Magneto-Optical Tracking of Flexible Laparoscopic Ultrasound: Model-Based Online Detection and Correction of Magnetic Tracking Errors

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 28, NO.
6, JUNE 2009 951
Magneto-Optical Tracking of Flexible Laparoscopic

Ultrasound: Model-Based Online Detection and
Correction of Magnetic Tracking Errors
Marco Feuerstein*, Tobias Reichl, Jakob Vogel, Joerg Traub, and Nassir Navab
Abstract—Electromagnetic tracking is currently one of the most the constantly changing imaging plane, and a lack of
promising means of localizing flexible endoscopic instruments
such as flexible laparoscopic ultrasound transducers. However,
awareness of the transducer tip location (the tip needs to be
electromagnetic tracking is also susceptible to interference from constantly observed in the laparoscopic camera images in
ferromagnetic material, which distorts the magnetic field and order to avoid inadvertent injury) [1]–[3].
leads to tracking errors. This paper presents new methods for Several groups have tried to address some of these issues by
real-time online detection and reduction of dynamic providing navigated laparoscopic ultrasound: the position and
electromagnetic tracking errors when localizing a flexible
orientation (“pose”) of the ultrasound transducer is estimated
laparoscopic ultrasound transducer. We use a hybrid tracking
setup to combine optical tracking of the transducer shaft and so that its shape and B-scan images can be visualized in rela-
electromagnetic tracking of the flexible transducer tip. A novel tion to the patient, other surgical instruments, or preoperative
approach of modeling the poses of the transducer tip in relation and intraoperative imaging data. This can provide great
to the transducer shaft allows us to reliably detect and support to surgeons using laparoscopic ultrasound in cancer
significantly reduce electromagnetic tracking errors. For
staging, radio frequency ablation, and other diagnostic and
detecting errors of more than 5 mm, we achieved a sensitivity
and specificity of 91% and 93%, respectively. Initial 3-D rms therapeutic procedures.
error of 6.91 mm were reduced to 3.15 mm. To estimate the pose of a transducer with a rigid tip, a robot
Index Terms—Electromagnetic tracking, hybrid tracking,
or an optical tracker (OT) may be used [4]. In the latter case, a
image-guided surgery, laparoscopic surgery, optical tracking. rigid optical marker can be attached to the transducer handle to
assure its continuous visibility. Several groups have also tried
to localize rigid laparoscopic instruments in laparoscopic
I. INTRODUCTION images by using advanced image processing techniques [5]–
[9]. However, the laparoscopic transducers most commonly
U LTRASONOGRAPHY is an appealing technology to used and preferred by surgeons have a flexible tip that can be
physicians because of its noninvasiveness, wide avail- steered left, right, up or down. The tip can also be bent by
ability, flexible handling, and low cost. Having been used external pressure from organ surfaces. Due to the missing line
primarily for diagnosis in the past, intraoperative and laparo- of sight to the flexible transducer tip, an OT cannot be used
scopic ultrasonography today play a greater role in abdominal exclusively to localize this tip. A robot could only be used if
surgery. The liver, biliary tract, and pancreas are the main the flexible ultrasound probe is fully integrated into the end-
application areas of intraoperative and laparoscopic effector. To the authors’ knowledge no such system exists.
ultrasound, for instance to detect liver lesions such as Promising alternatives are the use of an electromagnetic
metastases. Unfortu- nately, the success of laparoscopic tracker (EMT) localizing a sensor attached to the tip [10]–[12]
ultrasonography is operator dependent; for novice surgeons in or fully incorporated into the tip [13], or a magneto-optical
particular it is often difficult or even impossible to detect tracker, i.e., the combination of an OT and an EMT [14]–[16].
target objects such as common bile duct stones When using an EMT clinically, a significant problem is the
intraoperatively [1]. The major reasons given for this are the distortion of the magnetic field, which leads to tracking errors.
missing tactile feedback, the difficulty of in- terpreting This distortion can be caused by metallic or electrically
laparoscopic ultrasound images, a limited degree of powered objects inside or close to the working volume, for
positioning through the trocar access, disorientation caused by instance surgical instruments, an operating table, or imaging
devices such as a C-arm or a computed tomography scanner.
Depending on the operating room setup and instrumentation,
Manuscript received June 09, 2008; revised October 13, 2008. First
published February 10, 2009; current version published May 28, 2009. tracking errors of several millimeters or even centimeters can
Asterisk indicates corresponding author. occur [17], [18]. To compensate for erroneous measurements
*M. Feuerstein is with the Department of Media Science, Graduate School caused by stationary objects, various calibration techniques
of Information Science, Nagoya University, Nagoya 464-8603, Japan (e-mail:
[email protected]). have been proposed [19]. They usually require the user to
T. Reichl, J. Vogel, J. Traub, and N. Navab are with the Chair for Computer acquire a set of well distributed (distorted) measurements
Aided Medical Procedures and Augmented Reality, Technische Univer- within the tracking volume and corresponding (undistorted)
sität München, 85748 Garching, Germany (e-mail: [email protected];
[email protected]; [email protected]; [email protected]). reference measurements to compute a field distortion function
Color versions of one of more of the figures in this paper are available that is based on lookup tables or poly- nomials. Unfortunately,
online at http://ieeexplore.ieee.org. this function can only compensate for
Digital Object Identifier 10.1109/TMI.2008.2008954
0278-0062/$25.00 © 2009 IEEE
Authorized licensed use limited to: T U MUENCHEN. Downloaded on May 27, 2009 at 19:48 from IEEE Xplore. Restrictions apply.
952 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 28, NO. 6, JUNE 2009
stationary distortion field errors, so that the calibration process

has to be repeated for every new operating room setup before
an intervention. In the best case, calibration could be
performed intraoperatively only for the region of interest
using a special distortion measurement apparatus that fits to a
trocar, as proposed by the authors of [16] for laparoscopic 3-
D ultrasound.
However, dynamic changes of the field distortion, caused
for example by the intraoperative relocation of the magnetic
field generator or movement of instruments, cannot be
compensated for by the previously computed distortion
functions. A first step towards the intraoperative detection of
erroneous measurements caused by metallic objects distorting
the field was presented by [20] and later by [21]. They
incorporate two magnetic sensors into a pointer so that
redundant measurements can be obtained. Measured
deviations of the fixed distance between the two sensors are
used as a plausibility value. Fig. 1. System setup used for calibration and evaluation. Two magnetic
Our paper introduces two new methods to detect and re- sensors are attached to the laparoscopic ultrasound transducer, one to its
duce dynamic electromagnetic tracking errors online, i.e., flexible tip (a) and one to the rigid shaft (b); see Fig. 2 for a close-up of (a)
and (b). The magnetic field generator representing our electromagnetic
intraoperatively without a precomputed distortion function. tracker (c) is itself tracked by an optical tracking system. The optical tracking
This is achieved by a hybrid magneto-optical tracking setup. system also localizes two optical markers attached to the ultrasound
The laparoscopic ultrasound transducer is tracked electromag- transducer, one (temporarily) to its flexible tip (d) and another to its rigid shaft
(e). In addition, the laparoscope (f) and its telescope (g) are tracked optically.
netically as well as optically, and the EMT is itself tracked
optically. After magneto-optical coregistration, i.e., alignment
of the EMT and OT coordinate frames, we can get redundant
tracking data, where the OT provides undistorted, more View, CA), a laparoscopic camera with a forward-oblique 30
reliable measurements than the EMT. We employ this HOPKINS telescope from KARL STORZ GmbH & Co. KG
redundancy in our first, tracking redundancy-based approach: (Tuttlingen, Germany), a standard workstation PC including
comparing measurements of an optical marker and a magnetic two analog frame grabbers (for capturing the transducer and
sensor, both rigidly attached to the transducer shaft, we are camera video in real time), and a hybrid magneto-optical
able to detect deviations between the two measurements. tracking setup. The OT consists of 4 ARTtrack2 cameras and
These deviations can be used to predict tracking errors of a book-size PC running the DTrack tracking software, both
another magnetic sensor mounted on the flexible transducer provided by Advanced Realtime Tracking GmbH (Weilheim,
tip. In a second, more sophisticated model-based approach, we Germany). The EMT in use is a 3-D Guidance unit from
model all possible tip movements relative to the optically Ascension Technology Corporation (Burlington, VT) equipped
tracked shaft using only two rotation angles and thus reduce with a midrange transmitter for magnetic field generation and
the degrees-of-freedom (DOF) of any tip movement from six 1.3 mm sensors in vinyl, which have a total diameter of 1.7
to two. By discarding the four redundant DOF, each distorted mm including the vinyl tubing. Time synchronization of all
measurement of the magnetic sensor on the transducer tip can input data streams and visualization is performed by our
be mapped onto the model for error correction. medical augmented reality software platform CAMPAR [25].
For our two online error detection and correction We denote a rigid arrangement of at least three noncollinear
approaches, we evaluate not only 3-D errors, but also 2-D and nonsymmetric retro-reflective spheres, which are localized
overlay errors during camera superimposition, which is useful by the OT, as an optical marker. Magnetic sensing coils local-
e.g., for medical augmented reality [22], where the in-plane ized by the EMT are denoted as magnetic sensors. To localize
error is most important. Therefore, we also incorporate and all instruments, several optical markers and magnetic sensors
calibrate a laparoscope camera that is localized by the OT. For are used. In addition to an optical marker, which is attached to
the sake of completeness, we also briefly describe our the transducer shaft (below referred to as ), two mag-
ultrasound calibration techniques to determine the location of netic sensors are attached to the transducer: one to the flex-
the ultrasound plane in OT coordinates, which is important ible tip , and the other to the rigid shaft , at
for the final application [23], [24]. We also compare our a distance of about seven centimeters (see Fig. 2). Another op-
results on error detection and correction to the findings of tical marker is mounted on the EMT . This setup al-
other groups [20], [21]. lows us to co-register the EMT and the OT and to obtain re-
dundant tracking information for the rigid transducer shaft (see
Sections III-A3 and III-B). Finally, two optical markers are at-
II. SYSTEM SETUP
tached to the laparoscopic camera, one to the head
The hardware setup we used for all calibration and evalua- and another to the telescope shaft , to adjust for
tion steps consists of the following components (see Fig. 1): a telescope rotations (see Section III-A1).
flexible laparoscopic linear array transducer (LAP8-4, 5 MHz,
10 mm diameter) connected to a SONOLINE Omnia ultra-
sound system from Siemens Medical Solutions (Mountain
FEUERSTEIN et al.: MAGNETO-OPTICAL TRACKING OF FLEXIBLE LAPAROSCOPIC 953
ULTRASOUND segmented in the B-scan images. Several poses of the
magnetic sensor as well as corresponding B-scan
Fig. 2. Close-up of the proximal part of the ultrasound transducer with a

magnetic sensor attached to the transducer tip (a) and one attached to the
transducer shaft (b).
III. METHODS
In the following section, we describe all required system
calibration steps in order to detect and correct magnetic
tracking errors when the flexible tip of the laparoscopic
ultrasound transducer is in the close vicinity of ferromagnetic
and/or electrically conductive material.
A. Offline System
Calibration
The main purpose of all system calibration steps is to bring
the local coordinate frame associated with each instrument,
optical marker, or magnetic sensor into the same world
coordinate system, chosen here to be the OT coordinate
frame (see Fig. 3). Furthermore, all tracking and video data
streams need to be tagged with time stamps and
synchronized; see our previous work incorporating temporal
calibration [23], [24] for further details. Finally, our new error
estimation methods require an exact modeling of the
transducer axis as well as possible transducer tip movements.
This is described in Sections III-A4 and III-A5.
1) Laparoscope Camera Calibration: During laparoscope
calibration, the camera projection geometry including
distortion coefficients and the transformation from
the camera frame to the coordinate frame of the optical
marker on the laparoscope (via the coordinate frame of the
optical marker on the telescope) are estimated. This can be
done by using standard camera and hand–eye calibration
techniques based on a planar checkerboard pattern and is
described in more detail in various publications, for instance
in our prior work [26]. For hand–eye calibration, see Section
III-A3. It is important to note that the rotational offset, which
is introduced when rotating the oblique telescope shaft around
its longitudinal axis against the camera head, needs to be
corrected for. An elegant solution to this problem was
proposed by [27]. Instead of attaching a ro- tary encoder to the
camera head to determine the current rotation angle of the
telescope, we use the OT to estimate the relative rotation
between the optical markers on the telescope and laparoscope,
similar to [28].
2) Laparoscopic Ultrasound Calibration: For the
determination of the pixel scaling of the ultrasound B-scan
plane and the transformation between the plane and
the coordinate frame of the magnetic sensor on the transducer
tip, a single-wall calibration as proposed by the Cambridge
group is performed [29]–[31]. We acquire ultrasound images
of a nylon membrane stretched over a planar frame and
immersed in a water bath [32], so that the nylon membrane is
imaged as a fairly straight line that can be automatically
ULTRASOUND
Fig. 3. All coordinate frames shown need to be brought into a common

world coordinate system during system calibration. The optical tracker
defines our world coordinate frame (OT). During laparoscope camera
calibration, we determine the relation between the coordinate frames of the
optical marker on the laparoscope , the optical marker on the
telescope , and the camera (C). The origin of C coincides with the
camera center of our pinhole camera model. During ultrasound calibration, we
compute the rigid transformation between the coordinate frame of the
ultrasound B-scan plane (US) and the magnetic sensor on the tip by
means of a temporarily used nylon membrane (N). During magneto-optical
coregistration, we estimate the rigid transformations between the coordinate
frame of the EMT and the optical marker attached to it as well as
between the coordinate frames of the magnetic sensor and the
optical marker on the shaft. The temporary optical marker, which
is attached to the transducer tip and used for evaluation, and its coordinate
frame are not shown here.
images of the membrane inside a water bath are then acquired,

following the calibration protocol of the Cambridge group to
ensure numeric stability for all six DOF of .
3) Magneto-Optical Coregistration: The process of aligning
two tracking coordinate frames is referred to as coregistration
or tracker alignment [33]. Similar to the authors in [34], we
use a hybrid tracking approach by combining magnetic and
optical tracking. The EMT is itself tracked using the OT so
that any data acquired by the EMT can be transformed into the
OT coordinate frame.
During magneto-optical coregistration we determine the
transformation . As
the optical marker attached to EMT is continuously local-
ized by the OT, we obtain in real time and only
need to determine the static transformation
from the coordinates of the EMT to the coordinates of the
optical marker attached to it. Additionally, for the tracking
redundancy-based error detection and correction methods
described in Section III-B, the fixed Euclidean transformation
between the coordinate frames of the magnetic
sensor and the optical marker attached to the shaft needs to be
estimated.
To compute , the transducer is moved to
positions with distinct rotation axes. At each position
, we acquire measurements of the poses of
transducer to be stationary and the EMT and OT coordinate
frames to perform the
Fig. 4. During magneto-optical coregistration, we compute the unknown

transformations and , visualized here by dashed lines. From
at least three transducer positions with distinct rotation axes (two positions
and are shown here), we compute all possible unidirectional motion pairs
between these positions, e.g., between positions and the pair
representing the motion of the optical marker and
representing the motion of the magnetic sensor. From these motion pairs, we
can generate an overdeter- mined system of linear equations to solve for both
desired transformations.
the optical marker and of the magnetic sensor

(both attached to the transducer shaft) in the optical coordinate
frame and in the electromagnetic coordinate frame,
respectively (see Fig. 4 for two exemplary positions and ).
All possible unidirectional motion pairs
be-
tween these positions can now be computed, e.g., between po-
sitions and the pair
representing the motion of the optical marker
and rep-
resenting the motion of the magnetic sensor. For all motion
pairs between positions and , the relation
with holds.
All motions of the optical marker are now stacked to a
matrix and all corresponding motions of the magnetic
sensor
are stacked to a matrix , so both and contain
homogeneous transformation matrices. They are related to
each other by the equation system .
This equation system actually originated from robotics,
where the rigid transformation between a camera (“eye”)
and a robot arm (“hand”) holding the camera was estimated
using “hand–eye calibration” [35], [36]. We use a dual quater-
nions approach [37] to solve for .
The same transducer positions are also used to estimate
the rigid transformation
between the coordinate frame of the optical marker attached
to the EMT and the EMT coordinate system, but this time
generating motion pairs with assuming the ultrasound
Fig. 5. Transducer axis calibration. The transducer axis is calibrated both in

relation to the local coordinate system of the magnetic sensor attached to the
tip (a) and, in a second step not shown here, in relation to the coordinate
system of the magnetic sensor on the shaft (b). To do this, a plastic cylinder
(d) is put over the transducer tip with an additionally attached sensor (c) on the
left end. The plastic cylinder is rotated at least 360 , flipped, and rotated
another 360 or more.
motions. Further details can be found in our previous work

[38], including a derivation of the exact transformations used.
In a final optimization step, the two coregistration ma-
trices and are optimized for
all recorded positions by the Levenberg–Marquardt algorithm
[39]. The 4 4 matrix resulting from the
transformation chain starting at “magnetic sensor to
optical marker to OT to optical marker
to EMT to magnetic sensor ,” which theoretically is
an identity matrix, represents the accumulated transformation
errors
(1)
For the Levenberg–Marquardt optimizer, we chose a cost func-

tion that weights translation to rotation errors 1:3:
(2)
where the translation error is the norm of and the rota-

tion error is the rotation angle of , decomposed into
axis- angle parameters. This 1:3 ratio reflects the root mean
square (rms) error ratio provided independently by the two
tracking system manufacturers: The rms measurement errors of
the OT are stated as 0.4 mm (position) and 0.12 (orientation),
the static rms errors of the EMT as 1.4 mm and 0.5 .
Specifications of the typical accuracy of both tracking systems
are available from AR Tracking1 and Ascension Technology2,
respectively.
4) Transducer Axis Calibration: For modeling the transducer
tip movements (see next Section III-A5), the transducer axis is
first estimated in relation to the magnetic sensors attached to
the shaft and to the tip. This can be done by putting a plastic
cylinder, which contains an additional magnetic sensor on one
end (see Fig. 5), over the transducer shaft. For each of the rigid
part and the flexible part of the transducer shaft, the plastic
1http://www.ar-tracking.de
2http://www.ascension-tech.com
ULTRASOUND
Fig. 6. Typical bending region usually found in endoscopic instruments. Four

Bowden cables are pairwise and alternately connected to every second link of
the bending region. They are often controlled by steering levers, where each
pair of cables is connected to the same lever.
cylinder is rotated at least 360 , flipped, and rotated another

360 or more. During each rotation, the measurements of the
additional magnetic sensor are transformed into the local coor-
dinate frames of the magnetic sensor on the shaft and tip,
respectively. Each acquired point cloud is
then fitted to a cylinder with radius around the transducer
axis, which is defined by the 3-D point on the axis closest
to the magnetic sensor on the shaft (or on the tip, for the Fig. 7. Cross-section of the extended cardioid of revolution, describing all
other point cloud) and a 3-D unit vector pointing along the possible rotations of the bending region about one axis. The bending links are
axis to the tip of the transducer. The following function is shown for several angles (green: 3 , 10 ; red: 25 , 48 , 77 ), not all of which
minimized by the Leven- berg–Marquardt optimizer: are actually attainable due to physical constraints (green ones are attainable,
red ones are only theoretically possible).
(3)
Finally, two transducer axes are defined, one in coordinates of

the magnetic sensor on the tip by the base point and the
unit directional vector , and the other in coordinates of
the sensor on the shaft by and .
5) Transducer Tip Modeling: Modeling the transducer tip
serves two purposes. First, the model can provide all 3-D
Fig. 8. Mathematical model of the tip of the flexible ultrasound trans-
information necessary to overlay a virtual model of the ducer—only the rotation about the axis is visualized here, the rotation
ultrasound transducer onto the measured location, which can about the axis is zero.
be useful for the development of a navigation system. Second,
and more im- portantly, if built relative to the optical marker
on the transducer shaft, it can be used to detect and partially The extended cardioid of revolution only models the
correct tracking inaccuracies (see Section III-B2). bending region of the transducer. To change from the
The ultrasound transducer contains not a single joint, but an coordinate frame of the magnetic sensor on the tip into that of
elongated bending region with several links extending over ap- the sensor on the shaft, a chain of transformations is modeled,
proximately 3 cm, similar to the one shown in Fig. 6. as shown in Fig. 8. Transformation Chain: First, the
Since the horizontal and vertical bending of the tip is con- coordinate frame of the magnetic sensor on the tip is
trolled by two steering levers and each lever offers seven posi- rotated and translated into the “link frame” in such a way
tions, 49 tip poses are manually selectable. The tip can also be that the axis points along the transducer axis. This
bent by external pressure, allowing us to model continuous transformation has six DOF, i.e., three rotations
motions on a curved surface to approximate all possible tip and one translation along the axis of the sensor to align the
poses. The motions of the bending region on this surface can sensor with the transducer axis, followed by
be ap- proximated using an “extended cardioid of revolution.” a rotation about and a translation along the transducer axis.
A cardioid is usually constructed in 2-D by revolving a circle Starting at the link frame, short links follow, which
around another one with the same radius . This corresponds to approxi-
an identical angular motion of two links of length and , mate the possible tip movements. In an alternating manner,
and the construction can be extended to a greater number of each
links (see Fig. 7 for six links). Spatial motions about two link rotates the transducer axis by about the axis or by
orthogonal rotation axes define a 3-D surface. This surface about the axis, respectively. No rotation about the axis is
can be constructed by rotating the extended cardioid around assumed. Each link introduces a translational offset along
its central axis, resulting in a surface of revolution. If we the axis. Including , the number of links, in total
look at the position of the transducer tip rather than the last has
link, we have to introduce another outward translation, so the four DOF
actual geometric shape of the attainable tip poses is similar to,
but not exactly the same as, our
extended cardioid of revolution. (4)
where The model has to be computed only once, when the magnetic
sensors are attached to the transducer.
This behavior of alternating links is inspired by the real nature

of the bending region, where four Bowden cables controlled
by the steering levers are pairwise and alternately connected to
every second link (see Fig. 6).
After the bending region, i.e., at the base frame, a final
translational and rotational offset is applied to
align the final tip coordinate frame with the magnetic sensor
on the shaft, which has five DOF. Actually, and
would both have six DOF. However, one DOF of the rotation
about the transducer axis is not needed for , as
it can be entirely compensated for by just altering the combi-
nation and magnitudes of and . Conversely, one more DOF
is needed for the rotation about the transducer axis included in
to adjust for the fact that the sensors can be mounted
at different angles about the transducer axis, i.e., assuming
and to be 0, the line through the origins of the two sensor
coordinate frames will not be parallel to the transducer axis.
The final model transformation, from coordinates of the
magnetic sensor on the transducer tip, into coordinates of the
sensor on the transducer shaft, can be described by
(5)
For error estimation (see Section III-B2), the model is actually

not built relative to the magnetic sensor on the transducer
shaft, but relative to the optical marker. Therefore, during
modeling, only the additional static transformation
needs to be applied to magnetic sensor coordinates to
determine optical marker coordinates.
Parameter Optimization: All parameters except angles
and remain constant for a given configuration and can be
computed offline. Given that the transducer axes in relation
to the sensors on the shaft and tip have been calibrated, only
five model parameters have to be estimated offline. These are
the translation along the rigid shaft axis (contained in
), the length of the bending region , the number of links
in the bending region (which can be initialized or fixed to
, because many surgical instruments have 12 links like the
one in Fig. 6), the angle of rotation about the tip axis, and the
length of translation along the tip axis (both contained in
).
For every selectable position of the two control levers, the
pose of the magnetic sensor on the tip in relation to the one
on the shaft is recorded in a distortion-free environment, i.e.,
without ferromagnetic or electrically powered materials in the
close vicinity. Then the remaining model parameters are opti-
mized numerically using the Levenberg–Marquardt algorithm.
At runtime, i.e., intraoperatively, the remaining two DOF of
the rotations and about the and axis can be numerically
optimized by the Levenberg–Marquardt optimizer. Either the
translation to the magnetic sensor on the tip, rotation
difference, or a combination of both can be minimized. In our
experiments we used a 1:3 weighting of translation in
millimeters to rotation in degrees, as in the magneto-optical
coregistration [see (2)].
Applicability of the Model: We constructed the model of the
transducer tip according to our observations and studies of the
behavior of the Bowden cables pulling the bending region.
How- ever, our model is only an approximation inspired by
the gen- eral design of flexible endoscopic instruments. Of
course, other endoscopic instruments can also have different
mechanical tip designs, so it may not be sufficient to apply our
model to them without further testing and/or modification.
As seen in our evaluation of the transducer tip model (see
Section IV-D and Fig. 9), the transducer tip mostly follows
the model, independently of the order of selection of the two
steering levers. The model is less precise when the tip is bent
more than halfway between its neutral position and maximally
bent position (“neutral position” corresponds to and being
approximately zero). We suspect the reasons for this to be the
nonideal elasticity of the bending region and its exposure to
large mechanical forces that are irregularly distributed over the
links. However, our clinical partner reported that intraopera-
tively the transducer tip is seldom bent more than halfway, i.e.,
the transducer tip is rarely bent to the leftmost or rightmost
preset deflection setting, or beyond a single of the preset
deflec- tions up and down. In all, three preset defections are
possible in each of the four directions (up/down/left/right), in
addition to the neutral position.
B. Tracking Error Detection and Correction

We propose approaches for the detection and correction of
dynamic sources of electromagnetic tracking errors which
can be divided into redundancy-based, model-based, and
vision-based tracking, where the latter is described in more
detail in our previous work [24].
1) Tracking Redundancy-Based Approach: Each approach
to correcting static erroneous measurements that is based on
hybrid (and thus redundant) tracking of an object (e.g., [40] or
[41], that use magneto-optical tracking), can also be used for
error detection. Such an approach is however not intended to
detect and correct dynamic sources of errors, as it is performed
only once to create a static distortion function of the magnetic
field. Fur- thermore, magneto-optical tracking cannot be
applied to the tip of the ultrasound transducer due to its
flexibility and the lack of line of sight to the optical cameras.
Therefore, one idea for an error detection method is to esti-
mate errors at the magnetic sensor on the shaft instead and, for
a simple error correction, transfer them to the magnetic sensor
on the tip.
a) Error Detection: Similar to the authors of [20] and
[21], who connect two magnetic sensors rigidly to each other
and use the deviation of their distance to detect field
distortions, our simple error detection method is based on
tracking redundancy. However, we do not use two magnetic
sensors, but one optical
ULTRASOUND
Fig. 9. Validation of the transducer tip model. In relation to the distance of the tip from its neutral position (in millimeters), we compared 171 tip poses not
used for model generation to their corresponding poses computed via the model, in terms of (a) their translation difference in millimeters and (b) their rotation
difference in degrees. For example, for the tip pose represented by the first data point on the axis, the transducer tip is bent 3.7 mm away from its neutral
position, and we measured a deviation of 2.7 mm and 1.0 from the modeled pose.
marker and one magnetic sensor each rigidly attached to the
transducer shaft.
Intraoperatively, every measured pose of the magnetic
sensor on the shaft is transformed by applying (1). The error (7)
corresponding to the transformation is computed using (2), 2) Model-Based Approach: Dynamic field distortions,
consid- ering the translation error or the rotation error caused by e.g., moving instruments close to the sensors, are
only, or a combination of both . The respective error particularly difficult to reproduce, because the variation in
can be used as a distrust value, which is compared to a field strength is usually not shift invariant (i.e., disturbances
selectable threshold to reject data of the magnetic sensor on affect the field strength nonuniformly at different locations) as
the tip as probably distorted. The surgical staff can be already described by [20] and [21]. Hence the magnetic field
automatically warned of possible errors. at the sensor on the shaft is often distorted in a completely
b) Error Correction: As the magnetic sensor on the shaft different direction and magnitude than at the sensor on the tip,
is in close proximity to the one on the tip, both sensor mea- even though the sensors are placed close to each other. This
surements will most likely be affected by any magnetic field makes the simple, redundancy-based error estimation approach
distortions. As long as the main error component measured at unreliable (for experimental results, refer to Sections IV-E and
both sensors is the same, it may also be partially corrected like IV-F).
static field distortions. A simple approach to approximate a a) Error Detection: Another, more sophisticated approach
correction of erroneous measurements is to transfer for error estimation requires a previous calibration and
deviations detected at the magnetic sensor on the shaft to the modeling of the possible transducer tip movements (see
one on the tip: We can compute the distortion of the Section III-A5). In the absence of distortions, the
measurements on the shaft and undo this distortion on the tip measurements of the magnetic sensor on the tip can be
provided assumed to lie on the previously modeled cardioid of
that the distortion field is fairly constant for the two sensors. revolution-like shape, reducing their DOF from six to two (
As the sensors can be rotated relative to each other we need to and ). Because the true motion of the tip is con- strained to
compute this distortion relative to the fixed OT (world) those two DOF, we can exploit the redundancy of the
coordinate frame. remaining four DOF. If measurements do not lie exactly on the
Therefore, the rotational and translational parts of the devi- cardioid of revolution-like surface, we can deduce poses
ation between the previously calibrated (“calib”) and the cur- similar to the measurements but lying on the surface that are
rently measured (“meas”) transformation on the shaft can be likely to be closer to the true pose of the sensor.
applied to the measured transformation on the tip to obtain The model is built relative to the optical marker, which
corrected (“corr”) transformations is attached to the transducer shaft, and is not influenced by
magnetic field distortions. Therefore, only the previously
determined static transformation and the tip
model parameters (compare Section III-A5) are needed when
optimizing and by the Levenberg–Marquardt algorithm.
After optimization of the model angles and , we can
compute the remaining distance between the modeled and
(6)
the un-
corrected pose of the magnetic sensor attached to the tip. With TABLE I
a perfect model in an error-free environment it would always MAGNETO-OPTICAL CO-REGISTRATION RESIDUAL ERRORS BEFORE AND
AFTER OPTIMIZATION. FOR THE COMPUTATION OF SEE (1). MINIMUM,
be possible to find model angles and such that the MEAN, SD, RMS VALUE, AND MAXIMUM ARE GIVEN FOR EACH CASE
modeled pose is identical to the uncorrected pose. Thus, a
deviation between these two poses can be used as a distrust
value in order to warn the surgical staff about detected
distortions.
b) Error Correction: For error correction, we simply re-
place the measured, uncorrected pose of the sensor attached to
the tip with the pose on the modeled surface that was
previously computed to be closest to the distorted pose.
3) Vision-Based Correction: As the intrinsic and extrinsic
camera parameters of the laparoscope and hence the spatial lo- TABLE II
ROC KEY FIGURES FOR PREDICTION OF TRACKING ERRORS OF AT LEAST
cation of the image plane are known, another approach to im- 2.5, 5.0, AND 7.5 MM. THE AREA UNDER CURVE (AUC), MAXIMUM
prove the tracking accuracy of the magnetic sensor mounted YOUDEN INDEX , BEST THRESHOLD, TRUE POSITIVE RATE
on the tip is to automatically localize the transducer tip in the (TPR), SPECIFICITY (SPC), MINIMUM FALSE POSITIVE , AND
MAXIMUM FALSE NEGATIVE ARE GIVEN FOR EACH CASE
images of the laparoscope camera and adjust the measured
sensor pose by an in-plane transformation accordingly. This
correction transformation can further improve the results of
the model- based error correction method, especially in
correcting errors introduced by calibration inaccuracies. For
our work on this topic see [24].
In the next section, we describe our evaluation of all system
calibration and error estimation methods.
IV. EXPERIMENTAL EVALUATION

B. Magneto-Optical
The accuracy throughout system calibration and transducer Coregistration
modeling as well as error detection and correction was When coregistering the EMT and the OT we determine the
evaluated in a series of experiments. transformations from the magnetic sensor to
Our experimental setup is depicted in Fig. 1. As can be seen the optical marker on the transducer shaft and
on the figure, we temporarily mounted an additional optical from the EMT coordinate frame to its attached optical marker.
marker (made entirely of plastic so as not to influence the Because of tracking errors it is not possible to find an exact so-
EMT) on the flexible transducer tip, below referred to as “ lution for both transformations, only the optimal solution. At
.” Similarly to magneto-optical coregistration, we determined all poses which were used for registration, an rms residual
the transformation from magnetic sensor error of mm for translation and
coordinates to optical marker coordinates (see Section III-A3). degrees for rotation remained after optimization [see Section
Along with optical tracking, this transformation III-A3, (2)]. For detailed results before and after optimization
can be used as a reference method for determining the pose of see Table I.
the magnetic sensor on the tip. The method is of course not For verification of the coregistration accuracy, we
applicable to a clinical setting because of the additional optical conducted a series of independent registrations, each
marker at the flexible transducer tip, but it can be used well for performed with multiple transducer poses. At least 13 poses
comparison of accuracy. belonging to the same registration were used to verify the
accuracy of the respective previous registration. This was
A. Laparoscope Camera and Laparoscopic Ultrasound done while the transducer tip remained in an approximately
neutral position, with only minor movements due to
Calibration
gravitational force. In a second series, we fixed the transducer
Laparoscope camera calibration was studied extensively in shaft at one position and moved the transducer tip to all
our prior work [26], which yielded 2-D overlay rms error of selectable poses using the levers. This was repeated for
about 1.6 mm for C-arm CT data (for further results obtained multiple other transducer shaft poses.
in the current setup, refer to Section IV-F2). Laparoscopic ul- Applying (1) and (2), for each pose we calculated the trans-
trasound calibration was also assessed previously by imaging a lation error and the rotation error for the magnetic
magnetic sensor submerged in a water bath and comparing its sensor on the shaft. Replacing by and
appearance in the ultrasound plane with the tracking data [23]. by in these equations, we were also able to compute the
Because the spatial location of the ultrasound plane is known, errors for the magnetic sensor on the tip in a similar fashion.
we can compute a 3-D position for the appearance of the Table III shows the results.
sensor there and compute the difference to its tracked position. As can be seen in Table III, the error at the sensor on the
This re- sulted in 3-D rms error of about 1.7 mm. shaft
is significantly higher than at the sensor on the tip. We
attribute this to the larger distance between magnetic sensor
and optical marker: The mean orientation error of the sensor
on the shaft
ULTRASOUND
TABLE III
3-D TRANSLATION AND ROTATION ERRORS IN AN UNDISTORTED AND A DISTORTED MAGNETIC FIELD.
MINIMUM, MEAN, SD, RMS VALUE, AND MAXIMUM ARE GIVEN FOR EACH CASE. FOR A GRAPHICAL COMPARISON
SEE FIG. 12
is 1.10 . In practice, this error only partly results in an links of the bending region or a rotation around the instrument
additional position error, but theoretically, over the estimated axis can still occur, which is not modeled by the current
distance of design. Particularly when bending the transducer tip to extreme
366 mm from the magnetic sensor to the optical marker on the poses,
transducer shaft, for every 0.1 , there may be an additional po-
sition error of 0.6 mm. At the sensor attached to the tip the
same effect would only cause an additional position error of
0.07 mm because of the much smaller distance.
C. Transducer Axis
Calibration
During axis calibration the rms residual error of one calibra-
tion, i.e., the distance of the collected position measurements
from the optimized cylinder surface, ranges from 0.57 to
1.03 mm with a mean value of 0.74 mm using 1.3 mm sen-
sors (performing six calibrations). We also inserted a 1.8 mm
Ascension sensor into the end of the plastic cylinder instead
of a 1.3 mm sensor. This sensor yielded better results of
0.23–0.80 mm with a mean value of 0.42 mm (performing four
calibrations).
D. Transducer Tip
Modeling
We experimentally evaluated three important errors related
to the transducer tip motions: the residual error during
computation of the model, the residual error using the
computed model, and the repositioning error (independent of
the model).
First, we estimated the remaining residual error of the
poses used for computation of the model parameters (see
Section III-A5). For models determined from five pose sets the
rms error ranges from 0.58 to 1.28 mm with a mean value of
0.94 mm for translation and from 3.48 to 3.89 with a mean
of 3.66 for rotation.
We next verified the accuracy of the model for another 171
poses that were not used for computation of the model (see
Fig. 9). We measured the deviations of the poses from the the-
oretical model curve in relation to the distance of the tip from
its neutral position; a distance of about 75 mm corresponds to
at least one of the two steering levers pulled to its maximum
position. The rotation error stayed approximately the same
(rms was
3.22 ), but the translation error was higher (rms 3.50 mm).
We assume the higher translation error to be due to the fact
that the bending region of our ultrasound transducer proved to
be not ideally elastic, i.e., a slight translation within one of the
ULTRASOUND
the bending region is exposed to large mechanical pulling
forces that could cause an additional translation or rotation of
the tip. During our experiments, a Bowden cable broke twice
when we pulled the two levers to extreme positions, so it
seems that the bending region is not designed for such
positions. We also received a report from our clinical partner
that intraoperatively the transducer is maximally bent away
from its neutral position by only about 50 mm.
In a third experiment, we estimated the repositioning accu-
racy of the transducer tip. We repeatedly moved the two
steering levers to the same position and acquired
measurements of the transducer tip using the temporarily
attached optical marker. To find out whether the order of
selection of the control levers influ- ences the positioning
repeatability as well, we also alternated the order of selection
of the control levers, e.g., first pulled the left lever twice and
then pulled the right lever once, or first pulled the right lever
once and then pulled the left lever twice. For each of four
different final lever positions, we computed the mean tip
position and the distances from this mean position. Averaging
over the four series, we obtained a mean distance of 1.32 mm
from the mean tip position with a standard deviation (SD) of
1.02 mm, minimum of 0.36 mm, and maximum of 3.76 mm.
In- dependently comparing the two orders of selection of the
control levers, we obtained a mean distance of 0.56 mm
between the first and the second final tip position. So we
assume that the overall repositioning variance is much more
significant than the variance induced by an alternating order
of selection of the steering levers. The overall repositioning
variance obtained supports our assumption of the nonideal
elasticity of the bending region.
E. Error
Detection
For the evaluation of the error detection methods, we
recorded
517 distorted measurements, using a knife, a steel rod 10 mm
in diameter, and a power supply unit to create varying
distortions of the magnetic field. We could not identify their
exact grade of steel, but both the knife and the steel rod were
attracted to a magnet (in a similar way). The power supply
unit was heavily attracted despite being turned off. The knife
and steel rod were arbitrarily placed at a distance of about 3–
10 cm from the transducer tip, sometimes interposed
between the magnetic field generator and the magnetic
sensors, but most times not in the direct path between sensors
and generator. The power supply unit was placed next to the
field generator at a distance of about 30 cm, never in the
direct path between sensors and generator.
Fig. 10. Comparison of the detected magnitudes of error—using (a) the tracking redundancy-based approach and (b) the model-based approach—to the actual
error of the magnetic sensor on the tip (translation in millimeters in all cases). The actual error was determined via an additional optical marker temporarily
attached to the transducer tip only during our experiments. The identity function of the actual error is plotted as a dashed line to aid visual comparison. Our
comparison shows that errors detected by the model-based approach correlate better with the actual errors.
For each measurement we computed the translation from When predicting distortions, two sorts of errors are possible
the position of the magnetic sensor attached to the shaft (as (explained here for the redundancy-based method, the model-
measured via electromagnetic tracking, , and based methods follows similarly): A false positive (type 1
optical tracking) to its reference position (as computed via error) occurs when the distance between the calibrated pose
and optical tracking), which reflects the mag- and the measured pose of the magnetic sensor on the shaft is
nitude of error detected by our tracking redundancy-based
approach. We also determined the translation from the posi- above the threshold, but the measurement of the sensor on the
tion of the magnetic sensor on the tip (as measured via the tip is actually not distorted, i.e., we erroneously reject data.
electromagnetic tracking, , and optical tracking) A false negative (type 2 error) occurs when tracking data of
to its modeled position, which reflects the error detected by the sensor on the tip is distorted, but the distance between the
our model-based approach. In an ideal and error-free setting calibrated and measured pose of the sensor on the shaft is
both translations would be zero. Additionally, we computed below the threshold, i.e., we fail to predict the distortion.
the actual error of the magnetic sensor on the tip, i.e., the Similarly, we call correctly predicted errors “true positive” and
translation from the measured position to the reference correctly predicted absence of error “true negative.” Note that
position (as computed via and optical the error of magneto-optical coregistration is quite large (see
tracking), and compared it to both the errors detected by the Section IV-B) at the sensor on the shaft, so setting a low
redundancy-based approach and the model-based approach. threshold value will likely trigger false positives.
Fig. 10 depicts this comparison. For our set of distorted measurements we computed several
1) Correlation Coefficient: Both magnitudes of errors de- receiver operating characteristics (ROC) for predicting errors of
tected by the redundancy-based approach and the model-based
2.5, 5.0, and 7.5 mm.
approach correlate with the actual error of the magnetic sensor
Our error prediction is a classifier that computes a distrust
on the tip. As can be seen from Fig. 10, the magnitude of error
detected by the redundancy-based approach has a much lower value for the current measurement and then decides whether
correlation with the actual error (correlation coefficient 0.69) the measurement should be considered erroneous or not. The
than the one detected by the model-based approach computation of the distrust value is fixed [see (2)], but the
(correlation coefficient 0.95). threshold for comparison can be varied. Instead of picking only
2) Receiver Operating Characteristic: For both the redun- several exemplary thresholds, for all possible thresholds the
dancy-based approach and the model-based approach we can false positive rate and true positive rate are computed and
choose a certain threshold value and predict a distortion of the combined to form a curve. The ability of the redundancy-based
pose of the sensor on the tip if the detected magnitude of error and the model-based methods to predict erroneous
exceeds this threshold. measurements of the position of the magnetic sensor attached
to the tip is shown in Fig. 11.
ULTRASOUND
Fig. 11. ROC curves for prediction of tracking errors of at least 2.5, 5.0, and 7.5 mm. Each point on the curve represents the achievable performance for a
certain threshold, i.e., sensitivity (true positive rate) and specificity (1—false positive rate). The point with the highest sum of sensitivity and specificity (highest
Youden index) has been marked for each curve and may be considered the achievable performance of our error-detection method for each case. (a) 25 mm. (b)
5.0 mm. (c) 7.5 mm.
We considered the value with maximum Youden index [42]

to be the optimal choice for a threshold. The Youden index for
a given threshold is defined as follows:
(8)
where is the fraction of true positives, the fraction of false

negatives, the fraction of false positives, and the fraction of
true negatives. The possible range of values is from zero to
one inclusive.
In Table II we present the key figures for each ROC: For
both prediction methods the area under the ROC curve (AUC)
and the maximum Youden index have been computed.
Addi- tionally, corresponding to , the threshold value,
Fig. 12. 3-D translation errors in a distorted field. To distort the magnetic
sensitivity (true positive rate, TPR) and specificity ( field, we used either a knife, a steel rod, or a power supply unit, while the
, false positive rate), and both the smallest false positive value transducer was fixed in varying locations. In each column, minimum and
and the greatest false negative value are given. The maximum are represented by stars, the rms error by a circle, and mean and
SD by bars. See Table III for exact values including rotation and
latter define the most extreme cases where the classifier measurements in an undistorted field.
would have made a wrong decision.
For predicting an error of 2.5 mm or greater, redundancy-
based error prediction would in the best case have achieved a F. Error
sensitivity of 50% and a specificity of 75%. That is, it would Correction
have been able to correctly predict 50% of all errors of 2.5 mm Error correction accuracy was determined in two steps: first,
or greater and correctly predict 75% of the cases in our data we assessed the 3-D tracking accuracy of the system with and
set, where the error was below 2.5 mm. In comparison, the without our proposed error correction methods in comparison
model- based error prediction would have achieved a to reference measurements obtained via and op-
sensitivity of 91% and a specificity of 79% for the same data. tical tracking. Second, we measured the 2-D overlay accuracy
For predicting an error of 5.0 mm or greater, sensitivity and in an application-specific setting, i.e., superimposition of a la-
specificity would have been 62% and 75% (redundancy-based) paroscope camera image (see [23] and [24] for further details
versus 91% and 93% (model-based). For all values including on our application). Three-dimensional points can be projected
those for 7.5 mm, see Table II. onto the camera image plane once the projective geometry of
The model-based error prediction worked much more reli- the camera is known. The overlay error is the difference
ably. For example when trying to predict an error of 2.5 mm or between the computed 2-D location of the point projected onto
greater, it would have correctly identified all error-free the image plane and the actual location as extracted from the
measurements better than 1.26 mm and all errors above 4.67 video image. The two steps cover different aspects: the
mm. tracking error is especially relevant for applications like 3-D
ultrasound, while the
Fig. 13. Performance of (a) tracking redundancy-based error correction and (b) model-based error correction. The error remaining after correction is plotted
against the uncorrected error (translation in millimeters in all cases). The dashed line represents the original error without correction and has been added to aid
visual comparison.
overlay error is crucial e.g., for medical augmented reality performed the simple tracking redundancy-based approach
applications. The latter additionally includes the camera with an rms error of 6.67 mm. Model-based error correction
calibration error and may be partially corrected using the proves to be functional, as the uncorrected measurements of
vision-based correction approach presented in our previous the magnetic sensor on the tip had an rms error of 6.91 mm.
work [24], either exclusively or in combination with the error The orientation error, however, could not be improved (rms
correction approaches described here. 3.21 uncorrected and 3.37 corrected). This suggests that
1) 3-D Accuracy: When evaluating the performance of the either the orientation difference to the uncorrected sensor
magneto-optical co-registration, at the same time we evaluated should be weighted more heavily than it is now, or that the
the effectiveness of our proposed error correction approaches. model should be extended to better approximate rotation of the
For best possible comparison we assessed all errors simulta- transducer tip.
neously. We recorded 30 undistorted and 517 distorted For a comparison of both error correction approaches see
measurements, applied both the simple tracking redundancy- Fig. 13. Whereas the simple tracking redundancy-based error
based error correction (see Section III-B1) and the model- correction had a much wider variance as well as a poorer
based error correction (see Section III-B2), and recorded all overall performance, the model-based error correction in most
distances to a reference pose computed using cases could reduce the translation error to below 5 mm, even
and optical tracking. For the model-based error correction
for original errors in the centimeter range.
approach, the position difference of the modeled to the
However, probably due to the propagation of errors from
measured pose of the magnetic sensor on the tip (in
tracking, calibration, and modeling, for errors lower than ap-
millimeters) and the corresponding orien-
tation difference (in degrees) were weighted 1:3 [see (1)]. proximately 2.1 mm the model seems to give slightly worse
In the undistorted setting (see Table III) the magnetic sensor results than the uncorrected measurements of the sensor on the
on the tip had an rms error of 1.28 mm, the one on the shaft tip. The model-based error correction could thus be restricted
had an rms error of 2.92 mm, and the tracking redundancy- to cases where a relatively high error is predicted (for error
based error correction and the model-based error correction prediction see Section III-B2a).
had rms errors of 2.91 and 2.27 mm, respectively. 2) 2-D Overlay Accuracy: For assessing the overlay accu-
The tracking redundancy-based error correction performed racy in both the undistorted and distorted case, the ultrasound
almost exactly like the sensor on the shaft itself (for rotation it transducer was fixed in various poses and the laparoscope was
performed slightly worse). The model-based error correction used to observe the transducer tip from various angles and
performed better, although it was anchored to the optical distances.
marker on the shaft and part of the corresponding magneto- In the course of the experiments the transducer tip was
optical coregistration error presumably propagated into it. steered to different angles and the laparoscope was also rotated
In the distorted case (see Fig. 12 and Table III) the model- around its own axis. For distorting the magnetic field we again
based error correction with an rms error of 3.15 mm clearly used the steel rod with a diameter of 10 mm.
out-
ULTRASOUND
Fig. 14. Automatic segmentation of one retro-reflective sphere of the optical similar to an operating table. At the time of these
marker attached to the tip and projection of the position of the magnetic sensor
on the tip determined via various methods. Here, “Seg” refers to the position experiments, only the
of the segmented sphere, “OT” to the projection of the pose of the optical
marker attached to the tip centered in the sphere, “EMT” to the projection of
the uncorrected position of the magnetic sensor on the tip, and “CorrEMT”
to the projection of a corrected position of the sensor on the tip (corrected by
one of the above-mentioned methods). The metal rod was placed at a distance
of about
5 cm from the magnetic sensors to disturb the magnetic field. The image was
enhanced for readability.
Using the transformation from the magnetic

sensor to the optical marker on the tip, at each measurement
we transformed
1) the uncorrected position of the magnetic sensor attached
to the tip (“ ”);
2) the tracking redundancy-based corrected position of the
sensor on the tip (“CorrRedundancy”);
3) the model-based corrected position of the sensor on the
tip
(“CorrModel”).
The resulting positions were then projected into the image
plane, the spatial location of which was known from camera
calibration. The measured 3-D position of the optical marker
on the tip (centered in one of the retro-reflective spheres) was
also projected into the image plane, as shown in Fig. 14. The
distance in millimeters parallel to the image plane to the
segmented mid- point of the sphere was computed and taken
as the measure for overlay accuracy.
Only measurements within a distance of 5–20 cm between
the sphere and the camera center were accepted. This rather
large distance of up to 20 cm is required to observe both the
transducer tip (for axis segmentation) and the sphere at the
same time. Keeping such a large distance is however very
unlikely during surgery, for which the laparoscope was
calibrated for a maximum working distance of only about 10
cm. This reduces the overlay accuracy when the laparoscope is
further away than
10 cm, so theoretically the results obtained here could be
further improved.
In our previous experiments [23], we used the steel rod to
create dynamic distortions similar to those from surgical
instruments, and a metal plate to create static distortions
Fig. 15. 2-D overlay errors in a distorted magnetic field. In each column,
ULTRASOUND
minimum and maximum are represented by stars, the rms error by a circle,
and mean and SD by bars. See Table IV for exact (including undistorted)
values.
TABLE IV
OVERLAY ERRORS IN AN UNDISTORTED AND A DISTORTED
MAGNETIC FIELD. MINIMUM, MEAN, SD, RMS VALUE,
AND MAXIMUM ARE GIVEN FOR EACH CASE
tracking redundancy-based error correction method was avail-

able and the results were not yet promising: in an undistorted
environment, the accuracy of the corrected sensor was
sometimes poorer than the uncorrected measurements.
Dynamic distortions got worse in most cases and only static
distortions could be partially corrected; errors in the
centimeter range still remained.
To assess the overlay accuracy of the new, improved error
correction methods we took 207 undistorted and 935 distorted
measurements. For the results in undistorted and distorted
cases see Fig. 15 and Table IV. As illustrated, the simple
tracking redundancy-based error correction also performed
insuffi- ciently in the new 2-D experiments, whereas the
model-based approach yielded major improvements compared
to the uncorrected measurements.
Combining the model-based error correction with the
vision- based method described in our previous work [24]
(see also Section III-B3) yielded even better performance than
the model- based method alone. In the scatterplot of Fig. 16
you can see a comparison of the error correction performance
of
• the simple tracking redundancy-based error correction
method;
• the model-based error correction method;
Both [20] and [21] fix two magnetic sensors rigidly to
each other and determine the distance between them in a
Fig. 16. Performance of the different error correction methods, measured in

2-D perpendicular to the viewing direction. The error remaining after each
correction is plotted against the uncorrected error (translation in millimeters
in all
cases). The dashed line represents the original error without correction and has
been added to aid visual comparison.
• a combination of the model-based and vision-based error

correction methods.
V. DISCUSSION
As corroborated by our clinical partner, accuracy require-
ments in abdominal (laparoscopic) surgery are different than
in e.g., orthopedic surgery or neurosurgery. A discrimination
of about 5 mm is usually sufficient, since canalicular
structures such as vessels and bile ducts play a critical role if
they are 5 mm or thicker in width. A lymph node is considered
to be inflicted by a tumor if the diameter is more than 10 mm.
Accordingly, the measurements obtained for the undistorted
case as well as for the distorted case after model-based error
correction, both of which yielded rms error below 5 mm, are
certainly accept- able under clinical conditions.
Our results show that the model-based approach is clearly
superior to the redundancy-based method for error correction
as well as for error detection. This section presents a
comparison of our methods and results to related work. We
investigate the differences and suggest possible directions for
future research.
A. Error Detection
The authors in [20] and [21] mentioned that the variation
in field strength is usually not shift invariant, i.e., two sensors
mounted a certain distance apart are affected differently by
magnetic field distortions. While this is essential for their
error detection methods, it significantly impacts our
proposed redundancy-based error detection and correction
method.
[21]. We are able to achieve a higher sensitivity than [20] and
a higher correlation than [21].
TABLE V
DETECTION RATE OF ERRORS GREATER THAN 1 MM FOR DIFFERENT
THRESHOLDS , AS PRESENTED IN [20]. IF
IS EXCEEDED, THE SYSTEM IS CONSIDERED TO BE DISTORTED
TABLE VI
CORRELATION BETWEEN THE MEASURING ERROR AND THE PLAUSIBILITY
VALUE FOR DIFFERENT INSTRUMENTS, AS DESCRIBED IN [21]. IS THE
KNOWN DEVIATION OF THE MEASURED TOOL TIP POSITION FROM ITS
ACTUAL POSITION, THE PLAUSIBILITY VALUE REFLECTS THE CALCULATED
DEVIATION. DENOTES THE MAXIMUM KNOWN DEVIATION
OBTAINED IN ONE EXPERIMENT
distortion-free environment ([21] applies an additional fixed

transformation from each sensor to a calibrated pointer tip
before computing the distance, but otherwise the approaches
are identical). Then both groups induce various distortions,
measure the “movement” of one sensor (its position is actually
fixed) and compare it to the measured change in distance
between both sensors.
The results of [20] are presented in Table V, where
is the threshold value for the deviation of the
distance between both sensors. The system is considered to be
distorted if exceeds
, where
and denote the translational readings from sensor 1 and
2 and is the known distance between the sensors. Using
a threshold of 0.5 mm, a “detection rate” of 87% was achieved,
i.e., 87% of all errors greater than 1 mm were correctly
detected. Unfortunately a ratio of false positives is not
provided, so a direct comparison to all our results is not
possible.
[21] used different instruments to induce distortions of the
magnetic field, as shown in Table VI. For each instrument it
presents the correlation coefficient between the measuring error
and the plausibility value , where denotes the known devi-
ation of the measured tool tip position from the actual tool tip
position and is the difference between the
two tool tip positions and calculated from the
measurements of sensor 1 and 2. For example using a forceps
they caused a maximum measuring error of 4.7 mm with
a correlation coefficient of 94%. Again, a direct comparison to
all our results
is not possible given only the correlation coefficient.
However, we may assume that our redundancy-based error
detection method performed similarly to the results of both
[20] and [21], although we faced additional difficulties: the
distance between the magnetic sensor and the optical marker
on the shaft is rather large (366 mm). On the one hand, this
causes a magneto-optical coregistration error of 2.66 mm for
the magnetic sensor on the shaft in our setup (see Table III),
so low thresholds trigger false alarms easily. On the other
hand, our setup has the advantage that the optical tracking of
the optical marker on the transducer shaft is not affected by
magnetic field distortions. In contrast to the redundancy-based
error detection method, the model-based approach clearly
performs better for compa- rable measures than both [20] and
ULTRASOUND
For applications in minimally invasive surgery, the methods generator. One Sensor Setup: It is even possible to abandon
of [20] and [21] require instruments with a rigid tip where both the sensor
sensors can be mounted. It would be possible to place two sen- on the shaft entirely. The optical marker on the tip is then also
sors at the tip of our transducer, but the distance between those needed for construction of the model, but as in the current setup
must not be too small ([20] used 103 mm). Otherwise both
sensors might be within an area of similar field deformation
and affected too similarly to reliably detect deviations.
B. Error Correction
The redundancy-based error correction did not gain any im-

provements for the orientation error and only slight improve-
ments for the position error. This suggests that the magnetic
sensors on the tip and on the shaft are mounted too far apart,
so they are not affected by distortions in a similar enough way.
The model-based error correction did reduce the rms position
error from 6.91 to 3.15 mm. The orientation error stayed
approximately the same (rms 3.21 versus 3.37). The latter
suggests that either the model should be refined to better
match the real rotation, probably even including elasticity
constraints, or that the difference in rotation to the
uncorrected measurements of the magnetic sensor on the tip
should be weighted more strongly than it is now. However,
there is a trade-off between orientation error and position
error: a stronger weighting of the rotation difference to the
uncorrected measurements of the sensor can cause the model to
favor poses with lower orientation error but higher position
error (see Section III-A5). Finding an optimal combi-
nation requires further investigation.
There might be limits to the achievable error correction
capa- bilities of the model-based approach: errors are
propagated and accumulated through tracking, magneto-
optical coregistration, axis calibration, and modeling.
Coregistration also introduces additional errors. [14] mentions
that in their setup the coregistration of the EMT and the OT
alone introduced 0.1–0.2 mm of additional error compared to
using the EMT only. Probably even higher errors have to be
anticipated for the calibration of the laparoscopic ultrasound
transducer shaft, especially because the optical marker and
magnetic sensor on the shaft are mounted a large distance
apart.
Metal-Immune Electromagnetic Tracking Systems:
Magnetic field generators optimized for metal environments
were recently presented by NDI and Ascension. Ascension’s
flat transmitter is designed to overcome distortions from below
the field generator, caused e.g., by the operating table. Due to
its lower excitation, in our setup it unfortunately performed
worse than the midrange transmitter. During ultrasound
calibration, we obtained errors of 4–8 mm using the flat
transmitter versus 1.7 mm using the mid-range transmitter
[23], [24]. Bigger sensors could be used to improve accuracy
but this would require bigger trocars. Using
1.3 or 1.8 mm sensors, the total diameter of the transducer
could
be kept below 12 mm (including sterile cover), so it would still
fit a regular trocar. Furthermore, metal-immune systems will
not be “immune” from dynamic distortions caused e.g., by
moving ferromagnetic instruments close to the sensor or field
itULTRASOUND
can be removed afterwards. After determining the transfor-
mation from the magnetic sensor to the optical
marker on the tip, it is possible to compute the movement of
the magnetic sensor on the tip in relation to the optical marker
on the shaft. Instead of using and ,
the model would be constructed from
, and . For axis calibration, however, a
calibration phantom would have to be constructed to allow
axis calibration using the OT instead of the EMT.
A second option for model construction would be to trans-
form measurements of the magnetic sensor on the tip via
the EMT, its attached optical marker, and the OT into the
coordinate system of the optical marker on the shaft, i.e., by
, and
. For this method even the optical marker on the tip would not
be needed any more. But then the determination of the
transformation from the EMT to its attached
optical marker would require an additional calibration
phantom, because for magneto-optical coregistration one
magnetic sensor and one optical marker must be fixed in
relation to each other. In our opinion this approach would be
less favorable, because more tracking and calibration errors
may be accumulated during the determination of the model
parameters.
No matter which option is chosen, error detection using
the model-based approach would be possible in the same
manner—without a second sensor and still with superior per-
formance compared to the current state of the art. It should be
verified that these assumptions are correct and that our
proposed methods also work without an additional sensor
mounted on the shaft.
Sensor Fusion: The error correction approach we describe
here is based on the exploitation of redundancy in our tracking
data. For more complicated setups, where more and different
types of information are available, those could be fused using
more sophisticated methods like Kalman [43] or Monte Carlo
[44] filtering. The additional complexity of approaches like
these was not needed in our case but could easily be included,
e.g., by using a Kalman filter for smoothing our tracking data.
C. Future
Work
Future work should include a quantification of the
robustness of the error correction methods proposed in this
paper.
The current system has been evaluated thoroughly in a
laboratory setup. In future work the integration of the tracking
systems into a clinical operating room setup should also be
evaluated, e.g., regarding tracker installation and sterilization
of tracking bodies and sensors, along with their precise repo-
sitioning to laparoscopic instruments. The same holds for the
integration of our proposed procedures into the surgical work-
flow, such as practical calibration and verification procedures
and user interfaces for surgical staff. The first goal in
particular could be achieved using integrated devices: for
instance, the optical tracking cameras could be attached to
surgical lights or the operating room ceiling, as done in our
previous work [26], while the magnetic field generator could
be integrated into the surgical table, like the flat transmitter
from Ascension. Magnetic sensors should also be entirely
integrated into the laparoscopic instruments so that no
dedicated sterilization
procedure is required. Optical markers that can be used in [2] B. Rau, M. Hunerbein, and P. M. Schlag, “Is there additional infor-
autoclaves are already available from e.g., Brainlab. However, mation from laparoscopic ultrasound in tumor staging?,” Digestive
in order to simplify the intraoperative setup process, there is Surgery, vol. 19, pp. 479–483, 2002.
[3] S. Ganguli, J. B. Kruskal, D. D. Brennan, and R. A. Kane,
a need for either optical markers that can be very precisely “Intraopera- tive laparoscopic ultrasound,” Radiologic Clin. North
(re-)mounted to the instruments, e.g., using snap-on connec- Am., vol. 44, pp.
tions, or adequate routines for a fast intraoperative system 925–935, 2006.
[4] J. Leven, D. Burschka, R. Kumar, G. Zhang, S. Blumenkranz, X. D.
calibration. Dai, M. Awad, G. D. Hager, M. Marohn, M. Choti, C. Hasser, and R.
H. Taylor, “Davinci canvas: A telerobotic surgical system with
integrated, robot-assisted, laparoscopic ultrasound capability,” in
VI. CONCLUSION Proc. Int’l Conf. Medical Image Computing and Computer Assisted
We present new methods for detecting and partially cor- Intervention (MIC- CAI). New York: Springer-Verlag, 2005, vol.
3749, Lecture Notes in Computer Science, pp. 811–818.
recting static and dynamic magnetic tracking errors. They are [5] B. P. Lo, A. Darzi, and G.-Z. Yang, “Episode classification for the
applied to a flexible laparoscopic ultrasound transducer which analysis of tissue/instrument interaction with multiple visual cues,” in
is localized by a hybrid magneto-optical tracking system. Our Proc. Int’l Conf. Medical Image Computing and Computer Assisted
In-
new evaluation methodology uses an optical marker as a refer- tervention (MICCAI), R. E. Ellis and T. M. Peters, Eds. New York:
ence and closes the chain of transformations between optical Springer-Verlag, 2003, vol. 2878, Lecture Notes in Computer
Science, pp. 230–237.
tracking, electromagnetic tracking, laparoscope images, and [6] J. Climent and P. Mares, “Automatic instrument localization in la-
the marker. This provides a sound validation of the proposed paroscopic surgery,” Electron. Lett. Comput. Vis. Image Anal., vol.
methods against reference data obtained from optical tracking. 4, no. 1,
pp. 21–31, 2004.
The methods presented could improve navigation in a large [7] S. J. McKenna, H. N. Charif, and T. Frank, “Towards video under-
set of minimally invasive procedures where flexible laparo- standing of laparoscopic surgery: Instrument tracking,” Image and
scopic ultrasound or similarly constructed surgical instruments Vi- sion Computing New Zealand, 2005.
[8] S. Voros, J.-A. Long, and P. Cinquin, “Automatic localization of
with known mechanical properties of their bendable tips are laparoscopic instruments for the visual servoing of an endoscopic
used, e.g., for liver, biliary tract, and pancreas interventions. camera holder,” in Proc. Int’l Conf. Medical Image Computing and
A setup with these flexible instruments requires only a single Computer Assisted Intervention (MICCAI). New York: Springer,
2006, vol. 4190, Lecture Notes in Computer Science, pp. 535–542.
calibration routine, which can be done offline and remains [9] C. Doignon, F. Nageotte, and M. de Mathelin, “Segmentation and
valid until the sensors are repositioned. guidance of multiple rigid objects for intra-operative endoscopic
The novel model-based approach improves the error de- vision,” in Workshop on Dynamic Vision, European Conference on
Computer Vi- sion, Graz, Austria, 2006, pp. 314–327.
tection of [20] and [21] so that navigation systems based on [10] J. Ellsmere, J. Stoll, W. Wells, R. Kikinis, K. Vosburgh, R. Kane, D.
magnetic tracking can discover these uncertainties fully auto- Brooks, and D. Rattner, “A new visualization technique for laparo-
matically and reliably, and take them into account for proper scopic ultrasonography,” Surgery, vol. 136, no. 1, pp. 84–92, July
2004.
visualization and feedback during the intervention. [11] J. Krucker, A. Viswanathan, J. Borgert, N. Glossop, Y. Yanga, and B.
Many researchers have worked on the correction of static J. Wood, “An electro-magnetically tracked laparoscopic ultrasound
for multi-modality minimally invasive surgery,” in Comput. Assist.
electromagnetic errors and the detection of dynamic electro- Radiol.
magnetic errors. However, to the authors’ knowledge none of Surg., Berlin, Germany, Jun. 2005, pp. 746–751.
them have addressed the online and real-time reduction of dy- [12] M. Kleemann, P. Hildebrand, M. Birth, and H. P. Bruch,
“Laparoscopic ultrasound navigation in liver surgery: Technical
namic electromagnetic errors for tracking of laparoscopic aspects and accuracy,” Surgical Endoscopy, vol. 20, no. 5, pp. 726–
ultrasound, which does not need an on-site calibration of the 729, May 2006.
tracking volume. [13] J. Harms, H. Feussner, M. Baumgartner, A. Schneider, M.
Donhauser, and G. Wessels, “Three-dimensional navigated
This work is a step towards reliability within the usage of laparoscopic ultrasonography,” Surgical Endoscopy, vol. 15, pp.
magnetic tracking, confidence in the technology, and its 1459–1462, 2001.
integration into the localization and navigation of flexible [14] M. Nakamoto, Y. Sato, M. Miyamoto, Y. Nakamjima, K. Konishi, M.
Shimada, M. Hashizume, and S. Tamura, “3-D ultrasound system
instruments in operating theaters where dynamic magnetic using a magneto-optic hybrid tracker for augmented reality
field distortions cannot be entirely eliminated. visualization in
laparoscopic liver surgery,” in Proc. Int’l Conf. Medical Image Com-
puting and Computer Assisted Intervention (MICCAI), T. Dohi and R.
ACKNOWLEDGMENT Kikinis, Eds. New York: Springer-Verlag, vol. 2489, Lecture Notes
in Computer Science.
The authors would like to thank S. Holdstock for proof- [15] K. Konishi, M. Nakamoto, Y. Kakeji, K. Tanoue, H. Kawanaka,
reading of the paper. They would also like to thank H. S. Yam-aguchi, S. Leiri, Y. Sato, Y. Maehara, S. Tamura, and M.
Feussner and A. Schneider from Klinikum rechts der Isar for Hashizume, “A real-time navigation system for laparoscopic surgery
based on three-dimensional ultrasound using magneto-optic hybrid
their great support in all medical questions and for the tracking configuration,” Int. J. Comput. Assist. Radiol. Surg., vol. 2,
provision of the laparoscope and the ultrasound transducer. no. 1, pp. 1–10, Jun. 2007.
[16] M. Nakamoto, K. Nakada, Y. Sato, K. Konishi, M. Hashizume, and
S. Tamura, “Intraoperative magnetic tracker calibration using a mag-
REFERENCES neto-optic hybrid tracker for 3-d ultrasound-based navigation in
[1] J. J. Jakimowicz, “Intraoperative ultrasonography in open and laparo- laparo- scopic surgery,” IEEE Trans. Med. Imag., vol. 27, no. 2, pp.
scopic abdominal surgery: An overview,” Surgical Endoscopy, vol. 255–270, Feb. 2008.
20, pp. 425–435, Mar. 2006. [17] J. B. Hummel, M. R. Bax, M. L. Figl, Y. Kang, C. Maurer Jr., W.
W. Birkfellner, H. Bergmann, and R. Shahidi, “Design and
application of an assessment protocol for electromagnetic tracking
systems,” Med. Phys., vol. 32, no. 7, pp. 2371–2379, Jul. 2005.
[18] C. Nafis, V. Jensen, L. Beauregard, and P. Anderson, K. R. Cleary and R.
L. Galloway Jr., Eds., “Method for estimating dynamic em tracking
accuracy of surgical navigation tools,” in Med. Imag. 2006: Visual- izat.,
Image-Guided Procedures, Display, Mar. 2006, vol. 6141, pp.
61410K-1–61410K-16.
ULTRASOUND
[19] V. V. Kindratenko, “A survey of electromagnetic position tracker cal- [31] P.-W. Hsu, R. W. Prager, A. H. Gee, and G. M. Treece, “Rapid, easy
ibration techniques,” Virtual Reality: Res. Develop., Appl., vol. 5, no. and reliable calibration for freehand 3-D ultrasound,” Ultrasound
3, pp. 169–182, 2000. Med. Biol., vol. 32, no. 6, pp. 823–835, Jun. 2006.
[20] , I. W. M. Wells, A. C. F. Colchester, and S. L. Delp, Eds., “Con- [32] T. Lang0, “Ultrasound guided surgery: Image processing and naviga-
cepts and results in the development of a hybrid tracking system for tion,” Ph.D. dissertation, Norwegian Univ. Sci. Technol., Trondheim,
CAS,” in Proceedings of the First International Conference of Oct. 2000.
Medical Image Computing and Computer-Assisted Intervention [33] Y. Baillot, S. Julier, D. Brown, and M. Livingston, “A tracker align-
(MICCAI). New York: Springer, vol. 1496, pp. 343–351. ment framework for augmented reality,” in Proc. IEEE ACM Int.
[21] D. Mucha, B. Kosmecki, and J. Bier, “Plausibility check for error Symp. Mixed Augmented Reality (ISMAR), Oct. 2003, pp. 142–150.
compensation in electromagnetic navigation in endoscopic sinus [34] W. Birkfellner, F. Watzinger, F. Wanschitz, R. Ewers, and H.
surgery,” Int. J. Comput. Assist. Radiol. Surg., vol. 1, pp. 316–318, Bergmann, “Calibration of tracking systems in a surgical environ-
Jun. 2006. ment,” IEEE Trans. Med. Imag., vol. 17, no. 5, pp. 737–742, Oct.
[22] T. Sielhorst, M. Feuerstein, and N. Navab, “Advanced medical dis- 1998.
plays: A literature review of augmented reality,” IEEE/OSA J. [35] R. Tsai and R. Lenz, “Real time versatile robotics hand/eye
Display Technol., vol. 4, no. 4, pp. 451–457, Dec. 2008. calibration using 3-D machine vision,” in Proc. IEEE Int. Conf.
[23] M. Feuerstein, T. Reichl, J. Vogel, A. Schneider, H. Feussner, and Robot. Automat.,
N. Navab, “Magneto-optic tracking of a flexible laparoscopic ul- 1988, vol. 1, pp. 554–561.
trasound transducer for laparoscope augmentation,” in Proc. Int’l [36] Y. Shiu and S. Ahmad, “Calibration of wrist-mounted robotic sensors
Conf. Medical Image Computing and Computer Assisted Intervention by solving homogeneous transform equations of the form ,”
(MICCAI). New York: Springer-Verlag, vol. 4791, Proc. Int’l Conf. IEEE Trans. Robot. Autom., vol. 5, no. 1, pp. 16–29, Feb. 1989.
Medical Image Computing and Computer Assisted Intervention [37] K. Daniilidis, “Hand-eye calibration using dual quaternions,” Int. J.
(MICCAI), pp. 458–466. Robot. Res., vol. 18, pp. 286–298, 1999.
[24] M. Feuerstein, T. Reichl, J. Vogel, J. Traub, and N. Navab, “New ap- [38] T. Reichl, “Online error correction for the tracking of laparoscopic ul-
proaches to online estimation of electromagnetic tracking errors for trasound,” M.S. thesis, Technische Universitat , Miinchen, Jul. 2007.
laparoscopic ultrasonography,” Comput. Assist. Surg., vol. 13, no. 5, [39] P. E. Gill, W. Murray, and M. H. Wright, “The Levenberg-Marquardt
pp. method,” in Practical Optimization.. New York: Academic, 1981,
311–323, Sep. 2008. pp.
[25] T. Sielhorst, M. Feuerstein, J. Traub, O. Kutter, and N. Navab, 136–137, ch. 4.7.3.
“Campar: A software framework guaranteeing quality for medical [40] K. Nakada, M. Nakamoto, Y. Sato, K. Konishi, M. Hashizume, and S.
augmented reality,” Int. J. Comput. Assist. Radiol. Surg., vol. 1, pp. Tamura, “A rapid method for magnetic tracker calibration using a
29–30, Jun. 2006. magneto-optic hybrid tracker,” in Proc. Int’l Conf. Medical Image
[26] M. Feuerstein, T. Mussack, S. M. Heining, and N. Navab, “Intraoper- Com- puting and Computer Assisted Intervention (MICCAI), T. M.
ative laparoscope augmentation for port placement and resection Peters, Ed. New York: Springer, 2003, vol. 2879, Lecture Notes in
plan- ning in minimally invasive liver resection,” IEEE Trans. Med. Computer Science, pp. 285–293.
Imag., vol. 27, no. 3, pp. 355–369, Mar. 2008. [41] X. Wu and R. Taylor, “A direction space interpolation technique for
[27] T. Yamaguchi, M. Nakamoto, Y. Sato, K. Konishi, M. Hashizume, calibration of electromagnetic surgical navigation systems,” in Proc.
N. Sugano, H. Yoshikawa, and S. Tamura, “Development of a Int’l Conf. Medical Image Computing and Computer Assisted Inter-
camera model and calibration procedure for oblique-viewing vention (MICCAI), R. E. Ellis and T. M. Peters, Eds. New York:
endoscopes,” Comput. Aid. Surg., vol. 9, no. 5, pp. 203–214, 2004. Springer-Verlag, 2003, vol. 2879, Lecture Notes in Computer
[28] S. De Buck, F. Maes, A. D’Hoore, and P. Suetens, “Evaluation of a Science, pp. 215–222.
novel calibration technique for optically tracked oblique [42] W. J. Youden, “Index for rating diagnostic tests,” Cancer, vol. 3, no.
laparoscopes,” 1, pp. 32–35, 1950.
in Proc. Int’l Conf. Medical Image Computing and Computer [43] G. Welch and G. Bishop, “Course 8-an introduction to the kalman
Assisted filter,” SIGGRAPH 2001 [Online]. Available:
Intervention (MICCAI). New York: Springer-Verlag, vol. 4791, http://www.cs.unc.edu/
Lec- ture Notes in Computer Science, pp. 467–474. ~tracker/media/pdf/SIGGRAPH2001_CoursePack_08.pdf
[29] R. Prager, R. Rohling, A. Gee, and L. Berman, “Rapid calibration for [44] , A. Doucet, N. De Freitas, and N. Gordon, Eds., Sequential Monte
3-d freehand ultrasound,” Ultrasound Med. Biol., vol. 24, no. 6, pp. Carlo Methods in Practice.. New York: Springer,
855–869, 1998. 2001.
[30] G. M. Treece, A. H. Gee, R. W. Prager, C. J. C. Cash, and L. H.
Berman, “High-definition freehand 3-d ultrasound,” Ultrasound Med.
Biol., vol.
29, no. 4, pp. 529–546, 2003.

Magneto-Optical Tracking of Flexible Laparoscopic Ultrasound: Model-Based Online Detection and Correction of Magnetic Tracking Errors

Uploaded by

Copyright:

Available Formats

Magneto-Optical Tracking of Flexible Laparoscopic Ultrasound: Model-Based Online Detection and Correction of Magnetic Tracking Errors

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Magneto-Optical Tracking of Flexible Laparoscopic Ultrasound: Model-Based Online Detection and Correction of Magnetic Tracking Errors

Uploaded by

Copyright:

Available Formats

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 28, NO.

6, JUNE 2009 951

Magneto-Optical Tracking of Flexible Laparoscopic

0278-0062/$25.00 © 2009 IEEE

stationary distortion field errors, so that the calibration process

Fig. 2. Close-up of the proximal part of the ultrasound transducer with a

Fig. 3. All coordinate frames shown need to be brought into a common

images of the membrane inside a water bath are then acquired,

Fig. 4. During magneto-optical coregistration, we compute the unknown

the optical marker and of the magnetic sensor

Fig. 5. Transducer axis calibration. The transducer axis is calibrated both in

motions. Further details can be found in our previous work

For the Levenberg–Marquardt optimizer, we chose a cost func-

where the translation error is the norm of and the rota-

Fig. 6. Typical bending region usually found in endoscopic instruments. Four

cylinder is rotated at least 360 , flipped, and rotated another

Finally, two transducer axes are defined, one in coordinates of

This behavior of alternating links is inspired by the real nature

For error estimation (see Section III-B2), the model is actually

B. Tracking Error Detection and Correction

IV. EXPERIMENTAL EVALUATION

We considered the value with maximum Youden index [42]

where is the fraction of true positives, the fraction of false

Using the transformation from the magnetic

tracking redundancy-based error correction method was avail-

Fig. 16. Performance of the different error correction methods, measured in

• a combination of the model-based and vision-based error

distortion-free environment ([21] applies an additional fixed

The redundancy-based error correction did not gain any im-

You might also like