Nonlinear Viscoelastic Constitutive Model For Bovine Liver Tissue

Biomechanics and Modeling in Mechanobiology (2020) 19:1641–1662
https://doi.org/10.1007/s10237-020-01297-5
ORIGINAL PAPER
Nonlinear viscoelastic constitutive model for bovine liver tissue

Adela Capilnasiu1 · Lynne Bilston4,5 · Ralph Sinkus1,2 · David Nordsletten1,3
Received: 1 July 2019 / Accepted: 21 January 2020 / Published online: 10 February 2020
© The Author(s) 2020
Abstract
Soft tissue mechanical characterisation is important in many areas of medical research. Examples span from surgery training,
device design and testing, sudden injury and disease diagnosis. The liver is of particular interest, as it is the most commonly
injured organ in frontal and side motor vehicle crashes, and also assessed for inflammation and fibrosis in chronic liver dis-
eases. Hence, an extensive rheological characterisation of liver tissue would contribute to advancements in these areas, which
are dependent upon underlying biomechanical models. The aim of this paper is to define a liver constitutive equation that is
able to characterise the nonlinear viscoelastic behaviour of liver tissue under a range of deformations and frequencies. The
tissue response to large amplitude oscillatory shear (1–50%) under varying preloads (1–20%) and frequencies (0.5–2 Hz) is
modelled using viscoelastic-adapted forms of the Mooney–Rivlin, Ogden and exponential models. These models are fit to the
data using classical or modified objective norms. The results show that all three models are suitable for capturing the initial
nonlinear regime, with the latter two being capable of capturing, simultaneously, the whole deformation range tested. The
work presented here provides a comprehensive analysis across several material models and norms, leading to an identifiable
constitutive equation that describes the nonlinear viscoelastic behaviour of the liver.
Keywords Liver rheology · Biomechanics · Nonlinear mechanics · Viscoelasticity
1 Introduction in order to assess the presence of disease (Fovargue et al.

2018). In vivo magnetic resonance elastography (MRE) has
Biomechanical characterisation of tissues is essential shown that the existence of liver inflammation and fibrosis
in medical research. New surgery techniques, implants gives higher stiffness measurements (Huwart et al. 2006;
or devices are being tested in silico, in vitro and in vivo Sinkus et al. 2018). However, coexisting diseases (Mueller
(O’Toole et al. 1995; Marescaux et al. 1998; Rosen et al. et al. 2010) and the bias introduced by large deformations
2008; Clin et al. 2010; Gonzalez-Blohm et al. 2015). In on elastography measurements (Capilnasiu et al. 2019) can
these tests, it is critical to know the liver’s response to a further complicate the diagnosis. Comprehensive liver mod-
range of factors, such as puncturing, cutting, deformations els could also benefit other research areas like transportation
and displacements. For diagnosis purposes, elastography is safety (Viano et al. 1989; Yoganandan et al. 2000), where
a technique that depends on the underlying tissue properties biomechanical tests for abdominal injury tolerance limits
provide essential information in developing safer vehicles
(Kemper et al. 2010). With the liver being the most fre-
* Adela Capilnasiu
[email protected] quently injured organ in frontal and side impacts (Yoga-
nandan et al. 2000), a finite element (FE) model could be
1
Division of Biomedical Engineering and Imaging Sciences, used for prediction purposes, provided that it incorporates
King’s College London, London, UK local and global liver tissue response to mechanical testing.
2
Inserm U1148, LVTS, University Paris Diderot, University Hence, improved knowledge of the nonlinear viscoelastic
Paris 13, 75018 Paris, France behaviour of the liver is needed.
3
Department of Biomedical Engineering and Cardiac Surgery, Over the past decades, a range of rheological tests have
University of Michigan, Ann Arbor, USA been employed to characterise liver tissue, the most com-
4
Prince of Wales Clinical School, University of New South mon being uniaxial deformation (either as small sample
Wales, Sydney, Australia loading or indentation on the full organ) and shearing. Both
5
Neuroscience Research Australia, Sydney, Australia
13
Vol.:(0123456789)
1642 A. Capilnasiu et al.
oscillatory shear and uniaxial deformation tests show that, at contribute to our understanding of separate aspects of the
low strains, the liver exhibits quasi-linearity, with the non- viscoelastic behaviour of the liver, there remains a need for
linear behaviour being exposed at higher strains (Liu and a comprehensive 3D model which can describe the tissue
Bilston 2000; Gao et al. 2010; Tan et al. 2013). Addition- response under various deformation types and frequencies.
ally, loading–unloading tests reveal that hysteresis effects In this paper, we present a comparison of liver constitu-
are taking place (Jordan et al. 2011), with the response being tive models based on the tissue’s response to a range of large
rate dependent (Liu and Bilston 2000; Miller 2000). Multi- deformations and frequencies. A cross testing of uniaxial
frequency soft tissue measurements of the shear modulus preloads (1–20%), shear strains (1–50%) and frequencies
G∗ indicate a fractional-order dependence on the angular (0.5–2 Hz) is considered (Tan et al. 2013), thus emphasis-
frequency in the form of G∗ ∝ 𝜔𝛼 (Holm and Sinkus 2010), ing the rate-dependent, nonlinearly viscoelastic behaviour
with 𝛼 ∈ [0.2, 0.35] [e.g. 𝛼 ≈ 0.23 (Liu and Bilston 2000), of the liver. The testing protocol displays a strain softening
𝛼 ≈ 0.26 (Jordan et al. 2011; Sinkus et al. 2018), 𝛼 ≈ 0.32 effect, which is addressed by proposing a new error norm
(Asbach et al. 2008)]. Other biomechanical properties of the that allows for some degree of flexibility in fitting the linear
liver have also been investigated, such as relaxation (Liu and parameters of the models. This analysis and model-fitting
Bilston 2006; Chatelin et al. 2011) and creep (Wang et al. procedure lead to the identification of simplified constitutive
1992). models, which retain the essential components needed for
Liver tissue rheology measurements have lead to a range characterising the above-mentioned properties of the liver
of biomechanical models. Hyperelasticity is often assumed, exhibited under combined deformation and various fre-
with polynomial, exponential and logarithmic forms being quencies. To the authors’ knowledge, this is one of the first
employed for compression and elongation data (Chui et al. liver studies that investigates combined large uniaxial and
2004; Gao et al. 2010). The general findings indicate that shear loading, at various frequencies, and the first study that
the exponential, logarithmic and power law models offer proposes a three-dimensional nonlinear viscoelastic model
more flexibility in capturing the different regions of the which can capture the large amplitude oscillatory response
stress–strain curves. In order to probe viscoelasticity, cyclic across a range of preloads and frequencies.
deformations or relaxation tests usually need to be investi- In what follows, a brief introduction to kinematics
gated. Some studies employed relaxation (Liu and Bilston (Sect. 2.1) precedes the outlining of the experimental
2006), shear oscillations (Nicolle and Palierne 2015) or design and modelling assumptions (Sect. 2.2). Three differ-
cyclic indentation (Jordan et al. 2011) over a range of fre- ent constitutive models are proposed, which are then fit to
quencies, thus offering a broader picture of the biomechani- the data using a set of error norms that infer different model
cal behaviour of liver. Among these, the K-BKZ model was properties. Throughout Sect. 3, the results of the model-
proposed due to its awareness of the complete past time fitting process are going to be presented grouped by the
history and was validated against small amplitude oscil- norm investigated (Sects. 3.1–3.3), with the three models
latory shear and strain ramp (Nicolle and Palierne 2015). being compared within each subsection. This is followed
Alternatively, viscoelasticity was modelled by introducing by a discussion reviewing the findings and potential future
a Maxwell element. A complex differential model, with applications (Sect. 4).
ten model parameters, was investigated by Liu and Bilston
(2006) against the relaxation behaviour at four strain levels.
Ayyildiz et al. (2015) also proposed a Maxwell-based model 2 Materials and methods
with 13 parameters for capturing the viscoelastic behaviour
of liver at a range of uniaxial preloads, frequencies and strain The aim of this work is to characterise the constitutive
rates. There, large preloads (20%) and shear strains (5%) behaviour of the liver under a range of combined deforma-
were employed simultaneously in the testing protocol. How- tions and frequencies. In order to achieve this, Sect. 2.1
ever, the results focus on the effect of preload, strain rate and outlines the kinematics metrics that are needed throughout
frequency on the normal force and torque response, while this paper (see Taber 2004; Bonet and Wood 2008). Sec-
the combined effect of large preloads and shear strains is tion 2.2 explains the testing protocol and its modelling char-
not addressed. Jordan et al. (2011) considered increasingly acteristics. Three different types of models are proposed in
complex networks of springs and dashpots, arranged both in Sect. 2.3, which are due to be fit to the data using the meth-
series and in parallel, in order to model the liver behaviour ods described in Sect. 2.4.
under cyclic indentation at different strain rates and relaxa-
tion. A power law model considering solid-phase compress- 2.1 Kinematics background
ibility was employed by Perepelyuk et al. (2016), who, at
large preload strains and small oscillatory shear, measured Let the region 𝛺0 ⊂ ℝ3 define a solid body which can be
the storage modulus G′ . While these models significantly deformed in space and time using a displacement field
13
Nonlinear viscoelastic constitutive model for bovine liver tissue 1643
U ∶ 𝛺0 × [0, T] → ℝ3 . A point in the reference domain, t
∫
1 1
X ∈ 𝛺0 , corresponds at time t ∈ [0, T] to a point in the
D𝛼t Sv = 𝜕 S (z) dz, (6)
𝛤 (1 − 𝛼) 0 (t − z)𝛼 t v
physical domain 𝛺t , by the mapping
with 𝛼 = 0 leading to a hyperelastic contribution and 𝛼 = 1
x = U(X, t) + X, x ∈ 𝛺t . (1) leading to a purely viscous contribution in the form 𝜕t Sv . In
order to separate the deviatoric and hydrostatic stress com-
The deformation gradient tensor F relates the reference and
ponents, we introduce the deviatoric operator
physical domains via
A ∶ C −1
𝜕x 𝜕xi Dev[A] = A − C , (7)
F = ∇X x =
𝜕X
, Fij =
𝜕Xj
. (2) 3
which ensures that Dev[A] ∶ C = 0. Having defined the PK2
The volumetric changes between the two states are quan- tensor in the context of constitutive modelling, the Cauchy
tified by the determinant J = det F , with J = 1 implying stress tensor can be related to the PK2 tensor using
incompressibility. From the deformation gradient F , the
right and left Cauchy Green strains are defined as 1
𝝈= FSFT . (8)
J
T T
C = F F, B = FF ,
or, in their isochoric form (Bonet and Wood 2008), here 2.2 Nonlinear viscoelastic characterisation of liver
indicated by the “ ̂ ” symbol, as tissue
T T
F̂ = J −1∕3 F, Ĉ = F̂ F,
̂ B̂ = F̂ F̂ . In this paper, combined loading experiments are used to
Some tensor quantities that remain unchanged under rota- investigate the behaviour of liver tissue. The data presented
tions are the first and second invariants (Bonet and Wood here have been previously published in Tan et al. (2013).
2008), obtained using the double contraction “:” Here, we focus on the large amplitude oscillatory strain
(LAOS) tests and briefly review the protocol.
IA = A ∶ I, IIA = A ∶ A, (3)
where A is a general m × m tensor. For clarity, in index nota- 2.2.1 Sample preparation
tion, this is equivalent to
Fresh healthy bovine liver was collected from an abattoir,
∑ ∑
m m
∑ ∑
m m
with the samples being tested within 6 hours post-mortem.
IA = Aij 𝛿ij , IIA = Aij Aij . (4) During transport, the livers were wrapped in saline-soaked
i=1 j=1 i=1 j=1
gauze and transported on ice in a sealed container. Cylindri-
Constitutive equations can be used to describe a material’s cal samples were cut to approximately 10 mm radius and
behaviour under deformation. Let W(C) denote a strain 3 mm height. During testing, in order to ensure hydration,
energy function which depends on strain metrics, here in the samples were maintained in a 100% humidity chamber,
particular on C . The corresponding second Piola–Kirch- which is a part of the rheometer. The temperature was con-
hoff (PK2) tensor is obtained by taking the derivative of trolled to be 25 °C. For more complete preparatory details,
the strain energy function with respect to C , as S = 2∇C W see the original protocol published in Tan et al. (2013).
(Bonet and Wood 2008). For a viscoelastic material descrip-
tion, fractional-order viscoelastic models have been success- 2.2.2 Rheological experimental design
fully employed in modelling soft tissue behaviour (Kiss et al.
2004). Thus, here let S be defined as the sum of an elastic, Tissues were tested using a rotational rheometer (Kinexus
fractional viscoelastic and hydrostatic part, as Pro KNX 2100, Malvern, United Kingdom), as illustrated
S = Se + D𝛼t Sv + Sp . in Fig. 1. The cylindrical samples were fit in between two
(5)
serrated plates of 20 mm diameter, to avoid slipping. The
The elastic and viscoelastic parts are derivatives of elastic bottom plate was fixed, while the upper plate was vertically
and viscoelastic strain energy functions (i.e. Se = 2∇C We adjusted and oscillated around the cylindrical axis. Torque
and Sv = 2∇C Wv ), and the hydrostatic part is defined as measurements were acquired in combined loading tests, with
Sp = JPC−1, where P is the hydrostatic pressure. The frac- shear strains of 1, 10, 25 and 50% being investigated under
tional-order derivative, as defined by Caputo (1967), is different uniaxial preload levels—1, 10 and 20%, at a strain
rate of 1 Hz. Additionally, shear strains of 1, 10 and 25%,
at a preload of 10%, were investigated at strain rates of 0.5
13
Fig. 1 Illustration of the experi-

mental setup. (Top left) Liver
tissue in the oscillatory rheo-
logical instrument. The plates
were serrated, with the lower
plate being fixed, while the
upper plate could move verti-
cally and rotate around the axis.
(Bottom left) Axis and bounda-
ries defined with respect to the
liver sample. (Right) Example
of angular displacement trace
(sinusoidal) and torque response
(non-sinusoidal) at CS 10%, 𝛾
50%, 1 Hz
Table 1 Testing protocol across frequencies, compression and shear deformed by compression. Let the ratio of the deformed to
deformations undeformed height be 𝜆 = h∕H , which corresponds to each
SS 1% SS 10% SS 25% SS 50% compression strain (CS) level via CS = 1 − 𝜆. Here, ideal
compression is assumed,
√ which leads to the radius being
CS 1% ○ ○ ○ ○ deformed as r = R∕ 𝜆.
⨁ ⨁ ⨁
CS 10% ○
The shear strain, 𝛾 , is defined as the ratio between top
CS 20% ○ ○ ○ ○ plate rotational part of the displacement and inter plate gap,
Legend | 0.5 Hz ○ 1Hz − 2Hz 𝛾 = d∕h, with the rotational displacement depending on the
angular displacement and radius. Having a predefined 𝛾 level,
at frequency f the angular displacement on the top of the sam-
and 2 Hz. Table 1 summarises the loading, shearing and ple is given by
frequency protocols employed.
sin(2𝜋ft)h𝛾
A total of 18 tests were carried out—12 at 1 Hz and 3 at 𝜓(t) = , (9)
0.5 and 2 Hz, respectively. For each test, the data were aver- r
aged between at least four liver samples. For each test, the with the maximum angular displacement being reached
first eight cycles were used for preconditioning purposes, at 𝜓 = h𝛾∕r . Let 𝛹 (t, X3 ) define the angular displacement
with the data being recorded from the ninth cycle. Precon- throughout the sample, as
ditioning was carried out sequentially and not directly at
the maximum deformation level in order to avoid damaging 𝜓(t)
𝛹 (t, X3 ) = X.
tissue (e.g. eight preconditioning cycles at shear strain 1% H 3
and then three data cycles, followed by eight preconditioning Hence, the compression and shearing lead to a body defor-
cycles at shear strain 10% and then three data cycles, etc). mation (Taber 2004) defined by
Originally, two more shear strain levels were acquired—80
and 100% (Tan et al. 2013), but these data were excluded ⎡ X1 � � X2 � �⎤
⎢ √ cos 𝛹 (t, X3 ) − √ sin 𝛹 (t, X3 ) ⎥
here due to potential tissue damage. 𝜆 𝜆
⎢ � � X2 � � ⎥.
x(t) = ⎢ X1 (10)
2.2.3 Modelling the experiment ⎢ √ sin 𝛹 (t, X3 ) + √ cos 𝛹 (t, X3 ) ⎥⎥
⎢ 𝜆 𝜆 ⎥
⎣ 𝜆X3 ⎦
Compressive and shear deformations were imposed onto the
liver samples in order to investigate their 3D biomechanical Incorporating the above form into Eq. 2, the corresponding
response. Thus, denoting by H and R the undeformed height deformation gradient takes the form
and radius of a sample, let h and r denote the height and radius
13
⎛ cos 𝛹 (t, X3 ) sin 𝛹 (t, X3 ) 2.3.1 Viscoelastic modified Mooney–Rivlin model

r sin 𝜃(t) ⎞
𝜓(t)
⎜ √ − √ −
⎟
𝜆 𝜆 H x
⎜ ⎟ The simplest model considered here is a modified
F(t) = ⎜ sin 𝛹 (t, X3 ) cos 𝛹 (t, X3 ) 𝜓(t)
rx cos 𝜃(t) ⎟
,
⎜ √ √ ⎟ Mooney–Rivlin strain energy function, which comprises
𝜆 𝜆 H
⎜ ⎟ two parts: W1 = (IĈ − 3)∕2 and W2 = (IIĈ − 3)2 ∕8. The first
⎝ 0 0 𝜆 ⎠
(11) part is a linear neo-Hookean term, whereas the second term,
compared to the original Mooney–Rivlin model, is quadratic,
where angle 𝜃( is related
) to spatial position and is given by in order to trigger a more accentuated nonlinear response.
𝜃(t) = arctan X2 ∕X1 + 𝛹 (t, X3 ) and rx is the radial position
This modified form has been previously employed in captur-
throughout the sample. On the top surface, the deformation
ing polymer hyperelasticity (Capilnasiu et al. 2019), as the
metrics can be found by replacing 𝛹 (t, X3 ) with 𝜓(t).
classical Mooney–Rivlin form was found to be unsuitable
In every test, torque measurements are acquired at the top
to model liver tissue hyperelasticity at large strains (Chui
plate level. Here, the torque on the top surface 𝛤t (as identi-
et al. 2004). In this form, the PK2 tensors are derived to be
fied in Fig. 1) is computed as
( )
1 1 IC −1
(14a)
∫𝛤t ∫𝛤t
Se = 2∕3 I − C ,
𝝉= r × td𝛤 = r × (𝝈 ⋅ n) d 𝛤t , (12) J 3
( )
where n = [0, 0, 1]T is the normal to the top surface. Since 1 ( ) IIC −1
the rotational forces are acting in plane, symmetrically S2e = 4∕3 IIĈ − 3 C − C . (14b)
J 3
around the Z-axis, then the only nonzero torque component
is The above form provides a purely elastic part, with S1e
leading to a linear response in shear and S2e leading to a
∫𝛤t
𝜏3 = r1 𝜎23 − r2 𝜎13 d𝛤t , (13) nonlinear response. Viscoelasticity is introduced by taking
a fractional-order derivative on the S1e tensor, as described
in Eq. 5. Initially, a more extensive Mooney–Rivlin-based
where the Cauchy stress components can be found from
model was considered (“Appendix 1”, Eq. 31), but it did not
Eq. 8. Although the hydrostatic pressure P does not affect
perform much better than a model with fewer parameters.
the torque computation, note that its value can be retrieved
Hence, the total PK2 tensor for the vMR∗ law considered
due to the zero normal traction on the wall boundary 𝛤w. By
here is
combining Eqs. 5 and 8 into
S = CS2e + 𝛿D𝛼t (S1e ) + Sp , (15)
𝝈 = F(Se + D𝛼t Sv )FT + JPI,
with C, 𝛿 (Pa), and 𝛼 (unitless) being material parameters.
and knowing that t = 𝝈 ⋅ n = 0, P can be determined by bal-
Note that C and 𝛿 act as linear scalings on the model compo-
ancing out the elastic and viscoelastic components in the
nents, whereas 𝛼 triggers a nonlinear response (hence it will
traction normal on the wall, 𝛤w.
be referred to as a nonlinear parameter, with the understand-
ing that it leads to a nonlinear torque response).
2.3 Constitutive modelling of liver tissue
2.3.2 Viscoelastic Ogden‑based model
In this study, the nonlinear liver behaviour is investigated
The second type of model considered is the Ogden
under combined large compressions and shear strains. An
model, described by the strain energy function
example of angular displacements employed is shown in
W = (𝜆b1 + 𝜆b2 + 𝜆b3 − 3)∕(2b), and the corresponding PK2
Fig. 1. For modelling purposes, it is assumed that in the ref-
tensor
erence configuration the bovine liver samples are stress free
and isotropic. The observed torque behaviour is modelled ∑
3
using a viscoelastic adaption of three hyperelastic models Sbe = 𝜆b−1

i
vi ⊗ vi . (16)
commonly applied in soft tissue mechanics, with the aim i=1
of drawing a comparison between their suitability to model Here, power b is a nonlinear parameter (i.e. the torque
the data: polynomial (a modified form of the Mooney–Rivlin depends nonlinearly on it), 𝜆i are the three principal
model, which will be indicated by vMR∗ ), Ogden (vOG) and stretches, and vi are the corresponding eigenvectors of ten-
exponential (vEXP). sor C . A more comprehensive Ogden-based model was ini-
tially considered (“Appendix 1”, Eq. 32); however, it did not
13
perform significantly better than a single-component model 2.4.1 Minimisation problem

and made the parametrisation non-unique. Thus, the total
PK2 stress considered here includes only a viscoelastic and Let y denote a set of M parameters that matches a model
a hydrostatic part: to the data. In order to find y, a minimisation problem of
the form
S = 𝛿D𝛼t (Sbe ) + Sp , (17)
𝜃 = arg min J(y, 𝝉 d )
(20)
where 𝛿 (Pa) is a linear scaling parameter, while the frac- y∈ℝM
+
tional-order derivative 𝛼 and the eigenvalue power b (unit-

less) parameters act nonlinearly (and hence will be referred is posed, where J is a function to be minimised. For the
to as nonlinear parameters). This model will be referred to models presented here, y comprises the nonlinear parameter
as the viscoelastic Ogden model vOG. 𝛼 and, where applicable, b, C and 𝛿 . However, the m linear
parameters can simply be found by inverting a system of
linear equations, as it will be seen shortly. This leads to a
2.3.3 Viscoelastic exponential model simplified minimisation problem
𝜃 ∗ = arg min J∗ (y∗ , 𝝉 d ),

The last type of model considered here is the isotropic y∗ ∈ℝM−m
(21)
+
exponential Fung-type model, described by the strain
energy function W = (exp(b(IIC − 3)) − 1)∕(4b) . The cor- where y∗ spans the nonlinear parameters only ( 𝛼 and, if
responding elastic PK2 tensor is derived to be applicable, b).
In this work, the minimisation problem 𝜃 ∗ is solved
Sbe = exp(b(IIC − 3))C, (18) by carrying out a parameter sweep over the nonlinear
parameters. Specifically, the fractional order 𝛼 is iterated
with power b being a nonlinear scaling parameter. As before, between 0.05 and 1 (with a step of 0.05), to ensure that
viscoelasticity is introduced by the fractional-order deriva- the whole spectrum from elastic to viscous is captured.
tive D𝛼t Se . For this model, in order to ensure that the devia- In the interval [0.15, 0.4], which is close to the literature
toric and hydrostatic parts are separated after applying the range estimated for 𝛼 , a finer step of 0.01 was used. Simi-
time derivative, let the PK2 tensor be defined as larly, power b is examined over a range—[0.1, 14] for the
vOG model and [0.1, 3] for the vEXP model, with coarse
S = 𝛿Dev[D𝛼t Sbe ] + Sp , (19)
refinements, to see trends. Then, we focused on the range
where 𝛿 (Pa) is a linear scaling parameters and b (unit- [1, 14] for the vOG model, with refinements of 0.5, and
less) is a parameter that acts nonlinearly. Hence, Eq. 19 on [1, 1.5] for the vEXP model, with refinements of 0.1,
defines the viscoelastic exponential model vEXP. A more to isolate parameter values.
comprehensive form of the exponential model was also For each combination of 𝛼 and b considered, the best
considered (“Appendix 1”, Eq. 33). However, similarly to linear parameters C and 𝛿 can be found by solving a linear
the Ogden-based model, the extensive form did not per- system of equations
form significantly better and it led to non-unique parameter Ax = b, (22)
identification.
with the understanding that A = A(𝛼, b) and x = x(𝛼, b). The
matrix A and vector b
2.4 Data analysis and model fitting � m,e m,v � � d�
⎡ �𝜏1 𝜏1 � ⎤ ⎡ �𝜏1 � ⎤
⎢ 𝜏 m,e 𝜏2m,v ⎥ ⎢ 𝜏d ⎥
Parameters from all models were tuned to match the exper- A=⎢ 2 ⎥ b=⎢ 2 ⎥ (23)
imental data. In this case, torque measurements ( 𝜏 d ) were ⎢ � ⋮m,e m,v
⋮ �⎥ ⎢ � ⋮d � ⎥
⎣ 𝜏N 𝜏N ⎦ ⎣ 𝜏N ⎦
compared against the torque corresponding to the models
( 𝜏 m ). In order to obtain 𝜏 m , the PK2 stress tensor for each comprise the elastic (if applicable) and viscoelastic torque
individual model (vMR ∗ —Eq. 15, vOG—Eqs. 17, and model components and the torque data measurements,
vEXP—Eq. 19) was used in Eq. 8 in order to quantify the respectively. Subscripts 1 to N indicate the tests considered,
Cauchy stress and in Eq. 13 to quantify the torque. Then, [ ]
which are vertically concatenated. Each block 𝜏im,e 𝜏im,v
the model parameters were adjusted to match the data by comprises multiple time points. Vector x = [C 𝛿]T contains
solving a minimisation problem. Three different objective the unknown linear parameters to be found. For further use
functions were employed, which will be described later throughout the section, subscript i denotes the block cor-
in this section. responding to test i, i.e.
13
[ ] [ ] � �1∕2
Ai = 𝜏im,e 𝜏im,v and bi = 𝜏id . ∑� �
Ai x − bi ;bi pw
The model components 𝜏 m were computed across a circular
i
J ∗pw = � �1∕2 . (26)
surface of radius r, at time points corresponding to the data ∑
(bi ;bi )pw
readings. Spatial integration of the modelled torque across i
the top face of the cylindrical sample was carried out using
a triangular mesh with 765 elements. For the time integra- This adapted norm leads to potentially higher L2 norm
tion, a discrete time step was set so that it matches the data errors, but it also ensures that the curve trends (e.g. the non-
points, using dt = fT1 , where f is the frequency and T is the sinusoidal torque response in Fig. 1) are better matched,
irrespective of their amplitude.
number of points in an oscillatory period. Specifically, dt =
0.0059, 0.0049 and 0.0049 s for the samples at 0.5, 1 and 2
2.4.4 Parameter scaling norm
Hz, respectively.
Three different norms were employed in order to quan-
Parameter variability can be encountered when analysing sam-
tify the model fit to the data. Firstly, the classical L2 norm
ples collected from different livers or different locations in the
is investigated. Further on, a point-wise scaling norm is
liver. Furthermore, the parametrisation process can be affected
introduced, which is an adaption of the L2 norm. Lastly, a
by shear softening—an effect which might be observed when
parameter scaling norm is designed, where some constraints
a material is sheared at successive increasing levels (Perepe-
on the linear parameters are relaxed. The norms presented
lyuk et al. 2016). Here, in order to accommodate for param-
are constructed such that the error is 0% for a perfect fit
eter variability due to sample location or shear softening, we
and 100% when the linear parameters are set to 0 Pa. The
introduce a norm that allows for flexibility in the linear scaling
contrasting nature of the norms leads to gathering different
parameters.
insights about the data and thus contributes to an overall
Since a single frequency was investigated per sample (either
better understanding of the liver tissue.
0.5, 1 or 2 Hz), this is not sufficient to allow for variability
in the fractional order 𝛼 . Moreover, it is observed that the
L2 norm
2.4.2 shape of the data curves tends to be preserved across samples,
whereas the amplitude scaling changes. Hence, presuming
The first norm relies on the MATLAB-implemented linear that the nonlinear parameters govern the shape of the torque
solver “lsqnonneg”. The linear parameters are found by min- curves, 𝛼 and b (if applicable) are assumed to be consistent
imising the remainder in a least squares (lsq) sense, using across samples and shear softening. By contrast, the linear
the L2 norm (|| ⋅ ||2 ), i.e. scaling parameters C and 𝛿 are assumed to vary, and their dis-
tribution will be examined in order to understand the ampli-
||Ax − b||2 tude behaviour of the data.
J∗lsq = . (24)
||b||2 Firstly, consider the normalised
While straightforward, this norm favours the tests that ⎡ 𝜏 m,e 𝜏im,v ⎤ ⎡ 𝜏d ⎤
employ larger deformations and thus attain higher torque A∗i = ⎢ ⎥ and b∗ = ⎢ ⎥
i i
⎢ n ��𝜏 d �� n ��𝜏 d �� ⎥ i ⎢ n ��𝜏 d �� ⎥
amplitudes. ⎣ i �� i ��2 i �� i ��2 ⎦ ⎣ i �� i ��2 ⎦
which make up normalised matrix A∗ and vector b∗. Here, ni

2.4.3 Point‑wise norm is the number of data points in test i. Additionally, let
x ∗ = xi 𝛽 i (27)
An alternative to the classic L2 norm is to scale values to
ensure that all points carry equal importance in the fitting be a set of unique linear parameters [C∗ 𝛿 ∗ ] which can be
process. To achieve this, each point in the data and model is transformed into the test-specific parameters xi by employ-
scaled by its corresponding amplitude in the data (point-wise ing the scaling 𝛽i . In order to find x∗ and 𝛽 , an iterative
scaling). In general form, this can be written as process is employed, starting with each 𝛽i = 1:
( )2 ( )1∕2
∑ ak
(a;b)pw = , (25) ∑
N
max(tol, |bk |) ∗
x = arg min ||A∗ x∗ ∕𝛽 − b∗i ||||2
2
, (28)
k || i i
x∗ i=1
where tol is a nonzero tolerance level, to avoid division by
zero. Thus, the error measure, which is also the function to
be minimised, becomes
13
𝛽i = arg min ||||A∗i x∗ − b∗i 𝛽i ||||2 . (29)

versus shear strain decreases as the maximum strain ampli-
𝛽i tude increases.
The above sequence of minimisation equations is repeated
3.1 Viscoelastic models tailored with the L2 norm
until there is no change in the error J∗sc,
�∑ �1∕2 Figure 3 presents the L2 error norm for each 𝛼 value inves-
N ��A∗ x∗ ∕𝛽 − b∗ ��2
�� i i ��2 tigated for the optimal set of all remaining parameters,
(30)
i=1 i
J ∗sc = .
��b∗ �� with larger markers denoting the minimum. The data were
�� 2
grouped by compression strain and frequency, with each
Although the set of linear parameters xi is different for every group comprising the corresponding shear strain tests (e.g.
test considered, note that this is different than solving each CS 10%, 1 Hz, 𝛾 = 1, 10, 25, 50%). The best parameter set
test individually, as this norm ensures that the nonlinear for each data group is presented in the corresponding tables.
parameters are fixed across all tests. Solving each test indi- Figure 4 illustrates examples of the three models—
vidually would yield the similar results only if 𝛼 and b were vMR∗ , vOG and vEXP—fit to the data tests at 1 Hz and
known a priori. CS 10%. The three model curves were produced using the
best fit parameters obtained when fitting the whole shear
strain range (1–50%). Figure 5 shows the Lissajous plots
(torque readings depending on the angular displacement)
3 Results corresponding to CS 10%, 1Hz. The continuous lines show
the data readings, while the dashed lines show the vEXP
This work is based on data acquired at a range of compres- model fit.
sion preloads, oscillatory shear strains and frequencies. At
compression preloads of 1%, 10% and 20%, the averaged 3.2 Viscoelastic models tailored with the point‑wise
measured normal force is 0.07 N, 0.2 N and 0.7 N. At the norm
smallest shear strain (1%), the data are close to linear vis-
coelasticity, with the nonlinearity becoming more evident The point-wise error norm behaviour with 𝛼 is shown in
at increased shear strains. Increasing the frequency also Fig. 6 for the vMR∗ (top), vOG (middle) and vEXP (bottom)
enhances the nonlinear behaviour of the liver tissue samples. models. The larger error markers identify the overall best
Hysteresis effects manifest during all tests employed, and parameter fit, obtained for the parameters presented in the
they tend to increase slightly with increasing shear strain. corresponding tables.
Strain rate dependence is observed, as both nonlinearity and Figure 7 presents the best models fit to the data acquired
hysteresis increase with frequency. These observations point at 1 Hz, CS 10%. The models’ parameters can be found in
towards a strain rate dependent, nonlinearly viscoelastic liver Fig. 6. The same dataset and the vEXP model fit are also
tissue behaviour. conveyed in Lissajous curves in Fig. 8.
When investigating the acquired data, a strain softening
effect is observed. That is, the more strained the tissue sam- 3.3 Viscoelastic models tailored with the parameter
ple is, the less force is required to produce a strain incre- scaling norm
ment. This is exemplified in Fig. 2 for shear strain (CS 10%,
1Hz), but a similar trend is exhibited for compressive strain Figure 9 shows the minimum parameter scaling error norm
as well. The curves capture the mean loading and unloading corresponding to the models (vMR∗ , vOG, vEXP), for every
response, and it can be seen that the tangent of the torque fractional order 𝛼 investigated. All 18 tests were considered
simultaneously, hence a single curve corresponding to each
model, compared to the previous analogue Figs. 3 and 6,
where subgroups of the data tests were considered. The
smallest error across the 𝛼 range is identified by the larger
markers, and the set of parameters corresponding to these
minima are found in the corresponding table.
Figures 10 and 11 show all data tests acquired, and the
three model fits. Figure 10 shows all datasets acquired
at a frequency of 1 Hz, while Fig. 11 shows datasets
acquired under a compression strain of 10%. Example Lis-
Fig. 2 Mean loading and unloading torque response versus angular sajous curves are shown, for the data and vEXP model at
displacement at 1 Hz, CS 10% 1 Hz, in Fig. 12. The top row shows all four shear strains
13
Fig. 3 The minimum error for

the L2 norm (Eq. 24), obtained CS 1% 10% 20% 10% 10% mean
freq 1Hz 1Hz 1Hz 0.5Hz 2Hz ± SD
for each data group, for frac-
tional order 𝛼 values between err(%) 24.40 21.35 18.92 18.36 29.99
0.05 and 1. Plots show minimal α 0.33 0.26 0.28 0.19 0.31 0.27±0.05
error obtained for vMR∗ (Top),
C(Pa) 10.16 35.28 38.96 0 67.68 30.42±26.56
vOG (Middle) and vEXP (Bot-
tom) models, with the larger- δ(Pa) 80.2 95.78 62.88 328.12 166.44 146.68±108.79
sized markers being obtained
for the parameters presented in
the corresponding tables
CS 1% 10% 20% 10% 10% mean
err(%) 21.41 18.46 16.83 18.20 28.70
α 0.33 0.24 0.21 0.19 0.3 0.25±0.06
b 3.5 4 4.5 2.5 6.5 4.2±1.48
δ(Pa) 17.86 24.19 23.02 243.44 13.84 64.47±100.14
CS 1% 10% 20% 10% 10% mean

err(%) 21.48 19.27 17.73 18.20 28.44
α 0.33 0.23 0.21 0.19 0.3 0.25±0.06
b 0.1 0.1 0.1 0.1 1.2 0.32±0.49
δ(Pa) 82.26 149.85 177.70 438.22 184.56 206.52±135.69
Fig. 4 Example of the three models fit to the data acquired at 1 Hz, compression strain 10%, using the L2 norm. The parameters employed to pro-
duce the models’ curves can be found in column 2 (CS 10% 1 Hz) of the corresponding tables in Fig. 3
(1–10–25–50%), while the bottom row zooms in on the 3.4 Viscoelastic models tailored with the parameter
1 and 10% shear strains. Compression strain increases scaling norm in the case of non‑ideal
from the left column (1%) to middle (10%) and right compression
(20%). In these figures, the models’ curves were obtained
by employing the starred parameters from Fig. 9, scaled In this study, for rapid model evaluation, the analysis
according to Eq. 27. The test-specific linear parameters assumed ideal compression. However, the testing condi-
can be seen, for all models, in Fig. 13, and their mean and tions (serrated plates) actually led to non-ideal compression
standard deviation are shown in Fig. 9. and hence a barrelling of the tissue samples. “Appendix 2”
13
presents details on the differences observed between ideal

and non-ideal compression, and it reveals that, for torque,
a scaling factor needs to be applied in order to correct the
linear parameters. For the vEXP model, for the data at 1 Hz,
the correcting factors are presented in Table 2, and the cor-
rected parameters are shown in Fig. 14.
4 Discussion
When comparing the three models (vMR∗ , vOG and vEXP)

within the L2 norm performance, it can be seen that the
errors presented in Fig. 3 are very similar, with the vMR∗
model errors being slightly higher (at most ≈ 3%). This is
likely due to the fact that the other two models are more
suitable to capture the nonlinear trends. However, it can be
seen from Fig. 4 that all three models match the closest data
peaks at shear strain 50%. This is because the data in this
study can vary even by two orders of magnitude, with the
Fig. 5 Lissajous curve exemplifying the vEXP model fit to the data
acquired at 1 Hz, compression strain 10%. The parameters employed higher amplitude data having the largest impact on the L2
to produce the model’s curves can be found in column 2 (CS 10% norm. As a result, amplitudes in the other tests have lesser
1 Hz) of the bottom table in Fig. 3. The curves corresponding to shear fidelity, as a consequence of the L2 norm, which penalises
strains 1–10–25–50% are shown in the top quadrant, while the lower less discrepancies at smaller amplitudes. This mismatch is
quadrant zooms on shear strains 1–10%
further accentuated by the strain softening effect. It can be
seen that the data exhibit a shallower increase in amplitude
Fig. 6 Minimum error for

CS 1% 10% 20% 10% 10% mean
the point-wise scaling norm freq 1Hz 1Hz 1Hz 0.5Hz 2Hz ± SD
(Eq. 26), obtained for each data
group, for fractional-order 𝛼 val- err(%) 87.51 83.31 78.90 69.95 90.12
ues between 0.05 and 1. Plots α 0.4 0.31 0.29 0.25 0.4 0.33±0.07
show minimal error obtained for C(Pa) 0 0 0 18.93 0 3.79±8.47
vMR∗ (Top), vOG (Middle) and
vEXP (Bottom) models, with δ(Pa) 22.30 34.36 34.24 184.45 35.30 62.13±68.59
the larger-sized markers being
obtained for the parameters
presented in the corresponding
CS 1% 10% 20% 10% 10% mean
tables
err(%) 63.23 56.05 53.55 51.76 73.82
α 0.29 0.2 0.19 0.18 0.31 0.23±0.06
b 7 8 8.5 12 14 9.9±2.97
δ(Pa) 2.48 3.93 3.56 5.58 1.26 3.36±1.62
CS 1% 10% 20% 10% 10% mean

err(%) 61.25 54.56 52.91 52.55 77.88
α 0.28 0.21 0.19 0.19 0.33 0.24±0.06
b 1 1.2 1.4 3 3 1.92±1.00
δ(Pa) 55.08 89.73 63.98 235.39 71.20 103.07±75.05
13
Fig. 7 Example of the three models fit to the data acquired at 1 Hz, compression strain 10%, using the point-wise scaling norm. The parameters
employed to produce the models’ curves can be found in column 2 (CS 10% 1 Hz) of the corresponding tables in Fig. 6
with shear strain (1–50%) than the models are. The bias of
the L2 norm towards the higher amplitude points can also be
observed in Fig. 5, where the vEXP model fit well the peak
at shear strain 50%, yet the rest of the model torque curve
looks idealised (almost perfectly elliptic) and does not fol-
low the data curve.
In order for the models to better follow the data curve
trends throughout the entire deformation, a point-wise scal-
ing norm (Eq. 26) was designed, which penalises discrepan-
cies between data and models, across all points considered,
evenly. For this norm, the errors for the vOG and vEXP mod-
els are similar, with the vMR∗ model performing signifi-
cantly worse (≈ 20%, as shown in Fig. 6). This is because the
vMR∗ model cannot recreate the nonlinear trends exhibited.
Hence, this norm drives the model peak points to match
closely the lower amplitude data points (Fig. 7), by forcing
a delayed response (higher 𝛼 values, as shown in the top
table of Fig. 6, compared to the middle and bottom tables).
Fig. 8 Lissajous curve exemplifying the vEXP model fit to the data
Figure 7 shows that the vOG and vEXP models are able
acquired at 1 Hz, compression strain 10%. The parameters employed to capture the data trends; however, there is a data-model
to produce the model’s curves can be found in column 2 (CS 10% amplitude discrepancy for all tests. This discrepancy is more
1 Hz) of the bottom table in Fig. 6. The curves corresponding to shear accentuated than for the L2 norm (Fig. 4) because the models
strains 1–10–25–50% are shown in the top quadrant, while the lower
quadrant zooms on shear strains 1–10%
employ a higher nonlinear parameter b—8 to 4 for vOG, as
shown in the middle tables of Figs. 3 and 6, and 1.2 to 0.1
err(%) α b C ∗ (P a) δ ∗ (P a)
vMR∗ 19.92 0.22 N/A 97.38 130.85
mean±SD 145.55±106.26 195.43±142.79
vOG 17.04 0.19 10 N/A 2.7
mean±SD 6.74±6.09
vEXP 17.52 0.2 1.5 N/A 113.99
mean±SD 231.02±198.43
Fig. 9 Minimum model error (vMR∗ , vOG and vEXP) for the param- is enhanced by the larger-sized marker, obtained for the parameters
eter scaling norm (Eq. 30), obtained for the tests altogether, for frac- presented in the corresponding table
tional-order 𝛼 values between 0.05 and 1. The minimal error across 𝛼
13
Fig. 10 Three models fit to the data acquired at 1 Hz, using the ters employed to produce the models’ curves can be found in the table
parameter scaling norm. The first row illustrates the tests at CS 1%, corresponding to Fig. 9
the second row at CS 10% and the third row at CS 20%. The parame-
for vEXP, as shown in the bottom tables of the same figures. Comparing the three models, it can be observed that error
This determines a steeper increase in the models’ amplitude. behaviour with 𝛼 (Fig. 9) is almost identical for the vOG
Figure 8 clearly shows that this norm leads to a better model and vEXP models. For the vMR∗ model, the minimum error
match of the data curve trend, at the cost of discrepancy in curve is shallow for 𝛼 values between 0.15 and 0.4, leading
the peaks. to a less precise parameter identifiability.
In the models, the nonlinear parameter 𝛼 determines the With the three models reaching a similar minimum
phase delay between the shear strain input and torque output. error, it is expected that they behave similarly, as shown in
The nonlinear parameter b controls the shape and amplitude Figs. 10 and 11. The vOG and vEXP models are better at
of the torque curve, where b = 0 leads to a linear response capturing the data nonlinearity, which becomes apparent at
(i.e. sinusoidal output as a result of a sinusoidal input) and shear strains above 10%, as the curves start deviating from
higher values lead to an increased nonlinear response. The a pure sinusoidal wave, looking like a combination of trian-
linear parameters C and 𝛿 amplify the contribution of the gle and sine waves. Although it appears that an increased
model components. Comparing the results of the L2 and compression strain simplifies the appearance of the shear
point-wise norms, it can be concluded that the data curve strain torque curve, in reality it further complicates the non-
trends can only be captured by accentuating the nonlinear linear behaviour exposed due to shear straining. This com-
characteristics of the models (e.g. increasing the b param- plex behaviour is generally better captured by the vOG and
eter). However, this deteriorates the peak amplitude match, vEXP models, although one notable test where the vMR∗
as the data exhibit strain softening (Fig. 2). model performs better is at compression strain 1%, shear
Due to the fact that the data could not be described using strain 50%, 1 Hz (top right panel in Fig. 10). There, vMR∗
a set of fully consistent parameters, the parameter scaling captures the data amplitude and does not exhibit the exag-
norm was designed so that the set of linear parameters can gerated nonlinearity of the other two models; however, it
be adjusted according to each test, by scaling the models’ performs inferiorly in other tests (e.g. compression strain
amplitude response. Hence, it is expected that the errors, 10%, shear strain 25%, 0.5 Hz—top right panel in Fig. 11).
computed using Eq. 30, are small. Indeed, as per Fig. 9, The models’ curves in Figs. 10, 11 and 12 were produced
the parameter scaling norm leads to the smallest quantified using the parameters presented in Fig. 9. All three models
errors, compared to the L2 and point-wise scaling norms. employ a similar fractional-order derivative, between 0.19
13
Fig. 11 Three models fit to the data acquired at CS 10%, using the employed to produce the models’ curves can be found in the table
parameter scaling norm. The first row illustrates the tests at 0.5 Hz, corresponding to Fig. 9
the second row at 1 Hz and the third row at 2 Hz. The parameters
and 0.22, which is in good agreement with the literature. linear parameter variability introduced by sample location
It is notable that for the vOG and vEXP models, the non- and strain softening.
linear parameters 𝛼 and b presented in Fig. 9 are similar The effect of barrelling, due to non-ideal compression,
to the parameter means presented in Fig. 6, obtained using was investigated in “Appendix 2”. For the vEXP model
the point-wise norm. This reinforces the suitability of the simulating the data at 1Hz, the corrected parameters can
parameter scaling norm to be employed with this study’s be seen in Fig. 14. The parameters at CS 1% remain similar
data, as its main effect is facilitating the amplitude match. between ideal and non-ideal compression, which is expected
The linear parameters 𝛿 ∗ and, if applicable, C∗ , are given the small compressive strain level. The parameters at
scaled according to each data test, leading to the test-spe- CS 10% are downscaled by a factor of ≈ 1.5, according to
cific parameters depicted in Fig. 13. It can be observed Table 2, and maintain a similar softening trend. The param-
that for all three models, the linear parameters decrease eters at CS 20% are downscaled according to the factors pre-
with increasing shear strain, which reflects the shear strain sented in Table 2. Although the non-ideal parameters trend
softening effect. Interestingly, it appears that three data appears flatter, this is because of the scale used (0 to 230),
groups can be identified, which correspond to the fre- and in reality the softening trend is accentuated.
quency groups—0.5, 1 and 2 Hz. Since the linear param-
eters do not have a monotonic trend with frequency, with 4.1 Liver biomechanical considerations
the parameters at 0.5 and 2 Hz being larger than the ones in the context of existing literature
at 1 Hz, it is unlikely that this reflects a parameter–fre-
quency coupling. Instead, this might be explained by a In this study, polynomial, Ogden and exponential model
dependency on the sample location, as different samples forms were used. These types of models are commonly
were tested across the frequency tests. Additionally, the employed for capturing the nonlinear hyperelastic behav-
standard deviation of these parameters reaches up to 90% iour exhibited by tissue at large strains (Veronda and West-
of their mean, as presented in Fig. 9, indicating a high mann 1970; Zobitz et al. 2001; Chui et al. 2004; Gao et al.
13
Fig. 12 Lissajous curve exemplifying the vEXP model fit to the data shows the curves corresponding to shear strains 1–10–25–50%, while
acquired at 1 Hz. The parameters employed to produce the model’s the bottom row zooms on shear strains 1–10%. Compression strain
curves can be found in the table corresponding to Fig. 9. The top row increases from left (1%) to middle (10%) and right (20%)
Fig. 13 Absolute linear param-

eters for the models, obtained
by transforming the starred∗
relative parameters into their
absolute counterparts using
Eq. 27. The top panels show
the C and 𝛿 parameters scaling
the elastic and viscoelastic
components of the vMR∗ model.
The bottom panels show the 𝛿
parameter scaling the viscoelas-
tic component of the vOG and
vEXP models, respectively
13
Table 2 Multiplying factors between ideal and non-ideal torque and

normal force measurements
Torque SS 1% SS 10% SS 25% SS 50%
CS 1% 1.0050 1.0046 1.0031 1.0000

0.2041% 0.2026% 0.2043% 0.1730%
CS 10% 1.5560 1.5568 1.5460 1.5123
0.6844% 0.7326% 0.9007% 1.1374%
CS 20% 5.9433 6.0363 6.3176 6.8061
3.0972% 2.7304% 1.9302% 1.5789%
Fig. 14 Absolute linear parameters for the vEXP model, correspond-

2010). Polynomial models introduce nonlinearity through ing to the data at 1 Hz. The solid lines (blue, orange, yellow) indicate
the value of the 𝛿 parameter when using ideal deformation assump-
quadratic or higher-order terms, yet Ogden and exponential tion and are the same as the ones presented in Fig. 13, bottom right
models are generally better suited in capturing the nonlinear- panel; the dashed lines (purple, green, cyan) indicate the 𝛿 parameter
ity. However, loading–unloading tests reveal that biological when using non-ideal compression. The non-ideal parameters were
tissues are viscoelastic (Jordan et al. 2009) and strain rate obtained by scaling the ideal parameters by the values presented in
Table 2
dependent (Liu and Bilston 2000; Miller 2000). Here, frac-
tional derivative adaptions of the Mooney–Rivlin, Ogden
and exponential models were introduced in order to capture 4.2 Study overview
the viscoelastic, strain rate-dependent behaviour of liver tis-
sue. The fractional order 𝛼 can vary from 0 (purely elastic In this work, three constitutive equations were proposed
response) to 1 (purely viscous response), with intermedi- for modelling the behaviour of liver tissue under a range
ate values leading to viscoelasticity. For fractional models, of deformations and frequencies, being parametrised by
a frequency dependence of the type 𝜔𝛼 Holm and Sinkus three error norms. For the data employed here, the L2 norm
(2010) is achieved. Considering the above, the vMR∗ , vOG proved to be unsuitable, as it was biased by the larger ampli-
and vEXP models are appropriate for modelling the torque tude points. However, this is a straightforward norm that
data in this study, which exhibits a frequency-dependent, is convenient to use when analysing data of comparable
nonlinear viscoelastic behaviour. magnitude. The point-wise norm was introduced in order to
In a study similar to the present one, conducted by Ayy- equalise the weight of each data point in the models’ fitting
ildiz et al. (2015), cylindrical liver samples of similar dimen- process. The vMR∗ model performed particularly poor with
sions were investigated at a range of compression strains, this norm; however, the vOG and vEXP models followed
shear strains and frequencies. Specifically, the authors the data curve trends better than for the L2 norm. Hence,
investigated the effect of strain rate on compressions up to this norm offered better insights into the nonlinearity of the
20%, the influence of compressive preload (5–20%) on shear data. With the shear strain softening leading to amplitude
strains of 0.1–5% at a fixed frequency of 10Hz, and the com- mismatches, the parameter scaling norm was designed to
bined effect of compressive preload (5–20%) and frequency allow for flexibility in the set of linear parameters and thus
sweep (0.1–10 Hz) on the shear modulus measured using facilitate the model to data amplitude match. In this case, the
0.5% shear strain. It was found that the compressive preload models’ fit was significantly improved, with all three models
affects the measured shear modulus, which is consistent with performing similarly in capturing the data particularities.
our findings [and the original findings in Tan et al. (2013)], Based on the results of this study, the choice of constitu-
since the compression strain influences the torque response tive law for modelling the liver behaviour depends on the
measured, as shown in Fig. 10. Similarly, the frequency complexity desired. The vMR∗ model provided a marginally
increase accentuates the nonlinear shear response (Fig. 11), less suitable fit compared to the vOG and vEXP models;
which would correspond to an increased shear modulus, as however, it is very practical due to its simple form, requiring
indicated by Ayyildiz. As the detailed results on the shear only the computation of invariants and identifying a single
strain sweep (0.1–5%) are not presented, the studies cannot nonlinear parameter. Nevertheless, for the data presented
be compared in this regard. To model the data, Ayyildiz used here, the shallow minima of the error norm across the non-
a generalised Maxwell model with 13 parameters, which linear parameter 𝛼 reveal a non-ideal parameter identifiabil-
captures the strain-dependent, nonlinear viscoelastic behav- ity process. As an alternative, the vOG and vEXP models
iour of the liver. Here, we used fractional viscoelasticity, can be employed if more accuracy is required, as they lead
which helps reducing the number of rheological elements to the identification of a clear minima and also provide
considered and hence the number of model parameters. an improved fit. They are more complex, involving two
13
nonlinear parameters. Furthermore, the vOG model relies 4.4 Study limitations

on the computation of eigenvalues, which is more compu-
tationally expensive. The aim of this paper is to define a frequency-dependent
Considering this study’s findings, the vEXP model pro- nonlinear viscoelastic law that is able to characterise the
vides a good balance between model complexity and data liver tissue under a range of combined deformations and
fit. The use of the parameter scaling norm was essential in frequencies. The results presented advocate for the suitabil-
ensuring a good overall model fit. Nevertheless, note that ity of all three models employed, with vEXP in particular,
this norm was employed in order to overcome strain soften- to capture the liver behaviour, yet some limitations remain
ing effect. Ideally, this effect would be mitigated by precon- to be addressed by further research.
ditioning the samples at the highest strain applied, but here, One of the main difficulties in this study was to overcome
due to the very large strains, is avoided in order to avoid strain softening effect encountered, which was emphasised
tissue damage to the samples during early testing. by the testing procedure. In a study performed by Cheng
et al. (2009) on spinal cord, it was shown that the response of
4.3 Extension and impact in vivo and in silico tissue loaded to 2% uniaxial strain differs significantly if the
tissue is preconditioned to a maximum of 2% or 5% strain.
The transportation industry is particularly concerned with Nonetheless, the tissue behaviour that was loaded up to 2%
the study of liver’s response to sudden large impacts, as it or 5% did not differ if all samples were preconditioned at 5%
is the most exposed abdominal organ in frontal and side strain. This indicated that the strain softening phenomena
crashes. Having a model that is able to capture its behaviour can be addressed by preconditioning at the highest strain
at large deformations as well as rate-dependent effects ena- used in the testing protocol. In our study, given that large
bles the determination of injury tolerance limits and hence deformations were investigated, the preconditioning protocol
can guide the design of seatbelts, airbags, children car seats, had to be applied sequentially with the measurements. That
etc. Although the models designed here are more appropri- is, at each deformation level, eight oscillations were carried
ate for phenomena occurring repetitively, they can provide out before acquiring the data, which was enough in order to
a starting point or complement the investigation of sudden obtain a steady response in the data, as shown in the Lis-
high impacts experienced in injury crashes. sajous-type curves in Fig. 12. However, the tissue samples
In the medical research field, in silico simulations are were not strained to the maximum deformation level before
used in order to test new devices, provide surgery training, starting the data acquisition, as that could have damaged
predict diagnosis and treatment response, etc. These simu- the samples’ microstructure and led to non-representative
lations need to be able to gauge the nonlinear viscoelastic measurements at low strains. In order to address this effect,
tissue response to a range of mechanical stimuli, e.g. cutting, an initial preconditioning would need to be done at the high-
puncturing, pressing. The models discussed here are suitable est deformation level, as suggested by Cheng et al. (2009),
for in silico work, with the vEXP model performing better or measurements would need to be acquired in reverse, from
than the vMR∗ model and being computationally cheaper the highest to the lowest strain. This would also require
than vOG. determination of the maximal deformation allowed avoid-
A different possible application of the models discussed ing material damage. An alternative approach in dealing
is in the medical elastography area, which provides non- with the strain softening would be to devise models that are
invasive in vivo measurements of an organ’s stiffness in able to capture this phenomenon. However, such models are
order to assess disease severity. These measurements rely on usually heuristic (e.g. Perepelyuk et al. 2016), and their use
assumptions about an underlying biomechanical model. In is thus limited. Hence, the best approach in mitigating the
general, simple models are considered (e.g. Neo-Hookean), strain softening remains preconditioning to the highest strain
but frequency-dependent nonlinear viscoelastic models, like level applied in the testing protocol, despite that this level
the ones presented here, could lead to a better identifica- might be variable across studies.
tion and staging of pathology. With the liver being one of A limitation of the vOG and vEXP models, compared
the most imaged organs using elastography, the 3D liver to vMR∗ , is the absence of a purely elastic term. Over an
models proposed here could be translated and investigated infinite time, the viscoelastic term modelled using the frac-
in the context of small amplitude shear waves probing the tional-order derivative D𝛼t would approach 0. Nonetheless,
liver at various deformation states imposed by the respira- the time required for this decay is very large, as the normal
tory motion (Kang et al. 2012). force component in uniaxial compression relaxes proportion-
ally to the power law F ∝ Fmax t−𝛼 , where F is the force at
time t and Fmax is the maximum force reached due to defor-
mation. For a fractional order similar to the one identified
in this study, i.e. 𝛼 = 0.2, the time required for the force to
13
decay to 1% of the maximum value is 1010 s (≈ 317 years). was selected with care to avoid large vessels that may
This is an impractical time duration for both experiments bias measurements; however, smaller vessels were pre-
and simulations. Thus, despite the eventual decay of the sent throughout the test samples. While the presence of
force, here it was considered that a single viscoelastic term vasculature can exhibit behaviour similar to viscoelas-
is appropriate for modelling the data. Adding a purely elastic phenomena, studies neglecting viscoelastic response
tic term Se to the vOG and vEXP models was also consid- in cardiac tissue, arterial wall and articular cartilage (at
ered, as exemplified in “Appendix 1”; however, this leads to strain rate > 0.0001s−1 ) have not been able to replicate
parameter coupling. A broader frequency interval may be experimentally observed hysteresis or relaxation (Yang
beneficial in identifying two separate elastic and viscoelas- and Taber 1991; Zakerzadeh et al. 2016; DiSilvestro
tic terms, but for the data available in this study there was a et al. 2001). Further, as the constituents of the liver have
trade-off between adding a purely elastic term and identify- been shown to exhibit viscoelastic material response (e.g.
ing a fractional order 𝛼. hepatic cells are viscoelastic (Zhang et al. 2002; Yang
One of the assumptions made in the modelling process et al. 2019), extracellular collagen needs to exhibit viscoe-
is the idealised compression employed. In the experiments, lasticity in order to ensure accurate hepatic cell behaviour
however, the samples bulged under compression. Likely, the (Mattei et al. 2018) and isolated collagen fibrils exhibit
viscoelastic model response would change when considering viscoelasticity (Shen et al. 2011)), we believe the observed
a non-idealised compression. Investigations in simulated 3D behaviour stems from viscoelasticity within the material.
models indicate that there is a torque amplitude difference Further studies considering poro-viscoelasticity (particu-
between the idealised and non-idealised compression. How- larly at the whole-organ level) would provide an interest-
ever, this would have an impact only on the linear param- ing avenue for further investigations.
eters of the models, which would need to be upscaled. The
normal force, nonetheless, would be the most affected, as
simulations show a qualitative difference between ideal and
non-ideal compression. “Appendix 2” presents a detailed in 5 Conclusions
silico comparison of normal forces and torque between ideal
and non-ideal compression. This paper investigates the suitability of three models—
The liver samples used in this study were fixed in the vMR∗ , vOG and vEXP—to capture the nonlinear viscoelastic
rheometer using serrated plates. To ensure sample gripping, liver response at a range of large deformations. Their perfor-
an initial compressive strain and force had to be applied mance is gauged using three norms—the L2 norm, a point-
before running the experiments, and hence, the total com- wise scaling norm, which ensures even contributions from
pressive force at 1, 10 and 20% compression strain was 0.07, each data point, and a novel parameter scaling norm, which
0.2 and 0.7N, respectively. Despite the initial preload, other allows for flexibility in determining the linear parameters.
fixing methods like glue could have led to changes in tissue This norm was introduced in order to overcome the strain
properties (Nicolle et al. 2010) and artificial changes in the softening effect, and its implementation leads to significant
samples’ height. improvements in fitting the models to the data, with the vOG
Although the range of frequencies investigated here is and vEXP models being able to better capture the nonlinear
small (0.5–2Hz), the models presented in this study are particularities. Among the forms presented here, the vEXP
applicable for other frequencies as well. The fractional model is able to recreate the data trends, while identifying a
order 𝛼 is related to the frequency response, and the value clear minima across the fractional order 𝛼 . Furthermore, it
identified in this study (𝛼 = 0.2) is in good agreement with relies on strain metrics that are cheap to compute, making it
literature findings (Liu and Bilston 2000; Jordan et al. 2011; advantageous to use.
Sinkus et al. 2018; Asbach et al. 2008). Thus, by choosing
the nonlinear parameters presented in Fig. 9 and the appro- Acknowledgements The authors would like to thank Kirsty Tan for
acquiring the data.
priate linear parameters corresponding to a certain deforma-
tion level (Fig. 14), the models can be used for predicting Funding D.N. acknowledges funding from the Engineering and
phenomena occurring repetitively, at various frequencies. Physical Sciences (EP/N011554/1 and EP/R003866/1). L.B. acknowl-
(e.g. for MRE purposes). However, when a sudden force edges funding from the Australian Research Council Discovery
is experienced (e.g. crash injury), the response is hard to Grant (DP160100061) and an NHMRC senior research fellowship
(APP1077934). This work is funded by the King’s College London
predict, and hence, the use of these models in an area like and Imperial College London EPSRC Centre for Doctoral Train-
transportation industry is limited. ing in Medical Imaging (EP/L015226/1), and CDT industrial spon-
The current model presented considers liver tissue as sor Siemens Healthcare. This work was supported by the Wellcome
viscoelastic, neglecting porous media effects due to the EPSRC Centre for Medical Engineering at King’s College London
(WT 203148/Z/16/Z) and by the National Institute for Health Research
vasculature of the liver. Preparation of tested samples
13
(NIHR) Biomedical Research Centre award to Guy and St Thomas’ where Seb is characterised by Eq. 18. For all three extended
NHS Foundation Trust in partnership with King’s College London.
The views expressed are those of the authors and not necessarily those
models, the material parameters C and 𝛿 are linearly scaling
of the NHS, the NIHR or the Department of Health. the elastic and viscoelastic contributions, respectively.
Compliance with ethical standards Appendix 1.1: Results employing the extended
models
Conflict of interest The authors declare that they have no conflict of
interest. Employing the extended models with the parameter scaling
norm leads to smaller errors compared to the initial models.
Open Access This article is licensed under a Creative Commons Attri-
bution 4.0 International License, which permits use, sharing, adapta-
Figure 15 summarises the minimum errors and best param-
tion, distribution and reproduction in any medium or format, as long eter set for each model. The vMR∗ext model leads to a better
as you give appropriate credit to the original author(s) and the source, fit (≈ 4% error improvement) compared to vMR∗ . For the
provide a link to the Creative Commons licence, and indicate if changes other two extended models, the error improvement is less
were made. The images or other third party material in this article are
included in the article’s Creative Commons licence, unless indicated
than ≈1%. However, in Fig. 15 it can be seen that the minima
otherwise in a credit line to the material. If material is not included in are shallow, whereas when using the initial models vOG
the article’s Creative Commons licence and your intended use is not and vEXP (Fig. 9), the minima are more clearly identified.
permitted by statutory regulation or exceeds the permitted use, you will This is because the vOGext and vEXPext models offer more
need to obtain permission directly from the copyright holder. To view a
copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
flexibility in fitting the data, by adding a purely elastic com-
ponent. This becomes dominant, while the viscoelastic part
is used for fine-tuning of the nonlinear behaviour, which is
Appendix 1: Extended models why the error curve is shallow across 𝛼 values. Given the
data range used in this study, the extended models do not
The models presented in this study may be extended to lead to a clear parameter identifiability. Nevertheless, they
include more components. This could allow for more flex- might be suitable in modelling more complex experimental
ibility in the modelling process. For example, consider the data or different materials.
extended version of the vMR∗ model (vMR∗ext ) to be
S = C1 S1e + C2 S2e + 𝛿1 D𝛼t (S1e ) + 𝛿2 D𝛼t (S2e ) + Sp , (31) Appendix 2: Ideal and non‑ideal
compression
where Se1,2 are defined as in the set of Eqs. 14. Similarly, the
extended vOG model (vOGext ) would be characterised as The modelling part of this study assumes ideal compression
conditions, where the cylindrical samples can slide under
S = C1 Sbe 1 + 𝛿1 D𝛼t (Sbe 2 ) + Sp , (32) compression and maintain a cylindrical shape. However, due
to the nature of the rheological experiment, which employs
with Seb given by Eq. 16. Lastly, the extended vEXP model
serrated plates, the top and bottom faces of the samples
(vEXPext ) would be defined as
were prevented from sliding, hence leading to non-ideal
S = C1 Dev[Sbe 1 ] + 𝛿1 Dev[D𝛼t Sbe 2 ] + Sp , (33) compression and a barrelling effect. This effect has been
previously investigated in the literature (Mendis et al. 1995;
α b1 b2 C1∗ C2∗ δ1∗ δ2∗

vMR∗
ext 0.29 N/A N/A 47.58 49.02 9.12 140.11
(err 15.99%)
vOGext 0.4 7 11 5.39 N/A 0.73 N/A
(err 15.95%)
vEXPext 0.4 1.5 2 54.23 N/A 33.01 N/A
(err 16.67%)
Fig. 15 Minimum model error (vMR∗ext , vOGext and vEXPext ) for the across 𝛼 is enhanced by the larger-sized marker, obtained for the
parameter scaling norm (Eq. 30), obtained for the tests altogether, parameters presented in the corresponding tables
for fractional-order 𝛼 values between 0.05 and 1. The minimal error
13
Miller 2005; Wu et al. 2004; Spilker et al. 1990; Roan and compression, the sample is barrelling. Simulations of the
Vemaganti 2007), and it was shown that the compressive normal force on the top surface and torque are presented in
force was significantly higher when a no-slip boundary was Figs. 17, 18 and 19.
in place, leading to barrelling. This was more pronounced The normal force simulations show the most striking
for samples where height/diameter ratio is small, as in the differences between the two compression types, which is
current study. In order to address discrepancies that might consistent with literature findings. In order to facilitate
arise, here, we investigate the differences between ideal and the comparison of the force between ideal and non-ideal
non-ideal compression in an in silico setting. For this, C compression, the force corresponding to the ideal case was
heart (Lee et al. 2016) was used to solve the mechanical upscaled by the parameters presented in Table 3. These
problem posed in Sect. 2.2.3. were obtained by ensuring the best point-wise fit between
In order to replicate the experiment as accurately as pos- the two cases, with the errors being reported for each test.
sible, a cylindrical sample of radius R = 10 mm and height It can be seen that for small compression and shear strains
H = 3 mm was compressed by 1, 10 and 20% and sheared up (1%), an almost perfect match can be recovered. The factor
to 1, 10, 25 and 50% level, at a frequency of 1 Hz. The expo- used in this case (6.5821) matches the correction indicated
nential model (Eq. 18) was employed, with the linear scaling by Gent (2012) for short cylindrical samples (i.e. 6.5556,
𝛿 = 130, the nonlinear parameter b = 1.5 and the fractional corresponding to the specific sample size considered here).
power 𝛼 = 0.2. The uniform compression was prescribed at The correction proposed by Gent stands for small strains,
the top and bottom surfaces as and hence, it can be seen that for the larger strains utilised
in this study, the parameters are different and the match
⎡ X1 cos(𝜓(t)X ) − X2 sin(𝜓(t)X ) − X ⎤ deteriorates.
⎢ √𝜆(t) 3 √ 3 1⎥
⎢ 𝜆(t) ⎥ Despite the discrepancy in the normal force, the ideal
u(t) = ⎢ X1 X compression consideration does not invalidate this study.
sin(𝜓(t)X3 ) + √ 2 cos(𝜓(t)X3 ) − X2 ⎥
,
√
⎢ 𝜆(t) ⎥
𝜆(t) In Figs. 18 and 19, it can be seen that, although the torque
⎢ ⎥
⎣ X3 (𝜆(t) − 1) ⎦ amplitude differs between ideal and non-ideal compression,
(34) the trends are preserved and a good match can be obtained
while the non-ideal compression was prescribed as by upscaling the ideal compression case by the factors pre-
sented in Table 2. For 1% compression strain, the factor is
⎡ X1 cos(𝜓(t)X3 ) − X2 sin(𝜓(t)X3 ) − X1 ⎤ almost unity, which is expected since very little barrelling
u(t) = ⎢ X1 sin(𝜓(t)X3 ) + X2 cos(𝜓(t)X3 ) − X2 ⎥. (35) is experienced under 1% compression. The errors when
⎢ ⎥
⎣ X3 (𝜆(t) − 1) ⎦ matching the two cases is generally very small. For 20%
compression, the errors can reach 3%, but this is likely due
Note that the other boundaries were left unrestricted. Here, 𝜆 to the first oscillatory cycle, as the remaining cycles look
is the compression and 𝜓X3 is the shear angle, where better matched.
Although there is a scale factor that needs to be
𝛾𝜆3∕2 (t) sin(2𝜋f ̂t)
𝜓(t) = √ . applied to the torque simulations between ideal and non-
(36)
X12 + X22 ideal compression, this only affects the identification of
the linear parameters presented in Figs. 9 and 13. Since
The compression was set to take place linearly over the the parameter scaling norm (Eq. 30) allows for the linear
course of 0.5 s, i.e. 𝜆(t) = 1 − CS min(t, 0.5), and the angu- parameters to be adjusted according to each test, the ideal
lar displacement was applied after the desired compression compression assumption does not impact the model fit to
level was reached, i.e. ̂t = max(0, t − 0.5). the data, but it introduces additional uncertainty on the
Note that, while ideal compression was assumed in linear parameter 𝛿 . Note that, as the torque scaling factor
order to ensure rapid model evaluation, correction to is approximately constant across shear strain (for com-
non-ideal compression could be carried out for each case pression strain 1% and 10%) or increases (for the CS 20%
individually. However, the non-ideal model is time-con- group), consideration of non-ideal compression would
suming and costly, especially when analysing different not have resolved the shear strain softening phenomena
model forms over a broad parameter space. observed in the study, as shown in Fig. 14. Overall, it can
be concluded that the exponential model (18) with the
Appendix 2.1: Results comparing ideal nonlinear parameters presented in Fig. 9 and the linear
and non‑ideal compression parameters presented in Fig. 14 is suitable for capturing
the constitutive behaviour of liver tissue.
The liver under ideal and non-ideal 10% compression is
illustrated in Fig. 16. It can be seen that, under non-ideal
13
Fig. 16 Simulations of liver sample compression; (left) liver in undeformed state; (middle) liver under 10% ideal compression; (right) liver
under 10% non-ideal compression
Fig. 17 Normal force simulations during ideal and non-ideal com- The black curves show the non-ideal compression, the blue curves
pression. The simulations were carried out at CS 1% (top row), 10% show the ideal compression, and the red curves show the scaled ideal
(second row) and 20% (third row), and shear strain 1% (first column), compression, using the factors presented in Table 3
10% (second column), 25% (third column) and 50% (fourth column).
Fig. 18 Torque simulations during ideal and non-ideal compression. black curves show the non-ideal compression, the blue curves show
The simulations were carried out at CS 1% (top row), 10% (second the ideal compression, and the red curves show the scaled ideal com-
row) and 20% (third row), and shear strain 1% (first column), 10% pression, using the factors presented in Table 2
(second column), 25% (third column) and 50% (fourth column). The
13
Fig. 19 Lissajous curves of torque simulations during ideal and non- (fourth column). The black curves show the non-ideal compression,
ideal compression. The simulations were carried out at CS 1% (top the blue curves show the ideal compression, and the red curves show
row), 10% (second row) and 20% (third row), and shear strain 1% the scaled ideal compression, using the factors presented in Table 2
(first column), 10% (second column), 25% (third column) and 50%
by transient elastography: comparison with dynamic mechanical

Table 3 Multiplying factors between ideal and non-ideal torque and
analysis. Biorheology 48(2):75–88
normal force measurements
Cheng S, Clarke EC, Bilston LE (2009) The effects of preconditioning
Force SS 1% SS 10% SS 25% SS 50% strain on measured tissue properties. J Biomech 42(9):1360–1362
Chui C, Kobayashi E, Chen X, Hisada T, Sakuma I (2004) Com-
CS 1% 6.5821 6.0163 4.0603 2.2828 bined compression and elongation experiments and non-linear
0.2199% 6.4098% 21.9181% 29.6016% modelling of liver tissue for surgical simulation. Med Biol Eng
Comput 42(6):787–798
CS 10% 15.5741 15.3684 14.2890 11.6911 Clin J, Aubin CE, Parent S, Sangole A, Labelle H (2010) Com-
8.6189% 8.1965% 6.8576% 13.1985% parison of the biomechanical 3D efficiency of different brace
CS 20% 122.6414 121.2497 113.4665 93.5268 designs for the treatment of scoliosis using a finite element
16.0961% 15.6841% 14.1744% 8.3436% model. Eur Spine J 19(7):1169–1178
DiSilvestro MR, Zhu Q, Suh JF (2001) Biphasic poroviscoelastic
simulation of the unconfined compression of articular cartilage:
II–effect of variable strain rates. J Biomech Eng 123(2):198–200
Fovargue D, Nordsletten D, Sinkus R (2018) Stiffness reconstruc-
References tion methods for MR elastography. NMR Biomed 31(10):e3935
Gao Z, Lister K, Desai JP (2010) Constitutive modeling of liver tis-
Asbach P, Klatt D, Hamhaber U, Braun J, Somasundaram R, Hamm sue: experiment and theory. Ann Biomed Eng 38(2):505–516
B, Sack I (2008) Assessment of liver viscoelasticity using mul- Gent Alan (2012) Engineering with rubber, 3rd edn. Hanser Publish-
tifrequency MR elastography. Magn Reson Med 60(2):373–379 ers, Munich
Ayyildiz M, Cinoglu S, Basdogan C (2015) Effect of normal com- Gonzalez-Blohm SA, Doulgeris JJ, Lee WE, Shea TM, Aghayev K,
pression on the shear modulus of soft tissue in rheological Vrionis FD (2015) The current testing protocols for biomechani-
measurements. J Mech Behav Biomed Mater 49:235–243 cal evaluation of lumbar spinal implants in laboratory setting: a
Bonet J, Wood RD (2008) Nonlinear continuum mechanics for review of the literature. Biomed Res Int 1–15:2015
finite element analysis, 2nd edn. Cambridge University Press, Holm S, Sinkus R (2010) A unifying fractional wave equa-
Cambridge tion for compressional and shear waves. J Acoust Soc Am
Capilnasiu A, Hadjicharalambous M, Fovargue D, Patel D, Holub O, 127(3):1828–1828
Bilston L, Screen H, Sinkus R, Nordsletten D (2019) Magnetic Huwart L, Peeters F, Sinkus R, Annet L, Salameh N, ter Beek LC,
resonance elastography in nonlinear viscoelastic materials under Horsmans Y, Van Beers BE (2006) Liver fibrosis: non-invasive
load. Biomech Model Mechanobiol 18(1):111–135 assessment with MR elastography. NMR Biomed 19(2):173–179
Caputo M (1967) Linear models of dissipation whose Q is Jordan P, Socrate S, Zickler TE, Howe RD (2009) Constitutive mod-
almost Frequency Independent II. Geophys J R Astron Soc eling of porcine liver in indentation using 3D ultrasound imag-
13(5):529–539 ing. J Mech Behav Biomed Mater 2(2):192–201
Chatelin S, Oudry J, Périchon N, Sandrin L, Allemann P, Soler L, Jordan P, Kerdok AE, Howe RD, Socrate S (2011) Identifying a
Willinger R (2011) In vivo liver tissue mechanical properties minimal rheological configuration: a tool for effective and
13
efficient constitutive modeling of soft tissues. J Biomech Eng Rosen J, Brown JD, De S, Sinanan M, Hannaford B (2008) Biome-
133(4):041006 chanical properties of abdominal organs in vivo and postmortem
Kang N, Lee MW, Rhee T (2012) Simulating liver deformation dur- under compression loads. J Biomech Eng 130(2):021020
ing respiration using sparse local features. IEEE Comput Graph Shen ZL, Kahn H, Ballarini R, Eppell SJ (2011) Viscoelastic properties
Appl 32(5):29–38 of isolated collagen fibrils. Biophys J 100(12):3008–3015
Kemper AR, Santago AC, Stitzel JD, Sparks JL, Duma SM (2010) Sinkus R, Lambert S, Abd-Elmoniem KZ, Morse C, Heller T, Guenth-
Biomechanical response of human liver in tensile loading. Ann ner C, Ghanem AM, Holm S, Gharib AM (2018) Rheological
Adv Automot Med 54:15–26 determinants for simultaneous staging of hepatic fibrosis and
Kiss MZ, Varghese T, Hall TJ (2004) Viscoelastic characterization of inflammation in patients with chronic liver disease. NMR Biomed
in vitro canine tissue. Phys Med Biol 49(18):4207–4218 31(10):e3956
Lee J, Cookson A, Roy I, Kerfoot E, Asner L, Vigueras G, Sochi Spilker RL, Suh JK, Mow VC (1990) Effects of friction on the uncon-
T, Deparis S, Michler C, Smith NP, Nordsletten DA (2016) fined compressive response of articular cartilage: a finite element
Multiphysics computational modeling in CHeart. SIAM J Sci analysis. J Biomech Eng 112(2):138–146
Comput 38(3):C150–C178 Taber LA (2004) Nonlinear theory of elasticity: applications in bio-
Liu Z, Bilston L (2000) On the viscoelastic character of liver tissue: mechanics, vol 25. World Scientific Publishing Co., Singapore
experiments and modelling of the linear behaviour. Biorheology Tan K, Cheng S, Jugé L, Bilston LE (2013) Characterising soft tissues
37(3):191–201 under large amplitude oscillatory shear and combined loading. J
Liu Z, Bilston LE (2006) Large deformation shear properties of liver Biomech 46(6):1060–1066
tissue. Biorheology 39(2002):735–742 Veronda DR, Westmann RA (1970) Mechanical characterization of
Marescaux J, Clément JM, Tassetti V, Koehl C, Cotin S, Russier Y, skin-finite deformations. J Biomech 3(1):111–24
Mutter D, Delingette H, Ayache N (1998) Virtual reality applied Viano DC, King AI, Melvin JW, Weber K (1989) Injury biomechan-
to hepatic surgery simulation: the next revolution. Ann Surg ics research: an essential element in the prevention of trauma. J
228(5):627–634 Biomech 22(5):403–417
Mattei G, Magliaro C, Pirone A, Ahluwalia A (2018) Bioinspired liver Wang BC, Wang GR, Yan DH, Lid YP (1992) An experimental study
scaffold design criteria. Organogenesis 14(3):129–146 on biomechanical properties of hepatic tissue using a new measur-
Mendis KK, Stalnaker RL, Advani SH (1995) A constitutive relation- ing method. Technical report 3
ship for large deformation finite element modeling of brain tissue. Wu JZ, Dong RG, Schopper AW (2004) Analysis of effects of friction
J Biomech Eng 117(3):279–285 on the deformation behavior of soft tissues in unconfined com-
Miller K (2000) Constitutive modelling of abdominal organs. J Bio- pression tests. J Biomech 37(1):147–155
mech 33(3):367–373 Yang M, Taber LA (1991) The possible role of poroelasticity in the
Miller Karol (2005) Method of testing very soft biological tissues in apparent viscoelastic behavior of passive cardiac muscle. J Bio-
compression. J Biomech 38(1):153–158 mech 4(7):587–597
Mueller S, Millonig G, Sarovska L, Friedrich S, Reimann FM, Pritsch Yang B, Li L, Yang C, Li R, Wang J (2019) Measuring viscoelastic
M, Eisele S, Stickel F, Longerich T, Schirmacher P, Seitz HK properties of living cells. Acta Mech Solida Sin 32(5):599–610
(2010) Increased liver stiffness in alcoholic liver disease: dif- Yoganandan N, Pintar FA, Gennarelli TA, Maltese MR (2000) Pat-
ferentiating fibrosis from steatohepatitis. World J Gastroenterol terns of abdominal injuries in frontal and side impacts. Annu Proc
16(8):966–972 Assoc Adv Automot Med 44:17–36
Nicolle S, Vezin P, Palierne JF (2010) A strain-hardening bi-power law Zakerzadeh R, Bukac M, Zunino P (2016) Computational analysis of
for the nonlinear behaviour of biological soft tissues. J Biomech energy distribution of coupled blood flow and arterial deforma-
43(5):927–932 tion. Int J Adv Eng Sci Appl Math 8(1):70–85
Nicolle S, Palierne J-F (2015). Comportement viscoélastique non liné- Zhang G, Long M, Zhe-Zhi W, Wei-Qun Y (2002) Mechanical prop-
aire des organes pleins de l’abdomen. In: Congrès Français de erties of hepatocellular carcinoma cells. World J Gastroenterol.
Mécanique 8(2):243–246
O’Toole RV, Jaramaz B, DiGioia AM, Visnic CD, Reid RH (1995) Zobitz ME, Luo Z-P, An K-N (2001) Determination of the compres-
Biomechanics for preoperative planning and surgical simulations sive material properties of the supraspinatus tendon. J Biomech
in orthopaedics. Comput Biol Med 25(2):183–191 Eng 123(1):47
Perepelyuk M, Chin L, Cao X, Van OA, Shenoy VB, Janmey PA, Wells
RG (2016) Normal and fibrotic rat livers demonstrate shear strain Publisher’s Note Springer Nature remains neutral with regard to
softening and compression stiffening: a model for soft tissue jurisdictional claims in published maps and institutional affiliations.
mechanics. PLoS One 11(1):e0146588
Roan E, Vemaganti K (2007) The nonlinear material properties of liver
tissue determined from no-slip uniaxial compression experiments.
J Biomech Eng 129(3):450–456
13

Nonlinear Viscoelastic Constitutive Model For Bovine Liver Tissue

Uploaded by

Copyright:

Available Formats

Nonlinear Viscoelastic Constitutive Model For Bovine Liver Tissue

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Nonlinear Viscoelastic Constitutive Model For Bovine Liver Tissue

Uploaded by

Copyright:

Available Formats

Biomechanics and Modeling in Mechanobiology (2020) 19:1641–1662

Nonlinear viscoelastic constitutive model for bovine liver tissue

Keywords Liver rheology · Biomechanics · Nonlinear mechanics · Viscoelasticity

1 Introduction in order to assess the presence of disease (Fovargue et al.

U ∶ 𝛺0 × [0, T] → ℝ3 . A point in the reference domain, t

Fig. 1 Illustration of the experi-

⎛ cos 𝛹 (t, X3 ) sin 𝛹 (t, X3 ) 2.3.1 Viscoelastic modified Mooney–Rivlin model

using a viscoelastic adaption of three hyperelastic models Sbe = 𝜆b−1

perform significantly better than a single-component model 2.4.1 Minimisation problem

tional-order derivative 𝛼 and the eigenvalue power b (unit-

𝜃 ∗ = arg min J∗ (y∗ , 𝝉 d ),

which make up normalised matrix A∗ and vector b∗. Here, ni

𝛽i = arg min ||||A∗i x∗ − b∗i 𝛽i ||||2 . (29)

Fig. 3 The minimum error for

CS 1% 10% 20% 10% 10% mean

presents details on the differences observed between ideal

When comparing the three models (vMR∗ , vOG and vEXP)

Fig. 6 Minimum error for

CS 1% 10% 20% 10% 10% mean

Fig. 13 Absolute linear param-

Table 2 Multiplying factors between ideal and non-ideal torque and

CS 1% 1.0050 1.0046 1.0031 1.0000

Fig. 14 Absolute linear parameters for the vEXP model, correspond-

nonlinear parameters. Furthermore, the vOG model relies 4.4 Study limitations

α b1 b2 C1∗ C2∗ δ1∗ δ2∗

by transient elastography: comparison with dynamic mechanical

You might also like

⎛ cos 𝛹 (t, X3 ) sin 𝛹 (t, X3 ) 2.3.1 Viscoelastic modified Mooney–Rivlin model

perform significantly better than a single-component model 2.4.1 Minimisation problem

nonlinear parameters. Furthermore, the vOG model relies 4.4 Study limitations