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

1, JANUARY 2014

Image-Based Mechanical Analysis of Stent

Deformation: Concept and Exemplary
Implementation for Aortic Valve Stents
Michael Gessat∗ , Raoul Hopf, Thomas Pollok, Christoph Russ, Student Member, IEEE, Thomas Frauenfelder,
Simon Harald Sündermann, Sven Hirsch, Edoardo Mazza, Gábor Székely, Member, IEEE, and Volkmar Falk

Abstract—An approach for extracting the radial force load on I. INTRODUCTION

an implanted stent from medical images is proposed. To exem-
TENTS are used in various fields of medicine as a means
plify the approach, a system is presented which computes a ra-
dial force estimation from computer tomography images acquired
from patients who underwent transcatheter aortic valve implanta-
S to dilate and stabilize vessels or other body cavities. Most
stents show a tube-like geometry (or a geometry that is com-
tion (TAVI). The deformed shape of the implanted valve prosthesis’ posed of several tubes) and have a mesh structure that allows
Nitinol frame is extracted from the images. A set of displacement
vectors is computed that parameterizes the observed deformation.
radial compression (crimping) and expansion. Stents are man-
An iterative relaxation algorithm is employed to adapt the infor- ufactured from metal or polymers, can be self-expanding or
mation extracted from the images to a finite-element model of the are expanded by the means of an inflatable balloon introduced
stent, and the radial components of the interaction forces between into the stent before crimping, and, depending on the appli-
the stent and the tissue are extracted. For the evaluation of the cation, vary in diameter as well as length. Across all sizes,
method, tests were run using the clinical data from 21 patients.
Stent modeling and extraction of the radial forces were successful
shapes, materials, and ways of the deployment, all stents rely on
in 18 cases. Synthetic test cases were generated, in addition, for a mechanical equilibrium between the forces trying to embolize
assessing the sensitivity to the measurement errors. In a sensitivity (dislocate) them and a radial force that provides the fixation of
analysis, the geometric error of the stent reconstruction was below the stent inside the vessel. While in many cases this radial force
0.3 mm, which is below the image resolution. The distribution of and the dilatative effect it has on the surrounding tissue is an
the radial forces was qualitatively and quantitatively reasonable.
intended therapeutic necessity that aims at treating a disease-
An uncertainty remains in the quantitative evaluation of the radial
forces due to the uncertainty in defining a radial direction on the induced stenosis, it also gives rise to complications and adverse
deformed stent. With our approach, the mechanical situation of effects. An excessive radial force leads to the irritations of en-
TAVI stents after the implantation can be studied in vivo, which dothelial cells, which can evoke the thrombogenic processes and
may help to understand the mechanisms that lead to the complica- thereby cause restenosis [1], impede the conductivity of cardiac
tions and improve stent design.
cells [2], or in extreme cases may lead to vascular rupture [3].
Index Terms—Biomedical image processing, biomedical imag- Transcatheter aortic valve implantation (TAVI) is a technique
ing, biomechanical simulation, finite element analysis, implants. for the treatment of aortic stenosis in high-risk patients with
severe aortic stenosis. The TAVI procedure uses a stented valve
Manuscript received March 25, 2013; revised June 7, 2013; July 1, 2013;
replacement that is implanted inside the natural, dysfunctional
accepted July 5, 2013. Date of publication July 16, 2013; date of current version aortic valve under beating-heart conditions [4]. The procedure
December 16, 2013. This work was supported in part by the Swiss National is well established and clinical trials have proven its effective-
Science Foundation under Grant CR32I3_135044, in part by the Swiss Heart
Foundation, and in part by the Baden-Württemberg Stipendium. Asterisk indi-
ness and safety [5]. Nevertheless, TAVI is still associated with
cates corresponding author. a number of adverse effects, of which paravalvular leaks (di-
∗ M. Gessat is with the Hybrid Laboratory for Cardiovascular Technologies,
astolic blood flow from the aorta to the left ventricle through
University of Zurich, 8091 Zurich, Switzerland, and also with the Computer
Vision Laboratory, Swiss Federal Institute of Technology (ETH) Zurich, 8092
gaps between the prosthesis and the surrounding tissue) [6] and
Zurich, Switzerland (e-mail: mgessat@ethz.ch). conduction abnormalities, which require treatment with a per-
R. Hopf and E. Mazza are with the Center of Mechanics, Swiss Fed- manent pacemaker [7], are the most common.
eral Institute of Technology (ETH) Zurich, 8092 Zurich, Switzerland (e-mail:
hopf@imes.mavt.ethz.ch; edoardo.mazza@imes.mavt.ethz.ch).
T. Pollok is with the Berlin Institute of Technology, 10623 Berlin, Germany
(e-mail: tom.pollok@googlemail.com).
A. Study Aim
C. Russ, S. Hirsch, and G. Székely are with the Computer Vision Laboratory, We propose means to analyze the biomechanical status of
Swiss Federal Institute of Technology (ETH) Zurich, 8092 Zurich, Switzerland
(e-mail: russc@ethz.ch; hirsch@vision.ee.ethz.ch; szekely@vision.ee.ethz.ch). an implanted stent (or other deformable implant) in vivo by
T. Frauenfelder is with the Institute of Diagnostic and Interventional Radi- combining methods from medical-image analysis, image-based
ology, UniversityHospital Zurich, 8091 Zurich, Switzerland (e-mail: thomas. modeling, and biomechanics. The method makes use of the
frauenfelder@usz.ch).
S. H. Sündermann and V. Falk are with the Division of Cardiovascular stent’s superelastic properties, which define a relation between
Surgery, UniversityHospital Zurich, 8091 Zurich, Switzerland (e-mail: simon. the mechanical load imposed by the surrounding tissue and the
suendermann@usz.ch; volkmar.falk@usz.ch). state of deformation of the implanted stent, and is based on
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. medical images, which could be, depending on the application,
Digital Object Identifier 10.1109/TBME.2013.2273496 generated during or after an implantation. The deformed shape
0018-9294 © 2013 IEEE
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 5

of the implanted stent is extracted from the images, and the me- care, when it comes to slender structures that are subjected to
chanical forces acting on the interface between the device and large rotation bending. Depending on the selection of the exact
the surrounding tissues are estimated as a means to quantita- element type with its assigned integration scheme and on the
tively assess the quality of a treatment and to predict mid- and grade of the meshing itself, undesired numerical side effects
long-term effects such as stent healing, restenosis, or aneurism such as shear locking or hourglassing, may occur. Necessary
regression. Moreover, the method can be applied in order to countermeasures often include mesh refinement until conver-
create in vivo ground truth for the evaluation of predictive simu- gence is reached. This, however, increases model complexity,
lation systems and to provide stent developers with in vivo data raising computational cost, and jeopardizing numerical stabil-
about the conditions under which the stents are employed. ity. Detailed investigations concerning element type selection
Here, we present an exemplary implementation of the con- for the simulation of stent expansion processes were performed
cept applied to the stent of the Medtronic CoreValve revalving by Hall et al. [17], who recommended beam elements for rea-
System, which is one of the two most frequently used TAVI sons of stability and efficiency as used by [15]. This modeling
prostheses. Like any stent, TAVI stents are fixated against the approach has not yet been applied to CoreValve stents, which
hydrodynamic and hydrostatic forces acting on it in their lon- are one order of magnitude larger and not cylindrical.
gitudinal direction through the radial forces at the interface be- Most literature in the context of biomechanics, FEM, and
tween the stent and host tissues [8]–[10]. The valve which is TAVI, is aiming at the development of predictive models, i.e.,
carried by the stent is intended to seal the left ventricle against models for preinterventional in-silico prediction of intervention
blood flowing back from the aorta during diastole. A sufficient outcome. Our approach, however, uses FEM for retrospective
radial force along the complete perimeter of the aortic annulus analysis of the observed postinterventional situation of the aortic
is required to ensure this seal and to prevent paravalvular leaks. root. A similar concept has been proposed by Ganguly et al.
Excessive radial force may lead to rupture of the annulus or in [18]. They applied rotational angiography to patients with
the ascending aorta and is suspected to cause conduction ab- stents in the femoral artery to acquire 3-D image data of the
normalities (atrioventricular node blocks and left bundle branch implanted stents and to extract the centerlines of the stents from
blocks) through the stress imposed onto the conductive fibers. these images. They studied the accuracy of this method for
analyzing stent deformation and were able to show a very good
agreement of the extracted data with the physically measured
B. Related Work values in in vitro experiments. Their work does not include
Biomechanical modeling and analysis of stenting procedures the integration of the measured displacement with a mechanical
have been identified as key components in attempts to optimize simulation. They studied vascular stents with a cylindrical shape
the stents and stent deployment strategies. In order to better and a very fine-grid structure. Therefore, their analysis is limited
understand and be able to predict the behavior of stents inside the to a polynomial representation of the stents’ centerlines rather
human body, various models of the stents as a mechanical device than the wires of the stents. For large and complex shaped stents,
and of the vascular and surrounding tissues as biomechanical such as the CoreValve, this approach would not yield sufficient
components have been presented [11]–[14]. information to describe the deformation required to compute the
Mummert et al. [10] investigated the deployment of braided radial forces.
Nitinol stents in explanted porcine and ovine hearts. Crimping
diameters versus radial force, pullout, aortic annulus deforma- II. METHOD
tion, and friction between the stent and the aortic tissue were
The shape of the CoreValve stent is shown in Figs. 1 and 2. It
looked at. They found a radial dilatation of 2.5 mm necessary
consists of a nickeltitanium alloy known as Nitinol. As reported
to anchor the stent against embolization. The differences in
in [19], the 165 intersection points of the stent’s mesh structure
stent shape and structure (they used braided stents whereas the
can be categorized into 11 layers consisting of 15 points, so
CoreValve like most vascular stents is manufactured from one
that the points in each layer describe a circle around the stent’s
Nitinol tube by laser cutting and shape setting) and implanta-
longest axis. These points are connected by 30 strings running
tion environment (dead, yet nonpathologic animal valves versus
from the proximal to the distal end. At the distal end of the stent,
living, yet heavily calcified human valves) make their findings
two hooks are attached, which are required to pull the stent into
not directly applicable to clinical scenarios.
the delivery system through a funnel-shaped device (see Fig. 2).
Unlike the direct measurement of the stent-interaction forces
Together with the peaks of these hooks, a labeled set of 167
described in [10], our approach employs an indirect measure-
landmarks can be defined.
ment. For that purpose, a numerical model of the stent’s supere-
lastic behavior is required. Most finite-element stent models that
A. Method Overview
deal with TAVI focus on simulating balloon expandable stain-
less steel stents, which exhibit elasto-plastic material behavior An overview of the proposed method is given in Fig. 1.
(e.g., [9], [11], [15]). Mechanical models of the CoreValve stent, It consists of four principal steps. In image preprocessing
which this paper focuses on, have been presented previously, step, image analysis algorithms are employed to detect the
e.g., by Tzamtzis et al. [16]. They used a tetrahedral mesh- region of interest that contains the stent in a CT image and
ing procedure in order to obtain a finite-element model for the to suppress high-intensity voxels which do not belong to the
stent. Such solid element models must be treated with special stent. The subsequent stent reconstruction step extracts the grid
6 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 1, JANUARY 2014

work steps and the algorithms used in the proposed method are
explained in Section II-B –E.

B. Image Preprocessing
The image processing algorithms were implemented in Java
1.6 using the open-source IDE Eclipse. For the visualization
of the images and extracted models, the open-source software
framework OpenMAF 2.0 was used.
1) Imaging: ECG-gated Chest CTs were collected from pa-
tients who had undergone CoreValve implantation and given
their consent for accessing their data. The images were recon-
structed at a spatial resolution of 0.7 mm or finer and at 10
time-steps over the cardiac cycle, but not in all the cases were
all reconstructions still available on the PACS when the images
were retrieved for this study. Of all the available reconstructions
with sufficient resolution, we chose the best systolic one. The
datasets contained a field of view ranging from the abdomen to
the level of the shoulders in caudo-cranial direction, covering
the full depth and width of the chest and were acquired between
2 and 841 days (median: 328) days after the implantation.
2) Suppression of High-Intensity Objects: The CoreValve
stent is, in most cases, not the only metallic object seen in the
images–extracorporeal ECG wires, pacemakers, and other kinds
Fig. 1. Method overview. of implants have a similar X-ray density. In a preprocessing step,
all the high-intensity objects which do certainly not belong to
the stent are detected and suppressed. The selection of these
objects is based on cluster-size analysis. The stents’ wires are
about 0.3 mm thick. At the intersection points, two contacting
wires reach a thickness of about 0.7 mm. When applying a high
threshold of 2000 HU (Hounsfield Units) to the image, the stent
will break into many small voxel clusters due to partial-volume
effects. After thresholding, the connected clusters of the high-
intensity voxels are created and the number of voxels in each
cluster is counted. The maximum number of voxels in a cluster
depends on the image resolution, but will never exceed 300
(based on a theoretical resolution of 0.25x0.25x0.25 mm3 and
the volume of the largest cluster at the distal hooks of the stent
which is ≈ 5 mm3 ). A safety margin of 100 voxels is added to
Fig. 2. Shape and structure of the Medtronic CoreValve stent. The balls mark this value. All clusters larger than 400 voxels are erased.
the 167 grid points which are appointed to 11 layers, each visualized as a circle. 3) Stent Segmentation: After the suppression of all irrelevant
On these circles, the landmarks have a distance of 24◦ , neighboring layers are metal objects, presegmentation is performed in order to extract a
shifted by 12◦ . The regularity of the structure allows for estimating the positions
of the intersection points in the Layers (j − 1) and (j + 1) from the known volume of interest (VOI) that contains the stent. A high threshold
intersection points in Layer j. (1700 HU) is applied to generate the seed voxels for detecting
of the remaining objects with high intensity in the image. In
Fig. 3(a), all voxels in the region of the stent which remain after
points of the stent and a model of the deformed stent is created applying this threshold are marked red. These voxels are used
using Bezier interpolation. The displacements of the grid points as seed points for a region growing step that connects all the
are employed as kinematic boundary conditions to a mechanical neighboring voxels with an intensity of 650 HU or larger. The
model of the stent. A relaxation step is included, which allows result of this second step on the stent region can be seen in
the grid points to migrate within a small region toward a mini- Fig. 3(b).
mum of the deformation energy in the stent, to compensate for 4) Calcium Suppression: The segmentation method de-
landmark localization errors. This relaxation step is alternated scribed to this point is based on a combination of thresholding
with a radial straining step in order to bias the relaxation to the and region growing. Both methods are prone to oversegmen-
normal and the axial direction. After finding a force equilibrium, tation errors, which in this case would lead to an inclusion of
the reaction forces between the stent and the surrounding tissues the nearby calcifications into the stent segmentation–the image
are read out along the radial direction at each grid point. The intensity of calcium can reach values of 1600 HU and more.
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 7

among those voxels which are the most likely to represent the
intersection points of the stent in its deformed configuration.
This step is relying on the fact that, due to the partial-volume
effects, the intensity of the voxels at the intersection of two
strings is higher than anywhere else in the stent. The minimal
distance between two intersection points is in any case larger
than 2 mm. A nonmaximum suppression with a kernel size of
2 mm is applied to identify the local maxima within the stent
volume. These maxima are considered as candidates for the
identification of landmarks that mark the intersection points.
Noise, incompletely suppressed calcium, and other effects can
yield false positives in the list of landmark candidates. Moreover,
the variances in the image quality can result in undersegmenta-
tion, i.e., to incomplete sets of landmark candidates. Both the
errors need to be corrected in order to create a complete and cor-
rect set of 167 landmarks. At this point, prior knowledge about
the topology and the geometry of the stent is taken into consid-
eration as a lead in the identification of false candidates, and for
the generation of search spaces to find the missing landmarks.
Fig. 3. (a) Seed points for the region growing after thresholding in the stent
region (red). (b) Result of the region growing in the stent region. A reference model of the stent in its undeformed (unloaded)
state was created based on microCT images of CoreValves that
were acquired under no mechanical load (other than gravity).
Principal component analysis is applied to find the center of
gravity (mean) and the longest axis of the landmark candidates.
The reference model is registered according to this axis and the
mean. The remaining degree of freedom, rotation around the
long axis, is determined according to the clamp points at the
distal end of the stent.
For an identification of the correct landmark candidates and
in order to assign the landmarks to the correct layer, a slid-
ing plane is applied, which scans the candidate cloud along the
longest axis of the stent. The plane through the second highest
layer of the undeformed stent model (Layer 10) is computed.
All the landmark candidates which are within a vertical distance
less than 3.8 mm (closer to this plane than the two neighboring
planes) are considered the candidates for this layer. The candi-
dates are sorted according to their angle around the central axis
Fig. 4. Separation of the stent voxels and the calcium voxels. of the stent. In the unloaded state, the angular distance between
two neighboring intersection points is 24◦ . The selection algo-
The final step of the image preprocessing is the suppression rithm picks those 15 landmark candidates which best fulfill this
of those voxels represent the calcium clusters at the stent’s criterion and assigns them to Layer 10. Starting from Layer 10,
perimeter. The method applied for the calcium suppression is an iterative search is performed to identify the correct landmarks
summarized in Fig. 4. The VOI containing the stent is sliced in all layers based on the knowledge of the distance between all
in 4 mm steps along an approximation of the stent’s centerline layers in the unloaded state and the alternating rotation of ap-
computed as the longest axis of the VOI. Each slice is divided proximately 12◦ between the intersection points in neighboring
into 15 sectors of 24◦ around the central axis. In each sector, layers (see Fig. 2). Thus, the likelihood of a candidate to rep-
the local intensity maxima are found and the maximum closest resent an intersection point of the stent can be rated. In each
to the center is selected. Based on the maxima in all sectors, layer, those 15 landmarks with the highest level of confidence
the stent boundaries are approximated according to the known are selected.
thickness of the strings. The voxels outside of these bound- Once all the landmarks are localized, identified, and labeled,
aries are labeled as calcium and not considered for the stent the Bezier control polygons are constructed that define the 30
reconstruction. centerlines of each of the stent’s strings from the proximal to
the distal end according to the constraints defined in [19].
C. Stent Reconstruction
The objective of the image preprocessing performed to this D. Relaxation
point was the identification of the voxels which very likely be- The image processing pipeline generates an estimate for the
long to the stent. In the following, these candidates are identified position of each intersection point of the stent. The relation
8 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 1, JANUARY 2014

between neighboring intersection points is used to separate the

true and the false landmark candidates. Nevertheless, up to this
point in the algorithm, the position of each landmark is estimated
independently of its neighbors. Image resolution, noise, and ar-
tifacts will inevitably create an uncertainty in these positions,
which we quantify
 in SectionIII-C. Thus, the observed posi-
tions P0 = pi,0 , i ∈ [1, 165] ⊂ R3 delivered by the process
so far will differ from the true positions Ṕ = {ṕi , i ∈ [1, 165]},
and for each landmark an unknown landmark positioning error
ei,0 = pi,0 − ṕi ∈ R3 is to be considered. The error vector ei,0
is superimposed on the actual displacement vector d́i of that
intersection point which locally describes the deformation of
the stent from its unloaded state to the deformed state, so that
the measured displacement is di,0 = d́i + ei,0 .
A mechanical model of the stent implemented in Abaqus
6.10-EF1 (Dassault Systemes, Vélizy-Villacoublay, France) is
applied as a means to reduce the landmark positioning error ei,0
in order to get a more accurate estimate of the node displacement
d́i . Fig. 5. (a) A nonlinear spring is fixed between a intersection point q i and a
virtual pivot point q̂ i . (b) q̂ i is pulled to a position p̂ i , 0 lying 0.12 mm inside
1) Beam Model: A finite-element model of the stent is set up of the measured landmark position p i . (c) Spring allows the intersection point
using structural beam elements. MATLAB R2013a (The Math- to move to a position p i , 1 of minimal strain within a radius of 0.2 mm.
Works, Natick, MA, USA) is used to generate the Abaqus input
file. First the control polygon of each string of the stent in the with a linear-elastic material model, characterized by the two
stress-free reference configuration is sampled equidistantly. The parameters Young’s modulus E and Poisson number ν. The
neighboring nodes within a string are used to define the local latter is commonly set to 0.33. Young’s modulus E of Nitinol
end points of a two point beam element. Additionally, the beam is reported between 40 and 75 GPa in the literature [16], [21].
elements require the definition of a cross-sectional geometry For our beam model of the CoreValve stent, a value of 60 GPa
and its orientation. For the CoreValve stent this cross-sectional delivered the best experimental results (see Section III-A).
geometry is not constant over each string. This variation is im- 2) Soft Kinematic Constraints: Due to the uncertainty of the
plemented and controlled by using linear functions of the arc landmark extraction, the deformed coordinates cannot directly
length over every spline segment. The element type chosen for be applied as the kinematic boundary conditions. This would
this application is the Timoshenko-based B31 element [17]. This lead to a strong noise amplification in the axial and the tangen-
selection is not motivated by a strong need of the transverse shear tial directions of the reaction forces and render the desired radial
compliance but by the fact that these particular elements are well components highly unreliable. To allow the stent’s internal me-
suited to simulate situations involving large deflections in 3-D chanics smoothing out landmark localization errors, 1-D force
space. Assigning beam elements in this way, leads to a continu- elements with nonlinear force laws are used to communicate
ous beam structure for each of the 30 strings. The points at the between the measured displacements and the grid points of the
intersections of neighboring strings are locally approximated as stent model as shown in Fig. 5.
rigid bodies. For this purpose, both nodes of each intersection For each intersection point qi of the undeformed stent model,
are coupled kinematically in the corresponding manner. This is a virtual pivot point q̂i is created at a distance of 0.1 mm along
implemented using the beam-type constraint of Abaqus. Cou- the local normal of the stent surface. A force element (spring)
pling of the strings completes the structural model of the stent. fi is defined between q̂i and qi with the force law
Convergence analysis was employed to define the optimal res- di = q
olution of the model [20]. The final mesh consists of 252 linear ⎧i −N q̂i 2
⎨ 0 /mm , d < 0.1 mm
beam elements per string, resulting in 7560 elements for the s(d) = 0.1 N /mm , 0.1mm ≤ d < 0.2 mm
whole stent. ⎩
50 N /mm , d ≥ 0.2 mm
A constitutive model of the stent material is added to the fi = s(di ) · (qi − q̂i )
model. Nitinol exhibits superelastic properties, which are re-
quired for the crimping and deployment of the stenting proce- To softly constrain the model with the measured displace-
dure. At this stage, we are interested in the final equilibrium ments di,0 , the pivotal point q̂i is pulled to a location
between the tissue and stent. In the configurations that we found
in the actual patient data, the strain rates in the stent remain p̂i,0 = pi,0 − 0.12 · ri,0
within the linear low-strain regime of Nitinol. By taking the
along a radial vector ri,0, that is computed as the normalized
stent model from a stress-free reference state directly to the end
projection of the vector pointing from pi,0 to the centerline of
configuration, we omitted the high-strain states present in the
the stent onto the plane,
 that equalizes the 15 intersection points
crimped stent. This allowed us to describe the stent material
on layer li = i−1 15 ∈ [1,11].
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 9

The implicit solver of Abaqus is used to find a minimum of The results of each experiment and discussion of these findings
the strain energy in the stent and the virtual spring elements. The is given in Section IV.
force law is selected in such a way that the physical intersection
points have the freedom to move inside a 0.2 mm sphere around A. Mechanical Model Verification
the displaced pivotal points p̂i,0 to positions pi,1 . This allows
Most experimental protocols for stent testing, involve a two-
the stent to find a global minimum of its potential energy, thus
plate crush procedure as defined by ISO 25539-2:2008 [16].
smoothing out the high-force spikes in undesired directions.
Performing and simulating the crush tests is challenging by
3) Incremental Stent Relaxation: An incremental routine us-
default, since nonlinearities arise from solving the mechanical
ing alternating straining and relaxation steps is employed to
contact. The CoreValve’s curved geometry leads to additional
move the landmark coordinates P = {pi , i ∈ [1, 165]} from the
difficulties. To avoid these issues, we employed two-point ten-
initial positions P0 to more accurate positions Pn . Starting with
sile tests. The stent was clamped at two points and then sub-
the initial landmark coordinates P0 , the first relaxation step is
jected to tension. Internally, the stent reacted with a complex
computed as described previously to yield the landmark coordi-
mode of deformation involving bending, tension, and torsion of
nates P1 , and the spring forces fi,1 are reported. P1 is fed back
its struts. Externally, it acted as a 1-D, nonlinear spring connect-
into the relaxation algorithm as the new initial landmark coor-
ing the two clamping points. The characteristic curve of that
dinates, yielding the landmark coordinates P2 and the spring
spring was measured experimentally for three different pairs of
forces fi,2 . The relaxation of step k = 2 is computed as
clamping points and compared to the simulated reaction to the
165 same tension as computed with our beam model.
ρk = fi,k 2 − fi,k −1 2 .
i=1 B. Stability of the Image Processing Pipeline
The iteration is continued until ρk < 0.01 · ρ2 and Pn = Pk 21 CT datasets where available for testing. The automatic
is reported as the final intersection point set. The rationale for preprocessing and landmark extraction algorithms were run on
this stop condition lies in the aspect ratio between the forces pro- all 21 images in order to test whether they are able to locate all
duced by the stent in response to the radial, axial, and tangential the landmarks in all the images.
displacement of the nodes. When we pull the pivot points in-
ward on a 2-D plane through the measured coordinate and then
C. Estimation of Reconstruction Accuracy
allow the intersection points to migrate inside a sphere around
the pivot point, the axial and tangential components of the in- The accuracy of the automatic landmark localization was
ternal forces in the stent will dominate the direction of the point estimated by comparing the results with manually extracted
migration, as the axial and tangential stiffness of the stent ex- landmark positions, which were defined by three human ob-
ceeds its radial stiffness by several orders of magnitude (see servers using a graphical interface. Due to the very time-
Section IV-B). After a couple of iterations (in practice 3 to 5), consuming nature of the manual landmark localization, the
all nodes moved to a local minimum of the overall stent energy, comparison was limited to five datasets in which all the three
with respect to the axial and tangential deformation, and the observers localized all 167 landmarks. The three different loca-
annealing of force (energy) between two steps starts to decline. tions pij , i ∈ {1, 2, . . . , 167}, j ∈ {1, 2, 3} given by the three
From that point on, the stent will primarily use the radial degree observers for each landmark were used to define Si (p̂i , ri ), the
of freedom to relieve itself of the radial displacement. smallest sphere which contains the triangle pij . The center and
radius of this sphere are identical to the center mci and radius
E. Mechanical Analysis rci of the circumscribed circle for the triangles, where all the
interior angles are < 90◦ . For the blunt-angled triangles, the
At each intersection point pi,n , a local coordinate system is sphere is defined by the center of the longest edge and half its
defined through a radial, tangential, and axial vector. In Fig. 1, length:
these directions are indicated by the red (axial), green (radial), ⎧
and blue (normal) arrows at each node. The reaction forces fi,n ⎪
⎪
⎪ S p i 0 +p i1
, p i 0 +p i1 
, ∠ (pi0 , pi2 , pi1 ) ≥ 90◦
in the springs are transformed to these coordinate systems and ⎪
⎪
2 2
⎨
only the radial component is used for the analysis of the radial S p i 0 +p
2
i2
, p i 0 +p
2
i2 
, ∠ (pi0 , pi1 , pi2 ) ≥ 90◦
Si =
⎪
⎪
force between the stent and the tissue. The average radial force ⎪
⎪ S p i 1 +p i2
, p i 1 +p i2 
, ∠ (pi1 , pi0 , pi2 ) ≥ 90◦
Fj per layer is computed as ⎪
⎩
2 2
S (mci , rci ) , otherwise.
15
1 For each automatically identified landmark, an acceptance
Fj = f15j +i,n 2 .
15 i=1
function A is defined depending on a parameter α which accepts
a landmark if it is not further away from any of the three manual
III. EXPERIMENTAL SETUP landmarks pij than the radius of the sphere Si multiplied with
the factor α:
The performance of each system component as well as of the
overall system was investigated in a series of experiments. In 1, min (p, pij ) ≤ α · ri
Ai,α (p) = j
the following sections, the different experiments are described. 0, otherwise.
10 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 1, JANUARY 2014

With this, for a complete set of automatically localized land- Besides the geometric accuracy, we also analyzed the differ-
marks pi , i ∈ {1, . . . , 167}, the acceptance rate Ā for one stent ence between the radial interaction forces in the ground truth
is defined as a function of α: and the output after each increment and computed force errors
167 per node, per layer, and for the whole stent in each step k:
(Ai,α (pi ))
Ā (α) = i=1 .
167
 
To better relate these accuracy measurements to the radial  
φi,k = fi,k − f´i 
force, the effect of the localization error onto the effective radial 2
strain was considered. We computed for each layer k of the stent 15
the cross-sectional area Ak as the area circumscribed by a closed ϕj,k = e15j +i,k
polygon formed by the 15 intersection points on that layer in i=1
the undeformed configuration. For each of the five patients in 11
this arm of the study, four different versions of the intersection Φk = ϕj,k .
point sets in the deformed configuration were available, three j =1
were defined manually (j ∈ {1, 2, 3}), and one automatically
(j = 4). For each of these cases j and on each layer k we
computed the deformed cross-sectional area Aj k . The radial
IV. RESULT AND DISCUSSION
strain j k is defined for each layer k and the dataset j by relating
the root of Aj k to the root of Ak : A. Computational Cost
√ 
Ak − Aj k Depending on the image size and the amount of high-intensity
j k = √ . voxels and voxel regions that needed classification, the image
Ak
preprocessing step took between 120 and 240 s per dataset.
D. Validation of Error Reduction Using Incremental The extraction of landmarks for the stent reconstruction took
Relaxation between 30 and 60 s. Both steps were conducted on a work-
station with an Intel Xeon E5647 processor with six cores (12
The ability of the incremental method to reduce the landmark threads) at 2.4 GHz equipped with 12 GB of RAM running Mi-
localization error was assessed using realistic, yet artificial, dis- crosoft Windows 7. The implementation was not optimized with
placement fields as ground truth and perturbing them by random respect to the computation times or memory usage. The relax-
noise. Random displacement vectors were added to the intersec- ation step took 90 s on an Intel Core i7-3770K CPU with four
tion points in deformation cases Ṕ to create noisy landmark sets cores (8 threads) at 3.50 GHz equipped with 8 GB of RAM run-
P0 . The direction of these vectors was randomized according to ning Linux. The force readout and postprocessing took around
a uniform distribution; the norm of the vectors follows a normal 1.5 s on the same machine.
distribution with a mean of 0.74 mm and a standard deviation
of 0.16 mm.
The position of each node was traced over the relaxation
process. The node error ei,k , mean layer error j,k , and overall B. Mechanical Model Verification
mean error Ek were computed for each step k: Three tension experiments were conducted. In the first exper-
ei,k = pi,k − ṕi 2 iment, the two distal hooks were clamped for the tension test,

15 creating tension along a radial direction. This yielded relatively
1
j,k = /15 e15j +i,k small forces, as primarily bending of the thin stent wires oc-
i=1

11 curred. The second experiment was an axial test, with the stent
1
Ek = /11 j,k . being clamped at the distal and proximal end of one string. This
j =1 resulted primarily in the axial tension of struts, yielding rela-
As an additional measure for the geometric accuracy, the tively large reaction forces. Finally, a superimposition of both
projected circumference in each layer was computed for the effects was achieved by clamping one distal hook and a proximal
ground truth Ṕ as well as for each Pi . Therefore, the landmarks point underneath the opposing hook (see Fig. 6).
in each layer were projected to the equalizing plane and the The blue line in Fig. 6 shows for the third configuration the
circumferences of the polygons spanned by the projected points tension force that was measured continuously in one clamp
were computed. After each step k, the circumferential error while slowly increasing the distance between the two clamps
γj,k was computed for each layer as the difference between by 3.8 mm. The red line shows the simulation result for the
the projected circumference ćj in the noise-free dataset and the same setup. For larger deflections the numerical model leads to a
projected circumference cj,k . As a global measure for accuracy, higher global stiffness of the stent, which is to be expected for an
the mean Γk of the absolute values of γj,k was computed for implementation with a linear-elastic material model. However,
each increment k as the estimations of the occurring physiological strains which
11 are present in the implanted equilibrium state after TAVI are
1
Γk = |γj,k | . not expected to exceed the limit imposed by this choice of the
11 j =1 constitutive model.
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 11

1.6 voxels (0.68 mm) away from the nearest manual landmark.
The maximum error was 3.24 voxels (1.3 mm). Thus, the land-
mark localization error compensation included in the mechani-
cal analysis has to compensate an error of that magnitude.
Fig. 7(b) shows the acceptance function plotted for α ∈ [0, 7].
With an α-value of ≥ 2, more than 90% of landmarks were
accepted. Fig. 8 shows the impact the landmark position uncer-
tainty has on the radial strain.
Over all five datasets and eleven layers per dataset, differences
in the radial strain between the models based on the point clouds
found by the algorithms, and the point clouds defined by the
human observers lay between −2% and 2%. The differences
Fig. 6. Validation of the beam model. The image shows how the stent model between the three study participants (interobserver variance)
was loaded to create a superimposition of the radial, axial, and tangential strain. were in the same region.
The plot shows the simulated (red) and the experimentally measured (blue) reac-
tion force in the clamp over the linear deflection that was applied by increasing
the distance between the clamps. E. Validation of Error Reduction Using Incremental
Relaxation
C. Stability of the Image Processing Pipeline Fig. 9(a) and (b) shows the development of the mean node
The identification of the stent among all the high-intensity error Ek , after adding random noise to two different patient
objects in the image was successful in all 21 cases. The bound- datasets (“input”) and after 10 increments of the relaxation. The
ing boxes derived from each of the 21 datasets were visually initial mean error of 0.76 mm is reduced to 0.2 mm after five
approved to be consistent with the area that contained the stent. steps in case one, and to 0.27 mm after four steps in case two.
Thresholding and calcium suppression in all cases delivered a re- The same plots show the relaxation ρk for increments 2 . . . 10.
sult which passed qualitative visual assessment. At this point no The convergence criterion was met for k = 5 steps in case 1
accuracy measurement was performed, the assessment was only and k = 4 in case 2. The projected circumferential errors for the
looking for extreme over- or under-segmentation errors. The same test cases are plotted in Fig. 9(c) and (d).
landmark identification was able to extract all 167 landmarks The additional radial strain, which is added by the virtual
(including the distal hooks) in 10 datasets. In eight additional springs, lead to a reduction of the circumferences in the early
datasets, one or both of the landmarks representing the distal phase of the incremental relaxation. In that phase, the degrees
hooks could not be localized but all 165 intersections points of freedom were predominantly used to reduce the axial and
were found. In three datasets, landmark localization failed. Two tangential noise. This lead to negative values of γj,k for small
of these cases showed motion artifacts in the form of double- values of k. Only in the last step before the convergence cri-
images of the stent in the aortal region; one image was cut off terion is reached, the radial error was reduced significantly for
to not contain the complete stent. These three datasets were not each layer. After reaching the convergence criterion, the stent
considered for the subsequent analyses, leaving 18 cases for started an approximately linear radial expansion, which lead to
testing the relaxation step and extraction of forces. Adding the an increasingly large circumferential error.
one or two missing hook landmarks in those ten datasets where
automatic landmark localization did not find them, would re- F. Radial Forces
quire user interaction of 1 or 2 minutes, including the loading of
cases and landmark clouds, the manual localization of the miss- Fig. 9(e) and (g) shows on a logarithmic scale the development
ing landmarks, and the export of the fixed landmark cloud. Since of the radial force errors ϕj,k and Φk over the 10 increments
the radial forces administered to the aorta through these hooks k of the relaxation algorithm. Below each plot, in Fig. 9(f) and
on the distal end of the stent are negligible for the overall equi- (h), the decline of the tangential and axial force components is
librium, they were omitted in the mechanical model described plotted. The vertical red line indicates the step after which the
in Section II-D. convergence criterion was reached for the relaxation. In both
cases, the initial radial force error was reduced by several orders
of magnitude, the same is true for the axial and tangential force
D. Validation of Reconstruction Accuracy components. The minimum in the global radial force error Φk
The mean landmark localization error over all five datasets and for most layer-wise radial force errors ϕj,k was seen after
was 0.37 mm ± 0.16 mm. Fig. 7(a) shows for each of the pa- two steps in the case shown on the left and after three steps in
tients the accumulated histogram of the distance between the the case shown on the right, while the axial and tangential forces
automatically localized landmarks and the nearest manually an- required four steps in the first and five steps in the second case.
notated landmark. To level out differences in the image reso- Thus, the radial error was slightly higher in the final result of the
lution between the different datasets, the distances were mea- relaxation algorithm than it had been earlier in the incremental
sured in number of voxels. The voxel spacing lay between 0.4 routine. Nevertheless, an error reduction of more than 90% is
and 0.6 mm. More than 90% of the landmarks were less than reached.
12 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 1, JANUARY 2014

Fig. 7. (a) Accumulated histogram of the distance between automatically found landmarks to the nearest manual reference. Distances are measured in voxels.
P1 – P5 denote the five patients of which the manual data was available for the comparison. (b) Value of the acceptance function for α ∈ [0, 7] for the five patient
datasets. The solid red line shows the mean value over all the five datasets.

Fig. 8. Radial strain computed based on the three manually defined and one automatically extracted landmark cloud for the five different patients. The pictures
shown next to each plot depict the extracted stent shape for the respective patient.

G. Analysis Result When we computed the hoop force for the patient cases with
annular deformations in that range, the calculated hoop forces
The described procedure was applied to the CT image datasets
of 21 patients, stent reconstruction, and derivation of the radial fell into the same range of [3.5, 5.0] N.
forces was possible in 18 cases (see Section IV-C).
The radial forces measured at the stent nodes ranged from V. CONCLUSION AND OUTLOOK
0.0 N (in areas without contact to the vascular wall) to 2.06 N We propose an image-based system for the estimation of the
in one node. The average radial force per node was in the range in vivo forces on the interface of an implanted Nitinol stent
of [0.3, 0.4] N at the aortic annulus, ≤ 0.1 N in the medial area, and the surrounding vascular and cardiac tissues. This proce-
and between 0.2 and 0.4 N at the distal end. Visual inspection dure takes advantage of the superelastic behavior of the Nitinol
showed a good correlation between the local maxima of the stent. The system implements a two-staged approach, where the
radial force and calcium clusters (see Fig. 10) as well as regions first stage extracts a discrete set of displacement vectors which
of large radial strain or ellipticity. For visualization, the point parameterize the deformation of the stent from a CT image.
forces given in the intersection points are distributed along the The second stage includes a mechanical simulation, utilizing
splines interconnecting the intersection points using a linear the finite-element method as a means to estimate the radial
interpolation scheme. forces that are required to yield the observed deformation. Both
For comparison with the works of Tzamtzis et al., we com- stages were verified and validated using real patient data and
puted from our nodal-force data a hoop force similar to the artificial test data. The extraction of landmarks from the CT im-
definition found in [16]. They reported hoop forces in the range ages was shown to be not worse than the gold standard, which
of [3.5, 5.0] N for radial compressions in the range of 5 mm. was represented by the manual annotation of the test datasets by
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 13

Fig. 9. (a) and (b) Mean and standard deviation of node error E k in mm (linear, left axis) and relaxation ρk in Newtons (logarithmic, right axis) for two test
cases with random noise over 10 increments k. (c) and (d) Projected circumferential errors per slice γj, k (bars) and mean of absolute values Γ k (red dotted line)
for two test cases with the artificial noise over 10 increments k. (e) and (g) Radial force errors ϕ j, k and Φ k over the 10 increments k of the relaxation algorithm
for the two patient cases with artificial noise. (f) and (h) Axial and tangential force component in the same increments.

Fig. 10. Influence of calcifications onto force distribution: the calcium cluster in the foreground of the left image causes a local maximum of the radial force,
highlighted in the center. The blue plane indicates the level of the native aortic annulus. The images on the right show three cross-sectional visualizations of the
calcium cluster as seen in the CT image.
14 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 61, NO. 1, JANUARY 2014

three human observers. The employed mechanism for the FEM- [4] E. Grube, J. C. Laborde, B. Zickmann, U. Gerckens, T. Felderhoff,
based relaxation of the landmark positions in order to derive a B. Sauren, A. Bootsveld, L. Buellesfeld, and S. Iversen, “First report on a
human percutaneous transluminal implantation of a self-expanding valve
more natural stent shape was shown to significantly reduce the prosthesis for interventional treatment of aortic valve stenosis,” Catheter.
measurement uncertainty. This was shown using different error Cardiovasc. Interv., vol. 66, no. 4, pp. 465–469, Dec. 2005.
metrics based on the geometric measurements as well as on the [5] M. B. Leon, C. R. Smith, M. Mack, D. C. Miller, J. W. Moses,
L. G. Svensson, E. M. Tuzcu, J. G. Webb, G. P. Fontana, R. R. Makkar,
analysis of the interaction forces. D. L. Brown, P. C. Block, R. A. Guyton, A. D. Pichard, J. E. Bavaria,
With a mean node error below 0.2 mm and a final mean cir- H. C. Herrmann, P. S. Douglas, J. L. Petersen, J. J. Akin, W. N. Anderson,
cumferential error below 1 mm, the geometric accuracy of the D. Wang, and S. Pocock, “Transcatheter aortic-valve implantation for
aortic stenosis in patients who cannot undergo surgery,” Engl. J. Med.,
reconstruction algorithm can be judged as highly accurate, ex- vol. 363, no. 17, pp. 1597–1607, Oct. 2010.
ceeding the resolution of the underlying image data by a factor [6] N. Buzzatti, F. Maisano, A. Latib, M. Cioni, M. Taramasso, M. Mussardo,
of two. A large uncertainty remains in the final estimates of the A. Colombo, and O. Alfieri, “Computed tomography-based evaluation
of aortic annulus, prosthesis size and impact on early residual aortic re-
radial force. This uncertainty is caused by the three effects. First, gurgitation after transcatheter aortic valve implantation,” Eur. J. Cardio-
the structure of the stent results inherently in an amplification Thoracic Surg., vol. 43, no. 1, pp. 43–51, Jan. 2013.
of the tangential and axial measurement errors, which results in [7] T. Schroeter, A. Linke, M. Haensig, D. R. Merk, M. A. Borger, F. W. Mohr,
and G. Schuler, “Predictors of permanent pacemaker implantation af-
very large tangential and axial force vectors in comparison to the ter medtronic corevalve bioprosthesis implantation,” Europace, vol. 14,
acutally output radial forces. Second, no real contact simulation no. 12, pp. 1759–1763, Dec. 2012.
is employed to yield the true reaction forces normal to some [8] M. Padala, E. L. Sarin, P. Willis, V. Babaliaros, P. Block, R. A. Guyton,
and V. H. Thourani, “An engineering review of transcatheter aortic valve
contact surface. Instead, the radial direction is estimated based technologies,” Cardiovasc. Eng. Technol., vol. 1, no. 1, pp. 77–87, Mar.
on a normal vector computed for each layer and a tangential 2010.
vector computed for each landmark from the two neighboring [9] Q. Wang, E. Sirois, and W. Sun, “Patient-specific modeling of biomechan-
ical interaction in transcatheter aortic valve deployment,” J. Biomechan.,
landmarks in the same layer. Third, as a result of the first two, vol. 45, no. 11, pp. 1965–1971, Jul. 2012.
the projection of the reaction force in the spring element which [10] J. Mummert, E. Sirois, and W. Sun, “Quantification of biomechanical
is used to approximate the radial force is not accurately aligned interaction of transcatheter aortic valve stent deployed in porcine and
ovine hearts,” Ann. Biomed. Eng., vol. 41, no. 3, pp. 577–586, Mar.
with the real surface normal. Thus, the radial force is a linear 2013.
combination of the radial, normal, and axial force components, [11] F. Auricchio, M. Conti, S. Morganti, and A. Reali, “Simulation of tran-
whereby the aspect ratio between radial and axial/tangential scatheter aortic valve implantation: A patient-specific finite-element ap-
proach,” Comput. Methods Biomechan. Biomed. Eng., Feb. 13, 2013.
components already results in large errors, even for small de- [Epub ahead of print].
viations from the true surface normal. To compensate for these [12] H. Zahedmanesh, H. van Oosterwyck, and C. Lally, “A multi-scale
effects, the actual contact surface needs to be explicitly modeled mechanobiological model of in-stent restenosis: Deciphering the role
of matrix metalloproteinase and extracellular matrix changes,” Comput.
and simulated. Methods Biomechan. Biomed. Eng., Sep. 12, 2012. [Epub ahead of print].
Our data shows that calcium is the one key factor in the [13] D. Roy, C. Kauffmann, S. Delorme, S. Lerouge, G. Cloutier, and G.
distribution of the radial force on the stent. In future studies Soulez, “A literature review of the numerical analysis of abdominal aortic
aneurysms treated with endovascular stent grafts,” Comput. Mathemat.
we will quantify the impact of the other influences such as the Methods Med.: 820389. Epub 2012 Sep. 6.
natural annulus shape onto the stent shape and force distribution, [14] S. Morlacchi, C. Chiastra, D. Gastaldi, G. Pennati, G. Dubini, and F. Migli-
and will investigate the changes of both over-the-heart cycle as avacca, “Sequential structural and fluid dynamic numerical simulations of
a stented bifurcated coronary artery,” J. Biomechan. Eng., vol. 133, no.
an effect of the varying blood pressure. 12, 121010, 11 pp., Dec. 29, 2011.
[15] S. Schievano, C. Capelli, D. Cosentino, G. M. Bosi, and A. M. Taylor,
“Finite-element analysis to study percutaneous heart valves,” in Proc.
ACKNOWLEDGMENT Finite Element Anal. Biomed. Appl. Ind. Devel., New York, NY, USA:
InTech, 2011.
The authors would like to thank their colleagues at the Di- [16] S. Tzamtzis, J. Viquerat, J. Yap, M. Mullen, and G. Burriesci, “Numerical
vision for Cardiovascular Surgery, the Institute for Diagnostic analysis of the radial force produced by the medtronic-corevalve and
Edwards–Sapien after transcatheter aortic valve implantation (TAVI),”
and Interventional Radiology, and the Division for Cardiology Med. Eng. Phys., vol. 35, no. 1, pp. 125–130, Jan. 2013.
of UniversityHospital Zurich. [17] G. J. Hall and E. P. Kasper, “Comparison of element technologies for
modeling stent expansion,” J. Biomechan. Eng., vol. 128, no. 5, pp. 751–
756, Oct. 2006.
REFERENCES [18] A. Ganguly, J. Simons, A. Schneider, B. Keck, N. R. Bennett,
R. J. Herfkens, S. M. Coogan, and R. Fahrig, “In-vivo imaging of femoral
[1] J. W. Freeman, P. B. Snowhill, and J. L. Nosher, “A link between stent artery nitinol stents for deformation analysis,” J. Vasc. Intervent. Radiol.,
radial forces and vascular wall remodeling: The discovery of an optimal vol. 22, no. 2, pp. 244–249, Feb. 2011.
stent radial force for minimal vessel restenosis,” Connect. Tissue Res., [19] M. Gessat, L. Altwegg, T. Frauenfelder, A. Plass, and V. Falk, “Cubic
vol. 51, no. 4, pp. 314–326, Aug. 2010. hermite Bezier spline based reconstruction of implanted aortic valve stents
[2] G. Latsios, U. Gerckens, L. Buellesfeld, R. Mueller, D. John, S. Yuecel, from CT images,” in Proc. IEEE Eng. Med. Biol. Soc. Conf., Aug.–Sep.
J. Syring, B. Sauren, and E. Grube, “Device landing zone calcification, 2011, pp. 2667–2670.
assessed by MSCT, as a predictive factor for pacemaker implantation [20] R. Hopf, M. Gessat, V. Falk, and E. Mazza, “Stent induced loading of the
after TAVI,” Catheter. Cardiovasc. Interv., vol. 76, no. 3, pp. 431–439, aortic valve complex,” presented at the MICCAI Stent-Workshop, Nizza,
Sep. 2010. France, 2012.
[3] X. Chaufour, G. White, B. Hambly, W. Yu, J. May, J. Harris, and [21] L. Petrini, F. Migliavacca, P. Massarotti, S. Schievano, G. Dubini, and
M. Stephen, “Evaluation of the risks of using an oversized balloon catheter F. Auricchio, “Computational studies of shape memory alloy behavior in
in the human infrarenal abdominal aorta,” Eur. J. Vasc. Endovasc. Surg., biomedical applications,” J. Biomechan. Eng., vol. 127, no. 4, pp. 716–
vol. 16, no. 2, pp. 142–147, Aug. 1998. 725, Aug. 2005.
GESSAT et al.: IMAGE-BASED MECHANICAL ANALYSIS OF STENT DEFORMATION: CONCEPT AND EXEMPLARY IMPLEMENTATION 15

Michael Gessat received the Graduate degree in in- Simon Harald Sündermann received the Graduate
formatics from the University of Karlsruhe, Germany, degree in medicine and the M.D. degree from Uni-
and the Ph.D. degree from the Institute of Computer versität Leipzig, Germany, in 2007 and 2012.
Science, Universität Leipzig, Germany, in 2005 and From 2008 to 2009, he was a Trainee for Car-
2010, respectively. diac Surgery at the Heart Center Leipzig, Germany.
Since 2009, he has been with the UniversityHos- Since 2009, he has been a trainee at the University-
pital Zurich, Switzerland and the Swiss Federal In- Hospital Zurich (USZ), Switzerland. He is part of an
stitute of Technology (ETH) Zurich, Switzerland as a interdisciplinary project group between USZ and the
Senior Scientist. His research interest includes com- Swiss Federal Institute of Technology (ETH) Zurich,
puter assisted cardiovascular surgery. Switzerland, focussing on computer asssited cardiac
surgery.

Raoul Hopf received the M.Sc. degree in mechani-

cal engineering, focussing on continuum mechanics Sven Hirsch received the Graduate degree in physics
and dynamics of multibody and non-smooth systems, from the Ruprecht-Karls University, Heidelberg, Ger-
from the Swiss Federal Institute of Technology (ETH) many and the Ph.D. degree in laser physics from
Zurich, Switzerland, in 2011, and is currently work- the University of Dusseldorf, Germany, in 1998 and
ing toward the Ph.D. degree at the Center of Mechan- 2007, respectively.
ics, ETH Zurich. He is currently with the Biomedical Simulation
He is working on a joint project with the Group, the Computer Vision Laboratory, Swiss Fed-
UniversityHospital Zurich, Switzerland, on mod- eral Institute of Technology (ETH) Zurich, Switzer-
eling stent mechanics. His research interest in- land, His research interests include physiological
cludes the mechanical behavior of biological modeling of the cardiovascular system as well as
membranes. image-based simulation of tumor growth.

Edoardo Mazza received the Dr. sc. techn. degree

Thomas Pollok received the B.Sc. degree in com- from the Swiss Federal Institute of Technology (ETH)
puter science from the University of Applied Sci- Zurich, Switzerland, in 1997.
ences, Karlsruhe, Germany in 2012 and is currently He has been a Full Professor of Mechanics at the
working toward the Master’s degree in computer sci- Institute of Mechanical Systems, ETH Zurich since
ence at the Berlin Institute of Technology, Berlin, January 1, 2010. After the Ph.D. degree, he has been
Germany. working in industry for four years. In 2001, he was an
During his studies, he intensively collaborated Assistant Professor, in 2006 an Associate Professor
with the Fraunhofer IOSB in the field of image-based of Mechanics at ETH Zurich. Since 2006, he has
real-time systems. been leading the laboratory ”Mechanics for modeling
Mr. Pollok received a scholarship in 2011 by the and simulation”, the Swiss Federal Laboratories for
State of Baden-Württemberg to support his Bachelor Material Science and Technology (EMPA), Dubendorf, Switzerland.
Thesis at the Computer Vision Laboratory, Swiss Federal Institute of Technol-
ogy (ETH) Zurich, Switzerland.

Gábor Székely (M’94) received the Graduate de-

gree in chemical engineering, the Graduate degree in
applied mathematics, and the Ph.D. degree in analyt-
Christoph Russ (S’13) received the Graduate de-
ical chemistry from the Technical University of Bu-
gree in computational visualistics from the Otto-von-
dapest, Budapest, Hungary, in 1974, 1981, and 1985,
Guericke University, Magdeburg, Germany, in 2010,
respectively.
and is working toward the Ph.D. degree in the field of
He is currently a Full Professor of Medical Image
biomedical simulation for computer-assisted surgery
Analysis and Visualization at the Computer Vision
planning from the Swiss Federal Institute of Tech-
Laboratory, Swiss Federal Institute of Technology
nology (ETH) Zurich, Switzerland, in a joint project
(ETH) Zurich, Switzerland, where he is involved in
with the University Hospital Zurich, Switzerland.
the development of image analysis, visualization, and
He worked with the Australian E-Health Research
simulation methods for computer support of medical diagnosis, therapy, train-
Centre between 2007 and 2009, before he joined
ing, and education.
Siemens Corporate Research, Princeton, NJ, USA.

Volkmar Falk received the Graduation degree in

Thomas Frauenfelder received the M.D. degree medicine and the M.D. degree from the University of
from the University of Zurich, Zurich, Switzerland, Bonn in 1992 and 1993, respectively, and the Ph.D.
in 1999. degree from Universität Leipzig, Germany, in 1999.
From 2002 to 2003, he was Research Fellow in In 2006, he was the Codirector of the Heart Cen-
the field of computational fluid dynamics. In 2009, he ter Leipzig, Germany and Member of the Medical
became an Associate Professor of the University of Faculty of Universität Leipzig, Germany. He serves
Zurich. Since 2011, he has been the vice director of the European Association of Cardiothoracic Surgery
the Institute of Diagnostic and Interventional Radiol- as the Chairman of the Techno College. In 2009,
ogy at the University Hospital Zurich. His research he was the Director of the Division of Cardiovascu-
interests include postprocessing in the field of radi- lar Surgery at University Hospital Zurich, Switzer-
ological imaging, especially cardiovascular imaging. land and a Full Professor for Cardiac Surgery at the University of Zurich,
He collaborated with the ETH Zurich in various projects. Switzerland.