Three-dimensional heart motion estimation using endoscopic monocular vision system: From artificial...
-
Upload
independent -
Category
Documents
-
view
3 -
download
0
Transcript of Three-dimensional heart motion estimation using endoscopic monocular vision system: From artificial...
www.elsevier.com/locate/bspc
Biomedical Signal Processing and Control 2 (2007) 199–207
Three-dimensional heart motion estimation using endoscopic
monocular vision system: From artificial landmarks
to texture analysis
Mickael Sauvee a,b, Aurelien Noce b,*, Philippe Poignet b,Jean Triboulet b, Etienne Dombre b
a Sinters, 5 rue Paul Mesple, 31000 Toulouse, Franceb LIRMM, UMR 5506 CNRS UM2, 161 rue Ada, 34292 Montpellier, France
Received 28 February 2007; received in revised form 2 July 2007; accepted 9 July 2007
Available online 27 August 2007
Abstract
In robot-assisted beating heart surgery, motion of the heart surface might be virtually stabilized to let the surgeon work as in on-pump cardiac
surgery. Virtual stabilization means to compensate physically the relative motion between the instrument tool tip and the region of interest on the
heart surface, and to offer surgeon a stable visual display of the scene. To this end, motion of the heart must be estimated. This article focusses on
motion estimation of the heart surface. Two approaches are considered in the paper. The first one is based on landmark tracking allowing 3D pose
estimation. The second is based on texture tracking. Classical computer vision methods, as well as a new texture-based tracking scheme has been
applied to track the heart motion, and, when possible, reconstruct 3D distance to the heart surface. Experimental results obtained on in vivo images
show the estimated motion of heart surface points.
# 2007 Elsevier Ltd. All rights reserved.
Keywords: Medical applications; Motion estimation; Vision; Robot control; Spectral analysis
1. Introduction
One of the most widely spread intervention in cardiac
surgery is coronary artery bypass grafting (CABG). Currently
most of them are performed using heart–lung machine
and stopped heart, which allows the surgeon to achieve
complex and fine sutures on motionless heart surface. However
cardiopulmonary bypass (CPB) have deleterious effects. For
instance, systemic inflammatory response have been observed
[16]. To avoid the CPB problems, solutions for operating on the
beating heart should be proposed. Among proposed solutions,
passive mechanical stabilizers are used on the heart surface to
cancel the motion (e.g. OctopusTM by Medtronic [8]). The idea is
to apply a mechanical constraint on the heart surface to stabilize
the working area. To ensure contact between heart surface and
mechanical device, vacuum suction or suture technique have
* Corresponding author. Tel.: +33 4 67 14 85 64; fax: +33 4 67 14 85 00.
E-mail address: [email protected] (A. Noce).
1746-8094/$ – see front matter # 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.bspc.2007.07.006
been developed. However remaining motion inside the stabilized
area still exists. Lemma et al. [9] have observed excursion up to
2.4 mm on the stabilized area. So surgeon has to manually cancel
the movement of the target coronary. Another way to perform
beating heart surgery is then to compensate for the organ motion
with a robotized assistive device.
Telerobotic systems (e.g. Da VinciTM system by Intuitive
Surgical [6]) should contribute to minimize the invasiveness of
surgery. They also offer the possibility to compensate for
physiological motion. Thus, ideally the surgeon can concen-
trate on his task, his desired movement being superimposed to
the trajectory generated by the motion compensation algorithm.
Hence the surgeon can work on a virtually stabilized heart as in
on-pump surgery. In this robotized context, the key issues are
listed below:
(1) H
eart motion estimationMovement must be extracted from visual feedback
provided by an endoscope of a highly deformable and non
structured surface, which requires the development of
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207200
specific algorithms. It is intended that the estimated motion
will be used in a bio-mechanical model. This model will
improve the robustness of the robot control architecture; it
will also make it possible to model the displacement of the
tracked area.
(2) C
ontrol system designRobot manipulator must compensate for high-bandwidth
motions (up to 5 Hz), with high precision (suture tasks are
performed on 2-mm diameter vessels). Moreover, stability
must be guaranteed for motion in both free space and
constrained space when the robot interacts with the
environment.
(3) V
isual stabilizationA stable view of the suturing area must be displayed on
the surgeon screen.
Fig. 1. Projection model: frame description.
A first solution was proposed by Nakamura et al. [12] making
use of a high speed camera which tracked artificial landmarks
placed on the heart surface. Visual feedback control was used to
control a lightweight mini-robot. In vivo experiments performed
on porcine model showed good 2D trajectory tracking, but errors
of about 1 mm have been observed. More recently, motion
estimation based on natural landmarks have been presented by
Ortmaier et al. [15] who have developed a prediction algorithm
based on ECG signals and respiratory pressure signal to improve
robustness of landmark detection. Nevertheless, experimental
evaluation was restricted to tracking of landmarks inside a
mechanically stabilized area and results were expressed in image
coordinate. In ref. [5], active landmarks placed on the heart
surface and a laser spot are observed by a 500 Hz video sensor,
which allows to compute the distance between the instrument
tool tip and the reference surface. This information is processed
in an adaptive model predictive controller. In vivo evaluation on
porcine model exhibits motion cancellation with a residual
tracking error up to 1.5 mm. However, due to the large distance
between the markers, this estimation does not reflect accurately
enough the local behavior of the Region of Interest (ROI).
Recently, Cavusoglu et al. [3] proposed to combine biological
signals (ECG signal, arterial and ventricular blood pressures) and
heart motion measurement in a model-based predictive control
algorithm to add feedforward path to robot motion control. Heart
motion measurements are obtained by a Sonomicrometric
system (manufactured by Sonometric). This technique is based
on ultrasound signals, transmitted by small piezoelectric crystals
fixed to the heart surface. To perform this experiment, the
pericardial sac has to be filled with a saline solution, which is not
possible during CABG.
In this paper, we address the issue of heart motion estimation
in terms of displacement and acceleration. We use available
information, i.e. endoscopic image, to perform heart motion
estimation in three dimensions, using at first calibrated
landmarks placed on the heart surface, then a texture based
approach to avoid the use of such landmarks. In Section 2 we
present a method based on artificial landmarks. The 3D
reconstruction procedure is based on metric information
obtained from landmarks with calibrated dimension. Section
3 introduces our texture-based approach to track motion of the
heart without landmark. Experimental evaluation of these
methods on artificial and real image sequences are detailed in
Section 4. A conclusion is drawn in Section 5, exploring
possible developments of the methods.
2. Pose estimation using landmarks
Our approach for extracting visual information and
estimating motion of the ROI consists in three steps:
(1) I
mage acquisition using calibrated endoscopic visionsystem.
(2) T
racking of a geometrically known landmark, small enoughto be assumed planar (5 mm � 5 mm).
(3) P
ose estimation of the template using metric of the object.2.1. Endoscopic vision model
In classical vision systems, a pinhole model, based on thin
lens hypothesis, is applied to describe image projection.
Although endoscope is a long rigid tube composed of a
succession of lenses that provide light from extremity of the
tube to camera lens, the pinhole model is assumed in the
literature to reasonably model endoscopic image [2,11].
Nevertheless, the pinhole model is a linear approximation of
the real camera projection. Therefore, with endoscopic system
that induces high radial distortion, it is necessary to improve
accuracy of the model by adding nonlinear compensation [18].
The pinhole model is split into two parts. The first part (see
Fig. 1) maps the coordinates of point M defined in world frame
Rw (attached to the plane P corresponding to the landmark) to
the coordinates of point m on plane F at a distance Z ¼ 1 from
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207 201
the camera projection center Oc (center of camera frame Rc)
(Eq. (1)). The coordinates ðxn; ynÞT
of point m are the
normalized coordinates, obtained from perspective projection
without considering camera intrinsic parameters. The second
part of the model takes into account the intrinsic parameters of
the vision system by applying frame transformation from
camera frame to image frame (Eq. (2)).
sxn
yn
1
0@
1A ¼ PcTw
wXwYwZ1
0BB@
1CCA (1)
uv1
0@
1A ¼ K
xn
yn
1
0@
1A (2)
The pinhole model is composed of:
� s
up
m
, the scale factor induced by the perspective projection;
� K
, the (3� 3) intrinsic parameter matrix; composed of theoptical center coordinates ðu0; v0Þ and the focal length in X
and Y directions ð f c1; f c2Þ,
� P , a (3� 4) projection matrix,� c
Tw, the rigid transformation matrix defining the world frame(attached to the object plane) w.r.t. the camera frame.
Distortion model adds extra displacements on the normal-
ized coordinates of point m. The model is composed of two
components. The first one takes into account radial distortion
(Eq. (3)), and the second one approximates tangential distortion
(Eq. (4)):
xd
yd
� �¼ ð1þ k1r2 þ k2r4Þ xn
yn
� �þ dx (3)
dx ¼ 2k3 xnyn þ k4ðr2 þ 2x2nÞ
k3ðr2 þ 2y2nÞ þ 2k4xnyn
� �(4)
The intrinsic parameters and coefficients of distortion have
been computed using Matlab Calibration Toolbox1 (based on
algorithm proposed by Zhang [19]). Since the model is an
approximation of reality, the calibration procedure must con-
sider several points of the ROI, inside the workspace.
2.2. Pattern tracking algorithm
For each image acquired, we need to track the pattern
resulting from the perspective projection of the landmark onto
the image plane. We choose a pattern-based algorithm that
tracks the whole image of the square object. We used Efficient
Second order Method (ESM) proposed by Benhimane and
Malis [1]. The application of the ESM algorithm2 to visual
1 Available on: http://www.vision.caltech.edu/bouguetj/calib_doc/ (last
dated: 2005).2 Available on author’s website: http://www-sop.inria.fr/icare/personnel/
alis/software/ESM.html.
tracking allows an efficient homography estimation admitting
large inter-frame displacements. In this algorithm, an
iteratively estimation procedure finds the optimal homography
which minimizes the Sum of Square Differences (SSD)
between the reference pattern (defined off line) and the current
pattern (which has been reprojected in the reference frame
using the current homography). Because initial prediction of
the homography is not available, we start with an initial
estimation equal to the identity matrix. The image derivatives
of both template and the current pattern are used to obtain an
efficient second-order update. It is an efficient algorithm since
only the first derivatives are used and the Hessians are not
explicitly computed.
Once we have computed the homography matrix GðkÞ, at
time k, between the reference and the current patterns, we
extract the image coordinates of the four-corner points of the
square pattern ðuiðkÞ; viðkÞÞT using image coordinates in the
reference pattern ðuið0Þ; við0ÞÞT through Eq. (5). The coordi-
nates of the four corners in the reference pattern are manually
selected during the off line procedure and automatically refined
by detecting nearest corner location at subpixel level.
l
uiðkÞviðkÞ
1
0@
1A ¼ GðkÞ
uið0Þvið0Þ
1
0@
1A 8 i ¼ 1; . . . ; 4 (5)
At the end of this step, we have computed information in the
image space. We now have to integrate metric information
available on the pattern to estimate 3D pose.
2.3. Pose estimation
The method is inspired from Zhang[19] and assumes that the
intrinsic parameters are known. It is implemented with
OpenCV Library.3 Assuming that the world frame is attached
to the object plane, the 4 points satisfy wZ ¼ 0; thus, without
loss of generality, we can rewrite Eq. (1) as:
xn
yn
1
0@
1A ¼ s�1½ r1 r2 t �
wXwY1
0@
1A ¼ HðkÞ
wXwY1
0@
1A (6)
where r1 and r2 are the first two columns of the rotation matrix
and t is the translation vector from the rigid transformationcTwðkÞ at time k.
Image plane coordinates expressed in the world frame
coordinates are used to compute the estimation HðkÞ of the
projection matrix between ðxn; yn; 1ÞT and ðwX; wY ; 1ÞT (see
Appendix A of ref. [19]). Scale factor s is retrieved by
computing the mean of norms of the first two columns of matrix
HðkÞ (s ¼ 0:5ðjjh1jj þ jjh2jjÞ�1). Then, Eq. (6) yields:
r1 ¼ sh1; r2 ¼ sh2; t ¼ sh3 (7)
3 Available on: http://www.intel.com/technology/computing/opencv/
index.htm (last updated: 2005).
Table 1
Chosen texture features
Approach Feature
Coocurrence matrix Energy
Coocurrence matrix Contrast
Coocurrence matrix Cluster shade
Coocurrence matrix Cluster prominence
Run-length matrix Non-uniformity
Run-length matrix Short low grey
Level run emphasis
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207202
The third vector of the rotation matrix is obtained form the
orthogonality property of this matrix:
r3 ¼ r1 � r2 (8)
3. Motion tracking without landmarks
The previous approach and its practical set-up raised several
issues:
First order statistics Skewness
First order statistics Kurtosis
� P lacing landmarks is practically difficult and time-consum-ing;
� la
ndmarks hide portions of the heart surface;� b
io-compatibility problems may arise.To overcome these limitations, we developed a new
method that avoids the use of landmarks [13,14]. Such an
approach is promising and recent works focus on applying it
to visual servoing and robotics [4]. Our method relies on
tracking small regions of the heart based on texture
information instead of artificial landmarks, which should
provide a robust tracking.
3.1. Texture characterization
We characterize texture of the tracked pattern using a set
of statistical descriptors known as texture features. These
features can describe the pattern in terms of contrast,
luminosity (gray-levels), granularity, etc. Each attribute is
normalized, so the values are double precision numbers
between 0 and 1. For example, a luminosity value of 0
corresponds to a black image while a value of 1 correponds to
a white image. Naturally, the settings of the endoscope can
influence the quality of the texture-based description. But the
features has been selected in a way (see ref. [14]) that good
visual feedback corresponds to good texture characterization.
Furthermore, the use of several features in the texture vector
limits the risk of an inefficient feature bringin down overall
performance.
The texture features Ti; i2 1; . . . ; n from the texture vector Vwhich is characteristic of the ROI:
V ¼
T1
T2
..
.
Tn
0BBB@
1CCCA (9)
This characterization is, as experiments tend to prove, more
robust to changes in illumination and deformations than
classical similarity computation algorithm using correlation or
SSD. The 8 most efficient features (in Eq. (9) we use n ¼ 8) that
are recalled in Table 1 have been selected through Principal
Components Analysis and Discriminant Function Analysis
[14]. Using such a reduced set of features has several
advantages: faster computation, more efficient characteriza-
tion, and easier interpretation of the results.
3.2. Texture-based tracking
To integrate texture characterization in the pattern tracking
procedure, we introduce a new composite metric based on both
SSD and texture features. The resulting Composite Tracking
Algorithm, represented on Fig. 2, can be divided into three
steps: (1) compute normalized SSD and texture-based distance,
(2) merge those results in the Composite Metric and (3) run the
minimization procedure. Each of these steps is described
below:
� T
he Texture Distance (TD) is the distance between the texturevector (Eq. (9)) of the tracked pattern VI and the one of the
ROI V 0I . TD tðI; I0Þ is computed using Euclidean metric
(Eq. (10)), Ti and T 0i corresponding to texture feature
components in V p and Vr:
tðI; I0Þ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiX8
i¼1
ðTi � T 0i Þ2
vuut (10)
The distance is then normalized:
tNðI; I0Þ ¼tðI; I0Þ
max 8ROIðtÞ(11)
� O
nce TD tðI; I0Þ and SSD dðI; I0Þ have been computed, theComposite Metric, Eq. (12), is computed and will be used as
the similarity measurement between tracked pattern and ROI.
sðI; I0Þ ¼ ldðI; I0Þ þ gtNðI; I0Þlþ g ¼ 1
(12)
The parameters l and g give the balance (in percent) between
texture characterization and correlation-based pattern match-
ing. Those parameters can be set to fixed values, but better
performance is obtained through the use of dynamic values.
Several possible solutions for those parameters have been
evaluated on experimental seqences. From this evaluations,
we found good solutions for tuning l and g, as given in
Eqs. (13) and (14):
g ¼ dðI; I0Þ (13)
g ¼ e�dðI;I0Þ=ð1�dðI;I0ÞÞ
l ¼ 1� g(14)
Fig. 2. Outline of the texture based tracking procedure.
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207 203
� F
Fig. 3. Endoscopic image of the heart with landmarks.
inally, we use a minimization procedure to retrieve the
homography matrix images. This part is similar to the one
exposed in Sections 2.2 and [13].
4. Experiments
4.1. Precision evaluation
A first step of the experiments concerns the 3D pose
estimation from landmarks. But before applying this approach
to estimate heart motion, accuracy of the method has been
evaluated using calibrated measurement system to take into
account different sources of error such as camera model
approximation, pattern tracking accuracy, 3D reconstruction
algorithm.
The endoscopic vision system is composed of a rigid
endoscope (Hopkin’s II from Karl Storz Inc.), mounted on a
35 mm focal lens optic and a CMOS camera Dalsa 1 m75. A
300 W xenon light source is connected to the rigid endoscope.
The frame rate is adjusted at 125 Hz, and a 512� 512-pixel
image size is selected.
Considering that precision is better along Xc and Yc than
along Zc, which is normal to the image plane, we evaluate
precision along Zc using a laser measurement device
(precision ¼ 0:5 mm) in a dynamic way. An object was
manually moved in plane parallel to the image plane. A
maximum error of 0.8 mm was observed w.r.t. laser measure-
ments (standard deviation ¼ 0:34 mm). Thus, we consider that
the complete system (cameraþ pattern trackingþ pose esti-
mation) gives an estimation of motion with a precision up to
1 mm.
4.2. In vivo experimentation
The proposed approach without landmarks has been applied
on an anesthetized pig of 25 kg with the assistance of a trained
medical team, in accordance with ethical and regulatory issues
related to animal experiments. The pig was assisted by a
pulmonary ventilator with constrained respiratory cycle to 20
cycles per minute. A thoracotomy has been performed to
facilitate heart access. In the open chest, the heart is not
anymore constrained by the ribs. Then, the cardiac cycle
induces non natural motion of the heart within the thorax that
should be much less important in MIS conditions. Images have
been acquired using endoscopic vision system presented
in Section 4.1. The endoscopic vision system is calibrated
for a workspace of 20 mm � 15 mm � 15 mm, centered at a
distance of 55 mm from the endoscope tip. Intrinsic parameters
are reported in Table 2.
Artificial passive landmarks have simply been laid on the
heart surface as shown in Fig. 3. Each landmark is assumed to
be a planar square of 5 mm side, drawn on white paper. To
avoid lack of information for the tracking algorithm, variable
gray levels have been painted on the pattern. Sequences of 1900
images have been acquired with a sampling time of 8 ms.
4.3. Motion analysis
Estimation of heart motion is presented on Fig. 4, over the
entire acquired sequence of 15 s. A spectral analysis has been
performed to evaluate frequency components of the estimated
signal (see Table 3). The first two components ( f 1 ¼ 0:34 Hz,
and f 2 ¼ 0:68 Hz) must be related to the respiratory activity.
f 1 is equal to the frequency imposed by the pulmonary
ventilator (20 cycles per minute). f 2 may be considered as an
Fig. 4. Motion estimation of patch 1: (a) X direction, (b) Y direction, (c) Z
direction.
Table 2
Intrinsic parameters (in pixel)
f c1 f c2 u0 v0
Pinhole model
593.6 594.5 279.6 223.7
k1 k2 k3 k4
Distortion model
�0.241 0.207 0.005 0.008
Table 3
Spectral analysis
Frequency (Hz)
0.34 0.67 1.19 2.38 3.57 4.76
Density in Xc direction (%) 8.7 0 21.8 60 2.8 6.7
Density in Yc direction (%) 3.9 1.9 45.6 33 3.9 11.7
Density in Zc direction (%) 70.9 11.5 5.9 8.7 1.2 1.8
Table 4
Maximum acceleration of the estimated motion
x (m s�2) y (m s�2) z (m s�2)
Landmark 1 1.73 1.09 3.33
Landmark 2 1.38 0.94 3.67
Table 5
Error in translation tracking (sequence of 100 images at 100 Hz)
Method Mean error (Pixels) Standard deviation
Correlation 0 0
SSD 0.04 0.19
Optical flow 0.90 0.73
Composite 0 0
ESM 0 0
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207204
harmonic component of the respiratory activity ( f 2 ¼ 2 f 1).
Heart activity provides 4 other frequencies. f 3 ¼ 1:19 Hz
represents the heart beat cycle (around 70 beats per minute),
and the three others are harmonic components of f 3. The
spectral analysis shows that the estimated motion are
dominated by cardiac and respiratory activities. From power
spectral density analysis, we can see that motion in Xc and Yc
directions (parallel to image plane) are mostly governed by
cardiac activity, whereas amplitude of the motion along Zc
direction is greatly induced by respiratory activity. By applying
a low pass filter of 1 Hz cutoff frequency, we can extract the
respiratory component from the estimated motion along Zc
direction (see bold plot in Fig. 4c). Thus maximum amplitude
observed in Zc direction is about 11 mm and is composed of
respiratory displacement of 6 mm.
Accelerations have been computed as the second derivatives
of the position, after backward and forward lowpass filtering.
We obtain accelerations up to 3.67 m s�2 as reported in
Table 4.
4.4. Offline evaluation of motion tracking without
landmarks
To assess the precision and robustness of the texture-based
tracking and compare the result with other approaches, we used
both artificial and natural test sequences.
Fig. 5. Sample artificial sequences: translation.
Fig. 6. Sample trajectory for vertical translation. Motion tracking error is projection on several axes, underlining the inadequacy of classical approaches, such as
optical flow, with the considered application.
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207 205
Fig. 7. Sample trajectory for circular translation and rotation of the pattern. This shows superior robustness of the Composite Approach over other classical
algorithms.
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207206
� A
rtificial sequences were generated by moving the trackedpattern along a heart image, as illustrated on Fig. 5, in which a
small pattern (the target) is moved among the heart sequence
to challenge the tracking algotithms. Each image of the
sequence is obtained by superimposing on the original
experimental sequence a small textured patch and apply
displacements and transformations (such as translation,
rotation) on it. These sequences enabled us to quantify the
performance of our approach in terms of tracking precision,
and to compare it with other tracking schemes, namely SSD,
correlation [7], optical flow [17] and ESM [10]. Figs. 6 and 7
show sample trajectories and results (our method is termed as
‘‘Composite’’). On Fig. 6 we observe repetitive tracking
errors with the optical flow that can be related to the presence
of specularities. Fig. 6(a and b) shows that with the SSD and
optical flow approaches, the tracking motion is disturbed by
small oscillations. These oscillations come from the fact that
the algorithms loose temporally the target (due for instance to
specularities). Fig. 7 shows the robustness of our approach
Fig. 8. Displacements in pixels during motion.
during a composite circular trajectory involving translation
and rotation of the pattern: a scale factor up to �10% have
been applied to the pattern along the trajectory. The SSD and
the optical flow algorithms are unable to track the trajectory
while the correlation algorithm behaves well at the beginning
of the trajectory, then diverges. Two causes are possible:
specularity, or scale factor and rotation changes too high.
The precision of tracking has been computed in the image
for each sequence in order to quantify the performance of the
approaches. Table 5 gives sample results obtained on the
sequence corresponding to Fig. 6 in terms of mean and
standard deviation of the tracking error. As mentioned earlier,
correlation, ESM and Composite Tracking (our approach)
score better than SSD and optical flow. Other sequences have
been generated, that confirm the improvements in tracking
robustness using textures.
� Q
ualitative evaluation was also performed using registeredbeating heart sequence to tune the algorithm and to perform
application-focused comparison with other approaches.
Overall performance is very encouraging, and tracking is
visually correct, even if very homogeneous regions of the
heart are still difficult to track whichever approach is used and
even if we did not succeed in obtaining precise 3D metric of
the displacement of the heart. Fig. 8 shows the vertical and
horizontal motions obtained without landmark, which can be
related to the curves of Fig. 4(with landmarks), proving the
relevance of the approach.
5. Conclusion and perspectives
We have presented in this paper estimation of heart motion
using available vision system in operating room, using both
landmark-based and texture-based approaches. The latter
showed that evolution towards full vision-based motion
compensation is realistic and promising.
The approach with landmarks exhibits results with precision
better than 1 mm. In vivo experiments have been done on porcine
model. Estimated heart motion is clearly governed by cardiac and
M. Sauvee et al. / Biomedical Signal Processing and Control 2 (2007) 199–207 207
heart cycle activities. Depending on the location on heart surface,
accelerations up to 3.67 m s�2 have been observed.
Several improvements can be expected. Tests of a new
approach based on texture tracking on experimental sequences
revealed that the approach is quite robust in terms of tracking
but, due to the high deformation of the heart, it is still subject
to tracking losses in some extreme cases. Further work is
necessary to make up for the lack of metric in image and
achieve performance required by medical context.
Acknowledgement
We gratefully thank Roland Demaria4 for his help in
experimental applications.
References
[1] S. Benhimane, E. Malis, Real-time image-based tracking of planes using
efficient second-order minimization, in: Proceeding of IEEE IROS,
Sendai, Japan, (2004), pp. 943–948.
[2] J.J. Caban, W.B. Seales, Reconstruction and enhancement in monocular
laparoscopic imagery, in: Proceedings of Medicine Meets Virtual Reality,
vol. 12, 2004, pp. 37–39.
[3] M.C. Cavusoglu, J. Rotella, W.S. Newman, S. Choi, J. Ustin, S.S. Sastry,
Control algorithms for active relative motion cancelling for robotic
assisted off-pump coronary artery bypass graft surgery, in: Proceedings
of IEEE ICAR, Seattle, WA, USA, 2005.
[4] A.I. Comport, E. Marchand, M. Pressigout, F. Chaumette, Real-time
markerless tracking for augmented reality: the virtual visual servoing
framework, IEEE Trans. Visual. Comput. Graphics 12 (2006) 289–298.
[5] R. Ginhoux, J. Gangloff, M. de Mathelin, L. Soler, M.M. Arena Sanchez,
J. Marescaux, Active filtering of physiological motion in robotized surgery
using predictive control, IEEE Trans. Robotics 21 (1) (2005) 67–79.
[6] G.S. Guthart, J.K. Salisbury, The intuitive telesurgery system: overview
and application, in: Proceedings of IEEE ICRA, 2000, pp. 618–621.
4 Roland Demaria is a surgeon at the ‘‘Service de chirurgie Cardiovasculaire,
CHU Arnaud de Villeneuve, 371 av. G. Giraud, 34295 Montpellier, France’’.
[7] S. Hutchinson, G.D. Hager, P. Corke, A tutorial on visual servo control,
IEEE Trans. Robotics Automat. 12 (5) (1996) 651–670.
[8] E.W.L. Jansen, P.F. Grundeman, C. Borst, F. Eefting, J. Diephuis, A.
Nierich, J.R. Lahpor, J.J. Bredee, Less invasive off-pump CABG using a
suction device for immobilization: the octopus method, Eur. J. Cardio-
Thoracic Surg. 12 (1997) 406–412.
[9] M. Lemma, A. Mangini, A. Reaelli, F. Acocella, Do cardiac stabilizers
really stabilize? experimental quantitative analysis of mechanical stabi-
lization, Interact. Cardiovasc. Thoracic Surg. 4 (2005) 222–226.
[10] E. Malis, Improving vision-based control using efficient second-order
minimization techniques, in: Proceedings of IEEE ICRA, New Orleans,
USA, 2004.
[11] G. Marti, V. Bettschart, J.-S. Billiard, C. Baur, Hybrid method for both
calibration and registration of an endoscope with an active optical marker,
in: Proceedings of Computer Assisted Radiology and Surgery, 2004, pp.
159–164.
[12] Y. Nakamura, K. Kishi, H. Kawakami, Heartbeat synchronization for
robotic cardiac surgery, in: Proceedings of IEEE ICRA, Seoul, Korea,
(2001), pp. 2014–2019.
[13] A. Noce, J. Triboulet, P. Poignet, Composite visual tracking of the
moving heart using texture characterization, in: Proceedings of IEEE
MICCAI Work. Medical Robotics, Copenhagen, Denmark, (2006), pp.
54–65.
[14] A. Noce, J. Triboulet, P. Poignet, E. Dombre, Texture features selection for
visual servoing of the beating heart, in: Proceedings of IEEE/RAS-EMBS
BIOROB’06, number 183, 2006.
[15] T. Ortmaier, M. Groger, D.H. Boehm, V. Falk, G. Hirzinger, Motion
estimation in beating heart surgery, IEEE Trans. Biomed. Eng. 52 (10)
(2005) 1729–1740.
[16] A.L. Picone, C.J. Lutz, C. Finck, D. Carney, L.A. Gatto, A. Paskanik, B.
Searles, K. Snyder, G. Nieman, Multiple sequential insults cause post-
pump syndrome, Ann. Thoracic Surg. 67 (1999) 978–985.
[17] T. Suzuki, T. Kanade, Measurement of vehicle motion and orientation
using optical flow, in: Proceedings of IEEE ITSC, 1999, pp. 25–
30.
[18] X. Zhang, S. Payandeh, Application of visual tracking for robotic-assisted
laparoscopic surgery, J. Robotics Syst. 19 (7) (2002) 315–328.
[19] Z. Zhang, A flexible new technique for camera calibration. Technical
Report MSR-TR-98-71. Microsoft Research, 1998.