1998 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO.

7, JULY 2014

Which Spring is the Best?

Comparison of Methods for Virtual Stenting
Katerina Spranger∗ and Yiannis Ventikos

Abstract—This paper presents a methodology for modeling the stroke, thromboembolism, and luminal restenosis, that cause
deployment of implantable devices used in minimally invasive vas- reinterventions and, in the worst case, morbidity and mortal-
cular interventions. Motivated by the clinical need to perform ity [6], [7]. Similarly, endovascular repair of cerebral aneurysms
preinterventional rehearsals of a stent deployment, we have de-
veloped methods enabling virtual device placement inside arteries, can sometimes provoke an occasional hemorrhagic event that
under the constraint of real-time application. This requirement of follows an initially successful treatment. In approximately 1% of
rapid execution narrowed down the search for a suitable method cases, the subsequent rupture of the treated cerebral aneurysms
to the concept of a dynamic mesh. Inspired by the idea of a mesh of is reported between a few days and 3–4 months after the treat-
springs, we have found a novel way to apply it to stent modeling. ment [5]. With respect to stroke, the new rate of 6% has been
The experiments conducted in this paper investigate properties of
the stent models based on three different spring types: lineal, semi- recently established by the “Pipeline Embolization Device in the
torsional, and torsional springs. Furthermore, this paper compares Intracranial Treatment of Aneurysms (PITA)” trial [8], which
the results of various deployment scenarios for two different classes is unacceptably high and could threaten the general use of flow
of devices: a stent graft and a flow diverter. The presented results diverters (FDs) in the future.
can be of a high-potential clinical value, enabling the predictive It is known that a number of factors play an important role in
evaluation of the outcome of a stent deployment treatment.
the outcome of the stent deployment procedure; the key players
Index Terms—Flow diverter, modeling, stent, virtual stenting. being the target vessel’s geometry and hemodynamics, design
and the placement position of the implant, incidence of an ar-
terial injury during the intervention, etc. [9]. For example, it
I. INTRODUCTION has been reported that neointimal hyperplasia can initiate from
INIMALLY invasive vascular interventions have be- arterial injury during the procedure and cause significant steno-
M come a popular alternative to conventional open surgery
in the treatment of many vascular disorders, such as aneurysms,
sis [10]. In general, vessel remodeling can play an important part
in these processes [11]. Additionally, several studies have sug-
aortic dissections, and atherosclerotic stenoses [1], [2]. Mini- gested that injuries can also occur after the intervention, caused
mally invasive endovascular repair is performed through small by the nonuniform expansion of the device or its excessive axial
incisions, usually made in the femoral artery, from where a contraction [9]. On the other hand, incomplete contact between
prosthesis or stent1 is pushed to a target location, under fluoro- the stent-filaments and the arterial wall can increase the risk of
scopic guidance and using a catheter, where it is subsequently intimal hyperplasia [12], and stent underexpansion can lead to
deployed inside the diseased vessel. Compared to open surgery, in-stent thrombosis [13].
endovascular repair leads to better treatment outcomes, includ- The dangerous complications caused by arterial injury and in-
ing less complications, less blood loss, shorter surgery times, adequate device apposition could be mitigated by better choice
and hospital stays, etc., mainly attributed to its minimally inva- of the stent design [14] and by the appropriate positioning of the
sive nature [3]–[5]. implant inside the vessel [15], [16]. However, there is currently
However, the high initial success rates of endovascular repairs no possibility for a clinician to predict the resulting position
have been shadowed by reported complications, for example, and the expanded shape of the device for a given patient before
the intervention in a clinical setting. Moreover, knowing the ac-
curate configuration of the device, its attachment to the vessel
wall, especially the level of covering of the aneurysmal neck
or dissection’s entries, would enable the subsequent study of
Manuscript received October 10, 2013; revised January 20, 2014; accepted the postinterventional alterations in the hemodynamic environ-
March 3, 2014. Date of publication March 14, 2014; date of current version June
14, 2014. The work of Katerina Spranger was supported by the Centenary Year ment of the vessel. Subsequent hemodynamic analysis [17] can
Graduate Scholarship, Dept. of Engineering Science, University of Oxford. The provide early indications for such complications as stent graft
work of Yiannis Ventikos was supported by the Wellcome Trust/EPSRC, Centre (SG) migration, endoleakage [18], endotension [19], thrombus
of Excellence in Personalized Healthcare, under Grant WT 088877/Z/09/Z.
Asterisk indicates corresponding author. formation, etc.
∗ K. Spranger is with the Department of Engineering Science, University of
Although there have been numerous attempts to study the
Oxford, Oxford, OX1 2JD, U.K. (e-mail: [email protected]). process of stent deployment computationally, they all feature
Y. Ventikos is with the Department of Mechanical Engineering, University
College London, London, WC1E 6BT, U.K. (e-mail: [email protected]). considerable limitations when judged upon their suitability for
Color versions of one or more of the figures in this paper are available online clinical practice. The approaches reported in the literature can be
at http://ieeexplore.ieee.org. roughly divided into finite-element analysis (FEA) methods and
Digital Object Identifier 10.1109/TBME.2014.2311856
1 “Stent” will be used to denote stent grafts and flow diverters, for brevity. fast virtual stenting techniques. With respect to FEA methods,

0018-9294 © 2014 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution
requires IEEE permission. See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 1999

there has been a lot of work carried out to analyze the mechanical
behavior of stents as well as simulating their implantation. The
themes range from studies on the expansion mechanisms [20], to
deployment of the device within patient-specific anatomies [21],
interactions with the vessel wall [22], and implications on the
local hemodynamics [23]. Although highly accurate, these FEA
models were not aimed for usage in clinical practice on a day-to-
day basis, where the speed of computations becomes a crucial
factor.
In contrast to FEA studies, the second category of the fast
virtual methods feature severe simplifications in crucial aspects
of the process. Literature survey has revealed only three ap- Fig. 1. Reconstruction of stents. (a) Gore TAG thoracic endoprosthesis in the
load-free and partially expanded configuration (image from [29]). (b) Compu-
proaches to stent deployment simulations that aim at evaluating tational representation of the SG (4 cm×10 cm). (c) PED from ev3, Irvine, CA,
different stent and release scenarios on a patient-specific, clin- USA (image from [30]). (d) Reconstructed 3-D struts mesh representing the
ically relevant basis. In the methodology proposed in [24], the PED (3 mm×14 mm).
model is initiated by placing a cylinder along the centerline of
a vessel and then deforming it to follow the centerline curva-
ture. Such mapping of the stent geometry as a texture over a and are of a great relevance due to their simplicity of implemen-
cylinder may have implications on the accuracy of stent config- tation and relatively low computational complexity.
urations, especially in regions of high curvature of the vessel. In this paper, we investigate the suitability of the dynamic
Another study, [25], models a stent as a deforming cylinder, mesh approaches to the stent deployment problem and report
which is first positioned inside the vessel and thereafter serves the results of comparison of three different springs analogy
as a basis for the reconstruction of the struts’ configuration. methods. We start by describing the overall methodology (see
Similar to the previous approach, this method merely deforms Section II) involved in the virtual stent deployment. This sec-
a generic cylindrical structure and does not sufficiently account tion includes the preprocessing of the devices and vessels data as
for the geometrical properties and the design of the device, and well as the modeling of the expansion process. The three springs
its interaction with the vessel. The third method, [26], explicitly analogy methods investigated are 1) the lineal, 2) semitorsional,
models the stent design by capturing the design elements as soft and 3) torsional springs, in order of increasing complexity. Their
constrains; however, complex vessel geometries challenge the results are compared for two different types of devices: a SG and
accuracy of this computational method. a FD, representing the most widely used classes of endovascular
Motivated by clinical need and lack of suitable methods, this implants (see Section III). The comparison is performed on dif-
study aims at developing a methodology for modeling the virtual ferent deployment scenarios: 1) in the case of free expansion, 2)
deployment of implantable devices inside patients’ vessels that in idealized straight vessels, 3) in idealized bent vessels, and 4)
features fast computational times and can be used in the clinical in real patient vasculature. The primary goal of the comparisons
practice. Ultimately, such a capability can play a transformative is to determine whether the use of the more computationally
role in aiding clinical decision making in intravascular inter- intensive version is eventually justified by the increasing real-
ventions on a personalized basis and also give rise to overall ism of the results. We conclude by discussing the results and
improvements in the implant design and deployment procedure. limitations of this study and outline planned future directions
We will target interventions performed under radiological guid- (see Section IV).
ance both in the abdominal and intracranial areas. Hence, the
methodology should be general enough to be able to include II. METHODOLOGY
a broad range of devices, such as SGs and FDs. With this in
mind, we have narrowed down the investigated computational A. Preprocessing of Endovascular Stents
techniques to those that are both sufficiently generic and com- This section outlines the initial steps required to create the
putationally inexpensive. model of the stent and prepare it for the expansion process.
Our attention was drawn to the class of methods called de- 1) Geometry Reconstruction: SGs are usually comprised of
formable models that are widely used in many engineering ap- an external self-expanding nitinol wire structure (struts) that is
plications [27]. One version particularly suitable for medical attached to the surface of the graft [see Fig. 1(a)]. We recon-
applications is called the elastic analogy, where a deforming structed the design of the struts from an X-ray computed tomog-
body is viewed as an elastic structure, whose deformation is raphy scan of the device with the spatial resolution of around
then guided by elastic forces. The elastic analogy in its discrete 150 μm. The strut wires are approximated by their centerlines,
version is sometimes referred to as the spring analogy; here, i.e., they are initially modeled without an explicit thickness to
mesh edges are treated as artificial springs generating springs’ enable fast deployment (the thickness is restored for the subse-
forces, and deformations are governed by the condition of static quent CFD analysis). Afterward, the struts’ centerline represen-
equilibrium [28]. In the literature, these approaches are often tation is supplemented with a background mesh that emulates
subsumed under the category of the dynamic mesh algorithms the device’s sheath and prevents an excessive expansion and
2000 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

response to small deformations of up to 10%, whereas larger

strains can be achieved superelastically due to its shape-memory
effect [31].
Thus, we have to start the deployment simulation with the
crimping of the device. In order to compute the crimped config-
uration, we prescribe the displacements to the stent nodes with
respect to the load-free configuration. Sometimes the crimping
leads to a foreshortening effect, whereby the stent is becom-
ing shorter when deployed, which is observed in real stents,
especially, in the case of FDs.
Fig. 2. Initial positioning of the crimped stent along the centerline of the 4) Initial positioning: After the stent has been crimped, the
vessel. (a) 30%-crimped generic stent. (b) 30%-crimped SG. (c) 20%-crimped next step in a minimally invasive procedure is the positioning of
FD in a patient-specific geometry. the catheter in a desired location inside the vessel with the help
of a guidewire. Both the guidewire and the catheter are flexible
enough to be able to follow the curvature of the vessel they are
being pushed through. Once the target destination is reached,
such as the neck of an aneurysm or desirable entry in the case
of an aortic dissection, the catheter is pulled out to release the
stent which starts expanding under the elastic forces of nitinol
it is made of (for self-expanding devices).
In order to simulate the initial placement of the device before
Fig. 3. (a) SG mesh: thick lines show struts and thin lines—the background
mesh. (b) Schematic representation of the lineal spring analogy: each edge of its release out of the catheter, we deform the crimped device in a
the mesh is viewed as a lineal spring. (c) In the semitorsional spring analogy, way to let it follow the curvature of the vessel’s centerline. First,
artificial torsional springs are added to each node of a mesh [34]. all vertices of the stent are divided into layers depending on their
longitudinal position in the stent’s local coordinates. Afterward,
distortion of the stent during deployment. The merging of the we deform each layer of the stent according to the deviation
two meshes (struts and background) can be achieved in two of the corresponding vessel’s centerline point from the straight
ways: using an isotropic triangulation in which the stent ge- line—centerline of the stent. The prescribed displacements for
ometry has been given a “best-fit” representation within the both the crimping and bending procedure give rise to restoring
background mesh or creating an anisotropic mesh using trian- forces acting in the stent structure. The results of these prepara-
gulation constraints, i.e., requiring mesh elements to incorporate tory steps can be seen in Fig. 2, which shows the placement of
the stent struts (see Figs. 1(b) and 3(a) for a zoomed element). three different crimped stents along the centerline of the vessel.
Due to the high porosity, high pore density, and regular struts
architecture, the easiest way to model a typical FD (or devices
of a similar design) is to use the CAD software to replicate B. Modeling Expansion
its design. After a 2-D representation has been obtained with In this section, we explore and compare different fast com-
CAD, the sheet can then be wrapped around a cylinder to get putational methods in order to determine the most suitable ap-
a 3-D structure. Fig. 1(c) and (d) shows an example geometry proach for modeling the process of stent expansion. First, we
for the commercially available Pipeline Embolization Device briefly describe the overall theoretical framework of the dy-
(PED, ev3, Irvine, CA, USA). Since FDs have no sheath, there namic mesh approaches and then zoom into the details of the
is no need for an additional background mesh. three exemplar methods and show how to apply them to the stent
2) Local Coordinates: For later manipulations of the stent expansion problem. The three versions employed are 1) the lin-
geometry, it is convenient to establish a local coordinate system eal spring analogy, 2) the semitorsional spring analogy, and
for the device in order to decouple the mechanical component of 3) the torsional spring analogy. All three versions are approx-
computations from the stent positioning in the global coordinate imative methods; however, they feature increasing complexity
system. The local coordinates can be set along the principal and precision of the iterative equations. Our ultimate goal will
axes of the stent that can be obtained by a principal component be to investigate their suitability for modeling the stent expan-
analysis performed on the stent vertices. After the analysis, the sion process and finding out whether the more computationally
first component gives the centerline of the stent, whereas the intensive version is justified by a concomitant increase in pre-
second and third components lie in its circumferential plane. cision. In assessing the precision, we are mostly interested in
3) Crimped Configuration: During the minimally invasive the final configuration of the deployed device (and not in the
stent deployment in the clinical setting, the vascular prosthesis transient states) as this constitutes the clinically relevant com-
is first crimped and placed into a delivery system. Upon release ponent, especially with respect to the possible subsequent CFD
in a target vessel location, it “springs” back to a load-free config- analysis.
uration under the elastic forces of the material. Currently, most 1) Theoretical Framework: The methods we have focused
self-expanding devices are made of nitinol (nickel–titanium al- on fall under the category of dynamic meshes, which were orig-
loy), which is a superelastic material, that features an elastic inally proposed as a means of dealing with the moving boundary
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 2001

problems that involve the deformation of the computational do- We will be employing the Jacobi method in the sequel for
main during simulation. In our approach, we will be modeling solving the system of stent forces.
the stent structure as a dynamic mesh. 2) Lineal Springs: The idea behind the spring analogy meth-
A dynamic mesh is often viewed as a continuous or discrete ods entails replacing the mesh by fictitious springs, as schemat-
pseudostructural system with fictitious mass, damping, and stiff- ically depicted in Fig. 3(b). Each spring then behaves according
ness properties [32]. Then, the movement of such a mechanical to Hooke’s law, which can be written for the force F i acting on
system obeys the semidiscrete equations of dynamic equilibrium node i from its neighbors as follows:
Mq̈ + Dq̇ + Kq = 0; (1) 
ni
Fi = kij (δ j − δ i ), (9)
where M, D, and K are the system’s fictitious mass, damping, j =1
and stiffness matrices, respectively; q̇ indicates a time-derivative
of the displacement vector q which is defined by with δ i indicating the displacement of the vertex i and ni —the
number of vertices directly connected to i, kij stands for the
q(t) = ξ(t) − ξ(0), (2)
stiffness of the spring connecting the vertices i and j [33].
with t representing time and ξ standing for the coordinates Assuming that in equilibrium the force acting on each node is
vector of mesh vertices [32]. zero, the nodal displacement can be calculated at every time step
The displacement boundary condition is often accompanying as a weighted average of the displacements of the neighboring
the dynamic equilibrium equation vertices. This gives the following iterative equation (10), which
can be derived from (9) by setting the force to zero (k is the
q(t) = q̄(t) (3) number of iterations):
on Γm , with q̄(t) representing the known displacements of the v i (k )
moving boundary Γm . Usually, the boundary nodes of Γm re- j =1 kij δ j
(k +1)
δi =  vi . (10)
main at the same relative positions after the movement, whereas j =1 kij
the new positions of the internal nodes are found by iterating
(1) with the boundary condition (3). In the case of a moving boundary problem, the known dis-
In the discrete case, the dynamic mesh can be assembled by placements are prescribed to the boundary nodes and then the
adding to every node an artificial mass and to every edge an iterative equation (10) is solved for displacements for all internal
artificial dashpot and an artificial spring to get the damping and nodes on the vertex-to-vertex basis. After a sufficient number of
the stiffness matrices, respectively. iterations, the new nodal coordinates are determined by adding
Very often in applications, the system (1) is reduced to its the final displacement to the old coordinates.
quasi-static version resulting in a simplified form The value of the springs stiffness has implications on how
the boundary deformation is propagated to the inner parts of the
Kq = 0. (4) mesh. Usually the stiffness of lineal springs is set to be inversely
Numerical solution: Systems of linear equations in the form proportional to the segment length before the deformation [28].
Kq = f [such as (4)] are frequently solved with the Jacobi The intuition behind such a stiffness setting is that mesh vertices
method. This method is an iterative algorithm, where the matrix that are located in close proximity to each other should exert a
K is split into two matrices, diagonal matrix Kd and the residual stronger restoring force onto each other.
matrix Kr , so that K = Kd + Kr , under the assumption of Application to stent expansion: We have developed a method
dominant diagonal elements. Hence, that draws inspiration from the springs analogy and determines
mesh movements without known boundary deformations. Since
Kq = Kd q + Kr q. (5) crimping of the stent creates spring forces acting inside the
This implies structure, the idea is to allow the nodes in the mesh to move
under these forces. The movement takes into account the nodal
Kd q = Kq − Kr q = f − Kr q. (6) intertwinement with its neighbors by means of the nodal stiff-
ness, the value of which is set to be equal to the sum of all
Finally, we can derive an expression for q which can serve as a
the springs’ stiffness connecting it to the direct neighbors. The
base for iterative solutions q(k +1) :
point to note is that the background mesh of the SG device is
q = Kd−1 (f − Kr q) =⇒ q(k +1) = Kd−1 (f − Kr q(k ) ); (7) also viewed as consisting of springs which are contributing to
the expansion process.
where k + 1 is the number of the iteration. The algorithm is
In the expansion procedure, we need to distinguish between
iterated until it converges, i.e., q(k +1) is equal or close enough
the two different displacements (deltas). The first one is respon-
to the desired solution. Alternatively to the system approach,
sible for the current configuration of the stent, that is, it accounts
the solution can also be found on the element-to-element basis,
for the remaining crimping status of the structure. This delta, that
which takes the following form for an element i:
⎛ ⎞ we call crimping delta δ cr , is used to calculate restoring forces
 in the current (partially) crimped stent. The second delta, called
1 ⎝
kij qj ⎠ , i = 1, 2, . . . , n. (8)
(k +1) (k )
qi = fi − expansion delta δ exp , is the actual displacement with which the
kii stent in expanding in each iterative step. It is used to update the
j = i
2002 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

gular information into account. The idea is to provide additional

stiffening elements by placing a torsional spring at each node
of the mesh. The torsion spring would account for the angu-
lar displacement of mesh elements and enable controlling mesh
deformation. Based on the lineal spring analogy, the behavior
of the torsion spring is modeled indirectly by integrating the
element’s angular values into the stiffness of the spring edge,
giving rise to the semitorsional stiffness. Fig. 3(c) schemati-
cally demonstrates the idea behind the approach for a 2-D mesh
showing torsional springs placed at the corners between adjacent
edges of a triangular element.
There have been different proposals in the literature as to how
to define the semitorsional stiffness. We have used the definition
suggested by [34] that calculates the semitorsional stiffness of
sem itorsional
an edge (kij ) by gathering angular information from
all facing edges in a sum

nij
1
sem itorsional
kij =κ 2 ij
; (12)
m =1 sin θm
where nij is the number of all triangular elements that includes
the edge (i − j), θm ij
is the angle facing the edge (i − j) in
one of such triangular elements (m), and κ is a parameter for
controlling the dimensionality of the resulting stiffness.
Since the value of the sine approaches 0 when the angle
approaches 0 or π, the facing edge becomes very stiff when
neighboring edges are close to collision. Stiffening the edge
slows down the progression in the angle change, preventing
thereby the unwanted inversion of elements. Then, the total
stiffness of an edge is composed of the sum of the two types of
stiffness, the lineal and the semitorsional
lineal sem itorsional
kij = kij + kij . (13)
The procedure for free expansion is the same as in the lineal
case that has been reported in Section II-B2.
4) Torsional Springs: Another proposal to control the an-
gular deformations of the mesh elements has been to directly
stent vertices’ coordinates in the process of deployment. During implement the artificial torsional springs added to the nodes of
the deployment process, the two deltas feed into each other. the mesh (and not to merely simulate them, as in the previous
First, the initial δ cr is determined, which measures the ini- approach). Then, the torsional stiffness of a torsional spring
tial crimping deformation. The iterative process starts with the attached to a vertex i can be defined as
calculation of the restoring forces inside the mesh based on the 1 1 1
Ciij k = × = ; (14)
crimping delta δ cr . Based on this information and stiffness of the 1 + cos θiij k ij k
1 − cos θi 2 ij k
sin θi
vertices, each iterative step of the expansion process determines
the current displacement δ exp . Afterward, the current coordi- where Ciij k is the stiffness of the spring that is attached to the
nates of stent vertices are updated, as well as δ cr , so that they vertex i which is connected with vertices j and k and, thus,
both correspond to the current status of crimping in the stent associated with the angle θiij k (see Fig. 4), as proposed by
structure. The process is repeated until there is no change in the Farhat et al. [32]. Hence, this torsional spring will create torques
stent structure, that is, no restoring forces are acting on the struts opposing the change in the angle. We will show the derivation
(see the detailed pseudocode of the Incremental Displacement of the torsional method for the 2-D case and then extend it
Algorithm for Free Expansion in the box above). to the 3-D version which is employed for the stent expansion
3) Semitorsional Springs: The driving force of the lineal simulation.
spring analogy is the restoring force created by the stiffness of a) Kinematics: We use the kinematic formulation pro-
an edge. However, the process of stent expansion is mostly gov- posed by Farhat et al. [32] that was developed under the as-
erned by angular displacements of the stent struts and angular sumption that only small displacements and rotations occur be-
forces. Hence, we investigate an extension of the lineal spring tween two subsequent time steps. The assumption is necessary
analogy called “semitorsional spring analogy” that takes the an- during the derivation in order to get angular displacements to
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 2003

In order to attach the additional torsional springs to a lineal

spring mesh, Farhat et al. [32] have proposed the transformation
of the moments M ij k into a set of equivalent forces at each
vertex gathered in vector F ij k
torsion as

F ij k
torsion = [ Fi x Fi y Fj x Fj y Fk x Fk y ]torsion .

F ij k
torsion can be deduced from moments M
ij k
by a linear
ij k T
transformation with the help of the matrix R as follows
Fig. 4. Schematic representation of the torsional spring analogy: an artificial (the detailed derivation is reported in [32]):
torsional spring is placed at each node controlling angular deformation [32]. T
F ij k
torsion = R
ij k
M ij k .
be expressed in terms of vertices displacements in a linear way. Finally, Farhat et al. [32] define for every element Δij k the
Thus, the iterative algorithm based on this method has to suf- equivalent forces F ij k
torsion produced by the torsional springs as
ficiently break down the large deformation that might occur in T
the course of the simulation into small ones in order to respect F ij k
torsion = [R
ij k
C ij k Rij k ]q ij k = K ij k
torsion q
ij k
. (17)
the assumption.
Coming back to Fig. 4, we consider the displacement of a c) Dimensionality issues: In order to properly integrate
2-D triangle Δij k from its reference position, depicted by a both types of springs, some modifications to the original defini-
solid line, into a current configuration, shown by a dashed line tion of torsional forces proposed by Farhat et al. [32] had to be
in the image. The displacement components resulting from the made. The reason for the alteration is the discrepancy in dimen-
displacement of the triangle vertices i, j, and k are captured in sionality in both types of forces [36]. More precisely, the lineal
the vector q defined in the following way: spring forces are nondimensional quantities, since lineal stiff-
nesses kij is inversely proportional to segment lengths (9). In
q ij k = [ ui vi uj vj uk vk ] . contrast to that, the torsional forces are dimensional quantities.
This is because the torsional stiffness is defined in (17), where
The resulting rotation increment is denoted by the vector
C ij k is dimensionless and Rij k is inversely proportional to the

Δθij k = [ Δθi Δθj Δθk ] . length (16). Hence, K ij k
torsion is inversely proportional to length
square, resulting in torsional forces being inversely proportional
The rotation increment Δθij k can be defined in the matrix form to the length. Therefore, if we simply added the torsional spring
in the following way: forces to the lineal spring forces, the relative importance of the
ij k
Δθ = R q , ij k ij k
(15) lineal and torsional springs would depend on the choice of the
length units (for example, either meters or millimeters), which
where the matrix Rij k is composed as follows: Rij k = clearly should not be the case. In order to avoid this undesirable
⎡ ⎤ effect, we have redefined torsional spring forces as nondimen-
bik − bij aij − aik bij −aij −bik aik
⎢ ⎥ sional quantities by scaling the torsional stiffness according to
⎣ −bj i aj i bj i − bj k aj k − aj i bj k −aj k ⎦ a characteristic length
bk i −ak i −bk j ak j bk j − bk i ak i − ak j κ
(16) C ij k = (18)
and the components of the Rij k matrix are built as sin θiij k
2

xij yij where scaling constant κ = c can be defined through the local
aij = 2 , bij = 2 ∀(i, j).
lij lij characteristic length  of the mesh, such as the average edge
length of a triangular element, scaled with a chosen constant c.
b) Equilibrium: In order to express the moments pro-
By the way, the same applies to the semitorsional method and the
duced by torsional springs as forces acting on the mesh, we
overlay between the lineal and the semitorsional springs. There
first write these moments in a matrix form according to the
as well, the parameter κ was used to control the dimension [see
angular version of Hooke’s law
(12)].
M ij k = C ij k Δθij k , d) Numerical solution algorithm: We have used the Ja-
cobi scheme to solve the system Kδ = f on a vertex-by-vertex
where basis, making the following transformation to the original defi-
⎡ ⎤ ⎡ ij k ⎤
Mi Ci 0 0 nition given by (7)
ij k ⎢ ⎥ ij k ⎢ ij k ⎥
M = ⎣ M j ⎦; C = ⎣ 0 Cj 0 ⎦. δ (n +1) = K d −1 (f − K r δ (n ) )
Mk ij k
0 0 Ck = K d −1 (f − (K − K d )δ (n ) )
Substituting the expression for Δθij k defined in (15), we get = K d −1 f − K d −1 Kδ (n ) + K d −1 K d δ (n )
M ij k = [C ij k Rij k ]q ij k . = K d −1 f − K d −1 Kδ (n ) + δ (n ) .
2004 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

Noting that the system should approach static equilibrium, that be stored as an STL (stereolithography) or any other CAD-like
is, f = 0, and substituting in the above expression Kδ (n ) = 3-D geometric definition format. For example, the STL file will
F (n ) , the final iterative algorithm for the torsional springs model contain a description of the surface by listing unstructured tri-
can be summarized in the form as follows: angles that make up its mesh, where each triangle is described
by the 3-D coordinates of the three vertices that comprise it and
F (n ) = K lineal δ (n ) + K torsion δ (n )
the normal to the triangle plane.
δ (n +1) = δ (n ) − (K d torsion + K lineal−no dal )−1 F (n ) 2) Contact: The expansion process of the stent is performed
on a vertex-by-vertex basis, and the coordinates of a vertex are
δ (n +1) = ω δ (n +1) + (1 − ω) δ (n ) . (19) updated in each iteration as long as no contact between the vertex
and the vessel occurs. The check for the contact is performed
Note that in every iteration, torsional and lineal spring forces
after Step (b) in the incremental displacement algorithm, where
are computed separately and added afterward to obtain the
the expansion displacement is calculated for each node and
total force F (n ) = K lineal δ (n ) + K torsion δ (n ) due to differ-
before the nodal update in Step (c). Contact is defined to occur
ent structural organization of stiffness matrices K lineal and
when the distance between the vertex of the stent and any of
K torsion . Since the Jacobi method is operating on the inverse di-
the vessel’s vertices becomes smaller than a chosen parameter
agonal matrix in every iteration (see Section II-B1), the diagonal
. After contact is detected, the displacements for the vertex
components of both stiffness matrices have to be incorporated to
in contact are still calculated in future iterations. However, its
overlay the effect of lineal and torsional springs, which is equiva-
position can only change if it is located within the -boundary
lent to adding the two matrices K d torsion + K lineal−no dal . The
of the vessel inner surface.
relaxation parameter ω was permanently set to 0.5.
e) Extension to 3-D stent case: The main idea for the
modification is to use the 2-D torsional spring analogy in the III. RESULTS
plane of each triangular element and then transfer the results
We have performed deployment experiments for two differ-
to the 3-D stent world. More precisely, since the stent struts
ent devices in different vessel geometries in order to compare
are modeled without taking the thickness into account, each of
the three methods. The two devices were chosen to represent
the triangular elements lies in a 2-D plane, defined by its three
“typical” devices used in the two distinct areas of radiological
points. Thus, it is possible to perform computations in the lo-
procedures of interest—aortic and cerebral cases.
cal coordinate system of the triangular elements, applying the
1) The first device is a SG corresponding to the GORE TAG
2-D torsional method and then convert the obtained forces, dis-
Thoracic Endoprosthesis (model TGT4010) with dimen-
placements, etc., to the global coordinate system. Alternatively,
sions of 4 cm in diameter and 10 cm in length [see
we could convert the “local” stiffness matrix obtained for 2-D
Fig. 1(b)]. The device has 164 struts that are 13 mm in
case into a global 3-D stiffness matrix, which is then used for
length (on average) and 0.35-mm thick. It is one of the
calculation of forces and displacements in the global coordi-
most widely used devices in radiological procedures to
nate system. We have chosen the latter way, since it does not
treat thoracic aortic aneurysms and aortic dissections.
require storing in memory all the local bases and conversion of
2) A FD is similar to the commercially available PED, ev3,
displacements into these local coordinate bases.
Irvine, CA, USA) with the diameter of 3 mm and the length
Hence, after the 2-D stiffness was obtained, we have to trans-
of 14 mm [see Fig. 1(c)], which is one of the two devices
form this stiffness matrix to the global coordinate system. We
currently used in Europe to treat cerebral aneurysms. The
do that by multiplying it with the basis vectors of the local co-
device has 3200 struts that are 0.4 mm in length.
ordinate system, which is actually a rotation matrix (R) and can
The stents were crimped to simulate their positioning in the
be used to switch between the two coordinate systems:
catheter and aligned along the centerline of the target vessel at
K ij k ij k T
torsion−global = R K torsion−lo cal R . (20) the relevant positions using the methods described in previous
sections. For the first device, the initial crimping was to 20%
For the stent expansion process, we use the same iterative
of its labeled diameter; since in the clinical practice, the SG is
method that was used in the 2-D case; one iteration is sum-
compressed into a 24 FR (1FR = 0.333 mm) catheter system,
marized in (19). With respect to the forces’ vector, as before,
i.e., from 40-mm-outer diameter to 8 mm (24 × 0.333 = 8 mm).
it contains a sum of both types of forces: lineal overlaid with
With respect to the second device, the PED is usually delivered
torsional. This iteration substitutes the Step 2 (Iterate) in the in-
via a 3 FR microcatheter (3 × 0.333 = 1 mm); hence, it was
cremental displacement algorithm, described in Section II-B2,
crimped to the 33% of its load-free diameter.
which models the free expansion procedure.
In all simulations, dimensionality parameters were set to κ =
3
 for the semitorsional method (12), where  is the characteristic
C. Modeling the Vessel and the Contact length of the device element, measured as an average length of
1) Geometry Reconstruction: Clinical imaging data are usu- a strut. For the torsional case, the scaling was with κ = 3 (18).
ally obtained in the DICOM format. A sequence of the DICOM These settings were found to produce the best and most stable
images that capture different “slices” of the scanned part of the results, which were in good alignment with the results of the
body in 2-D can be combined to reproduce a 3-D representa- full FEA simulation performed in our validation study [36]. We
tions of the scanned organ in the process of segmentation. The can iterate here in brief that the error reported in that paper for
outcome of the segmentation is a 3-D representation that can the springs method was between 2% and 10%, depending on
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 2005

Fig. 5. Free expansion of stents with the lineal spring method: after 10, 100
iterations of layered expansion and at the end of the expansion process, (a)–(c)
for the SG and (d)–(f) for the FD, respectively.

the complexity of the case—values that are deemed perfectly

acceptable in this context.
To compare the methods and evaluate the results, we define
the following metrics of convergence.
1) Mean nodal  distance between load-free and expanding
N 
stent: δ = N1 i=1 ni − ni , where ni is the position of Fig. 6. Free expansion of the SG device for 300 iterations. Upper row: mean
the node i in the load-free stent and ni —in the current nodal distance (δ) and mean angle difference (θ) between load-free and expand-
expanding stent (δ is measured in millimeters). ing stent. Bottom row: mean strut length difference (L) and mean nodal force
2) Mean angle difference between load-free and expanding in the expanding stent (F ). All three methods were able to reach the load-free
M  configuration. The lineal springs method (in blue) displayed the fastest con-
stent: θ = M1 j =1 j − θj |, where θj is the value of the
|θ vergence, after 200 iterations. The semitorsional springs method (green) is the
angle j in the load-free stent, θj —in the current expanding second fastest, converging after 250 iterations. The torsional springs method
stent, and M stands for the number of angular elements in (red) is the slowest, reaching full expansion only after 500 iterations.
the stent structure (θ is measured in degrees).
3) Mean relative strut length difference between load-free its load-free configuration, the shape of which is also known, the
 |l −l  |
and expanding stent: L = S1 Sk=1 k l k k , where lk is the clear way to demonstrate the expansion process is to track the
length of the strut k in the load-free stent, lk —in the distance to the target load-free stent. Hence, we have measured
current expanding stent, and S stands for the number of the average distance from the vertices of the current expanding
struts (L is measured in % relative to the initial stent to their counterparts in the load-free device (mean nodal
Nlength).
i=1 F i ,
4) Mean nodal force in expanding stent: F = N1 distance), which is displayed in the first graph in the upper left
where F i is the reaction force at the node i with the index corner. All three spring methods are able to reach the load-free
i running from 1 to N . configuration; however, they do so with different speed, with
We have conducted experiments in four different deployment the lineal method being the fastest and the torsional method
scenarios, reported in the coming sections: being the slowest (completely converges only after about 500
a) free expansion; iterations).
b) deployment in a straight tube; The evolution of the angles is shown in the upper right corner
c) deployment in a bent tube; as the mean angle difference between the load-free and the cur-
d) deployment in real vessel geometries: dissection case for rent expanding stent. This measure is very important since the
SG and aneurysm case for FD. difference in angles reveals the degree of crimping in the struc-
All models were implemented in MATLAB and executed on ture and indirectly shows the average angular displacement in
an Intel Core 2 Duo with CPU 2.66 GHz with 4 GB of memory every iteration. Again, the lineal method displays faster conver-
and without using any parallelization—a modest setup easily gence, semitorsional is similarly fast, and the torsional is the
reproducible in any clinic. slowest. Equally, it is important to ensure the adequate struts’
length evolving throughout the iterative process. This is done
by measuring the mean strut length difference which gives the
A. Free Expansion average error length as a relative percentage of the initial length.
1) SG: From the initially crimped configuration (8 mm in The bottom left graph displays the behavior of struts’ length.
diameter which is 20% of the labeled diameter), the SG de- The initial iterations corrupt the average length up to 14% (for
vice was subject to free expansion until it reached its load-free lineal springs method), but it gets corrected in the course of the
configuration. Fig. 6 captures the evolution of the expansion expansion process.
process during 300 iterations. The four different convergence Finally, the bottom right graph displays the mean nodal force
metrics are plotted for each of the three springs methods: lin- in the expanding stent which gives account for the reaction
eal springs in blue, semitorsional springs in green, and torsional forces guiding the expansion process. The point to note is that
springs in red. Since the released SG is going to “spring back” to the forces for semitorsional and torsional methods include the
2006 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

TABLE I reached the vessel walls. The snapshots of the expansion with
VALUES OF CONVERGENCE METRICS AT THE END OF THE
DEPLOYMENT PROCESS
the semitorsional method after a certain amount of iterations are
captured in Fig. 7(a)–(c). The other two methods displayed a
comparable progression.
Table I (cases 7–9) shows the convergence of parameters
throughout the deployment process. The values in the table are
displayed in the format meanm ax
m in , with the average value in
bold. The lineal springs method displays the fastest convergence
rate again. However, torsional springs perform slightly better
this time with respect to the final nodal displacement (see δ),
which is surprising given their earlier performance in the free
expansion case. When the deployment is complete, after about
100 iterations on average, there is a very small but nonzero
restoring force that remains due to oversizing, which would
help keeping the device placement intact.
2) FD: The PED was crimped to 1 mm and released into an
idealized cylinder with the 2.3-mm diameter in order to account
for the oversizing. The snapshots of the expansion with the
torsional method after a certain amount of iterations are captured
in Fig. 7(g)–(i). The other two methods displayed a comparable
progression with a slightly faster convergence. The convergence
results can be seen in Table I (cases 10–12). The struts length
difference remains substantial since struts are not fixed due to
the woven design of the device; therefore, the struts can slide,
emulated by the flexible nature of the springs.

C. Ideal Vessel: Bent Tube

lineal component, for this is how it is calculated in the algo-
rithms. Consequently, it is to expect that the lineal reaction 1) SG: This time the geometry of the vessel was represented
forces are smaller in magnitude. On the other hand, the semitor- as a bent cylinder with a diameter of 34 mm. The lineal springs
sional forces are much larger, even after the lineal component method was applied in layers, starting from the center of the
has been subtracted from them (almost three times the lineal device, as performed in the clinical practice. Fig. 7(d)–(f) dis-
value). plays the snapshots of the layered lineal expansion after a certain
2) FD: The FD device was initially crimped to 33% of its amount of iterations.
labeled diameter and then underwent the free expansion test for Since the lineal springs method was applied in layers, its
500 iterations. As it was the case for the SG, the lineal springs convergence is slower than its semitorsional and torsional coun-
method displayed the fastest convergence, after about 250 itera- terparts (see Table I, cases 13–15). The latter two methods dis-
tions, closely followed by the semitorsional springs. Again, the played a progression comparable to the previous cases, although
torsional method converged only after about 500 iterations. The the semitorsional method is slightly more accurate with respect
overall behavior of the convergence metrics is similar to the SG to the struts’ length error. After the deployment is complete,
case, depicted in Fig. 6. However, there are some differences there is a small nonzero restoring force that remains present due
with respect to the end values, which are recorded in Table I to the oversizing of the device and the geometry of the vessel.
(cases 4–6). All convergence metrics display zero residual value Note that metrics that we have used to evaluate the perfor-
since the FD returned to the load-free configuration. The lineal mance of the methods are not suitable for the bent (and more
method was again the quickest, requiring 250 iterations (it) and complicated) geometries. For example, the nodal displacements
26 s (time in brackets). are calculated as the distance from the current expanding stent to
Fig. 5 illustrates the snapshots of the free expansion for two the load-free configuration and does not represent a meaningful
stents with the lineal spring analogy expansion method. value, since the load-free configuration is not the target of the
expansion this time. In fact, the target is unknown! Neverthe-
less, we chose to continue to use these familiar metrics merely
B. Ideal Vessel: Straight Tube to compare the difference in the convergence rates of the three
1) SG: In this set of experiments, the device was deployed methods.
inside an idealized vessel in the shape of a straight cylinder. 2) FD: The PED was crimped to 1 mm and released in the
The diameter of the cylinder was 34 mm, which is a usual idealized bent cylinder with the 2.6-mm diameter to account
vessel size for a 40-mm device, because the grafts are always for the oversizing. This time, we have applied the semitorsional
oversized in the clinical practice for about 30%. As before, the method in layers, the snapshots of which can be seen in Fig. 7(j)–
device was crimped to 8 mm and deployed, this time until it (l). The other two methods were applied simultaneously to all
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 2007

Fig. 7. Expansion of stents in different geometries: (a)–(c) SG expansion from

the crimped stent to fully deployed after 100 iterations; (g)–(i) FD expansion
from the crimped stent to fully deployed after 100 iterations; (d)–(f) and (j)–(l)
layered expansion of the SG and FD devices in bent vessels, respectively.

layers. The general convergence values are depicted in Table I

(cases 16–18).

D. Real Vessel Geometries

1) SG: We have tested the performance of the virtual de-
ployment algorithm in the case of the aortic dissection. As it
would be done in clinical practice, the graft was placed inside
the true lumen to cover the initial tear entry and prevent blood
from flowing into the false lumen. Fig. 8. Virtual SG deployment in the dissected patient’s vessel. Upper row:
The vessel diameter was varying at the deployment site and (a) 20%-crimped device initially placed on the centerline of the target vessel on
ranged between 2.9 and 2.1 cm in the distal and proximal loca- the site of the pathology; (b) and (c) after 50 and 100 iterations of the expansion
process, respectively, starting from the central part of the device and propagating
tion of the device in the true lumen, respectively. Therefore, we to its extremities. Lower row: final configuration of the deployed SG as in (c),
had to adjust the dimensions of the modeled SG to match those displayed from different rotation angles.
that would be required in clinical practice. The modified device
featured 3.2-cm diameter and 10-cm length.
in a clinical setting. Therefore, we had to adjust the modeled FD,
The opening of the device was started in the center of the
reconfiguring it to be 4.5 mm in diameter in order to account
device and was progressing to its extremities, as common with
for both vessel geometry and oversizing. We have deployed the
this type of devices. Fig. 8 shows the deployment snapshots
device over the neck of the aneurysm, applying the torsional
at different phases during the deployment process. The final
method in layers.
configuration was obtained after approximately 200 iterations
As it would be the case in real interventions, the opening of
of the virtual deployment algorithm and required approximately
the device was started from the distal extremity of the device and
30 s. The device was in a good opposition to the vessel wall and
progressed to its proximal end, with respect to the aneurysm sac.
has covered the initial tear completely. During the opening of
Fig. 9 shows the deployment snapshots at different phases during
the device, the angles and struts’ length progressively approach
the deployment process. The final configuration was obtained
the values corresponding to the load-free configuration. Table I
after approximately 500 iterations of the virtual deployment
(cases 19–21) shows the resulting metrics.
algorithm and required about 2.5 min. The device was in a good
2) FD: The final deployment experiment involved a case of
opposition to the vessel wall. Table I (cases 22–24) shows the
a giant aneurysm located in the left ICA artery of a patient
resulting metrics.
(see Fig. 9; note the small opthalmic artery originating on the
opposite vessel’s wall with respect to the aneurysmal sac). The
3-D vessel geometry featured a variation in vessel’s diameter IV. CONCLUSION
from 4.1 mm in the distal and 4.4 mm in the proximal location In this paper, we have reported on methods for virtual stent
with respect to the aneurysm. For such dimensions of the vessel, deployment. The inspiration for the presented algorithms came
a FD of 4.5-mm diameter and 14 mm in length could be utilized from the dynamic mesh methods that are widely used in many
2008 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

taking care of the torsional displacements and forces even in the

lineal method. In general, the difference in methods is slightly
more pronounced for the SG case as compared to the FD (see
Table I).
Another characteristic that was similar to all three springs
methods was the small residual error in the struts lengths, even in
the final stent configuration. This fact is due to the accumulation
of imprecisions along the entire deployment processing chain:
the crimping, positioning along the centerline, and the sequence
of iterations based on the approximative methods. In the future,
the methods should pay more attention to this factor with respect
to the struts length. It is possible, however, that a special handling
of the struts length is necessary for the springs-based methods,
for example, incorporation of an additional constraint into the
algorithm to preserve or correct the lengths.
With respect to the speed of the virtual simulation, it was
sufficiently fast; the longest simulations lasting about 2.5 min for
the patient’s case of the cerebral aneurysm with the high density
PED device on a very modest hardware platform. Gaining a
ten times speed up on a state-of-the-art core is certainly within
easy reach. Being implemented in MATLAB, so far, with no
Fig. 9. Virtual FD deployment in a cerebral aneurysmal vessel. Upper row: (a), parallelization and on slow and antiquated hardware, the models
(b), and (c) after 100, 230, and 500 iterations of the torsional expansion method,
respectively, starting from the distal to proximal extremity of the device with
are likely to obtain much better speed. Achieving such fast
respect to the aneurysm sac. Lower row: final configuration of the deployed FD running times is a very positive outcome, for high speed close to
as in (c), displayed from different rotation angles. real time constitutes one of the top priorities for the development
of deployment methods, as outlined in Section I.
Although the lineal spring analogy method proved suitable for
engineering applications. Using the spring mesh idea, we have modeling the stent expansion, there are still a number of limita-
developed algorithms that enable modeling the process of stent tions associated with it and this study. The first one to mention
expansion and its deployment in vessels. Furthermore, we have is the fact that the validation of the results was omitted here due
compared three different spring analogy methods featuring dif- to space considerations. The detailed validation is reported in
ferent properties of the springs and showed their performance another study that is comparing the results reported here with
in the given problem setting. The three spring analogy meth- the results of the FEA [36]. The FEA method, although being
ods that were compared are 1) the lineal spring analogy, 2) the very slow, can provide validation for the fast springs methods,
semitorsional spring analogy, and 3) the torsional spring anal- assess the eventual divergence in the results, and estimate the
ogy. The comparison was based on the results of expansion of error. The comparison is, as we discussed, very favorable to the
two different devices—a SG and a FD—in four different scenar- fast method.
ios: 1) in the case of free expansion, 2) in the idealized straight The second point worth mentioning is related to the forces.
vessels, 3) in the idealized bent vessels, and (4) in real patient Since the springs methods presented here are approximative,
cases. it is not clear what physical meaning is associated with the
When applied to the stent deployment problem, all three restoring forces measured by them. From our comparison (see
spring analogy methods converged; however, they did so at Table I), one can see that the lineal forces feature values that are
different rates. The convergence was assessed by different met- very similar to those reported by the torsional method; in con-
rics, measuring the displacement and force evolution as well as trast, the semitorsional method is displaying much higher val-
the evolution of angles between the struts and the struts length. ues. Although from the validation study, we know that the lineal
One of the interesting findings entailed the fact that the lineal forces lie close to the forces reported by the FEA comparative
springs method displays faster convergence rate in all of the four experiments; the full analysis of the force evolution is yet to be
metrics used. Additionally, it produced final results compara- undertaken. In the meantime, one could imagine that although
ble with those obtained by its more sophisticated semitorsional the lineal version is powerful enough to model the movable
and torsional counterparts. This is surprising since the lineal stent structure [and for now it seems like there is no need for
method constitutes the base for the other two spring methods the complex (semi-)torsional springs], the more sophisticated
and is associated with the simplest implementation. In the case methods could be able to provide better force approximations
of the first device, such outcome for the lineal version was fa- than the “easy” lineal approach. This would mean that they are
cilitated by the presence of the background mesh, which served actually worthwhile to employ for a more thorough analysis of
as an integrity-preserving element and simulated the angular the deployment process.
elements indirectly. For the second device, the fine mesh of the Related to the force is the physical meaning of the different
struts effectively produced the effect of the “background” mesh, types of stiffness and their settings. In this paper, the stiffness
SPRANGER AND VENTIKOS: WHICH SPRING IS THE BEST? COMPARISON OF METHODS FOR VIRTUAL STENTING 2009

values were assigned in an arbitrary way, disregarding any con- tomatic severe carotid stenosis,” New Engl. J. Med., vol. 355, no. 16,
siderations of the material properties. In the future, the stiffness pp. 1660–1671, 2006.
[8] D. Fiorella, P. Lylyk, I. Szikora, M. Kelly, F. Albuquerque, C. McDougall,
values have to be reverse engineered in advance in order to ac- and P. Nelson, “Curative cerebrovascular reconstruction with the pipeline
count for the mechanical behavior as closely as possible. We embolization device: The emergence of definitive endovascular therapy for
hope to gain more understanding of that in in vitro experiments intracranial aneurysms,” J. NeuroIntervent. Surg., vol. 1, no. 1, pp. 56–60,
2009.
with stents, planned for the future. Again, it could turn out that [9] D. Kiousis, A. Wulff, and G. Holzapfel, “Experimental studies and
the torsional methods are more suitable for the incorporation of numerical analysis of the inflation and interaction of vascular balloon
realistic stiffness values than their lineal counterpart. catheter-stent systems,” Ann. Biomed. Eng., vol. 37, no. 2, pp. 315–330,
2009.
Other limitations of this study include the treatment of the [10] R. Schwartz and T. Henry, “Pathophysiology of coronary artery resteno-
vessel as a rigid body and the simplified contact model em- sis,” Rev. Cardiovasc. Med., vol. 3, pp. S4–S9, 2002.
ployed. We plan to extend the current model by implementing [11] P. Watton, N. Raberger, G. Holzapfel, and Y. Ventikos, “Coupling
the hemodynamic environment to the evolution of cerebral aneurysms:
the stent’s self-contact and the possibility of the vessel to de- Computational framework and numerical examples,” J. Biomech. Eng.,
form under the forces exerted by the stent. This would require vol. 131, p. 101003, 2009.
handling the contact between the two bodies—the interplay of [12] M. Jamous, K. Satoh, S. Matsubara, J. Satomi, N. Nakajima, M. Uno,
and S. Nagahiro, “Ischemic basilar artery dissecting aneurysm treated by
forces and displacements—which would provide valuable infor- stenting only,” Neurologia Medico-Chirurgica, vol. 44, no. 2, pp. 77–81,
mation for the clinicians, in order to assess the possible short- 2004.
and long-term implications of the stent deployment treatment. [13] S. Cook, P. Wenaweser, M. Togni, M. Billinger, C. Morger, C. Seiler,
R. Vogel, O. Hess, B. Meier, and S. Windecker, “Incomplete stent apposi-
This further-elaborated model would then be validated against tion and very late stent thrombosis after drug-eluting stent implantation,”
real in vivo cases, for which we are gathering pre- and postpro- Circulation, vol. 115, no. 18, pp. 2426–2434, 2007.
cedural data at the moment. [14] B. Lieber, V. Livescu, L. Hopkins, and A. Wakhloo, “Particle image ve-
locimetry assessment of stent design influence on intra-aneurysmal flow,”
To sum up, this paper demonstrated the overall ability of Ann. Biomed. Eng., vol. 30, no. 6, pp. 768–777, 2002.
spring-based methods to model virtual stent expansion in a [15] M. Hirabayashi, M. Ohta, D. Rüfenacht, and B. Chopard, “A lattice Boltz-
computationally expedient manner, which constitutes a valu- mann study of blood flow in stented aneurism,” Future Generat. Comput.
Syst., vol. 20, no. 6, pp. 925–934, 2004.
able base for future extensions of the model, with the ultimate [16] Y. Ventikos, E. Holland, T. Bowker, P. Watton, N. Kakalis, M. Megahed,
goal to enable predictive simulations of the minimally invasive F. Zhu, P. Summers, and J. Byrne, “Computational modelling for cerebral
methods in a clinical setting. aneurysms: Risk evaluation and interventional planning,” Brit. J. Radiol.,
vol. 82, no. Special Issue 1, pp. S62–S71, 2009.
[17] D. Chen, M. Muller-Eschner, D. Kotelis, D. Bockler, Y. Ventikos, and
ACKNOWLEDGMENT H. von Tengg-Kobligk, “A longitudinal study of Type-B aortic dissection
and endovascular repair scenarios: Computational analyses,” Med. Eng.
The authors would like to thank Prof. J. Byrne (University Phys., vol. 35, pp. 1321–1330, 2013.
of Oxford) and Dr. H. von Tengg-Kobligk (Inselspital Bern, [18] J. Golzarian and D. Valenti, “Endoleakage after endovascular treatment of
abdominal aortic aneurysms: Diagnosis, significance and treatment,” Eur.
University of Heidelberg) for their valuable clinical input. Radiol., vol. 16, no. 12, pp. 2849–2857, 2006.
[19] M. Gawenda, G. Jaschke, S. Winter, G. Wassmer, and J. Brunkwall, “En-
REFERENCES dotension as a result of pressure transmission through the graft following
endovascular aneurysm repair—An in vitro study,” Eur. J. Vasc. Endovasc.
[1] A. Schanzer and L. Messina, “Two decades of endovascular abdominal Surg., vol. 26, no. 5, pp. 501–505, 2003.
aortic aneurysm repair: Enormous progress with serious lessons learned,” [20] C. Lally, F. Dolan, and P. Prendergast, “Cardiovascular stent design and
J. Amer. Heart Assoc., vol. 1, no. 3, pp. e0000752012. vessel stresses: A finite element analysis,” J. Biomech., vol. 38, no. 8,
[2] J. Byrne and I. Szikora, “Flow diverters in the management of intracra- pp. 1574–1581, 2005.
nial aneurysms: A review,” eJ. Eur. Soc. Minimally Invasive Neurolog. [21] D. Ma, G. F. Dargush, S. K. Natarajan, E. I. Levy, A. H. Siddiqui, and
Therapy, vol. 1225000057, 2012. H. Meng, “Computer modeling of deployment and mechanical expansion
[3] M. Adriaensen, J. Bosch, E. Halpern, M. Myriam Hunink, and G. Gazelle, of neurovascular flow diverter in patient-specific intracranial aneurysms,”
“Elective endovascular versus open surgical repair of abdominal aortic J. Biomech., pp. 2256–2263, 2012.
aneurysms: Systematic review of short-term results,” Radiology, vol. 224, [22] F. Auricchio, M. Conti, M. De Beule, G. De Santis, and B. Verhegghe,
no. 3, pp. 739–747, 2002. “Carotid artery stenting simulation: From patient-specific images to finite
[4] A. J. Molyneux, R. S. Kerr, L.-M. Yu, M. Clarke, M. Sneade, J. A. Yarnold, element analysis,” Med. Eng. Phys., vol. 33, no. 3, pp. 281–289, 2011.
and P. Sandercock, “International subarachnoid aneurysm trial (ISAT) [23] N. Rebelo, R. Fu, and M. Lawrenchuk, “Study of a nitinol stent deployed
of neurosurgical clipping versus endovascular coiling in 2143 patients into anatomically accurate artery geometry and subjected to realistic ser-
with ruptured intracranial aneurysms: A randomised comparison of effects vice loading,” J. Mater. Eng. Perform., vol. 18, no. 5, pp. 655–663, 2009.
on survival, dependency, seizures, rebleeding, subgroups, and aneurysm [24] L. Flórez Valencia, J. Montagnat, and M. Orkisz, “3d models for vascular
occlusion,” Lancet, vol. 366, no. 9488, pp. 809–817, 2005. lumen segmentation in MRA images and for artery-stenting simulation,”
[5] Z. Kulcsar, E. Houdart, A. Bonafe, G. Parker, J. Millar, A. Goddard, IRBM, vol. 28, no. 2, pp. 65–71, 2007.
S. Renowden, G. Gal, B. Turowski, K. Mitchell, F. Gray, M. Rodriguez, [25] S. Appanaboyina, F. Mut, R. Löhner, C. Putman, and J. Cebral, “Compu-
R. van den Berg, A. Gruber, H. Desal, I Wanke, and D. A. Rufenacht, tational fluid dynamics of stented intracranial aneurysms using adaptive
“Intra-aneurysmal thrombosis as a possible cause of delayed aneurysm embedded unstructured grids,” Int. J. Numer. Methods Fluids, vol. 57,
rupture after flow-diversion treatment,” Amer. J. Neuroradiol., vol. 32, no. 5, pp. 475–493, 2008.
no. 1, pp. 20–25, 2011. [26] I. Larrabide, M. Kim, L. Augsburger, M. Villa-Uriol, D. Rüfenacht, and
[6] P. I. D’Urso, G. Lanzino, H. J. Cloft, and D. F. Kallmes, “Flow diversion A. Frangi, “Fast virtual deployment of self-expandable stents: Method
for intracranial aneurysms a review,” Stroke, vol. 42, no. 8, pp. 2363–2368, and in-vitro evaluation for intracranial aneurysmal stenting,” Med. Image
2011. Anal., vol. 16, no. 3, pp. 721–730, 2010.
[7] J. Mas, G. Chatellier, B. Beyssen, A. Branchereau, T. Moulin, [27] J. Montagnat, H. Delingette, and N. Ayache, “A review of deformable sur-
J. Becquemin, V. Larrue, M. Lièvre, D. Leys, J. Bonneville, J. Watelet, faces: Topology, geometry and deformation,” Image Vis. Comput., vol. 19,
J.-P. Pruvo, J.-F. Albucher, A. Viguier, P. Piquet, P. Garnier, F. Viader, no. 14, pp. 1023–1040, 2001.
E. Touze, M. Giroud, H. Hosseini, J.-C. Pillet, P. Favrole, J.-P. Neau, [28] J. Batina, “Unsteady Euler airfoil solutions using unstructured dynamic
and X. Ducrocq, “Endarterectomy versus stenting in patients with symp- meshes,” AIAA J., vol. 28, no. 8, pp. 1381–1388, 1990.
2010 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 7, JULY 2014

[29] G. Wheatley, A. Gurbuz, J. Rodriguez-Lopez, V. Ramaiah, D. Olsen, Katerina Spranger holds the Dipl.-Inf. degree in
J. Williams, and E. Diethrich, “Midterm outcome in 158 consecutive Computer Science from the Humboldt University of
gore tag thoracic endoprostheses: Single center experience,” Ann. Thorac. Berlin, Germany.
Surg., vol. 81, no. 5, pp. 1570–1577, 2006. She is currently a postgraduate researcher at the
[30] L. Pierot, “Flow diverter stents in the treatment of intracranial aneurysms: Department of Engineering Science at the Univer-
Where are we?” J. Neuroradiol., vol. 38, pp. 40–46, 2011. sity of Oxford, U.K. Her research interests include
[31] D. Stoeckel, C. Bonsignore, and S. Duda, “A survey of stent designs,” biomedical engineering, simulations of surgical inter-
Minimally Invas. Ther. Allied Technol., vol. 11, no. 4, pp. 137–147, 2002. ventions, and predictive computations for medicine.
[32] C. Farhat, C. Degand, B. Koobus, and M. Lesoinne, “Torsional springs for
two-dimensional dynamic unstructured fluid meshes,” Comput. Methods
Appl. Mech. Eng., vol. 163, no. 1–4, pp. 231–245, 1998.
[33] F. Blom, “Considerations on the spring analogy,” Int. J. Numer. Methods
Fluids, vol. 32, no. 6, pp. 647–668, 2000.
[34] D. Zeng and C. Ethier, “A semi-torsional spring analogy model for updat- Yiannis Ventikos holds the Kennedy Chair at Uni-
ing unstructured meshes in 3d moving domains,” Finite Elements Anal. versity College London, U.K., where he is also the
Design, vol. 41, no. 11–12, pp. 1118–1139, 2005. Head of the Department of Mechanical Engineering.
[35] K. Spranger, C. Capelli, G. Bosi, S. Schievano, and Y. Ventikos, “Compar- His research interests include transport phenomena,
ison and calibration of a real-time virtual stenting algorithm using finite (bio)fluid mechanics, clinical computing, minimally
element analysis and genetic algorithms,” Comput. Methods Appl. Mech. invasive interventional devices design and optimiza-
Eng. (in review), 2014. tion, and biocomplexity. He has received the M.Eng.
[36] C. Bottasso, D. Detomi, and R. Serra, “The ball-vertex method: a new in Naval Architecture and Marine Engineering and
simple spring analogy method for unstructured dynamic meshes,” Comput. the Ph.D. in Hydrodynamics and Fluid Mechanics
Methods Appl. Mech. Eng., vol. 194, no. 39, pp. 4244–4264, 2005. from National Technical University of Athens.