Augmented Reality Based Navigation for Computer Assisted Hip Resurfacing: A Proof of Concept Study
Augmented Reality Based Navigation for Computer Assisted Hip Resurfacing: A Proof of Concept Study
HE LIU 2 3
EDOUARD AUVINET 1 2
JOSHUA GILES 0 2 3
FERDINANDO RODRIGUEZ Y BAENA 2 3
0 Department of Mechanical Engineering, University of Victoria , Victoria, BC , Canada
1 MSk Lab, Imperial College , London , UK
2 Laboratory, Imperial College , London , UK. Electronic mail:
3 Mechatronics in Medicine Laboratory, Imperial College , London , UK
-Implantation accuracy has a great impact on the outcomes of hip resurfacing such as recovery of hip function. Computer assisted orthopedic surgery has demonstrated clear advantages for the patients, with improved placement accuracy and fewer outliers, but the intrusiveness, cost, and added complexity have limited its widespread adoption. To provide seamless computer assistance with improved immersion and a more natural surgical workflow, we propose an augmented-reality (AR) based navigation system for hip resurfacing. The operative femur is registered by processing depth information from the surgical site with a commercial depth camera. By coupling depth data with robotic assistance, obstacles that may obstruct the femur can be tracked and avoided automatically to reduce the chance of disruption to the surgical workflow. Using the registration result and the pre-operative plan, intra-operative surgical guidance is provided through a commercial AR headset so that the user can perform the operation without additional physical guides. To assess the accuracy of the navigation system, experiments of guide hole drilling were performed on femur phantoms. The position and orientation of the drilled holes were compared with the pre-operative plan, and the mean errors were found to be approximately 2 mm and 2 , results which are in line with commercial computer assisted orthopedic systems today.
Computer assisted orthopedic surgery; Guide
Hip replacement has been proven to be an effective
procedure for end-stage hip joint problems like severe
osteoarthritis. Of all hip replacement procedures, hip
resurfacing can sometimes be a preferable alternative
for young and active patients due to the potential
advantages of improved bone preservation, lower risk
of hip dislocation, easier revision to total hip
replacement,13 and encouraging functional outcomes. Like
total hip replacement, hip resurfacing also contains
operations on the femur and the acetabulum to remove
the damaged bone and cartilage, after which femoral
and acetabular components are placed to reconstruct
the articular surface. In conventional femoral
preparation, a specially designed mechanical jig is used to
identify the axis of the femoral neck, then a guide wire
is inserted along the axis so that subsequent reamers
and cutters can be properly aligned with the femoral
neck in order to achieve ideal component positioning
and fit. Therefore, creating the central guiding axis
through the femoral neck is crucial to the success of
this surgical technique.
As with all joint replacements, implantation
accuracy, which is directly related to the accuracy of the
implantation technique and technology used, is one of
the most important considerations as it significantly
impacts joint functional recovery and longevity of the
replacement.9,28 Studies have shown that varus
malpositioning of the femoral component with a
stemshaft angle £ 130 can lead to higher risk of early
complications such as femoral neck fracture.4
Although the implantation positional accuracy has not
been reported to negatively affect clinical outcomes,
inaccurate positioning will require larger
femoral/ac2018 The Author(s)
etabular components to avoid notching, resulting in
less acetabular bone stock which can harm joint
stability.12 Currently, bone resection and implant
positioning are mainly achieved manually by the surgeon
with the assistance of mechanical jigs; thus, the
accuracy relies heavily on the surgeon’s skills and
experience, which can take significant time to develop.18
To provide more effective assistance that improves
surgeon accuracy, computer navigation can be
introduced into the procedure. Conventional computer
navigation displays information and guidance through
two-dimensional (2D) monitors, which is not intuitive
and may lead to eye–hand coordination problems and
loss of depth perception.1,16 In contrast, augmented
reality (AR) has the ability to overcome these
challenges by overlaying three-dimensional (3D)
information in situ, which has shown potential for
intraoperative navigation.6 In contrast to virtual reality,
which creates a completely virtual environment to the
exclusion of the real world, AR does not block the real
environment, but rather overlays the virtual
information on it so that intuitive guidance is provided while
the original task is not obscured. Therefore, in sensitive
applications like surgery, AR could be a better choice
to provide both safety and efficiency.25,27 In recent
years, the development of AR headsets has made AR
technology available for a wide range of tasks, and the
Microsoft HoloLens (Microsoft, Inc.) is an
outstanding example. As an optical see-through, self-contained
holographic computer, HoloLens achieves
self-tracking and information rendering together without any
other auxiliary equipment, which greatly simplifies its
setup and facilitates its integration into the operating
Apart from using optical see-through headsets, AR
can be implemented in a variety of ways. The most
basic one is monitor-based, where the surgical scene is
displayed on a monitor with the augmenting
information aligned and overlaid on it. This rudimentary
approach still suffers from the eye–hand coordination
problems that affect conventional surgical
navigation.6,25 In order to improve the display of AR in situ,
portable monitors like smart phones and tablets have
more recently been used to capture the scene, augment
it and display it to the user at the same time. However,
due to limitations in processing power and the need to
hold the display during use, their application remains
limited to simple cases.2,21 Video see-through headsets
are, in essence, monitor based AR devices with two
augmented video streams displayed right in front of
the eyes, which enable the inclusion of 3D perception.
However, this advantage is compromised by the safety
issue that once the headset stops working, the users
will completely lose sight of what they are performing,
which can be catastrophic during surgery.24 Other
popular AR approaches include projecting virtual
information directly on the real environment and
overlaying information on a semi-transparent mirror.
Although these approaches preserve the reality, the
imbalance of overlaying 2D information onto a 3D
scene may lead to lower accuracy and awkward
ergonomics.16,25 As an upgrade from video
seethrough headsets and semi-transparent mirrors, optical
see-through headsets such as the HoloLens possess
many advantages from the approaches described
earlier, such as portability, 3D in situ rendering, and
reality preservation.26,29 Admittedly, problems still
exist. Accuracy and latency issues are limited by the
computing and sensing abilities, occlusion is almost
inevitable, and the rendering techniques are still basic,
resulting in unnatural interactions with the virtual
objects. However, with the expected improvements in
sensor technology, registration algorithms, and
medical image processing arising from improved
manufacturing methods, cloud computing and machine
learning, optical see-through headsets have the
potential to eventually become ubiquitous, in medicine and
Consequently, the aim of our study is to explore the
possibility of applying AR based 3D computer
navigation to hip resurfacing, and more specifically, to
femoral preparation. Since most of the instruments for
femoral preparation are based around the premise of
creating a central guiding hole along the axis of the
femoral neck, in this paper, our goal is to simplify the
procedure of locating and drilling such a hole by
means of optical see-through AR assistance and an
automatic registration and limb tracking system. We
employ in vitro testing with a small user group to assess
both the accuracy and repeatability of guide hole
drilling with our setup.
MATERIALS AND METHODS
In this paper, we developed a navigation system
based on depth sensing technology and the HoloLens.
The AR guidance was generated according to a
preoperative plan, which was shown to the user through
the HoloLens. In order to accurately display the
guidance information in the HoloLens display, our
previously developed robotic registration system14 was
used to measure the pose of the target femur.
Previous Camera–Robot Registration System
At the basis of any navigation system is the need to
correctly register the geometry/s of interest, as this
process is integral to locating and tracking objects
within the surgical scene. In this paper, the registration
object is the femur. Instead of using optical markers,
our registration system is built on a depth camera with
the assistance of robotic technology, as shown in
Fig. 1, with technical details about the complete setup
available in our previous open access publication.14
The depth camera (Xtion Pro Live, Asustek Computer,
Inc., Taipei, Taiwan, China), which has the ability to
capture the environment’s geometry in the form of a
3D point cloud, is mounted on, and calibrated to, the
end-effector of a serial manipulator (LBR iiwa,
KUKA Aktiengesellschaft, Augsburg, Bavaria,
Germany). The point cloud of the environment is
processed to identify descriptors of the geometrical
features, which are then compared with descriptors of
the pre-scanned femur model to obtain a rough
estimate of the femur pose. This gross registration step is
then refined by a standard implementation of the
iterative closest point (ICP) algorithm5 so that the
femur pose can be ascertained with accuracy, to enable
surgical navigation. Once the first ICP succeeds, the
registration result of this iteration will be used as the
initial estimate for the subsequent one. This process
runs continuously at approximately 10 Hz in order to
track the target femur dynamically during the
procedure. After the target is located, a conical region is
defined based on a line connecting the target and the
camera, and objects that are detected by the depth
camera to be located inside the cone are tracked as
obstacles. The robot will move the camera to avoid
occlusion while keeping the target at the center of the
camera view. Meanwhile, admittance control is applied
to the robot so that direct manipulation by human
operators is allowed, thus making the system more
user-friendly and intuitive.
While clinical deployment of robotic-based obstacle
avoidance would require further consideration
(autonomous robotic motion would have to negotiate the
crowded operating theater space), in principle it would
address the line-of-sight problem that exists in optical
tracking and reduce the possibility of tracking failure.
These advantages would give more freedom to the
surgeon and alleviate the mental burden caused by the
continual need to adjust traditional optical navigation
systems to avoid occlusion, if an appropriate and safe
implementation of the system could be provided.
Consequently, our robot-based registration and
tracking system was employed in this work.
AR System Setup and Workflow Control
The AR system includes four parts: the depth
camera, the robot, the HoloLens, and a tracked
surgical drill. Each of these has its own coordinate system,
and appropriately establishing a correspondence
between them is the key to accurate AR performance.
The calibration between the depth camera and the
robot end effector has been described in our previous
study.14 After registration, the target femur geometry
is described in the depth camera coordinate frame. In
order for the HoloLens to use this data, the
transformation from the depth camera frame to the HoloLens
frame must be found. To do this, we use an image
marker (marker1) which can be recognized by the RGB
cameras on both the depth camera and the HoloLens.
Its pose measured by both cameras can be denoted by
two 4 9 4 time-varying matrices, Tcmaamrker1ðtÞ and
Thmoalroker1ðtÞ: Thus, the correspondence between the two
camera frames is obtained, and the target pose can be
transformed into the HoloLens frame by
TthaorlgoetðtÞ ¼ Thmoalroker1ðtÞ
is computed from Tthaorlgoetð0Þ] is recorded as a
proprietary HoloLens ‘‘spatial anchor’’ [i.e., anchor =
target(0)], such that subsequent target pose updates are
computed relative to this anchor [i.e., TtaanrcgheotrðtÞ]. Since
the robot base frame and the anchor frame are both
static relative to the world frame, obviously, we have
TtaanrcgheotrðtÞ ¼ TcbaamseðtÞ
where Tcbaamseð0Þ and Tcaanmchorð0Þ are constant matrices
recorded at t = 0, and TcbaamseðtÞ is calculated from the
forward kinematics of the robot.
From Eq. (
), we can have the real-time update of
the target pose
We can see from Eq. (
) that the rendered target is
updated continuously without the need for additional
markers, as long as the target is tracked by the depth
camera and the robot kinematics is computed.
Another image marker, marker2, is used and fixed
to the robot base in order to align the HoloLens with
the robot before registration is carried out, and the
transformation from marker2 to the robot base,
Tbmaasreker2; is constant and known by the time the system
is set up. Once marker2 is recognized by the HoloLens,
the correspondence between the depth camera and the
HoloLens is established as follows
TchaomloðtÞ ¼ TcbaamseðtÞ
where TchaomloðtÞ is the transformation matrix from the
HoloLens to the depth camera, TcbaamseðtÞ is the inverse of
TcbaamseðtÞ; and Thmoalroker2ðtÞ is the pose of marker2 in the
As for the surgical drill, a cube shaped marker that
can be tracked by the HoloLens is fixed onto it. After
calibration between the drill and the marker, the
coordinates of the tip and the axis of the drill are
known in the cube marker frame (i.e., Tcduriblle is constant
and known). These coordinates will also be
transformed into the HoloLens frame by
ThdroillolðtÞ ¼ TchuobloeðtÞ
where TchuobloeðtÞ is the cube marker pose tracked by the
HoloLens. The drill will then be compared with the
transformed pose of the target femur. If, after
calculation, the drill is properly aligned with the planned
guide hole, the HoloLens will inform the user of the
results so that the user can proceed with drilling. All
relationships between each of the coordinate systems
composing the whole system are shown in Fig. 2.
To minimize the learning burden on the user,
interaction with the system is achieved through
intuitive methods such as voice commands, gestures, and
manual manipulation. These human inputs are
preted automatically by the HoloLens or the robot so
that no additional control devices are necessary. The
AR integrated workflow is divided into several stages
by simple voice commands, with different tasks
performed in each stage:
After the system is started, the robot is force
compliant so that the surgeon can set its
configuration by hand to make the target
roughly in view of the depth camera.
The surgeon wearing the HoloLens defines the
HoloLens and robot correspondence by looking
at marker2 to obtain the matrix Thmoalroker2ðtÞ:
With the help of HoloLens gaze and gesture
features, a spherical ROI that contains the target
femur is selected by the surgeon with respect to
the HoloLens frame, then the ROI position is
transformed into the depth camera frame using
) and sent to the registration controller.
This ROI guides where the depth camera should
look, and the point cloud outside the ROI will
not be processed, thus increasing registration
The point cloud inside the ROI is processed to
register the target femur in the depth camera
frame and estimate the target pose TtcaarmgetðtÞ:
Marker1 is provided for the depth camera and
HoloLens to acquire Tcmaamrker1ðtÞ and Thmoalroker1ðtÞ
so that the target pose is available for HoloLens
using Eq. (
) and the 3D navigation can be
rendered, then updated by Eq. (
The surgeon holds the surgical drill and registers
it to the HoloLens, then follows the navigation
cues to perform the planned drilling.
The complete workflow for drilling of the guide hole
into the femur with AR navigation is shown in Fig. 3.
Experiments were performed to test the accuracy of
the proposed AR navigation system. In this paper, we
apply the system to the hip resurfacing procedural step
of guide hole drilling. Instead of requiring a
mechanical alignment jig to locate the axis of the femoral neck,
AR navigation provides visual guidance overlaid onto
the surgeon’s view, where the approach vector for the
guide hole is rendered in situ, such that drilling can be
performed free-hand and without having to look away
from the patient. Three groups of experiments were
carried out under different conditions, each consisting
of 30 trials. The first group was performed by the
author to test the repeatability of the system. The second
group was also performed by the author, but
disturbances were introduced during the procedure (e.g.,
obstacles appeared between the camera and the target
femur, leading to obstacle avoidance motion of the
robot). In the third group, five students without
medical background volunteered to perform six trials each,
and before the experiments they were trained to use the
whole system for about an hour. The actual setup for
the experiment of guide hole drilling is shown in Fig. 4.
The experiments were performed on foam femur
phantoms (SKU 1129, Sawbones, Pacific Research
Laboratories, Vashon, Washington, US), the surface
geometry of which was 3D scanned beforehand to
obtain the anatomical model used for the pre-operative
plan and the model-based registration. After the target
was registered in the HoloLens, a virtual red arrow was
shown to the user, indicating the planned entry point
and the drilling direction. Once the drill was also
registered in the HoloLens frame, its position and
orientation were compared with the arrow, and the arrow
head and shaft turning green indicated that the
position difference and the orientation difference were
within the thresholds, respectively (in the experiments,
the position error threshold was set to 3 mm and the
orientation threshold was 2.5 ). When the whole arrow
was green, the user could proceed with drilling. The
HoloLens guidance feedback is shown in Fig. 5. It was
ensured that the arrow was kept green during drilling.
Ten fiducial points with known positions on the
scanned femur model were used as the gold standard
measurements for comparison. After drilling, the
position and direction of the guide hole were measured
together with the positions of the 10 fiducial points
using a MicroScribe (G2X, Solution Technologies,
Inc., Oella, Maryland, US), then a closed form, paired
point registration method11 was used to calculate the
transformation from the measured fiducial positions to
the fiducial positions on the femur model. The
measured position and direction of the guide hole could
then be transformed to the femur model frame so that
they could be compared with the pre-operative plan to
obtain the errors. Time spent on each procedure was
also recorded during the experiments. An analysis of
variance (ANOVA) was adopted to compare the
results of different volunteers, and Student’s t-tests were
performed to compare results between groups.
An external PC running Ubuntu 16.04 was used to
process the images from the depth camera, and to
control the robot through KUKA’s Fast Robot
Interface (KUKA Roboter GmbH, Augsburg,
Bavaria, Germany). Image processing was based on the
Point Cloud Library23 and the AprilTags C++
Library,20 while the Eigen Library10 was used to
facilitate coding of the robotic controller. Unity (Unity
Technologies SF, San Francisco, California, US),
Vuforia (PTC, Inc., Needham, Massachusetts, US)
and HoloToolkit (GitHub, Inc., San Francisco,
California, US) were used to develop the program for the
HoloLens. The communication between the HoloLens
and the external PC was via the user datagram packet
over internet protocol. Result processing was done in
MATLAB (R2016a, MathWorks, Inc., Natick,
Massachusetts, US), and the Statistical Package for the
Social Sciences (SPSS, version 24, IBM, Armonk, New
York, US) was used to conduct the statistical analysis.
The time cost of the whole procedure, including
loading the program onto the HoloLens and the robot
controller, selecting the ROI, registering the target
femur with the depth camera, rendering AR navigation
via the HoloLens, and drilling the guide hole with AR
navigation, was recorded as a reference. The average
time was about 2 min (115 s) for the author and 4 min
(234 s) for other users. With more practice, all the
users were able to finish the procedure within 3 min.
Position of the entry point on the femur phantom
and direction of the guide hole were measured and
transformed into the femur model frame to compare
with the pre-operative plan. The angular errors were
measured in 3D as absolute values, then represented in
clinically relevant inclination and version errors. The
mean values and standard deviations of the errors of
each group are shown in Table 1. Figure 6 displays the
variations of the absolute errors, and Fig. 7 displays
the inclination errors and the version errors. Scatter
plots of the entry position errors and direction errors
of the measured guide holes are shown in Fig. 8.
Statistical analysis was conducted on the absolute
errors of the three groups. One-way between subjects
ANOVAs were used to compare the position and
direction errors of the five volunteers in the Volunteers
group, and no statistically significant differences were
found in either position errors (p = 0.370) or direction
errors (p = 0.900). The t-tests showed no significant
The errors are represented as mean ± standard deviation.
differences between the Author group and the
Volunteers group (position error: p = 0.493, direction error:
p = 0.188), indicating good usability of the AR
navigation system (i.e., with some practice, an
inexperienced user can achieve similarly accurate results). The
differences between the Author group and the
Author + Obstacle group were also insignificant
(position error: p = 0.200, direction error: p = 0.146),
demonstrating appropriate dynamic tracking of the
limb during the navigation of guide hole drilling.
In this study we tested the accuracy of drilling a
guide hole along the axis of the femoral neck for hip
resurfacing with our AR based navigation system.
Under different conditions, the position and direction
errors from the pre-operative plan were around 2 mm
and 2 , results which are promising. Although we did
not perform another set of experiments with other
techniques to compare the outcomes, several studies
have been done on the accuracy of guide wire insertion
with conventional jigs, patient specific instrumentation
(PSI) and navigation,3,8,12,19 and their results are
briefly listed in Table 2. Since the measurements and
experimental conditions are not the same, results
cannot be compared directly. However, the mean
errors of our experiments are comparable to those in
these studies, and our standard deviations are generally
smaller. According to the quantitative score table for
guide concepts proposed by Audenaert et al.,3 the
achieved accuracy is scored as ‘acceptable’ if the
position error £ 4 mm and the direction error £ 4 ,
and as ‘good’ if the position error £ 2 mm and the
direction error £ 2 . Therefore, the accuracy of this
AR navigation system under the somewhat artificial
test conditions described here is within the ‘acceptable’
range, and very close to the ‘good’ range.
In addition to its accuracy and precision, the AR
navigation system also shows promising advantages in
terms of surgical workflow. With some simple voice
commands and gestures, the user could easily interact
with the HoloLens, through which the depth camera
and the robot are controlled. In this way, a
user-centered workflow is achieved, where all other equipment
will serve the user rather than the user needing to
operate the equipment. The registration is automatic
and markerless, only requiring a model of the bone for
pre-operative planning purposes, as is customary in
image-based computer assisted orthopedic systems.
Importantly, it is not necessary to insert bone markers
or to manually select bone features to register the
markers, which has the potential to reduce
intra-opResults are in the form of mean ± standard deviation.
*Positional errors are measured in superior, medial and anterior directions, angular errors are measured for inclination and anteversion.
Positional errors are measured in posterior and inferior directions; angular errors are measured for inclination and anteversion.
Positional and angular errors are measured in 3D.
erative time. Finally, this AR navigation technique
also possesses usability advantages over conventional
navigation on computer monitors by allowing intuitive
visual guidance that does not distract from the surgical
field. The system’s somewhat higher complexity in
comparison to traditional navigation is warranted in
this proof-of-concept setup, as it is shown to have the
potential to provide a more informative surgical
environment, which may eventually contribute to higher
efficiency and better surgical accuracy. Significant
improvements in the commercial offering for AR
headsets and low-cost surgical robotic assistants are
also expected to alleviate this issue in the future.
Current limitations of the system, as given in the
feedback from the volunteers after the experiments,
can be grouped into the following aspects:
Depth perception of the directional guidance.
Although the guidance arrow is rendered in 3D,
it cannot be completely opaque to occlude objects
behind it. And objects in front of the arrow cannot
occlude it, either, as the arrow is projected to the
eyes directly. This may cause difficulties in
perceiving the depth information of the virtual object, so it
takes some practice to align the real drill axis with
the virtual arrow.
Deviations of the HoloLens display for the user.
HoloLens does not track the user’s eye positions
(e.g., eye width), so for different users the position
of the hologram may appear in a slightly different
location within the immediate visual scene.
Although this does not affect the actual position
of the hologram, which is pinned to the scene in a
pre-operatively defined pose, this projective
distortion can be misleading for the user during
The increased complexity of the system. Although
the AR navigation system may help simplify the
surgical workflow, the whole procedure may
become technically more complex because of the
use of different technologies such as robotics and
an AR headset. Other techniques like voice
commands and gaze control may also cause
interruptions or even unexpected responses of the system
when the commands are misinterpreted. This
added complexity and the robustness issues could
be address with further development (e.g., with a
bespoke robotic effector, a ceiling mounted option,
a multiple-camera setup, and less intrusive AR
glasses), but represent significant limitations of this
Stability of free-hand drilling. Although none of
our volunteers were medically trained, all found it
difficult to maintain good stability during
freehand drilling [inclination errors are more varus for
all the three groups (p < 0.05)], an issue which
might require further investigation.
The biggest obstacle to eventual clinical deployment
is registration in a realistic scenario. Specifically, the
challenge will be to identify suitable image processing
algorithms to segment the target femur from the
surgical scene in order to correctly register it. In our study,
the conditions are artificially simple, with a femur
phantom isolated from the leg, no soft tissue or blood,
and no drapes to contend with. Conventional
featurebased segmentation may encounter significant
difficulties in such a complex environment, but we believe
that the exponential growth in machine learning
techniques and technologies in recent years may eventually
come to the rescue. We are currently exploring this line
of research via the application of fully convolutional
networks,15 which have shown promising results in 2D
and 3D medical image registration,7,17,22 to our AR
He Liu was supported by China Scholarship
This article is distributed under the terms of the
Creative Commons Attribution 4.0 International
which permits unrestricted use, distribution, and
reproduction in any medium, provided you give
appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons license, and
indicate if changes were made.
1Abe, Y. , S. Sato , K. Kato , T. Hyakumachi , Y. Yanagibashi , M. Ito , and K. Abumi . A novel 3D guidance system using augmented reality for percutaneous vertebroplasty .
J. Neurosurg . Spine 19 : 492 - 501 , 2013 .
2Arth, C. , and D. Schmalstieg . Challenges of Large-Scale Augmented Reality on Smartphones . Graz: Graz University of Technology, pp. 1 - 4 , 2011 .
3Audenaert , E. , K. De Smedt , F. Gelaude , T. Clijmans , C.
4Beaule´ , P. E. , J. L. Lee , M. J. Le Duff , H. C. Amstutz , and E. Ebramzadeh . Orientation of the femoral component in surface arthroplasty of the hip: a biomechanical and clinical analysis . JBJS 86 : 2015 - 2021 , 2004 .
5Besl, P. J. , and N. D. McKay . A method for registration of 3-D shapes . IEEE Trans. Pattern Anal. Mach . Intell.
6Blackwell, M. , F. Morgan , and A. M. DiGioia , III. Augmented reality and its future in orthopaedics . Clin. Orthop.
Relat . Res. 354 : 111 - 122 , 1998 .
7C¸ i c¸ek , O¨ ., A. Abdulkadir , S. S. Lienkamp , T. Brox , and O.
Ronneberger. 3D U-Net: learning dense volumetric segmentation from sparse annotation . In: International Conference on Medical Image Computing and ComputerAssisted Intervention . Berlin: Springer, 2016 , pp. 424 - 432 .
8Cobb, J. P. , V. Kannan , W. Dandachli , F. Iranpour , K. U.
Brust , and A. J. Hart . Learning how to resurface cam-type femoral heads with acceptable accuracy and precision: the role of computed tomography-based navigation . JBJS 90 : 57 - 64 , 2008 .
9De Haan, R. , P. Campbell , E. Su, and K. A. De Smet .
Revision of metal-on-metal resurfacing arthroplasty of the hip . Bone Jt J . 90 : 1158 - 1163 , 2008 .
10Guennebaud, G. , B. Jacob , P. Avery , A. Bachrach , and S. Barthelemy . Eigen v3 . 2010 . http://eigen.tuxfamily.org.
11Horn , B. K. Closed-form solution of absolute orientation using unit quaternions . JOSA A 4 : 629 - 642 , 1987 .
12Kitada, M. , T. Sakai , T. Murase , T. Hanada , N. Nakamura , and N. Sugano . Validation of the femoral component placement during hip resurfacing: a comparison between the conventional jig, patient-specific template, and CT-based navigation . Int. J. Med. Robot. Comput. Assist. Surg . 9 : 223 - 229 , 2013 .
13Koutras, C. , S. A. Antoniou , M. A. Talias , and H. Heep . Impact of total hip resurfacing arthroplasty on health-related quality of life measures: a systematic review and metaanalysis . J. Arthroplast . 30 : 1938 - 1952 , 2015 .
14Liu, H., S. Bowyer , E. Auvinet, and F. R. Baena . A smart registration assistant for joint replacement: a concept demonstration . Bone Jt J . 99 : 58 , 2017 .
15Long J. , E. Shelhamer , and T. Darrell . Fully convolutional networks for semantic segmentation . In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2015 , pp. 3431 - 3440 .
16Marescaux, J. , M. Diana , and L. Soler . Augmented reality and minimally invasive surgery . J. Gastroenterol. Hepatol. Res . 2 : 555 - 560 , 2013 .
17Milletari , F. , N. Navab , and S.-A. Ahmadi . V-net: fully convolutional neural networks for volumetric medical image segmentation . In: 2016 Fourth International Conference on 3D Vision (3DV) . IEEE, 2016 , pp. 565 - 571 .
18Nunley, R. M. , J. Zhu , P. J. Brooks , C. A. Engh , S. J. Raterman , J. S. Rogerson , and R. L. Barrack . The learning curve for adopting hip resurfacing among hip specialists . Clin. Orthop. Relat. Res . 468 : 382 - 391 , 2010 .
19Olsen, M. , M. Chiu , P. Gamble , R. A. Boyle , N. Tumia , and E. H. Schemitsch . A comparison of conventional guidewire alignment jigs with imageless computer navigation in hip resurfacing arthroplasty . JBJS 92 : 1834 - 1841 , 2010 .
20Olson , E. AprilTag: a robust and flexible visual fiducial system . In: 2011 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2011 , pp. 3400 - 3407 .
21Papagiannakis, G. , G. Singh , and N. Magnenat-Thalmann . A survey of mobile and wireless technologies for augmented reality systems . Comput. Animat. Virtual Worlds 19 : 3 - 22 , 2008 .
22Ronneberger , O. , P. Fischer , and T. Brox . U-net: convolutional networks for biomedical image segmentation . In: International Conference on Medical Image Computing and Computer-Assisted Intervention . Berlin: Springer, 2015 , pp. 234 - 241 .
23Rusu, R. B. and S. Cousins . 3D is here: point cloud library (PCL) . In: 2011 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2011 , pp. 1 - 4 .
24Shuhaiber, J. H. Augmented reality in surgery . Arch. Surg . 139 : 170 - 174 , 2004 .
25Sielhorst, T. , M. Feuerstein , and N. Navab . Advanced medical displays: a literature review of augmented reality . J. Disp. Technol . 4 : 451 - 467 , 2008 .
26Van Krevelen , D. , and R. Poelman . A survey of augmented reality technologies, applications and limitations . Int. J. Virtual Real . 9 : 1 , 2010 .
27Va´ vra, P., J. Roman, P. Zon cˇa, P. Ihn a´t, M. Neˇ mec , J. Kumar, N. Habib , and A . El-Gendi . Recent development of augmented reality in surgery: a review . J. Healthc. Eng . 2017 . https://doi.org/10.1155/ 2017 /4574172.
28Williams , D. , M. Royle , and M. Norton . Metal-on-metal hip resurfacing: the effect of cup position and component size on range of motion to impingement . J. Arthroplast . 24 : 144 - 151 , 2009 .
29Zhou , F. , H. B.-L. Duh , and M. Billinghurst . Trends in augmented reality tracking, interaction and display: a review of ten years of ISMAR . In: Proceedings of the 7th IEEE/ ACM International Symposium on Mixed and Augmented Reality. IEEE Computer Society , 2008 , pp. 193 - 202 .