TYPE Original Research

PUBLISHED 26 August 2022

DOI 10.3389/fphys.2022.953702

Fast prediction of blood flow in

OPEN ACCESS stenosed arteries using machine
EDITED BY
Jung Hee Seo,
Johns Hopkins University, United States
learning and immersed
REVIEWED BY
Huseyin Enes Salman,
boundary-lattice Boltzmann
TOBB University of Economics and
Technology, Turkey
Yang Liu,
method
Hong Kong Polytechnic University,
Hong Kong SAR, China Li Wang, Daoyi Dong and Fang-Bao Tian*
*CORRESPONDENCE
Fang-Bao Tian, School of Engineering and Information Technology, University of New South Wales, Canberra, ACT,
[email protected] Australia

SPECIALTY SECTION
This article was submitted to
Computational Physiology and
Medicine, A fast prediction of blood flow in stenosed arteries with a hybrid framework of
a section of the journal
Frontiers in Physiology
machine learning and immersed boundary-lattice Boltzmann method (IB–LBM)
is presented. The integrated framework incorporates the immersed boundary
RECEIVED 26 May 2022
ACCEPTED 25 July 2022 method for its excellent capability in handling complex boundaries, the multi-
PUBLISHED 26 August 2022 relaxation-time LBM for its efficient modelling for unsteady flows and the deep
CITATION neural network (DNN) for its high efficiency in artificial learning. Specifically, the
Wang L, Dong D and Tian F-B (2022), stenosed artery is modelled by a channel for two-dimensional (2D) cases or a
Fast prediction of blood flow in
stenosed arteries using machine tube for three-dimensional (3D) cases with a stenosis approximated by a fifth-
learning and immersed boundary- order polynomial. An IB–LBM is adopted to obtain the training data for the DNN
lattice Boltzmann method.
which is constructed to generate an approximate model for the fast flow
Front. Physiol. 13:953702.
doi: 10.3389/fphys.2022.953702 prediction. In the DNN, the inputs are the characteristic parameters of the
COPYRIGHT
stenosis and fluid node coordinates, and the outputs are the mean velocity and
© 2022 Wang, Dong and Tian. This is an pressure at each node. To characterise complex stenosis, a convolutional
open-access article distributed under
neural network (CNN) is built to extract the stenosis properties by using the
the terms of the Creative Commons
Attribution License (CC BY). The use, data generated by the aforementioned polynomial. Both 2D and 3D cases
distribution or reproduction in other (including 3D asymmetrical case) are constructed and examined to
forums is permitted, provided the
original author(s) and the copyright
demonstrate the effectiveness of the proposed method. Once the DNN
owner(s) are credited and that the model is trained, the prediction efficiency of blood flow in stenosed arteries
original publication in this journal is is much higher compared with the direct computational fluid dynamics
cited, in accordance with accepted
academic practice. No use, distribution simulations. The proposed method has a potential for applications in clinical
or reproduction is permitted which does diagnosis and treatment where the real-time modelling results are desired.
not comply with these terms.

KEYWORDS

stenosis, immersed boundary-lattice Boltzmann method, machine learning, deep

neural network, convolutional neural network

Introduction
Coronary artery disease is one of the major health threats in the 21st century, and is
reported as one of the leading causes of death worldwide (World Health Organization,
2014). This disease, where a localised accumulation of plaque in the arteries surrounding
the heart, prevents sufficient blood supply to the heart muscle, can lead to ischemia,

Frontiers in Physiology 01 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

stroke, heart attacks, and ultimately death. The hemodynamics of (Lee et al., 2020) and patient quality-of-life after prostate
the blood flow has long been thought to play an important role in radiation therapy (Yang et al., 2020). An ML based model was
the pathogenesis and pathophysiology of atherosclerosis (Saxena applied to the diagnosis of coronary artery diseases by using the
et al., 2019). fraction flow reserve, without providing the detailed information
The cardiovascular system typically features low Reynolds of the flow field (Coenen et al., 2018). A physics-informed neural
number pulsatile flow due to the cyclic pumping motion of the network was introduced by Kissas et al. (2020) to solve
heart. Computational fluid dynamics (CFD) has proven to be an conservation laws in graph topologies to predict arterial blood
effective method to uncover the fluid dynamics of blood flow. For pressure from non-invasive unsteady flow MRI data. Husso et al.
example, Tu et al. (1992) studied the Newtonian blood flow through (2021) introduced an ML to estimate myocardial blood flows
a stenosis by using finite element simulations focusing on the effects from tissue impulse response signal in an animal model. A
of various degrees of stenosis, stricture length, Reynolds number and physical-informed neural netw006Frk was used by Arzani
Womersley number. Köhler et al. (2001) compared the results of et al. (2021) to obtain the near-wall hemodynamics and wall
blood flow in the realistic model of the human carotid bifurcation shear stress data from sparse velocity measurements and without
from CFD simulations with that from magnetic resonance imaging knowledge of the inlet/outlet boundary conditions. These
(MRI) measurements finding that CFD is a reliable tool for the flow applications inspire us to create a highly efficient predicting
prediction. A detailed numerical study of the blood flow through a tool for the accurate and detailed flow dynamics prediction in the
localized stenosis in an idealized 2D blood vessel was conducted by stenosed arteries by using machine learning and CFD, which can
Tian et al. (2013). Abuouf et al. (2020) studied the effects of be used not only for diseases diagnosis, but also for the treatment.
guidewire position on the measurements of fractional flow The comprehensive details of the flow properties provide more
reserve (FFR) by using CFD. In addition, the fluid-structure robust impressions of the disease and therefore better treatment
interaction of collapse tubes has been considered as a model of strategies. Deep neural network (DNN) has shown powerful
blood-induced deformation of arteries (see e.g., Tang et al. (2015); capability in various applications (e.g., LeCun et al. (2015);
Wang et al. (2021)). However, the high fidelity CFD simulations are Ren et al. (2018); Liang et al. (2020)). Compared with the
usually time consuming, especially for 3D arterial flows involving popular reduced-order method such as proper orthogonal
complex geometries, making CFD impractical for the clinical decomposition (Guibert et al., 2014) and dynamic mode
diagnosis and treatment where the real-time modelling results are decomposition (Habibi et al., 2020), a DNN is expected to
required. To address this challenge, an efficient tool without the provide higher fidelity results due to their powerful training
assistance of professional fluid dynamics knowledge is desired, capability which could be comparable with high-fidelity CFD
which is the focus of this paper. simulations. In this work, we explore the powerful capability of
Among several techniques to provide prediction of blood DNN for fast predicting the blood flows in stenosed arteries
flow, machine learning (ML) is a promising method which has following several relevant studies (Coenen et al., 2018; Liang
drawn considerable attentions recently in the field of CFD. It has et al., 2020; Arzani et al., 2021; Husso et al., 2021).
been successfully applied to solve a series of physical problems. This paper aims at providing accurate and detailed flow
For example, the data-driven machine learning algorithm has properties in stenosed arteries by using an ML, which can be
been adopted for turbulent flow modelling to obtain a closure used for both the stenosis diagnosis and design of treatment
subgrid scale stress model for the large eddy simulation or strategies. To construct the predicting tool based on the ML, an
Reynolds-averaged Navier-Stokes equations (Wang et al., idealised function, which could be arbitrary options with
2018; Zhu et al., 2019; Sun et al., 2022). A convolutional stochastic disturbance, is adopted to generate a large number
neural network was introduced by Mao et al. (2018) to of stenosis profiles. The learning datasets, the mean flow fields of
predict the unsteady wave forces on bluff bodies due to the the constructed stenosed arteries are obtained by using an
free-surface wave motion. Bukka et al. (2021) presented a immersed boundary-lattice Boltzmann method (IB-LBM)
recurrent neural network to construct a reduced-order model (Tian et al., 2011; Wang and Tian, 2018; Xu et al., 2018; Ma
for the unsteady flow field and fluid-structure interaction. An ML et al., 2020) which incorporates the immersed boundary method
method based on convolutional neural netwrok (CNN) was for its excellent capability in handling complex boundaries and
proposed by Jing et al. to predict the unsteady velocity field the LBM for its efficient modelling for unsteady flows (Wang
around a circular cylinder from the pressure coefficients et al., 2022). Finally, a DNN is trained and tested for the fast
measurements (Jin et al., 2018). Sekar et al. (2019) presented a prediction of the mean blood flow in stenosed arteries. Inspired
data-driven approach by using the combination of CNN and by the CNN-based feature extraction of various airfoils (Sekar
multilayer perceptron for the prediction of laminar flow around et al., 2019), a CNN is also built to extract the stenosis features,
NACA airfoil, which achieves the flow prediction of various which is used to represent the complex stenosis as the input of the
airfoils. ML has also been combined with CFD tools to study fish DNN. To the best of our knowledge, this is the first attempt to
swimming (Zhu et al., 2021, 2022). Moreover, ML has been predict the comprehensive flow properties in stenosed arteries by
studied in the field of biomedicine, such as prediction of malaria using a DNN model with the IB-LBM.

Frontiers in Physiology 02 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 1
Schematic of the stenosed arteries: (A) 2D asymmetric case, and (B) 3D axis-symmetric case.

The rest of this paper is organised as follows: the ML method y
a1 x5 + a2 x4 + a3 x3 + a4 x2 + a5 x + a6 , (1)

and IB-LBM used in this study are presented in Section 2; both
2D and 3D DNNs are trained and tested, with the results and with y (0) = y(D) = D, y′(0) = y′(D) = 0 and D being the diameter
discussions being presented in Section 3; final conclusions are of the artery at the inlet, as shown in Figure 1. The free
given in Section 4. parameters in Eq. 1 are randomly selected to generate a
database for the learning algorithm. Both axis-symmetric and
asymmetric 3D arteries are considered here. The axis-symmetric
Hybrid framework of machine artery has a similar stenosis governed by Eq. 1 as shown in
learning and immersed boundary- Figure 1, so that the 3D stenosis shares the same characteristic
lattice Boltzmann method parameters with its 2D counterpart. The asymmetric arteries will
be described later in. For simplicity, D is used as the length of the
The present approach includes three main parts, i.e., the stenosis. Although Eq. 1 is an idealised model, these free
characterisation of the stenosed artery, the IB-LBM solver for parameters can be arbitrary options with stochastic
the high-fidelity CFD simulations of the blood flows in disturbance. Therefore, the idealised model does not affect the
stenosed arteries, and the DNN for predicting the mean effectiveness of the method in patient-specific cases which could
flow fields. Specifically, the stenosis is first characterised be modelled with more parameters.
by either the polynomial factors or the CNN extracted To represent the stenosis shape, both the coordinates of the
features; then high-fidelity CFD simulations are conducted stenosis and the free parameters in Eq. 1 can be used as the input
by using the IB-LBM solver to obtain datasets, and finally an of the DNN for the mean flow prediction. However, the large
approximate DNN is trained by using these datasets. number of stenosis coordinates will definitely make the DNN
more complex. There are only six factors in Eq. 1, which seems to
be a good representation of the stenosis, and will be discussed
Characterization of the stenosed artery later. As the realistic stenosis could be even more complex, a
CNN (Sekar et al., 2019) is also built to extract the geometric
Stenosis is usually caused by the accumulation of lipids in features of stenosis. Here, the CNN will be trained based on the
the intima of artery, and it could have different shapes for dataset generated by using Eq. 1, which can be further extended
various patients and arteries (Varghese et al., 2007a,b; Huang for an arbitrarily complex stenosis. The current CNN includes an
et al., 2020). Here, a fifth-order polynomial is used to describe input layer, 3 convolutional layers, 5 fully connected layers and
the stenosis, i.e, an output layer. Rectified linear unit (ReLU) and the hyperbolic

Frontiers in Physiology 03 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

TABLE 1 Details of the 2D CNN.

Layer Size Filter size Pooling filter Output shape

Input 214 × 214 × 1 - - -

Conv1 32 4×4 3×3 214 × 214 × 32
Conv2 64 4×4 3×3 71 × 71 × 64
Conv3 128 4×4 3×3 23 × 23 × 128
Fully Connected × 2 128 - - 128
Fully Connected (extracted features) 16 - - 16
Fully Connected × 2 128 - - 128
Output - - - 201

FIGURE 2
An example of image generation for the input of the 2D CNN.

tangent functions are used as the activation functions for the artery is transferred into the image as the rest part of the artery is
convolutional and fully connected layers, respectively. The uniform. The convolutional layers and the first three fully-
parameters of the CNN are listed in Table 1. In the 2D CNN, connected layers are used to decode the training images into
the inputs are 2D images (an example is shown in Figure 2), 16 parameters representing the stenosis features, while the rest
which is generated by filling the pixels crossed by the stenosis as fully connected layers then encode the features to obtain the y-
one with the rest of the pixels as zero, and each image has a coordinates of the stenosis. The loss function of the CNN is
resolution of 214 × 214. Here, only the stenosis section of the defined as

Frontiers in Physiology 04 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

method is briefly reviewed in this section, and more details of

this method and its 3D version can be found in Refs. (Tian et al.,
2011; Wang and Tian, 2018; Xu et al., 2018; Ma et al., 2020). The
main reason of choosing IB–LBM is due to its high
computational efficiency and simplicity in handling arbitrarily
complex geometries, which are especially desirable for the large
data requirements in ML and more complex stenosed arteries.
In the multiple relaxation time (MRT)-based IB-LBM, the
evolution equation of the particle distribution function gi along
the i-th direction at position x is expressed as (He and Luo, 1997;
Guo and Zheng, 2008)
FIGURE 3 gi (x + ei Δt, t + Δt)
gi (x, t) − Ωi (x, t) + ΔtGi , (3)
Velocity profile in a 3D cylindrical pipe measured at a distance
of 5D from the inlet.
where i = 0, 1, . . . , 8, Δt is the time step, ei is the lattice speed, Ωi is
the collision operator, and Gi represents the body force effects on
the distribution function. Ωi and Gi are defined as

Ωi
−M−1 SMij gj (x, t) − gj (x, t),

eq
(4)
S
Gi

M−1 I − M Fj , (5)
2 ij

Where M is a 9, ×, 9 transform matrix for the two dimensional

nine-speed (D2Q9) model, and S is a non-negative diagonal
matrix. The details for the determination of S and M can be found
in Ref. (Peng et al., 2006). The lattice speed ei in 2D simulation is
defined as
⎪
⎧
⎪ (0, 0), i
0
⎪
⎪
⎪
⎪ Δx
⎨
ei
⎪ (cos(0.5π (i − 1)), sin(0.5π (i − 1))) Δt , i
1, 2, 3, 4 (6)
⎪
⎪
⎪
⎪ √ Δx
⎪
⎩ 2 (cos(0.5π (i − 4.5)), sin(0.5π (i − 4.5))) , i
5, 6, 7, 8
Δt
FIGURE 4
Architecture of the DNN. where Δx is the lattice spacing. The macro density and
momentum are given as follows,
8 8
1
1 2 ρ
 gi , ρu
 gi ei + f Δt. (7)
yloss
Σ
ypred − ylabel  , (2) i
0 i
0
2
N
eq
where N is the sample number, and ypred and ylabel are the The local equilibrium distribution function gi and the force
predicted and labelled y-coordinates, respectively. Further term Fi are calculated by
details of the CNN can be found in Refs. (Lee and You, 2019; eq ei · u uu: ei ei − c2s I 
gi
ωi ρ
1 + +  (8)
Sekar et al., 2019; Hasegawa et al., 2020; Bukka et al., 2021). c2s 2c4s
ei − u ei · u
Fi
ωi
+ 4 ei  · f (9)
c2s cs
Immersed boundary-lattice Boltzmann √
method solver for the computational fluid where the sound speed cs
Δx/( 3 Δt), f is the force acting on
dynamics simulation the fluid, and the weights ωi are given by ω0 = 4/9, ωi = 1/9 for
i = 1, 2, 3, 4 and ωi = 1/36 for i = 5, 6, 7, 8 in 2D domain. The
The flows in the stenosed artery are solved by using an relaxation time τ (one of the components of $S$) is related to
IB–LBM (Tian et al., 2011; Wang and Tian, 2018; Xu et al., 2018; the kinematic viscosity ] of the fluid according to
Ma et al., 2020). In this method, the fluid dynamics is solved by ]
(τ − 0.5)c2s Δt. The non-equilibrium extrapolation
the LBM, and the complex no-slip boundary conditions on the method is used for the boundary conditions at the out
fluid–structure interface are achieved by an immersed boundary boundaries of the computational domain (Guo et al., 2002).
(IB) method. Without loss of generality, the 2D numerical For 3D simulations, D3Q19 MRT-LBM is used (Wang and

Frontiers in Physiology 05 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

This is different from the realistic arterial flows that are

pulsatile. However, the conclusions obtained can be
extended to cases of higher Reynolds numbers and pulsatile
inlet boundary conditions which would generate more
complicated flow structures such as secondary vortex flows
and turbulence.
The IB-LBM solver and its previous version used here have
been validated in our previous work extensively considering
fluid–structure interaction in steady and unsteady flows
(Wang and Tian, 2018; Xu et al., 2018, 2019; Wang and Tian,
2019; Ma et al., 2020; Huang et al., 2021; Wang, 2021). To further
validate the IB-LBM solver in modelling internal flows (e.g.,
hemodynamics with immersed structures), a Newtonian flow in a
3D cylindrical pipe is conducted. In this problem, a constant inlet
velocity (U0) is applied on the left side of the computational
FIGURE 5 domain. The fluid domain has a dimension of 10D × 0.6D × 0.6D.
Comparison of the predicted and original y-coordinates of
2D stenosis. A mesh convergence study is conducted at a Reynolds numbers
of 100 with three mesh spacings of the fluid domain, i.e., Δs = D/
25, D/50 and D/100. The mesh spacing of the cylinder is roughly
half of the fluid mesh spacing. The velocity profiles at a distance
Tian, 2018). The feedback IB method (Goldstein et al., 1993; of 5D from the inlet are shown in Figure 3 along with the
Kim and Peskin, 2007; Tian et al., 2011; Huang and Tian, analytical solution
2019) is used to handle the no-slip boundaries between the 4r2
structure and fluid. The IB method is a type of Cartesian- u
2U0 1 − , (10)
D2
mesh-based methods which has excellent capability in
handling complex boundaries (Peskin, 2002; Mittal and where, r is the distance to the center of the pipe. As shown in
Iaccarino, 2005; Tian et al., 2012; Huang and Tian, 2019). Figure 3, the velocity profiles calculated by the present numerical
In this work, the Newtonian fluid flow in 2D and 3D arteries is method agree well with the exact solutions in Eq. 10. It shows that
considered at a Reynolds number (Re = U0D/], where D being the numerical results are in excellent agreement with the
the diameter of the artery and U0 being the average incoming analytical solutions when the mesh is refined to D/50. Note
velocity) of 100. It should be noted that this Reynolds number that the velocity profile is corrected by using D + 0.5Δx in the IB-
is considered, as it captures the major flow features in arteries, LBM simulations, with Δx being the mesh spacing of the fluid
while the computational cost is low. In addition, the steady domain containing the structure, as suggested in Refs. (Huang
uniform velocity boundary condition is applied at the inlet. et al., 2020, 2021).

FIGURE 6
Time histories of loss of the DNN in a 2D channel with a stenosis: (A) polynomial factors, and (B) CNN-extracted characteristic parameters.

Frontiers in Physiology 06 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 7
Comparison of the mean fluid fields for 2D channel flows from DNN prediction and CFD simulation: (A) um by CFD, (B) um by DNN prediction,
(C) vm by CFD, (D) vm by DNN prediction, (E) cp,m by CFD, (F) cp,m by DNN prediction.

Frontiers in Physiology 07 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 8
Absolute errors of the mean fluid field: (A) Δum, (B) Δvm, and (C) Δcp,m.

Deep neural network for the flow root-mean-square error is defined as the loss function for the
prediction DNN. The ReLU function is adopted as the activation function as
it is less susceptible to vanishing gradients that prevent deep
Once the characteristic parameters of the stenosis and the models from being trained compared with other functions such
mean flow fields (calculated over a dimensionless time of 50 after as Sigmoid and Tanh. The adaptive moment estimation (Adam)
the initial flow developing periods) are obtained, a DNN is then is used as the optimizer. During the training process, back
constructed to generate an approximate model for the fast propagation method is used to update the trainable
prediction of the mean flow field close to the stenosis, parameters in the NN according to the gradients of the loss
i.e., from D upstream and 4D downstream of the stenosis. function. All the NNs in this study are created by using the open
The architecture of the DNN used for the flow field source library Tensorflow (Abadi et al., 2015) because of its
prediction is shown in Figure 4. It consists of an input layer, simple implementations.
5 hidden layers (each of them has 128 perceptrons) and an output The major procedures of the present approach to predict the
layer. Here, the inputs are the stenosis characteristic parameters blood flow are as follows:
(i.e., the free factors of Eq. 1 or the characteristic parameters
extracted by the CNN) and the coordinates of fluid nodes, and 1) Generate stenosis samples by using Eq. 1, and the stenosis is
the outputs are the mean flow velocities (um, vm and wm) and then characterised by using the free parameters of Eq. 1;
pressure (cp,m) at each fluid node (the output shape is 3 for 2D 2) For the stenosis characterised by using CNN-extracted
and 4 for 3D), while the output in Ref. (Liang et al., 2020) is the features, the 2D images are first generated according to the
values of the whole flow field and thus the output shape is stenosis coordinates. The network is trained to get a model for
significantly larger and varies with the stenosis size. By using such the extraction of the stenosis features by feeding these image
an architecture, the DNN will have significantly less trainable samples to the CNN and training;
parameters due to the smaller output shape and thus may be 3) Conduct high fidelity CFD simulations of the randomly
more efficient in the training stage. Similarly to the CNN, the selected samples to generate a dataset for the DNN;

Frontiers in Physiology 08 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 9
Comparison of the mean velocities and pressure along the streamwise direction measured at y = 0.2D and the comparison of streamwise
velocity profile: (A) um along x-axis, (B) vm along x-axis, (C) cp,m along x-axis, and (D) streamwise velocity profile um along y-axis.

FIGURE 10
Histories of loss and mean absolute error of the DNN for 3D (axis-symmetric) flow prediction: (A) loss, and (B) mean absolute error.

4) Organise the dataset with the input being fluid node 5) Feed the organised dataset to the DNN and train the network
coordinates and stenosis features, and the output being the to obtain the approximate model for the fast prediction of the
mean flow velocity and pressure. blood flow field.

Frontiers in Physiology 09 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

extraction is because the arteries are usually constructed by using

MRI which can be directly used as the input of the CNN.

Results and discussions

2D flow prediction in a channel with a
stenosis

By feeding the image dataset generated by using Eq. 1 into the

CNN to characterise the arteries, a trained model can be
obtained. Here, 1,000 samples are used as the training dataset
and 200 samples are used as the validation dataset. The learning
rate is set as 1.0, ×, 10−4. Some validation samples are shown in
Figure 5. It is found that the predicted y-coordinates of the
stenosis agree well with the original data, which confirms that the
CNN is accurate in characterising the stenosis.
Here, 100 CFD simulations are conducted to obtain the training
dataset for the DNN to predict the blood flow, with another
20 simulations as the validation dataset. There are approximately
150,000 data points in each simulation, and 1.5 × 107 data points in
total are used to train the DNN. The initial learning rate is set as
1.0, ×, 10−3, and the learning rate is gradually decreased to around
4.0, ×, 10−5 in order to stabilize the learning process. Two DNNs are
trained by using the polynomial factors and the characteristic
parameters extracted by the CNN, respectively. Figure 6 shows
the time histories of the loss from the two DNNs. It is found that the
DNN using features extracted by the CNN achieves a significantly
smaller loss compared that using the polynomial factors. This
indicates that the CNN has a better ability in representing the
stenosis features. As a general approach, it also can be extended to
arbitrarily complex stenosis with the required dataset such as MRI or
CT images obtained in clinic practices.
The mean flow fields (i.e., um
u/U, vm
v/U and

cp,m
p/(0.5ρU 2
)) of one validation by using the CNN extracted
features are presented in Figure 7, which shows that the predictions
from the DNN is very close to the high fidelity CFD simulations. The
mean fluid field errors are further shown in Figure 8, which
demonstrates that the error of the predictions is mostly observed
close to the wall due to the drastic changes. The local error is less than
5% of the maximum velocity near the centre, acceptable in such fast
FIGURE 11 predictions. A more robust comparison of the mean velocities and
Comparison of the mean velocities and pressure with axis-
symmetric stenosis on the plane of z = 0: (A) um by CFD, (B) um by pressure measured along the streamwise direction at y = 0.2D are
DNN prediction, (C) vm by CFD, (D) vm by DNN prediction, (E) cp,m presented in Figure 9, which shows a good agreement of the predicted
by CFD, (F) cp,m by DNN prediction.
fluid values and those from CFD simulations with acceptable errors.

When the clinical patient-specific stenosis data are available, it 3D flow prediction in axis-symmetric
can be modelled by more parameters and fed into the CNN instead artery
of the data generated from the polynomial for the feature extraction.
It means that the current framework by combining the CNN and To demonstrate the ability of the DNN in considering more
DNN can be still effective for the patient-specific geometries. realistic blood flows, the DNN model is further trained based on 3D
Another important reason to use CNN for the stenosis feature simulation data to predict the blood flow. A similar DNN

Frontiers in Physiology 10 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 12
Comparison of the mean velocities and pressure along the streamwise direction measured on two lines in the 3D axis-symmetric case: (A) um at
z = 0 and y = 0.2D, (B) um at z = 0 and y = 0.8D, (C) vm at z = 0 and y = 0.2D, (D) vm at z = 0 and y = 0.8D, (E) cp,m at z = 0 and y = 0.2D, and (F) cp,m at
z = 0 and y = 0.8D.

Frontiers in Physiology 11 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 13
An example of 3D asymmetric stenosed artery: (A) 3D view, and (B) x-axis view.

TABLE 2 Details of the 3D CNN.

Layer Size Filter size Pooling filter Output shape

Input 61 × 61 × 61 × 1 - - -
Conv1 32 4×4×4 3×3×3 61 × 61 × 61 × 32
Conv2 64 4×4×4 3×3×3 71 × 71 × 71 × 64
Conv3 128 4×4×4 3×3×3 23 × 23 × 23 × 128
Fully Connected × 2 128 - - 128
Fully Connected (extracted features) 8/16/32 - - 8/16/32
Fully Connected × 2 128 - - 128
Fully Connected - - - 800
Reshaped - - - 400 × 2

FIGURE 14
Histories of loss and mean absolute error in the shape construction of the 3D CNN: (A) loss, and (B) mean absolute error.

Frontiers in Physiology 12 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

pressure distributions in the streamwise direction measured on

two lines, i.e., line 1 with z = 0 and y = 0.2D, and line 2 with y = 0
and z = 0.8D, is presented in Figure 12. It shows negligible
discrepancies and further confirms the accuracy of the present
machine learning approach.

3D flow prediction in asymmetric artery

The 3D flow in axis-symmetric artery presented in the last

section shows that the present method has good ability to predict
the flow properties in such stenosed arteries. However, the real
stenosed arteries are generally not symmetric, and thus the
simplified axis-symmetric model may not be applicable. Here,
to clarify the ability of the present method in predicting flow
FIGURE 15
Comparison of the predicted and original coordinates of a 3D properties in more practical stenosed arteries, we use the
asymmetric stenosis. combination of Eq. 1 and a normal distribution function to
describe the radius along the circumferential direction at the
stenosed section. Therefore, the stenosis with complex and
asymmetric shape can be represented. An example is shown
architecture of the 2D case is adopted for the 3D model, except that in Figure 13 to illustrate the asymmetric stenosed arteries
the input and output have an extra dimension. By feeding 100 3D generated by this model.
simulation data, around 1.6 × 108 data points, to the DNN, an As asymmetric arteries are considered, the established 2D CNN
approximate model is obtained by using Adams optimizer. Figure 10 is not applicable. Here, a 3D CNN is further built to extract the
shows the time histories of the loss and mean absolute error, it is stenosis features. The input of the 3D CNN is a 3D image with a
found that the training approaches to a small loss after training for dimension of 61 × 61 × 61, i.e., the resolution is 0.02D, which is
200 epochs, the mean absolute error is also small compared with the generated by filling the pixels crossed by the stenosis as one with the
absolute value of dimensionless velocities and pressure (generally in rest as zero. The coordinates of 400 points evenly distributed on the
a range more than 1.0). stenosis surface are used as the output of the 3D CNN, with the
Figure 11 shows a direct comparison of the mean flow fields mean-square-error as the loss function. ReLU and the hyperbolic
from DNN prediction and CFD simulation on the plane of z = 0. tangent functions are used as the activation functions for the
It is found that the predicted flow fields are very close to those convolutional and fully connected layers, respectively.
from high fidelity CFD simulation. The stenosis effects on the 800 examples are used as the training data with another 200 as
velocity and pressure change are well captured by the DNN. A the validation data. More details of the 3D CNN are shown in
more careful comparison of the dimensionless velocities and Table 2. Three numbers of the extracted features are tested, i.e., 8,

FIGURE 16
Histories of loss and mean absolute error of the DNN for asymmetric stenosis: (A) loss, and (B) mean absolute error.

Frontiers in Physiology 13 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

FIGURE 17
Comparison of the mean velocities and pressure with asymmetric stenosis on the plane of z = 0: (A) um by CFD, (B) um by DNN prediction, (C) vm
by CFD, (D) vm by DNN prediction, (E) wm by CFD, (F) wm by DNN prediction, (G) cp,m by CFD, (H) cp,m by DNN prediction.

TABLE 3 Details of the computation time. learning histories and good convergence, and the increase of features
from 8 to 32 does not improve the accuracy. Therefore, 8 features are
Type Wall-time
adopted for the further DNN training. An example to compare the
CNN-2D training 24 cpu hours
predicted (by using 8 features) and original coordinates of 3D
CNN-3D training 912 cpu hours
asymmetric stenosis is shown in Figure 15. It shows that the 3D
DNN-2D training 80 cpu hours
CNN model predicts the coordinates well and thus has the ability to
DNN-3D axis-symmetric training 500 cpu hours
extract the features from complex stenosed arteries.
DNN-3D asymmetric training 960 cpu hours
After the features of the stenosis are successfully extracted by the
CFD-2D 0.6 cpu hours per simulation
established 3D CNN, the DNN model built in the last section is
CFD-3D 50 cpu hours per simulation
adopted for the 3D asymmetric flow prediction. Similarly, 120 CFD
DNN-2D prediction 6s
simulations are conducted to obtain the training dataset, with 20 of
DNN-3D prediction 65 s
them serving as the validation data. The histories of loss and mean
absolute error are shown in Figure 16. It is found that the DNN
converges after around 150 epochs, the mean absolute error (MAE)
16 and 32, as an appropriate number of the stenosis feature can well is less than 1%, which is much lower than the best value (6.2%)
represent the stenosis and improve the learning efficiency of the achieved by Liang et al. (2020) considering flow in human thoracic
DNN. The histories of loss and mean absolute error versus epoch are aorta. An example of the mean velocities and pressure fields
shown in Figure 14. It is clear that all three networks show similar obtained by the DNN and CFD simulation are shown in

Frontiers in Physiology 14 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

Figure 17. It is clear that the DNN gives an excellent prediction Although the stenosis is generated by an analytical model in this work,
compared with the high fidelity CFD simulation in a significantly the versatility of the CNN makes it applicable to patient-specific
less expensive way. Almost all flow details such as the velocities and geometries that can be modelled with more parameters.
pressure change can be accurately captured. This work has successfully demonstrated the superiority to
The 2D and 3D stenosed arteries are considered, and the present use machine learning and CFD based data for the fast prediction
results show that the combination of CNN for extracting stenosis of mean flow fields in complex stenosed arteries. The future work
features and the fully connected DNN for the prediction of blood flows will be focused on more realistic problems to include the non-
works well in capturing the flow details in stenosed arteries. The success Newtonian fluid, the pulsatile flow and more complex stenosis to
of this combination can be attributed to the versatility of CNN in obtain a more versatile DNN model.
extracting features of complex images (Badrinarayanan et al., 2017),
which can be easily extended to arbitrarily complex stenosis once the
required dataset is obtained. The computation time by using full CFD Data availability statement
simulation and the DNN to predict the blood flow is further presented
in Table 3 for comparison. It shows that the blood flow in stenosed The raw data supporting the conclusions of this article will be
arteries can be obtained in a few seconds for 2D and about 1 minute for made available by the authors, without undue reservation.
3D after the model is trained. The prediction time is in seconds
excluding the time of initializing and file I/O, and it is comparable to
that of other similar methods e.g., Ref. (Liang et al., 2020). In some Author contributions
medical applications, we only need to reconstruct the wall shear stress
or the pressure drop due to the stenosis, which requires even less LW has made contributions to the conception of the work,
computational time. Therefore, the trained DNN method is methodology, software development, data analysis, and writing
significantly more efficient than the CFD simulation, and may have of the work. DD has made contributions to methodology and
a great potential in clinic applications. It should be noted that more revising of the work. F-BT has made contributions to the
training sets may be required when more factors are included. conception of the work, methodology, and revising of the work.
Specifically, if the stenosis length is varied, the training set will not
increase and the data size increases linearly. If the inlet velocity is varied,
the training dataset will increase linearly and each data will be the same Funding
size. Therefore, the training dataset will increase linearly when we
include two or more factors. New datasets may be required for cases This work was supported by the Australian Research Council
with different topologies. (project numbers DE160101098 and DP190101566), and
conducted with the assistance of resources from the National
Computational Infrastructure (NCI), which is supported by the
Conclusion Australian Government.

This paper has introduced a fast prediction method of blood flow

in stenosed arteries with a hybrid framework of machine learning and Conflict of interest
IB–LBM which incorporates the immersed boundary method, the
MRT LBM and the DNN, and takes advantages of their strengths. The authors declare that the research was conducted in the
Several validation cases have been conducted by training and testing a absence of any commercial or financial relationships that could
DNN for the fast flow prediction in stenosed arteries with the CFD be construed as a potential conflict of interest.
generated data. The results show that the DNN can predict the mean
flow fields accurately with the results being available within about
1 minute, which is improved at least 1,000 times compared with direct Publisher’s note
numerical simulation. In addition to parametrise the stenosis by using
an analytical polynomial, a 2D and a 3D CNNs are constructed to All claims expressed in this article are solely those of the authors
extract the stenosis features from 2D and 3D images, respectively. The and do not necessarily represent those of their affiliated organizations,
results show that the CNN has a good performance in representing or those of the publisher, the editors and the reviewers. Any product
the stenosis and can be straightforwardly extended to arbitrarily that may be evaluated in this article, or claim that may be made by its
complex stenosis, as demonstrated by the 3D asymmetric stenosis. manufacturer, is not guaranteed or endorsed by the publisher.

Frontiers in Physiology 15 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

References
Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., et al. (2015). Köhler, U., Marshall, I., Robertson, M. B., Long, Q., Xu, X. Y., and Hoskins, P. R.
TensorFlow: Large-scale machine learning on heterogeneous systems. [Dataset]. (2001). Mri measurement of wall shear stress vectors in bifurcation models and
Software available from tensorflow.org. comparison with cfd predictions. J. Magn. Reson. Imaging. 14, 563–573. doi:10.
1002/jmri.1220
Abuouf, Y., Ookawara, S., and Ahmed, M. (2020). Analysis of the effect of
guidewire position on stenosis diagnosis using computational fluid dynamics. LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521,
Comput. Biol. Med. 121, 103777. doi:10.1016/j.compbiomed.2020.103777 436–444. doi:10.1038/nature14539
Arzani, A., Wang, J.-X., and D’Souza, R. M. (2021). Uncovering near-wall blood Lee, S., and You, D. (2019). Data-driven prediction of unsteady flow over a
flow from sparse data with physics-informed neural networks. Phys. Fluids 33, circular cylinder using deep learning. J. Fluid Mech. 879, 217–254. doi:10.1017/jfm.
071905. doi:10.1063/5.0055600 2019.700
Badrinarayanan, V., Kendall, A., and Cipolla, R. (2017). Segnet: A deep Lee, Y. W., Choi, J. W., and Shin, E.-H. (2020). Machine learning model for
convolutional encoder-decoder architecture for image segmentation. IEEE predicting malaria using clinical information. Comput. Biol. Med. 129, 104151.
Trans. Pattern Anal. Mach. Intell. 39, 2481–2495. doi:10.1109/TPAMI.2016. doi:10.1016/j.compbiomed.2020.104151
2644615
Liang, L., Mao, W., and Sun, W. (2020). A feasibility study of deep learning for
Bukka, S. R., Gupta, R., Magee, A. R., and Jaiman, R. K. (2021). Assessment of predicting hemodynamics of human thoracic aorta. J. Biomech. 99, 109544. doi:10.
unsteady flow predictions using hybrid deep learning based reduced-order models. 1016/j.jbiomech.2019.109544
Phys. Fluids 33, 013601. doi:10.1063/5.0030137
Ma, J., Wang, Z., Young, J., Lai, J. C., Sui, Y., and Tian, F.-B. (2020). An immersed
Coenen, A., Kim, Y.-H., Kruk, M., Tesche, C., De Geer, J., Kurata, A., et al. (2018). boundary-lattice Boltzmann method for fluid-structure interaction problems
Diagnostic accuracy of a machine-learning approach to coronary computed involving viscoelastic fluids and complex geometries. J. Comput. Phys. 415,
tomographic angiography–based fractional flow reserve: result from the machine 109487. doi:10.1016/j.jcp.2020.109487
consortium. Circ. Cardiovasc. Imaging 11, e007217. doi:10.1161/CIRCIMAGING.
Mao, X., Joshi, V., Miyanawala, T. P., and Jaiman, R. K. (2018). “Data-driven
117.007217
computing with convolutional neural networks for two-phase flows: Application to
Goldstein, D., Handler, R., and Sirovich, L. (1993). Modeling a no-slip flow wave-structure interaction,” in Proceedings of the International Conference on
boundary with an external force field. J. Comput. Phys. 105, 354–366. doi:10.1006/ Offshore Mechanics and Arctic Engineering – OMAE.
jcph.1993.1081
Mittal, R., and Iaccarino, G. (2005). Immersed boundary method. Annu. Rev.
Guibert, R., Mcleod, K., Caiazzo, A., Mansi, T., Fernández, M. A., Sermesant, M., Fluid Mech. 37, 239–261. doi:10.1146/annurev.fluid.37.061903.175743
et al. (2014). Group-wise construction of reduced models for understanding and
Peng, Y., Shu, C., Chew, Y.-T., Niu, X., and Lu, X.-Y. (2006). Application of multi-
characterization of pulmonary blood flows from medical images. Med. Image Anal.
block approach in the immersed boundary–lattice Boltzmann method for viscous
18, 63–82. doi:10.1016/j.media.2013.09.003
fluid flows. J. Comput. Phys. 218, 460–478. doi:10.1016/j.jcp.2006.02.017
Guo, Z., and Zheng, C. (2008). Analysis of lattice Boltzmann equation for
Peskin, C. S. (2002). The immersed boundary method. Acta Numer. 11, 479–517.
microscale gas flows: relaxation times, boundary conditions and the knudsen
doi:10.1017/s0962492902000077
layer. Int. J. Comput. Fluid Dyn. 22, 465–473. doi:10.1080/10618560802253100
Ren, Z., Dong, D., Chen, C., Li, H., Ren, Z., Dong, D., et al. (2018). Self-paced
Guo, Z., Zheng, C., and Shi, B. (2002). Non-equilibrium extrapolation method for
prioritized curriculum learning with coverage penalty in deep reinforcement
velocity and pressure boundary conditions in the lattice Boltzmann method. Chin.
learning. IEEE Trans. Neural Netw. Learn. Syst. 29, 2216–2226. doi:10.1109/
Phys. 11, 366–374. doi:10.1088/1009-1963/11/4/310
TNNLS.2018.2790981
Habibi, M., Dawson, S., and Arzani, A. (2020). Data-driven pulsatile blood flow
Saxena, A., Ng, E. Y. K., and Lim, S. T. (2019). Imaging modalities to diagnose
physics with dynamic mode decomposition. Fluids 5, 111. doi:10.3390/
carotid artery stenosis: progress and prospect. Biomed. Eng. Online 18, 66–23.
fluids5030111
doi:10.1186/s12938-019-0685-7
Hasegawa, K., Fukami, K., Murata, T., and Fukagata, K. (2020). Machine-
Sekar, V., Jiang, Q., Shu, C., and Khoo, B. C. (2019). Fast flow field prediction over
learning-based reduced-order modeling for unsteady flows around bluff bodies
airfoils using deep learning approach. Phys. Fluids 31, 057103. doi:10.1063/1.
of various shapes. Theor. Comput. Fluid Dyn. 34, 367–383. doi:10.1007/s00162-020-
5094943
00528-w
Sun, X., Cao, W., Liu, Y., Zhu, L., and Zhang, W. (2022). High reynolds number
He, X., and Luo, L.-S. (1997). Theory of the lattice Boltzmann method: From the
airfoil turbulence modeling method based on machine learning technique. Comput.
Boltzmann equation to the lattice Boltzmann equation. Phys. Rev. E 56, 6811–6817.
Fluids 236, 105298. doi:10.1016/j.compfluid.2021.105298
doi:10.1103/physreve.56.6811
Tang, C., Zhu, L., Akingba, G., and Lu, X.-Y. (2015). Viscous flow past a
Huang, W.-X., and Tian, F.-B. (2019). Recent trends and progress in the
collapsible channel as a model for self-excited oscillation of blood vessels.
immersed boundary method. Proc. Inst. Mech. Eng. Part C J. Mech. Eng. Sci.
J. Biomech. 48, 1922–1929. doi:10.1016/j.jbiomech.2015.04.011
233, 7617–7636. doi:10.1177/0954406219842606
Tian, F.-B., Luo, H., Zhu, L., Liao, J. C., and Lu, X.-Y. (2011). An efficient
Huang, Q., Tian, F.-B., Young, J., and Lai, J. C. (2020). “A diffused interface
immersed boundary-lattice Boltzmann method for the hydrodynamic
immersed boundary–lattice Boltzmann method for simulation of channel flow,” in
interaction of elastic filaments. J. Comput. Phys. 230, 7266–7283. doi:10.
22nd Australasian Fluid Mechanics Conference (The University of Queensland).
1016/j.jcp.2011.05.028
Huang, Q., Tian, F.-B., Young, J., and Lai, J. C. (2021). Transition to chaos in a
Tian, F., Lu, X., and Luo, H. (2012). Onset of instability of a flag in uniform flow.
two-sided collapsible channel flow. J. Fluid Mech. 926, A15. doi:10.1017/jfm.
Theor. Appl. Mech. Lett. 2, 022005. doi:10.1063/2.1202205
2021.710
Tian, F.-B., Zhu, L., Fok, P.-W., and Lu, X.-Y. (2013). Simulation of a pulsatile
Husso, M., Afara, I., Nissi, M., Kuivanen, A., Halonen, P., Tarkia, M., et al. (2021).
non-newtonian flow past a stenosed 2d artery with atherosclerosis. Comput. Biol.
Quantification of myocardial blood flow by machine learning analysis of modified
Med. 43, 1098–1113. doi:10.1016/j.compbiomed.2013.05.023
dual bolus MRI examination. Ann. Biomed. Eng. 49, 653–662. doi:10.1007/s10439-
020-02591-0 Tu, C., Deville, M., Dheur, L., and Vanderschuren, L. (1992). Finite element
simulation of pulsatile flow through arterial stenosis. J. Biomech. 25, 1141–1152.
Jin, X., Cheng, P., Chen, W.-L., and Li, H. (2018). Prediction model of velocity
doi:10.1016/0021-9290(92)90070-h
field around circular cylinder over various reynolds numbers by fusion
convolutional neural networks based on pressure on the cylinder. Phys. Fluids Varghese, S. S., Frankel, S. H., and Fischer, P. F. (2007a). Direct numerical
30, 047105. doi:10.1063/1.5024595 simulation of stenotic flows. part 1. steady flow. J. Fluid Mech. 582, 253–280. doi:10.
1017/s0022112007005848
Kim, Y., and Peskin, C. S. (2007). Penalty immersed boundary method for an
elastic boundary with mass. Phys. Fluids 19, 053103. doi:10.1063/1.2734674 Varghese, S. S., Frankel, S. H., and Fischer, P. F. (2007b). Direct numerical
simulation of stenotic flows. part 2. pulsatile flow. J. Fluid Mech. 582, 281–318.
Kissas, G., Yang, Y., Hwuang, E., Witschey, W. R., Detre, J. A., and Perdikaris, P.
doi:10.1017/s0022112007005836
(2020). Machine learning in cardiovascular flows modeling: Predicting arterial
blood pressure from non-invasive 4D flow MRI data using physics-informed neural Wang, L., and Tian, F.-B. (2018). Heat transfer in non-Newtonian flows by a
networks. Comput. Methods Appl. Mech. Eng. 358, 112623. doi:10.1016/j.cma.2019. hybrid immersed boundary–lattice Boltzmann and finite difference method. Appl.
112623 Sci. 8, 559. doi:10.3390/app8040559

Frontiers in Physiology 16 frontiersin.org

Wang et al. 10.3389/fphys.2022.953702

Wang, L., and Tian, F.-B. (2019). Numerical simulation of flow over a parallel fluid–structure interactions at moderate and high Reynolds numbers. J. Comput.
cantilevered flag in the vicinity of a rigid wall. Phys. Rev. E 99, 053111. doi:10.1103/ Phys. 375, 22–56. doi:10.1016/j.jcp.2018.08.024
PhysRevE.99.053111
Xu, L., Wang, L., Tian, F.-B., Young, J., and Lai, J. C. (2019). “A geometry-
Wang, Z., Luo, K., Li, D., Tan, J., and Fan, J. (2018). Investigations of data-driven adaptive immersed boundary–lattice Boltzmann method for modelling
closure for subgrid-scale stress in large-eddy simulation. Phys. Fluids 30, 125101. fluid–structure interaction problems,” in IUTAM symposium on recent advances
doi:10.1063/1.5054835 in moving boundary problems in mechanics (Springer), 161–171.
Wang, D., Luo, X., and Stewart, P. (2021). Energetics of collapsible channel flow Yang, Z., Olszewski, D., He, C., Pintea, G., Lian, J., Chou, T., et al. (2020).
with a nonlinear fluid-beam model. J. Fluid Mech. 926, A2. doi:10.1017/jfm. Machine learning and statistical prediction of patient quality-of-life after
2021.642 prostate radiation therapy. Comput. Biol. Med. 129, 104127. doi:10.1016/j.
compbiomed.2020.104127
Wang, L., Liu, Z., and Rajamuni, M. (2022). Recent progress of lattice Boltzmann
method and its applications in fluid-structure interaction. Proc. Institution Mech. Eng. Zhu, L., Zhang, W., Kou, J., and Liu, Y. (2019). Machine learning methods for
Part C J. Mech. Eng. Sci.. in press, 095440622210775. doi:10.1177/09544062221077583 turbulence modeling in subsonic flows around airfoils. Phys. Fluids 31, 015105.
doi:10.1063/1.5061693
Wang, L., Sun, X., Bu, T., Wang, Q., Zhang, H., Jia, P., et al. (2021). Construction
of a photothermal hydrogel platform with two-dimensional PEG@zirconium- Zhu, Y., Tian, F.-B., Young, J., Liao, J. C., and Lai, J. C. (2021). A numerical study
ferrocene MOF nanozymes for rapid tissue repair of bacteria-infected wounds. of fish adaption behaviors in complex environments with a deep reinforcement
Acta Biomater. 235, 342–355. doi:10.1016/j.actbio.2021.08.022 learning and immersed boundary–lattice Boltzmann method. Sci. Rep. 11, 1691.
doi:10.1038/s41598-021-81124-8
World Health Organization (2014). Global status report on noncommunicable
diseases 2014. Geneva: World Health Organization. WHO/NMH/NVI/15.1. Zhu, Y., Pang, J.-H., and Tian, F.-B. (2022). Stable schooling formations emerge
from the combined effect of the active control and passive self-organization. Fluids
Xu, L., Tian, F.-B., Young, J., and Lai, J. C. (2018). A novel geometry-adaptive
7, 41. doi:10.3390/fluids7010041
cartesian grid based immersed boundary–lattice Boltzmann method for

Frontiers in Physiology 17 frontiersin.org