7. It can be sum-

marised into the following steps: (i) execution of transmission

reconstruction; (ii) segmentation using Otsu’s threshold; (iii)

execution of reﬂection forward solver to produce simulated

data; (iv) calculation of optimised reﬂection TOF by “mini-

mum distances”; (v) checking the convergence criterion and

if true, ﬁnding the misﬁt data TOF

opt

rf by solving the cost

function for TOF

obs

rf and TOFpr ed

rf . The whole method is

displayed in Algorithm 1.

V. T RIPLE MODALITY

The developed triple modality approach consists of three

sets of information, and the fusing method is described by a

speciﬁc pipeline which is depicted in Figure 9.First,theTOF

and AA transmission images are fused using a “wavelet trans-

form method” [46]. This intensity-based method was chosen

due to the similarities of the TOF and AA images as they both

come from transmission reconstruction. The two images were

normalised, before being merged, to have the same scale. The

result contains both the information coming from TOF and

Algorithm 1 Novel Reﬂection Signal Picking

1:Compute TOF

obs

rfl by using the “traditional” TOF

picking method.

2:Produce fused AA-TOF transmission image.

3:Create the acoustic domain by segmentation of fused

transmission image.

4:Compute TOFsim

rﬂ by solving the reﬂection forward

problem.

5:Detect all waveform’s peaks, P,above a

minimum threshold.

6:Calculate TOFpr ed

relf by locating the shortest

distance peaks from TOF

sim

rfl by using the minimum

distance method.

7:Calculate the average percentage of similarity, C.

8:If (C<0.1)

9:Solve the cost function of TOFpr ed

relf and

TOF

obs

refl.

10 : end

AA data subsequently proved as an optimised reconstructed

image. Then, the transmission image is combined with the

reﬂection image.

Because of the different nature of the transmission and

reﬂection images, a different fusion method was used. The

transmission images usually contain high “regions” in the

position where the objects are located, due to signiﬁcantly

high TOF-delays and amplitude attenuation due to object

introduction. On the other hand, the reﬂection image has

almost zero values to the locations of the objects, as all the

reﬂections are encountered in their boundaries and, according

to the ellipse algorithm, no ellipse interaction is happening

within the object. Therefore, a method that accounts for these

characteristics, by superimposing regions of the images, was

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20905

Fig. 9. Image fusion algorithm for triple modality USCT.

applied to fuse the transmission and reﬂection images. This is

described in eq. (30) method.

TM

i,j=Ti,jwhere Ri,j>0

0where Ri,j=0where,

i=[1,...,32]j=[1,...,32](30)

where Ti,jis the transmitted image, Ri,jis the reﬂection image

and TM

i,jis the triple-modality image; i,jrepresents the

rows and columns of the image that is 32 by 32.

VI. RESULTS AND ANALYSIS

The system was experimentally validated by applying sev-

eral single and multiple static inclusions tests with different

shapes and sizes. All the inclusions are made from plas-

tic (PVC) and are not compact; thus, the sound can only

be diffracted and reﬂected. Circular inclusions of 1cm, 2cm

and 3cm of diameter, square inclusion of 4cm side-length and

an equilateral triangle inclusion of 3cm were used to provide

various testing cases. These tests aimed to simulate dispersed

phases of a liquid mixture existing in industrial processes. The

change in the structural phase aims to simulate the change

happening during a crystallization or fermentation process.

Figure 10 presents results using different reconstruction meth-

ods of 10 different experimental conﬁgurations. Among the

reconstructed methods are TOF, AA, fused transmission, “tra-

ditional” reﬂection, “proposed” reﬂection and triple-modality

reconstructions. It is evident that transmission mode can be

used in object localisation, notably upon multiple inclusions.

On the other hand, reﬂection is signiﬁcantly better in detecting

the boundaries of the domain accurately. However, reﬂection

has a clear disadvantage in reconstructing regions that lie

between two objects. Therefore, in those cases, the trans-

mission mode aid more the triple-modality results than the

reﬂection.

To quantify the imaging quality of the proposed reconstruc-

tion approach, Correlation Coefﬁcient (CC) and Root Mean

Square Error (RMSE) were calculated, eq. (31) and eq. (32)

respectively. The segmentation method described in section IV.

Normalisation was applied to all the images to turn them into

a uniform form, aiming at quantitative similarities.

CC =

N

n=1

(σn−δ)σ∗

n−δ∗

N

n=1

(σn−δ)2N

n=1σ∗

n−δ∗2

(31)

RMSE =

N

n=1

(σ−σ∗)2

N(32)

where σis the calculated acoustic distribution by the recon-

struction algorithms and σ∗is the real one (true image), σn

and σ∗

nare nth elements of σand σ∗respectively, δand

δ∗are the mean values of σand σ∗

n, respectively. Figure 11

shows the CC and the RMSE of TOF, AA, fused trans-

mission, “traditional” reﬂection, “proposed” reﬂection and

triple-modality reconstructions. In almost all the cases, the

proposed algorithm proved to be more efﬁcient by acquiring

the overall highest CC and lowest RMSE value. CC was

higher and RMSE was lower in single inclusion cases com-

paring with the multiple inclusions cases due to the medium’s

complexity.

Concluding in the supremacy of the proposed reﬂection

algorithm, a triple modality approach was applied using TOF,

AA and reﬂection images. The MRSE of triple-mode images

is generally smaller, while CC is more prominent than all

the other methods. TOF and AA images converted to binary

form using a high threshold to segment the inclusions. Then

transmission and reﬂection images were fused in binary for-

mat. Regarding CC, in almost all cases, the ﬁnal image is

closer to the real geometry. The signiﬁcant aid of the triple

modality method can clearly be noticed as, in all cases, the

TOF, AA and reﬂection reconstructions’ accuracy differs, but

the triple-modality reconstruction is always higher.

The qualitative difference can be noticed in multiple inclu-

sions cases, as they consist of more complex nature. The

quantitative analysis indicates that the multi-modality method

provides more accurate reconstruction on both the area and

the location of the objects than a single modality of either

transmission or reﬂection mode.

To further test the performance of the proposed system and

the multi-modality approach, different setups of water/sucrose

solutions were used. These experimental scenarios simulate

miscible liquids and multi-phase ﬂow in industrial tanks and

pipes scenarios. A plastic cup of 1mm length, ﬁlled with the

solutions, was used as a static inclusion. The cup’s sound

transmissivity has been tested and approved. The system

acquired transmitted signals passed through the water/sucrose

solutions and reﬂected signals came from the cup’s surface.

Figure 12 displays experimental photos and reconstructions

of three different water/sucrose cases. Figure 12(a) shows

a 60.7% solution positioned in the centre, Figure 12(b) a

combination of 50% and 42.8%, and Figure 12(c) acom-

bination of 20% and 42.8%. The TOF mapping distin-

guished well between different concentrations in multiple

inclusions experiments. Furthermore, the tests TOF-delays

20906 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

Fig. 10. Image reconstructions of the Triple-Modality USCT.

scale follows the overall concentration increase, as shown

in Figure 12(b),(c). Additionally, six different single inclu-

sions cases with water/sucrose concentrations of 20%, 33%,

42.8%, 50%, 56.7% and 60.7% were reconstructed. Tab le I

shows the TOF delays caused due to the existence of the

solution. The presented TOF-delays, calculated as the object

was segmented and its mean value was calculated. Difference

imaging was used by subtracting the background from the

full TOF measurements. Since the sound velocity of the

concentrations is higher than the medium (water in 20◦C),

the produced difference data were negative. Small positive

values were caused by noise and therefore were neglected.

TOF delays showed good response, as they form an ascending

function over the increasing concentration of the solutions.

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20907

Fig. 11. (a) CC and (b) RMSE of several different reconstruction methods.

Fig. 12. Experimental photos and reconstructions of water/sucrose solutions of (a) 60.7% in the centre (b) 50% down-left and 42.8% up-right

(c) 20% down-left and 42.8% up-right.

TABLE I

TOF DEL AYS FROM THE EXPERIMENTAL PROCESS WITH

WATE R /SUCROSE SOLUTIONS

The solutions experiment proved efﬁcient in distinguishing

between low changes of concentration, showing the high

quantitative resolution that the system can provide.

VII. CONCLUSION

This work presents the advantages of triple-modality ultra-

sound tomographic imaging for real industrial processes.

Accurate results of multiple solid objects and various concen-

trated solutions could signiﬁcantly beneﬁt complex industrial

processes of two-phase media and multi-material interactions.

Reﬂection and transmission reconstruction methods can work

in a complementary way and provide optimal results. More-

over, acoustic attenuation measurements were proven effective,

facilitating the transmission of TOF reconstructions, especially

in more inhomogeneous media. So, there is great potential in

the combination of two types of transmission mode tomogra-

phy. This kind of rich full-waveform tomography proved to

work well in exploiting full-waveform information. Without

introducing heavy computational algorithms, it can beneﬁt

20908 IEEE SENSORS JOURNAL, VOL. 21, NO. 18, SEPTEMBER 15, 2021

from combining different reconstructions and at the same time

perform at a high temporal frequency. Therefore, it comprises

a potential solution to many industrial processes that need

inspection over time and a good temporal resolution.

The developed methods provided good qualitative and quan-

titative performance regarding the quality image measurements

and the correlation of TOF-delays with various solutions.

Static experiments showed good system performance in dis-

tinguishing objects of different sizes and shapes in single and

multiple objects. The solutions used in the experiments showed

that the triple-modality imaging could also use the TOF scale

to characterise small changes in the density of biphasic media,

which is a signiﬁcant addition to the system. The results of

this research show that this rich full-waveform USCT can

aid industrial processes and may be used for stirred tanks

chemical processes. Given the existence of biphasic media,

which include integration of liquid solutions and suspensions,

the added value of the multi-modality full-waveform system

will become apparent in our future studies. In case of cup

with multiple percentage solution both a reﬂection image due

to the cup and a quantitative transmission image due to particle

concentration can be produced.

ACKNOWLEDGMENT

The authors would like to thank Dr. Dana Beiki (University

of Bath) for proofreading the manuscript.

REFERENCES

[1] M. Yang, H. I. Schlaberg, B. S. Hoyle, M. S. Beck, and C. Lenn, “Real-

time ultrasound process tomography for two-phase ﬂow imaging using

a reduced number of transducers,” IEEE Trans. Ultrason., Ferroelectr.,

Freq. Control, vol. 46, no. 3, pp. 492–501, May 1999.

[2] T. Rymarczyk and J. Sikora, “Applying industrial tomography to control

and optimization ﬂow systems,” Open Phys., vol. 16, no. 1, pp. 332–345,

Jun. 2018.

[3] L. Xu, Y. Han, L.-A. Xu, and J. Yang, “Application of ultrasonic

tomography to monitoring gas/liquid ﬂow,” Chem. Eng. Sci., vol. 52,

no. 13, pp. 2171–2183, 1997.

[4] C. Tan, X. Li, H. Liu, and F. Dong, “An ultrasonic transmission/reﬂection

tomography system for industrial multiphase ﬂow imaging,” IEEE Trans.

Ind. Electron., vol. 66, no. 12, pp. 9539–9548, Dec. 2019.

[5] R. Seidl, “Full waveform inversion for ultrasonic nondestructive testing,”

Ph.D. dissertation, Technische Univ. München, Munich, Germany, 2018.

[6] N. N. Kishore, I. Sridhar, and N. G. R. Iyengar, “Finite element

modelling of the scattering of ultrasonic waves by isolated ﬂaws,” NDT

&EInt., vol. 33, no. 5, pp. 297–305, Jul. 2000.

[7] T. Rymarczyk, E. Kozłowski, G. Kłosowski, and K. Niderla, “Logis-

tic regression for machine learning in process tomography,” Sensors,

vol. 19, no. 15, p. 3400, 2019.

[8] H. I. Schlaberg, F. J. W. Podd, and B. S. Holye, “Ultrasound process

tomography system for hydrocyclones,” Ultrasonics, vol. 38, nos. 1–8,

pp. 813–816, Mar. 2000.

[9] J. Gu, H. Yang, F. Fan, and M. Su, “A transmission and reﬂection

coupled ultrasonic process tomography based on cylindrical miniaturized

transducers using PVDF ﬁlms,” J. Instrum., vol. 12, no. 12, p. P12026,

Dec. 2017.

[10] Q. Su, C. Tan, and F. Dong, “Measurement of phase fraction in

oil-water two-phase ﬂow using ultrasound attenuation method,” in

Proc. Chin. Autom. Congr. (CAC), Nov. 2015, pp. 1890–1895, doi:

10.1109/CAC.2015.7382812.

[11] X. Duan, P. Koulountzios, and M. Soleimani, “Dual modality EIT-UTT

for water dominate three-phase material imaging,” IEEE Access,vol.8,

pp. 14523–14530, 2020.

[12] M. Yang, H. I. Schlaberg, B. S. Hoyle, M. S. Beck, and C. Lenn,

“Real-time ultrasound process tomography for two-phase ﬂow imaging

using a reduced number of transducers,” IEEE Trans. Ultrason., Fer-

roelectr., Freq. Control, vol. 46, no. 3, pp. 492–501 , May 1999, doi:

10.1109/58.764834.

[13] G. T. Bolton and K. M. Primrose, “An overview of electrical tomo-

graphic measurements in pharmaceutical and related application areas,”

AAPS PharmSciTech, vol. 6, no. 2, pp. E137–E143, Jun. 2005.

[14] H. Liu, C. Tan, S. Zhao, and F. Dong, “Nonlinear ultrasonic trans-

missive tomography for low-contrast biphasic medium imaging using

continuous-wave excitation,” IEEE Trans. Ind. Electron., vol. 67, no. 10,

pp. 8878–8888, Oct. 2020.

[15] P. Koulountzios, T. Rymarczyk, and M. Soleimani, “A quantitative

ultrasonic travel-time tomography to investigate liquid elaborations in

industrial processes,” Sensors, vol. 19, no. 23, p. 5117, Nov. 2019.

[16] L. Huang et al., “Breast ultrasound waveform tomography: Using

both transmission and reﬂection data, and numerical virtual point

sources,” Proc. SPIE, vol. 9040, Mar. 2014, Art. no. 90400T, doi:

10.1117/12.2043136.

[17] G. Liang, S. Ren, and F. Dong, “Ultrasound guided electrical impedance

tomography for 2D free-interface reconstruction,” Meas. Sci. Technol.,

vol. 28, no. 7, Jul. 2017, Art. no. 074003.

[18] C. Li, L. Huang, N. Duric, H. Zhang, and C. Rowe, “An improved

automatic time-of-ﬂight picker for medical ultrasound tomogra-

phy,” Ultrasonics, vol. 49, no. 1, pp. 61–72, Jan. 2009, doi:

10.1016/j.ultras.2008.05.005.

[19] Y. Bao and J. Jia, “Improved time-of-ﬂight estimation method for

acoustic tomography system,” IEEE Trans. Instrum. Meas., vol. 69,

no. 4, pp. 974–984, Apr. 2020.

[20] L. Espinosa, J. Bacca, F. Prieto, P. Lasaygues, and L. Brancheriau,

“Accuracy on the time-of-ﬂight estimation for ultrasonic waves applied

to non-destructive evaluation of standing trees: A comparative exper-

imental study,” Acta Acustica United With Acustica, vol. 104, no. 3,

pp. 429–439, May 2018.

[21] Y. S. Huang, Y. P. Huang, K. N. Huang, and M. S. Young, “An accurate

air temperature measurement system based on an envelope pulsed

ultrasonic time-of-ﬂight technique,” Rev. Sci. Instrum., vol. 78, no. 11,

Nov. 2007, Art. no. 115102.

[22] H. I. Schlaberg, M. Yang, and B. S. Hoyle, “Ultrasound reﬂection

tomography for industrial processes,” Ultrasonics, vol. 36, nos. 1–5,

pp. 297–303, 1998.

[23] A. Macchi, H. Bi, J. R. Grace, C. A. McKnight, and L. Hackman, “Effect

of gas density on the hydrodynamics of bubble columns and three-phase

ﬂuidized beds,” Can. J. Chem. Eng., vol. 81, nos. 3–4, pp. 846–852,

May 2008.

[24] M. Pérez-Liva, J. L. Herraiz, J. M. Udías, E. Miller, B. T. Cox, and

B. E. Treeby, “Time domain reconstruction of sound speed and attenu-

ation in ultrasound computed tomography using full wave inversiona,”

J. Acoust. Soc. Amer., vol. 141, no. 3, pp. 1595–1604, Mar. 2017.

[25] F. E. Grubbs, “Procedures for detecting outlying observations in sam-

ples,” Technometrics, vol. 11, no. 1, pp. 1–21, 1969.

[26] C. Li, N. Duric, and L. Huang, “Comparison of ultrasound attenua-

tion tomography methods for breast imaging,” Proc. SPIE, vol. 6920,

Mar. 2008, Art. no. 692015.

[27] P. Koulountzios, T. Rymarczyk, and M. Soleimani, “Ultrasonic time-

of-ﬂight computed tomography for investigation of batch crystallisation

processes,” Sensors, vol. 21, no. 2, pp. 1–19, Jan. 2021.

[28] M. L. Buursink, T. C. Johnson, P. S. Routh, and M. D. Knoll, “Cross-

hole radar velocity tomography with ﬁnite-frequency Fresnel volume

sensitivities,” Geophys. J. Int., vol. 172, no. 1, pp. 1–17, Jan. 2008.

[29] J. M. Jensen, B. H. Jacobsen, and J. Christensen-Dalsgaard, “Sensitivity

kernels for time-distance inversion,” Sol. Phys., vol. 192, nos. 1–2,

pp. 231–239, 2000.

[30] T. M. Hansen, K. S. Cordua, M. C. Looms, and K. Mosegaard, “SIPPI:

A MATLAB toolbox for sampling the solution to inverse problems

with complex prior information: Part 2—Application to crosshole GPR

tomography,” Comput. Geosci., vol. 52, pp. 481–492, Mar. 2013.

[31] L. I. Rudin, S. Osher, and E. Fatemi, “Nonlinear total variation based

noise removal algorithms,” Phys. D, Nonlinear Phenomena, vol. 60,

nos. 1–4, pp. 259–268, Nov. 1992.

[32] T. Goldstein and S. Osher, “The split Bregman method for

L1-regularized problems,” SIAM J. Imag. Sci., vol. 2, no. 2, pp. 323–343,

2009.

[33] F. Li, J. F. P. J. Abascal, M. Desco, and M. Soleimani, “Total variation

regularization with split Bregman-based method in magnetic induction

tomography using experimental data,” IEEE Sensors J., vol. 17, no. 4,

pp. 976–985, Feb. 2017.

[34] J. F. P.-J. Abascal et al., “Fluorescence diffuse optical tomography using

the split Bregman method,” Med. Phys., vol. 38, no. 11, pp. 6275–6284,

Oct. 2011.

KOULOUNTZIOS

et al.

: TRIPLE-MODALITY USCT BASED ON FULL-WAVEFORM DATA 20909

[35] L. Yang, C. Xu, and X. Guo, “Ring-shaped array ultrasound imaging

using ellipse algorithm,” Comput. Model. New Technol., vol. 18, no. 12,

pp. 315–320, 2014.

[36] J. Tu and B. Yang, “The research of ellipse parameter ﬁtting algorithm

of ultrasonic imaging logging in the casing hole,” Appl. Math.,vol.5,

no. 8, pp. 1317–1321, 2014.

[37] L. Yang, Q. Pan, C. Xu, X. Guo, and K. Peng, “Immersion ultrasonic

reﬂection tomography by annular array system,” in Proc. Far East Forum

Nondestruct. Eval./Test., New Technol. Appl. (FENDT), Jun. 2013,

pp. 82–89.

[38] F. Wiegand and B. S. Hoyle, “Simulations for parallel processing of

ultrasound reﬂection-mode tomography with applications to two-phase

ﬂow measurement,” IEEE Trans. Ultrason., Ferroelectr., Freq. Control,

vol. 36, no. 6, pp. 652–660, Nov. 1989.

[39] L. Huang et al., “Breast ultrasound tomography with two parallel

transducer arrays,” Proc. SPIE, vol. 9783, Mar. 2016, Art. no. 97830C.

[40] L. Brancheriau, P. Lasaygues, E. Debieu, and J. P. Lefebvre, “Ultrasonic

tomography of green wood using a non-parametric imaging algorithm

with reﬂected waves,” Ann. Forest Sci., vol. 65, no. 7, p. 712, Jan. 2008.

[41] N. Otsu, “A threshold selection method from gray-level histograms,”

IEEE Trans. Syst., Man, Cybern., vol. SMC-9, no. 1, pp. 62–66,

Jan. 1979.

[42] R. M. Haralick and L. G. Shapiro, Computer and Robot Vision,vol.1.

Reading, MA, USA: Addison-Wesley, 1992, pp. 28–48.

[43] M. Perez-Liva, J. L. Herraiz, J. M. Udias, B. T. Cox, and B. E. Treeby,

“Full-wave attenuation reconstruction in the time domain for ultrasound

computed tomography,” in Proc. IEEE 13th Int. Symp. Biomed. Imag.

(ISBI), Apr. 2016, pp. 710–713.

[44] T. Alkhalifah and Y. Choi, “From tomography to full-waveform inver-

sion with a single objective function,” Geophysics, vol. 79, no. 2,

pp. R55–R61, Mar. 2014.

[45] R. Djebbi and T. Alkhalifah, “Traveltime sensitivity kernels for wave

equation tomography using the unwrapped phase,” Geophys. J. Int.,

vol. 197, no. 2, pp. 975–986, May 2014.

[46] H. Li and B. S. Manjunath, “Multisensor image fusion using the wavelet

transform,” Graph. Models Image Process., vol. 57, no. 3, pp. 235–245,

1995.

Panagiotis Koulountzios received the M.Eng.

degree in electrical and computer engineering

(ECE) from the Technical University of Crete

(TUC). He is currently pursuing the Ph.D. degree

with the Engineering Tomography Laboratory

(ETL), Electronic and Electrical Engineering

Department, University of Bath (UoB). Since

April 2018, he has been with ETL as a Research

Associate for Tomocon, an EU Training Network

Project. His Ph.D. is in ultrasound tomography for

industrial process applications.

Tomasz Rymarczyk (Member, IEEE) is the

Director of the Research and Development Cen-

tre, Netrix S.A., and also the Director of the

Institute of Computer Science and Innovative

Technologies, University of Economics and Inno-

vation, Lublin, Poland. He worked in many

companies and institutes developing innovative

projects and managing teams of employees. His

research interests include application of non-

invasive imaging techniques, electrical tomogra-

phy, ultrasound tomography, radio tomography,

image reconstruction, numerical modeling, image processing and analy-

sis, process tomography, software engineering, knowledge engineering,

artiﬁcial intelligence, and computer measurement systems.

Manuchehr Soleimani received the B.Sc.

degree in electrical engineering and the M.Sc.

degree in biomedical engineering, and the Ph.D.

degree in inverse problems and electromagnetic

tomography from The University of Manchester,

Manchester, U.K., in 2005. From 2005 to 2007,

he was a Research Associate with the School

of Materials, The University of Manchester.

He joined the Department of Electronic and Elec-

trical Engineering, University of Bath, Bath, U.K.,

in 2007, where he was a Research Associate and

became a Lecturer in 2008, a Senior Lecturer in 2013, a Reader in 2015,

and a Full Professor in 2016. He founded the Engineering Tomography

Laboratory (ETL), University of Bath, in 2011, working on various areas

of tomographic imaging, in particular multimodality tomographic imaging.

He has authored or coauthored well over 300 publications in the ﬁeld.