Professional Documents
Culture Documents
Textbook Intelligent Orthopaedics Artificial Intelligence and Smart Image Guided Technology For Orthopaedics Guoyan Zheng Ebook All Chapter PDF
Textbook Intelligent Orthopaedics Artificial Intelligence and Smart Image Guided Technology For Orthopaedics Guoyan Zheng Ebook All Chapter PDF
https://textbookfull.com/product/general-orthopaedics-and-basic-
science-nikolaos-k-paschos/
https://textbookfull.com/product/operative-orthopaedics-2nd-
edition-timothy-wr-briggs/
https://textbookfull.com/product/proprioception-in-orthopaedics-
sports-medicine-and-rehabilitation-defne-kaya/
https://textbookfull.com/product/millers-review-of-
orthopaedics-7th-edition-mark-miller/
Bio orthopaedics A New Approach 1st Edition Alberto
Gobbi
https://textbookfull.com/product/bio-orthopaedics-a-new-
approach-1st-edition-alberto-gobbi/
https://textbookfull.com/product/playing-smart-on-games-
intelligence-and-artificial-intelligence-1st-edition-julian-
togelius/
https://textbookfull.com/product/artificial-intelligence-and-
algorithms-in-intelligent-systems-radek-silhavy/
https://textbookfull.com/product/postgraduate-orthopaedics-viva-
guide-for-the-frcs-tr-orth-examination-2nd-edition-paul-a-
banaszkiewicz-editor/
https://textbookfull.com/product/tachdjians-pediatric-
orthopaedics-from-the-texas-scottish-rite-hospital-for-
children-6th-edition-john-a-herring/
Advances in Experimental Medicine and Biology 1093
Intelligent
Orthopaedics
Artificial Intelligence and Smart Image-
guided Technology for Orthopaedics
Advances in Experimental Medicine
and Biology
Volume 1093
Editorial Board
IRUN R. COHEN, The Weizmann Institute of Science, Rehovot, Israel
ABEL LAJTHA, N.S. Kline Institute for Psychiatric Research, Orangeburg,
NY, USA
JOHN D. LAMBRIS, University of Pennsylvania, Philadelphia, PA, USA
RODOLFO PAOLETTI, University of Milan, Milan, Italy
NIMA REZAEI, Tehran University of Medical Sciences, Children’s Medical
Center Hospital, Tehran, Iran
More information about this series at http://www.springer.com/series/5584
Guoyan Zheng • Wei Tian • Xiahai Zhuang
Editors
Intelligent Orthopaedics
Artificial Intelligence and Smart
Image-guided Technology for
Orthopaedics
123
Editor
Guoyan Zheng Wei Tian
University of Bern Beijing Jishuitan Hospital
Bern, Switzerland Peking University
Beijing, Beijing, China
Xiahai Zhuang
Fudan University
Shanghai, China
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore
189721, Singapore
Contents
v
vi Contents
Abstract Keywords
Introduced more than two decades ago, Computer-aided orthopaedic surgery
computer-aided orthopaedic surgery (CAOS) (CAOS) · Smart instrumentation · Medical
has emerged as a new and independent robotics · Artificial intelligence · Machine
area, due to the importance of treatment learning · Deep learning · Big data analytics ·
of musculoskeletal diseases in orthopaedics Intelligent orthopaedics
and traumatology, increasing availability of
different imaging modalities and advances in
analytics and navigation tools. The aim of
this chapter is to present the basic elements
of CAOS devices and to review state-of-the- 1.1 Introduction
art examples of different imaging modalities
used to create the virtual representations, The human musculoskeletal system is an organ
of different position tracking devices for system that includes the bones of the skeleton and
navigation systems, of different surgical the cartilages, ligaments, and other connective
robots, of different methods for registration tissues that bind tissues and organs together. The
and referencing, and of CAOS modules that main functions of this system are to provide form,
have been realized for different surgical pro- support, stability, and movement to the body.
cedures. Future perspectives will be outlined. Bones, besides supporting the weight of the body,
It is expected that the recent advancement work together with muscles to maintain body
on smart instrumentation, medical robotics, position and to produce controlled, precise move-
artificial intelligence, machine learning, and ments. Musculoskeletal disease is among the
deep learning techniques, in combination with most common causes of severe long-term disabil-
big data analytics, may lead to smart CAOS ity and practical pain in industrialized societies
systems and intelligent orthopaedics in the [1]. The impact and importance of musculoskele-
near future. tal diseases are critical not only for individual
health and mobility but also for social function-
G. Zheng () · L.-P. Nolte ing and productivity and economic growth on a
Institute for Surgical Technology and Biomechanics, larger scale, reflected by the proclamation of the
University of Bern, Bern, Switzerland Bone and Joint Decade 2000–2010 [1].
e-mail: guoyan.zheng@istb.unibe.ch
Both traumatology and orthopaedic surgery breviation, which is the target of the treatment),
aim at the treatment of musculoskeletal tissues. a virtual object (VO in abbreviation, which is
Surgical steps such as the placement of an im- the virtual representation in the planning and
plant component, the reduction and fixation of navigation computer), and a so-called navigator
a fracture, ligament reconstruction, osteotomy, that links both objects. For reasons of simplicity,
tumour resection, and the cutting or drilling of the term “CAOS system” will be used within this
bone should ideally be carried out as precisely as article to refer to both navigation systems and
possible. Not only will optimal precision improve robotic devices.
the post-operative outcome of the treatment, but The central element of each CAOS system is
it will also minimize the risk factors for intra- the navigator. It is a device that establishes a
and post-operative complications. To this end, global, three-dimensional (3-D) coordinate sys-
a large number of pure mechanical guides have tem (COS) in which the target is to be treated
been developed for various clinical applications. and the current location and orientation of the
The pure mechanical guides, though easy to use utilized end effectors (EE) are mathematically
and easy to handle, do not respect the individual described. End effectors are usually passive sur-
patient’s morphology. Thus, their general ben- gical instruments but can also be semi-active or
efit has been questioned (see for example [2]). active devices. One of the main functions of
Additionally, surgeons often encounter the chal- the navigator is to enable the transmission of
lenge of limited visibility of the surgical situs, positional information between the end effectors,
which makes it difficult to achieve the intended the TO and the VO. For robotic devices, the robot
procedure as accurately as desired. Moreover, itself plays the role of the navigator, while for
the recent trend towards increased minimally surgical navigation a position tracking device is
invasive surgery makes it more and more im- used.
portant to gain feedback about surgical actions For the purpose of establishment of a CAOS
that take place subcutaneously. Just as a Global system through coactions of these three entities,
Positioning System (GPS)-based car navigation three key procedural requirements have to be
provides visual instruction to a driver by display- fulfilled. The first is the calibration of the end
ing the location of the car on a map, a computer- effectors, which means to describe the end ef-
aided orthopaedic surgery (CAOS) module al- fectors’ geometry and shape in the coordinate
lows the surgeon to get real-time feedback about system of the navigator. For this purpose, it is
the performed surgical actions using information required to establish physically a local coordinate
conveyed through a virtual scene of the situs system at the end effectors. When an optical
presented on a display device [3, 4]. Parallel to tracker is used, this is done via rigid attach-
the CAOS module to potentially improve surgical ment of three or more optical markers onto each
outcome is the employment of surgical robots end effector. The second is registration, which
that actively or semi-actively participate in the aims to provide a geometrical transformation
surgery [5]. between the TO and the VO in order to display
Introduced more than two decades ago [3–5], the end effect’s localization with respect to the
CAOS has emerged as a new and independent virtual representation, just like the display of
area and stands for approaches that use computer- the location of a car in a map in a GPS-based
enabled tracking systems or robotic devices to navigation system. The geometrical transforma-
improve visibility to the surgical field and in- tion could be rigid or non-rigid. In literature,
crease application accuracy in a variety of sur- a wide variety of registration concepts and as-
gical procedures. Although CAOS modules use sociated algorithms exist (see the next section
numerous technical methods to realize individual for more details). The third key ingredient to a
aspects of a procedure, their basic conceptual CAOS system is referencing, which is necessary
design is very similar. They all involve three ma- to compensate for possible motion of the navi-
jor components: a therapeutic object (TO in ab- gator and/or the TO during the surgical actions
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 3
to be controlled. This is done by either attach- image data. Below detailed examples of different
ing a so-called dynamic reference bases (DRB) forms of VOs will be reviewed.
holding three or more optical markers to the When the VO is derived from medical image
TO or immobilizing the TO with respect to the data, these data may be acquired at two points in
navigator. time: either pre-operatively or intra-operatively.
The rest of the chapter is organized as follows. Two decades ago, the VOs of majority CAOS
Section 1.2 will review the state-of-the-art exam- systems were derived from pre-operatively ac-
ples of basic elements of CAOS systems. Section quired CT scans, and a few groups also tried to
1.3 will present clinical fields of applications. In use magnetic resonance imaging (MRI) [6, 7]. In
Sect. 1.4, future perspectives will be outlined, comparison with MRI, CT has clear advantages
followed by conclusion in Sect. 1.5. of excellent bone-soft tissue contrast and no ge-
ometrical distortion despite its acquisition induc-
ing radiation exposure to the patient. Soon after
1.2 Basic Elements of CAOS the introduction of the first CAOS systems, the
Systems limitations of pre-operative VOs were observed,
which led to the introduction of intra-operative
1.2.1 Virtual Object imaging modalities. More specifically, the bony
morphology may have changed between the time
The VO in each CAOS system is defined as a of image acquisition and the actual surgical pro-
sufficiently realistic representation of the mus- cedure. As a consequence, the VO may not nec-
culoskeletal structures that allows the surgeon to essarily correspond to the TO any more leading
plan the intended intervention, as exemplified in to unpredictable inaccuracies during navigation
Fig. 1.1a Intra-operatively, it also serves as the or robotic procedures. This effect can be particu-
“background” into which the measured position larly adverse for traumatology in the presence of
of a surgical instrument can be visualized (see unstable fractures. To overcome this problem in
Fig. 1.1b for an example). Though most of the the field of surgical navigation, the use of intra-
time VO is derived from image data of the pa- operative CT scanning has been proposed [8], but
tient, it can also be created directly from intra- the infrastructural changes that are required for
operative digitization without using anymedical the realization of this approach are tremendous,
Fig. 1.1 Example of CT-based navigational feedback. (b) of pedicle screw placement. (Courtesy of Brainlab AG,
These screenshots show a CT-based CAOS system during Munich, Germany)
pre-operative planning (a) and intra-operative navigation
4 G. Zheng and L.-P. Nolte
Fig. 1.2 Example of fluoroscopy-based navigation. This screenshot shows the fluoroscopy-based navigation for distal
locking of an intramedullary nail. (Courtesy of Brainlab AG, Munich, Germany)
often requiring considerable reconstruction of a mon coordinate system established on the target
hospital’s facilities. This has motivated the de- structure via the DRB technique. Such a system
velopment of navigation systems based on fluo- can thus provide visual feedback just like the use
roscopic images [9–11]. The image intensifier is of multiple fluoroscopes placed at different posi-
a well-established device during orthopaedic and tions in constant mode but without the associated
trauma procedures but has the limitations that the radiation exposure, which is a clear advantage
images generated with a fluoroscope are usually (see Fig. 1.2 for an example). This technique
distorted and that one-dimensional information is therefore also known as “virtual fluoroscopy”
gets lost due to image projection. To use these [11]. Despite the fact that in such a system, only
images as VOs therefore requires the calibration two-dimensional (2-D) projected images with
of the fluoroscope which aims to compute the im- low contrast are available, the advantages offered
age projection model and to compensate for the by a fluoroscopy-based navigation system pre-
image distortion [9–11]. The resultant systems ponderate for a number of clinical applications
are therefore known as “fluoroscopy-based nav- in orthopaedics and traumatology.
igation systems” in literature [9–11]. Additional In order to address the 2-D projection limi-
feature offered by a fluoroscopy-based navigation tation of a fluoroscopy-based navigation system,
system is that multiple images acquired from a new imaging device was introduced [12] that
different positions are co-registered to a com- enables the intra-operative generation of 3-D flu-
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 5
Fig. 1.3 Navigation using surgeon-defined anatomy ap- Although a very abstract representation, it provides suf-
proach. This virtual model of a patient’s knee is gen- ficient information to enable navigated high tibial os-
erated intra-operatively by digitizing relevant structures. teotomy
oroscopic image data. It consists of a motor- used to acquire a graphical representation of the
ized, isocentric C-arm that acquires series of 50– patient’s anatomy by intra-operative digitization.
100 2-D projections and reconstructs from them By sliding the tip of a tracked instrument on the
13 × 13 × 13 cm3 volumetric datasets which surface of a surgical object, the spatial location of
are comparable to CT scans. Being initially advo- points on the surface can be recorded. Surfaces
cated primarily for surgery at the extremities, this can then be generated from the recorded sparse
“fluoro-CT” has been adopted for usage with a point clouds and used as the virtual representa-
navigation system and has been applied to several tion of the surgical object. Because this model is
anatomical areas already [13, 14]. As a major generated by the operator, the technique is there-
advantage, the device combines the availability fore known as “surgeon-defined anatomy” (SDA)
of 3-D imaging with the intra-operative data ac- (Fig. 1.3). It is particularly useful when soft
quisition. “Fluoro-CT” technology is under con- tissue structures such as ligaments or cartilage
tinuous development involving smaller and non- boundaries are to be considered that are difficult
isocentric C-arms, “closed” C-arm, i.e. O-armTM to identify on CTs or fluoroscopic images [17].
design [15, 16], faster acquisition speeds, larger Moreover, with SDA-based systems, some land-
field of view, and also flat panel technology. marks can be acquired even without the direct
A last category of navigation systems func- access to the anatomy. For instance, the centre of
tions without any radiological images as VOs. In- the femoral head, which is an important landmark
stead, the tracking capabilities of the system are during total hip and knee replacement, can be
6 G. Zheng and L.-P. Nolte
Fig. 1.4 An example of bone morphing. Screenshots model; and (c) verification of final result. (Courtesy of
of different stages of an intra-operative bone morphing Brainlab AG, Munich, Germany)
process. (a) Point acquisition; (b) calculation of morphed
calculated from a recorded passive rotation of Early CAOS systems implemented paired-
the leg about the acetabulum. It should be noted point matching and surface matching [22]. The
that the generated representations are often rather operational procedure for paired-point matching
abstract and not easy to interpret as exemplified is simple. Pairs of distinct points are defined pre-
in Fig. 1.3. This has motivated the development operatively in the VO and intra-operatively in the
of the so-called “bone morphing” techniques [18, TO. The points on the VO are usually identified
19], which aim to derive a patient-specific model pre-operatively using the computer mouse, while
from a generic statistical forms of the target the corresponding points on the TO are usually
anatomical structure and a set of sparse points done intra-operatively with a tracked probe.
that are acquired with the SDA technique [20]. In the case of a navigation system, the probe
As the result, a realistic virtual model of the is tracked by the navigator, and for a robotic
target structure can be presented and used as a surgery, it is mounted onto the robot’s actuator
VO without any conventional image acquisition [23]. Although paired-point matching is easy to
(Fig. 1.4). solve mathematically, the accuracy of the resul-
tant registration is low. This is due to the fact that
the accuracy of paired-point matching depends
1.2.2 Registration on an optimal selection of the registration points
and the exact identification of the associated
Position data that is used intra-operatively to dis- pairs which is error prone. One obvious solution
play the current tool location (navigation system) to this problem is to implant artificial objects to
or to perform automated actions according to a create easily and exactly identifiable fiducials
pre-operative plan (robot) are expressed in the for an accurate paired-point matching [23].
local coordinate system of the VO. In general, However, the requirement of implanting these
this coordinate system differs from the one in objects before the intervention causes extra
which the navigator operates intra-operatively. In operation as well as associated discomfort and
order to bridge this gap, the mathematical rela- infection risk for the patient [24]. Consequently,
tionships between both coordinate spaces need none of these methods have gained wide clinical
to be determined. When pre-operative images acceptance. The other alternative that has been
are used as VOs, this step is performed interac- widely adopted in early CAOS systems is to
tively by the surgeon during the registration, also complement the paired-point matching with
known as matching. A wide variety of different surface matching [25, 26], which does not require
approaches have been developed and realized implanting any artificial object and only uses the
following numerous methodologies [21]. surfaces of the VO as a basis for registration.
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 7
Other methods to compute the registration a conventional probe which usually requires
transformation without the need for extensive an invasive exposure of the surfaces of the
pre-operative preparation utilize intra-operative target structures. Two different tracked mode
imaging such as calibrated fluoroscopic images ultrasound probes are available. A (amplitude)-
or calibrated ultrasound images. As described mode ultrasound probes can measure the
above, a limited number of fluoroscopic images depth along the acoustic axis of the device.
(e.g. two) acquired at different positions are cali- Placed on the patient’s skin, they can measure
brated and co-registered to a common coordinate percutaneously the distance to tissue borders,
system established on the target structure. A so- and the resulting point coordinates can be
called “2-D-3-D registration” procedure can then used as inputs to any feature-based registration
be used to find the geometrical transformation algorithm. The applicability of this technique has
between the common coordinate system and a been demonstrated previously but with certain
pre-operatively acquired 3-D CT dataset by max- limitations which prevent its wide usage [27, 28].
imizing a similarity measurement between the 2- More specifically, the accuracy of the A-mode
D projective representations and the associated ultrasound probe-based digitization depends on
digitally reconstructed radiographs (DRRs) that how well the probe can be placed perpendicularly
are created by simulating X-ray projections (see to the surfaces of the target bony structures,
Fig. 1.5 for an example). Intensity-based as well which is not an easy task when the subcutaneous
as feature-based approaches have been proposed soft tissues are thick. Moreover, the velocity of
before. For a comprehensive review of differ- sound during the probe calibration is usually
ent 2-D-3-D registration techniques, we refer to different from the velocity of sound when the
[21]. probe is used for digitization as the latter depends
Another alternative is the employment of on the properties of the traversed tissues. Such
intra-operative ultrasonography. If an ultrasound a velocity difference will lead to unpredictable
probe is tracked by a navigator and its inaccuracies when the probe is used to digitize
measurements are calibrated, it may serve as a deeply located structures. As a consequence,
spatial digitizer with which points or landmarks the successful application of this technique
on the surfaces of certain subcutaneous bony remains limited to a narrow field of application.
structures may be acquired. This is different In contrast to an A-mode probe, a B (brightness)-
from the touch-based digitization done with mode ultrasound probe scans a fan-shaped area.
Fig. 1.5 An example of CT-fluoro matching. Screenshots CT-fluoro matching. (Courtesy of Brainlab AG, Munich,
of different stages of a CT-fluoro matching process. (a) Germany)
Preregistration for CT-fluoro matching and (b) results of
8 G. Zheng and L.-P. Nolte
in the next sub-section. During the surgical pro- fore, magnetic tracking has been employed only
cedure, the system is under the direct surgeon in very few commercial navigation systems and
control and gives real-time tactile feedback to the with limited success.
surgeon. Other semi-active robots such as Spine- Recently inertial measurement unit (IMU)-
Assist (Mazor Robotics Ltd., Israel) can be seen based navigation devices have attracted more
as intelligent gauges that place, for example, cut- and more interests [47–51]. These devices at-
ting jigs or drilling guides automatically [41, 42]. tempt to combine the accuracy of large-console
CAOS systems with the familiarity of conven-
1.2.3.2 Tracker tional alignment methods and have been suc-
The navigator of a surgical navigation system is cessfully applied to applications including TKA
a spatial position tracking device. It determines [47, 48], pedicle screw placement [49], and pe-
the location and orientation of objects and pro- riacetabular osteotomy (PAO) surgery [50, 51].
vides these data as 3-D coordinates or 3-D rigid With such devices, the line-of-sight issues in
transformations. Although a number of track- the optical surgical navigation systems can be
ing methods based on various physical media, completely eliminated. Technical limitations of
e.g. acoustic, magnetic, optical, and mechanical such devices include (a) relatively lower accuracy
methods, have been used in the early surgical in comparison with optical tracking technique
navigation systems, most of today’s products rely and (b) difficulty in measuring translations.
upon optical tracking of objects using operating
room (OR) compatible infrared light that is either
actively emitted or passively reflected from the 1.2.4 Referencing
tracked objects. To track surgical end effectors
with this technology then requires the tools to be Intra-operatively, it is unavoidable that there will
adapted with reference bases holding either light- be relative motions between the TO and the
emitting diodes (LED, active) or light-reflecting navigator due to surgical actions. Such motions
spheres or plates (passive). Tracking patterns need to be detected and compensated to secure
with known geometry by means of video images surgical precision. For this purpose, the operated
has been suggested [43, 44] as an inexpensive anatomy is linked to the navigator. For robotic
alternative to an infrared-light optical tracker. surgery this connection is established as a phys-
Optical tracking of surgical end effectors re- ical linkage. Large active robots, such as the
quires a direct line of sight between the tracker early machines used for total joint replacement,
and the observed objects. This can be a critical is- come with a bone clamp that tightly grips the
sue in the OR setting. The use of electromagnetic treated structure or involve an additional multi-
tracking systems has been proposed to overcome link arm, while smaller active and semi-active
this problem. This technology involves a homo- devices are mounted directly onto the bone. For
geneous magnetic field generator that is usually all other tracker types, bone motion is determined
placed near to the surgical situs and the attach- by the attachment of a DRB to the TO [52],
ment of receiver coils to each of the instruments which is designed to house infrared LEDs, re-
allowing measuring their position and orientation flecting markers, acoustic sensors, or electromag-
within the magnetic field. This technique senses netic coils, depending on the employed tracking
positions even if objects such as the surgeon’s technology. Figure 1.6 shows an example of a
hand are in between the emitter coil and the DRB for an active optical tracking system that
tracked instrument. However, the homogeneity is attached to the spinous process of a lumbar
of the magnetic field can be easily disturbed by vertebra. Since the DRB is used as an indicator
the presence of certain metallic objects caus- to inform the tracker precisely about movements
ing measurement artefacts that may decrease the of the operated bone, a stable fixation throughout
achievable accuracyconsiderably [45, 46]. There- the entire duration of the procedure is essential.
10 G. Zheng and L.-P. Nolte
Fig. 1.7 Patient-specific instrumentation for pelvic tu- (a) A pre-operative X-ray radiograph, (b) the im-
mour resection surgery. These images show the plant; (c) the patient-specific guide; (d) a post-
application of patient-specific instrumentation for operative X-ray radiograph. (Courtesy of Prof.
pelvic tumour treatment. Implant and template Dr. K Siebenrock, Inselspital, University of Bern,
manufactured by Mobelife NV, Leuven, Belgium. Switzerland)
reintroduced to the market for total knee arthro- a recent study [73] suggested that surgeons did
plasty [33, 68, 69], hip resurfacing [34, 70], total not select them as major weaknesses. It has been
hip arthroplasty [35], and pelvic tumour resection indicated that barriers to adoption of surgical
[71, 72] (see Fig. 1.7 for an example). It should navigation are neither due to a difficult learning
be noted that most of the individual templates curve nor to a lack of training opportunities.
are produced using additive manufacturing tech- The barriers to adoption of navigation are more
niques, while most of the associated implants are intrinsic to the technology itself, including intra-
produced conventionally. operative glitches, unreliable accuracy, frustra-
tion with intra-operative registration, and line-
of-sight issues. These findings suggest that sig-
1.4 Future Perspectives nificant improvements in the technology will be
required to improve the adoption rate of sur-
Despite its touted advantages, such as decreased gical navigation. Addressing these issues from
radiation exposure to the patient and the sur- the following perspectives may provide solutions
gical team for certain surgical procedures and in the continuing effort to implement surgical
increased accuracy in most situations, surgical navigation in everyday clinical practice.
navigation has yet to gain general acceptance
among orthopaedic surgeons. Although issues • 2-D or 3-D image stitching. Long-bone frac-
related to training, technical difficulty, and learn- ture reduction and spinal deformity correc-
ing curve are commonly presumed to be major tion are two typical clinical applications that
barriers to the acceptance of surgical navigation, frequently use the C-arm in its operation.
12 G. Zheng and L.-P. Nolte
Such a surgery usually involves corrective • Image fusion. Fusion of multimodality pre-
manoeuvers to improve the sagittal or coronal operative image such as various MRI or CT
profile. However, intra-operative estimation of datasets with intra-operative images would
the amount of correction is difficult, especially allow for visualization of critical structures
in longer instrumentation. Mostly, anteropos- such as nerve roots or vascular structures
terior (AP) and lateral fluoroscopic images are during surgical navigation. Different imaging
used but have the disadvantage to depict only a modalities provide complementary informa-
small portion of the target structure in a single tion regarding both anatomy and physiology.
C-arm image due to the limited field of view The evidence supporting this complementarity
of a C-arm machine. As such, orthopaedic has been gained over the last few years
surgeons nowadays are missing an effective with increased interest in the development
tool to image the entire anatomical structure of platform hardware for multimodality
such as the spine or long bones during surgery imaging. Because multimodality images by
for assessing the extent of correction. Al- definition contain information obtained using
though radiographs obtained either by using different imaging methods, they introduce
a large field detector or by image stitching new degrees of freedom, raising questions
can be used to image the entire structure, they beyond those related to exploiting each single
are usually not available for intra-operative modality separately. Processing multimodality
interventions. One alternative is to develop images is then all about enabling modalities
methods to stitch multiple intra-operatively to fully interact and inform each other. It
acquired small fluoroscopic images to be able is important to choose an analytical model
to display the entire structure at once [74, 75]. that faithfully represents the link between
Figure 1.8 shows an image stitching example the modalities without imposing phantom
for spinal intervention. The same idea can be connections or suppressing existing ones.
extended to 3-D imaging to create a panoramic Hence it is important to be as data driven
cone beam computed tomography [76]. At this as possible. In practice, this means making
moment, fast and easy-to-use 2-D or 3-D im- the fewest assumptions and using the simplest
age stitching systems are still under develop- model, both within and across modalities.
ment, and as the technology evolves, surgical Example models include linear relationships
benefits and improved clinical outcomes are between underlying latent variables; use of
expected. model-independent priors such as sparsity,
Fig. 1.8 Image stitching for spinal interventions. Several small field-of-view C-arm images are stitched into one big
image to depict the entire spine
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 13
In addition these models will not only be limitations and models of the musculoskeletal
used pre-operatively but need to function apparatus that are not only anatomically but
also in near real time in the operating also functionally correct and accurate.
theatre. • Musculoskeletal imaging. Musculoskeletal
First attempts have been made to incor- imaging is defined as the imaging of bones,
porate biomechanical simulation and mod- joints, and connected soft tissues with an
elling into the surgical decision-making pro- extensive array of modalities such as X-
cess for orthopaedic interventions. For ex- ray radiography, CT, ultrasonography, and
ample, a large spectrum of medical devices MRI. For the past two decades, rapid but
exists for correcting deformities associated cumulative advances can be observed in
with spinal disorders. Driscoll et al. [81] de- this field, not only for improving diagnostic
veloped a detailed volumetric finite element capabilities with the recent advancement on
model of the spine to simulate surgical cor- low-dose X-ray imaging, cartilage imaging,
rection of spinal deformities and to assess, diffusion tensor imaging, MR arthrography,
compare, and optimize spinal devices. An- and high-resolution ultrasound but also for
other example was presented in [82] where enabling image-guided interventions with
the authors showed that with biomechanical the introduction of real-time MRI or CT
modelling the instrumentation configuration fluoroscopy, molecular imaging with PET/CT,
can be optimized based on clinical objectives. and optical imaging into operating room [87].
Murphy et al. [83] presented the development One recent advancement that has found
of a biomechanical guidance system (BGS) a lot of clinical applications is the EOS 2-
for periacetabular osteotomy. The BGS aims D/3-D image system (EOS imaging, Paris,
to provide not only real-time feedback of the France), which was introduced to the mar-
joint repositioning but also the simulated joint ket in 2007. The EOS 2-D/3-D imaging sys-
contact pressures. tem [88] is based on the Nobel Prize-winning
Another approach is the combined use work of French physicist Georges Charpak
of intra-operative sensing devices with on multiwire proportional chamber, which is
simplified biomechanical models. Crottet placed between the X-rays emerging from the
et al. [84] introduced a device that intra- radiographed object and the distal detectors.
operatively measures knee joint forces and Each of the emerging X-rays generates a sec-
moments and evaluated its performance and ondary flow of photons within the chamber,
surgical advantages on cadaveric specimens which in turn stimulate the distal detectors that
using a knee joint loading apparatus. Large give rise to the digital image. This electronic
variation among specimens reflected the avalanche effect explains why a low dose of
difficulty of ligament release and the need primary X-ray beam is sufficient to generate
for intra-operative force monitoring. A a high-quality 2-D digital radiograph, making
commercial version of such a device (e- it possible to cover a field of view of 175 cm
LIBRA Dynamic Knee Balancing System, by 45 cm in a single acquisition of about
Synvasive Technology, El Dorado Hills, 20s duration [89]. With an orthogonally co-
CA, USA) became available in recent years linked, vertically movable, slot-scanning X-
and is clinically used (see, e.g. [85]). It is ray tube/detector pairs, EOS has the benefit
expected that incorporation of patient-specific that it can take a pair of calibrated posteroan-
biomechanical modelling into CAOS systems terior (PA) and lateral (LAT) images simul-
with or without the use of intra-operative taneously [90]. EOS allows the acquisition
sensing devices may eventually increase the of images while the patient is in an upright,
quality of surgical outcomes [86]. Research weight-bearing (standing, seated, or squatting)
activities must focus on existing technology position and can image the full length of the
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 15
body, removing the need for digital stitch- open numerous and important perspectives in
ing/manual joining of multiple images [91]. CAOS research.
The quality and nature of the image gener- Another novel technology on 2-D/3-
ated by EOS system are comparable or even D imaging was introduced in [99], which
better than computed radiography (CR) and had the advantage of being integrated with
digital radiography (DR) but with much lower any conventional X-ray machine. A mean
radiation dosage [90]. It was reported by Illés reconstruction parameter of 1.06±0.20 mm
et al. [90] that absorbed radiation dose by was reported. This technology has been used
various organs during a full-body EOS 2-D/3- for conducting 3-D pre-operative planning
D examination required to perform a surface and post-operative treatment evaluation of
3-D reconstruction was 800–1000 times less TKA based on only 2-D long leg standing
than the amount of radiation during a typical X-ray radiographs [100].
CT scan required for a volumetric 3-D recon- • Artificial intelligence, machine learning, and
struction. When compared with conventional deep learning. Recently artificial intelligence
or digitalized radiographs [92], EOS system and machine learning-based methods have
allows a reduction of the X-ray dose of an gained increasing interest in many different
order 80–90%. The unique feature of simul- fields including musculoskeletal imaging and
taneously capturing a pair of calibrated PA surgical navigation. Most of these methods are
and LAT images of the patient allows a full based on ensemble learning principles that can
3-D reconstruction of the subject’s skeleton aggregate predictions of multiple classifiers
[90, 93, 94]. This in turn provides over 100 and demonstrate superior performance in
clinical parameters for pre- and post-operative various challenging problems [77, 101, 102].
surgical planning [90]. With a phantom study, A crucial step in the design of such systems
Glaser et al. [95] assessed the accuracy of EOS is the extraction of discriminant features
3-D reconstruction by comparing it with 3-D from the images [103]. In contrast, many
CT. They reported a mean shape reconstruc- deep learning algorithms that have been
tion accuracy of 1.1±0.2 mm (maximum 4.7 proposed recently, which are based on models
mm) with 95% confidence interval of 1.7 mm. (networks) composed of many layers that
They also found that there was no significant transform input data (e.g. images) to outputs
difference in each of their analysed parameters (e.g. segmentation), let computers learn the
(p > 0.05) when the phantom was placed in features that optimally represent the data for
different orientations in the EOS machine. the problem at hand. The most successful
The reconstruction of 3-D bone models allows type of models for image analysis to date are
analysis of subject-specific morphology in a convolutional neural networks (CNN) [104],
weight-bearing situation for different applica- which contain many layers that transform their
tions to a level of accuracy which was not input with convolution filters of a small extent.
previously possible. For example, Lazennec Deep learning-based methods have been
et al. [96] used the EOS system to measure successfully used to solve many challenging
pelvis and acetabular component orientations problems in computer-aided orthopaedic
in sitting and standing positions. Further ap- surgery [105–108]. Figure 1.10 shows an
plications of EOS system in planning total example of the application of cascaded fully
hip arthroplasty include accurate evaluation of convolutional networks (FCN) for automatic
femoral offset [97] and rotational alignment segmentation of lumbar vertebrae from CT
[98]. The low dose and biplanar information images [108]. It is expected that more and
of the EOS 2-D/3-D imaging system introduce more solutions will be developed based on
key benefits in contemporary radiologyand different types of deep learning techniques.
16 G. Zheng and L.-P. Nolte
Fig. 1.10 A schematic view of using cascaded fully convolutional networks (FCN), which consists of a localization
net and a segmentation net for automatic segmentation of lumbar vertebrae from CT images
9. Hofstetter R, Slomczykowski M, Bourquin Y, Nolte 23. Bargar WL, Bauer A, Börner M (1998) Primary and
LP (1997) Fluoroscopy based surgical navigation: revision total hip replacement using the Robodoc
concept and clinical applications. In: Lemke HU, system. Clin Orthop 354:82–91
Vannier MW, Inamura K (eds) Computer assisted 24. Nogler M, Maurer H, Wimmer C, Gegenhuber C,
radiology and surgery. Elsevier Science, Amster- Bach C, Krismer M (2001) Knee pain caused by
dam, pp 956–960 a fiducial marker in the medial femoral condyle: a
10. Joskowicz L, Milgrom C, Simkin A, Tockus L, clinical and anatomic study of 20 cases. Acta Orthop
Yaniv Z (1998) FRACAS: a system for computer- Scand 72:477–480
aided image-guided long bone fracture surgery. 25. Besl PJ, McKay ND (1992) A method for reg-
Comput Aided Surg 36:271–288 istration of 3-D shapes. IEEE Trans Pattern Anal
11. Foley KT, Simon DA, Rampersaud YR (2001) Vir- 14(2):239–256
tual fluoroscopy: image-guided fluoroscopic naviga- 26. Baechler R, Bunke H, Nolte L-P (2001) Re-
tion. Spine 26:347–351 stricted surface matching – numerical optimiza-
12. Ritter D, Mitschke M, Graumann R (2002) Mark- tion and technical evaluation. Comput Aid Surg 6:
erless navigation with the intra-operative imag- 143–152
ing modality SIREMOBIL Iso-C3D. Electromedica 27. Maurer CR, Gaston RP, Hill DLG, Gleeson MJ,
70:47–52 Taylor MG, Fenlon MR, Edwards PJ, Hawkes DJ
13. Grützner PA, Waelti H, Vock B, Hebecker A, Nolte (1999) AcouStick: a tracked A-mode ultrasonogra-
L-P, Wentzensen A (2004) Navigation using fluoro- phy system for registration in image-guided surgery.
CT technology. Eur J Trauma 30:161–170 In: Taylor C, Colchester A (eds) Medical image
14. Rajasekaran S, Karthik K, Chandra VR, Rajkumar computing and image-guided intervention – MIC-
N, Dheenadhayalan J (2010) Role of intraoperative CAI’99. Springer, Berlin, pp 953–962
3D C-arm-based navigation in percutaneous exci- 28. Oszwald M, Citak M, Kendoff D, Kowal J, Amstutz
sion of osteoid osteoma of lone bones in children. C, Kirchhoff T, Nolte L-P, Krettek C, Hüfner T
J Pediatr Orthop 19:195–200 (2008) Accuracy of navigated surgery of the pelvis
15. Lin EL, Park DK, Whang PG, An HS, Phillips after surface matching with an a-mode ultrasound
FM (2008) O-Arm surgical imaging system. Semin proble. J Orthop Res 26:860–864
Spine Surg 20:209–213 29. Kowal J, Amstutz C, Langlotz F, Talib H, Gonzalez
16. Qureshi S, Lu Y, McAnany S, Baird E (2014) Three- Ballester MA (2007) Automated bone contour de-
dimensional intraoperative imaging modalities in tection in ultrasound B-mode images for minimally
orthopaedic surgery: a narrative review. J Am Acad invasive registration in image-guided surgery – an
Orthop Surg 22(12):800–809 in vitro evaluation. Int J Med Rob Comput Assisted
17. Sati M, Stäubli HU, Bourquin Y, Kunz M, Nolte Surg 3:341–348
LP (2002) Real-time computerized in situ guidance 30. Schumann S, Nolte L-P, Zheng G (2012) Compen-
system for ACL graft placement. Comput Aided sation of sound speed deviations in 3D B-mode ul-
Surg 7:25–40 trasound for intraoperative determination of the an-
18. Fleute M, Lavallée S, Julliard R (1999) Incorporat- terior pelvic plane. IEEE Trans Inf Technol Biomed
ing a statistically based shape model into a system 16(1):88–97
for computer assisted anterior cruciate ligament 31. Wein W, Karamalis A, Baumgarthner A, Navab
surgery. Med Image Anal 3:209–222 N (2015) Automatic bone detection and soft tis-
19. Stindel E, Briard JL, Merloz P, Plaweski S, Dubrana sue aware ultrasound-CT registration for computer-
F, Lefevre C, Troccaz J (2002) Bone morphing: aided orthopedic surgery. Int J Comput Assist Ra-
3D morphological data for total knee arthroplasty. diol Surg 10(6):971–979
Comput Aided Surg 7:156–168 32. Radermacher K, Portheine F, Anton M et al (1998)
20. Zheng G, Dong X, Rajamani KT, Zhang X, Styner Computer assisted orthopaedic surgery with image
M, Thoranaghatte RU, Nolte L-P, Ballester MA based individual templates. Clin Orthop Relat Res
(2007) Accurate and robust reconstruction of a 354:28–38
surface model of the proximal femur from sparse- 33. Hafez MA, Chelule KL, Seedhom BB, Sherman KP
point data and a dense-point distribution model (2006) Computer-assisted total knee arthroplasty
for surgical navigation. IEEE Trans Biomed Eng using patient-specific templating. Clin Orthop Relat
54:2109–2122 Res 444:184–192
21. Zheng G, Kowal J, Gonzalez Ballester MA, Caver- 34. Kunz M, Rudan JF, Xenoyannis GL, Ellis RE (2010)
saccio M, Nolte L-P (2007) Registration technique Computer-assisted hip resurfacing using individual-
for computer navigation. Curr Orthop 21:170–179 ized drill templates. J Arthroplast 25:600–606
22. Lavallée S (1996) Registration for computer- 35. Shandiz MA, MacKenzie JR, Hunt S, Anglin C
integrated surgery: methodology, start of the art. (2014 Sept) Accuracy of an adjustable patient-
In: Taylor RH, Lavallée S, Burdea GC, Mösges R specific guide for acetabular alignment in hip re-
(eds) Computer integrated surgery. The MIT Press, placement surgery (Optihip). Proc Inst Mech Eng H
Cambridge, pp 77–97 228(9):876–889
18 G. Zheng and L.-P. Nolte
36. Honl M, Dierk O, Gauck C, Carrero V, Lampe F, 48. Huang EH, Copp SN, Bugbee WD (2015) Accu-
Dries S, Quante M, Schwieger K, Hille E, Mor- racy of a handheld accelerometer-based navigation
lock MM (2003) Comparison of robotic-assisted system for femoral and tibial resection in total knee
and manual implantation of a primary total hip arthroplasty. J Arthroplast 30(11):1906–1910
replacement. A prospective study. J Bone Joint Surg 49. Walti J, Jost GF, Cattin PC (2014) A new cost-
85A8:1470–1478 effective approach to pedicular screw placement. In:
37. Oszwald M, Ruan Z, Westphal R, O’Loughlin AE-CAI 2014, LNCS 8678. Springer, Heidelberg,
PF, Kendoff D, Hüfner T, Wahl F, Krettek C, pp 90–97
Gosling T (2008) A rat model for evaluating phys- 50. Pflugi S, Liu L, Ecker TM, Schumann S, Cullmann
iological responses to femoral shaft fracture re- JL, Siebenrock K, Zheng G (2016) A cost-effective
duction using a surgical robot. J Orthop Res 26: surgical navigation solution for periacetabular os-
1656–1659 teotomy (PAO) surgery. Int J Comput Assist Radiol
38. Oszwald M, Westphal R, Bredow J, Calafi A, Surg 11(2):271–280
Hüfner T, Wahl F, Krettek C, Gosling T (2010) 51. Pflugi S, Vasireddy R, Lerch T, Ecker TM, Tannast
Robot-assisted fracture reduction using three- T, Boemake N, Siebenrock K, Zheng G (2018) A
dimensional intraoperative fracture visualization: an cost-effective surgical navigation solution for peri-
experimental study on human cadaver femora. J acetabular osteotomy (PAO) surgery. Int J Comput
Orthop Res 28:1240–1244 Assist Radiol Surg 13(2):291–304
39. Jaramaz B, Nikou C (2012) Precision freehand 52. Nolte LP, Visarius H, Arm E, Langlotz F,
sculpting for unicondylar knee replacement: de- Schwarzenbach O, Zamorano L (1995) Computer-
sign and experimental validation. Biomed Tech aided fixation of spinal implants. J Imag Guid Surg
57(4):293–299 1:88–93
40. Conditt MA, Roche MW (2009) Minimally invasive 53. Foley KT, Smith MM (1996) Image-guided spine
robotic-arm-guided unicompartmental knee arthro- surgery. Neurosurg Clin N Am 7:171–186
plasty. J Bone Joint Surg 91(Suppl 1):63–68 54. Glossop ND, Hu RW, Randle JA (1996) Computer-
41. Ritschl P, Machacek F, Fuiko R (2003) Com- aided pedicle screw placement using frameless
puter assisted ligament balancing in TKR using the stereotaxis. Spine 21:2026–2034
Galileo system. In: Langlotz F, Davies BL, Bauer A 55. Kalfas IH, Kormos DW, Murphy MA, McKenzie
(eds) Computer assisted orthopaedic surgery – 3rd RL, Barnett GH, Bell GR, Steiner CP, Trimble MB,
annual meeting of CAOS-International (Proceed- Weisenberger JP (1995) Application of frameless
ings). Steinkopff, Darmstadt, pp 304–305 stereotaxy to pedicle screw fixation of the spine. J
42. Shoham M, Burman M, Zehavi E, Joskowicz Neurosurg 83:641–647
L, Batkilin E, Kunicher Y (2003) Bone-mounted 56. Merloz P, Tonetti J, Pittet L, Coulomb M, Lavallée
miniature robot for surgical procedures: concept S, Sautot P (1998) Pedicle screw placement using
and clinical applications. IEEE Trans Rob Autom image guided techniques. Clin Orthop 354:39–48
19:893–901 57. Amiot LP, Lang K, Putzier M, Zippel H, Labelle
43. de Siebenthal J, Gruetzner PA, Zimolong A, Rohrer H (2000) Comparative results between conven-
U, Langlotz F (2004) Assessment of video track- tional and image-guided pedicle screw installation
ing usability for training simulators. Comput Aided in the thoracic, lumbar, and sacral spine. Spine 25:
Surg 9:59–69 606–614
44. Clarke JV, Deakin AH, Nicol AC, Picard F (2010) 58. Laine T, Lund T, Ylikoski M, Lohikoski J, Schlen-
Measuring the positional accuracy of computer as- zka D (2000) Accuracy of pedicle screw insertion
sisted surgical tracking systems. Comput Aided with and without computer assistance: a randomised
Surg 15:13–18 controlled clinical study in 100 consecutive patients.
45. Meskers CG, Fraterman H, van der Helm FC, Ver- Eur Spine J 9:235–240
meulen HM, Rozing PM (1999) Calibration of the 59. Schwarzenbach O, Berlemann U, Jost B, Visarius
“Flock of Birds” electromagnetic tracking device H, Arm E, Langlotz F, Nolte LP, Ozdoba C (1997)
and its application in shoulder motion studies. J Accuracy of image-guided pedicle screw placement.
Biomech 32:629–633 An in vivo computed tomography analysis. Spine
46. Wagner A, Schicho K, Birkfellner W, Figl M, See- 22:452–458
mann R, Konig F, Kainberger F, Ewers R (2002) 60. Digioia AM 3rd, Simon DA, Jaramaz B et al (1999)
Quantitative analysis of factors affecting intraop- HipNav: pre-operative planning and intra-operative
erative precision and stability of optoelectronic navigational guidance for acetabular implant place-
and electromagnetic tracking systems. Med Phys ment in total hip replacement surgery. In: Nolte
29:905–912 LP, Ganz E (eds) Computer Assisted Orthopaedic
47. Nam D, Cody EA, Nguyen JT, Figgie MP, Mayman Surgery (CAOS). Hogrefe & Huber, Seattle, pp
DJ (2014) Extramedullary guides versus portable, 134–140
accelerometer-based navigation for tibial alignment 61. Croitoru H, Ellis RE, Prihar R, Small CF, Pichora
in total knee arthroplasty: a randomized, controlled DR (2001) Fixation based surgery: a new technique
trial: winner of the 2013 Hap Paul Award. J Arthro- for distal radius osteotomy. Comput Aided Surg
plast 29(2):288–294 6:160–169
1 Computer-Aided Orthopaedic Surgery: State-of-the-Art and Future Perspectives 19
91. Wade R, Yang H, McKenna C et al (2013) A Proceedings of MIAR 2016, LNCS 9805. Springer,
systematic review of the clinical effectivenss of Heidelberg, pp 404–414
EOS 2D/3D x-ray imaging system. Eur Spine J 22: 100. Hommel H, Alcoltekin A, Thelen B, Stifter J,
296–304 Schwägli T, Zheng G (2017) 3X-Plan: A novel
92. Deschenes S, Charron G, Beaudoin G et al (2010) technology for 3D prosthesis planning using 2D X-
Diagnostic imaging of spinal deformities – Reduc- ray radiographs. Proc CAOS 2017:93–95
ing patients radiation dose with a new slot-scanning 101. Glocker B, Feulner J, Criminisi A, Haynor DR,
x-ray imager. Spine 35:989–994 Konukoglu E (2012) Automatic localization and
93. Langlois K, Pillet H, Lavaste F, Rochcongar G, identification of vertebrae in arbitrary field-of-
Rouch P, Thoreux P, Skalli W (2015 Oct) Assessing view CT scans. In: Proceedings of MICCAI 2012;
the accuracy and precision of manual registration 15(Pt3). Springer, Heidelberg, pp 590–598
of both femur and tibia using EOS imaging system 102. Liu Q, Wang Q, Zhang L, Gao Y, Sheng D (2015)
with multiple views. Comput Methods Biomech Multi-atlas context forests for knee MR image seg-
Biomed Eng 18(Suppl 1):1972–1973 mentation. MLMI@MICCAI 2015:186–193
94. Ferrero E, Lafage R, Challier V, Diebo B, Guigui 103. Litjens G, Kooi T, Bejnordi BE, Setio AAA, Ciompi
P, Mazda K, Schwab F, Skalli W, Lafage V (2015 F, Ghafoorian M, van der Laak JAWM, van Gin-
Sept) Clinical and stereoradiographic analysis of neken B, Sánchez CI (2017) A survey on deep
adult spinal deformity with and without rotatory learning in medical image analysis. Med Image
subluxation. Orthop Traumatol Surg Res 101(5): Anal 42:60–88
613–618 104. Krizhevsky A, Sutskever I, Hinton GE (2012) Ima-
95. Glaser DA, Doan J, Newton PO (2012) Comparison geNet classification with deep convolutional neural
of 3-Dimensional spinal reconstruction accuracy. networks. Advances in Neural Information Pro-
Spine 37:1391–1397 cessing Systems 25, Curran Associates, Inc., 2012,
96. Lazennec JY, Rousseau MA, Rangel A, Gorin M, 1097–1105
Belicourt C, Brusson A, Catonne Y (2011) Pelvis 105. Prasoon A, Petersen K, Igel C, Lauze F, Dam
and total hip arthroplasty acetabular component ori- E, Nielsen M (2013) Deep feature learning for
entation in sitting and standing positions: measure- knee cartilage segmentation using a triplanar con-
ments reproductibility with EOS imaging system volutional neural network. MICCAI 2013 16(Pt2):
versus conventional radiographies. Orthop Trauma- 246–253
tol Surg Res 97:373–380 106. Zeng G, Yang X, Li J, Yu L, Heng P-A, Zheng
97. Lazennec JY, Brusson A, Dominique F, Rousseau G (2017) 3D U-net with multi-level deep super-
MA, Pour AE (2015) Offset and anteversion re- vision: fully automatic segmentation of proximal
construction after cemented and uncemented total femur in 3D MR images. MLMI@MICCAI 2017:
hip arthroplasty: an evaluation with the low-dose 274–282
EOS system comaring two- and three-dimensional 107. Li X, Dou Q, Chen H, Fu CW, Qi X, Belavý DL,
imaging. Int Orthop. 39(7):1259–1267 Armbrecht G, Felsenberg D, Zheng G, Heng PA
98. Folinais D, Thelen P, Delin C, Radier C, Catonne (2018) 3D multi-scale FCN with random modality
Y, Lazennec JY (2011) Measuring femoral and voxel dropout learning for intervertebral disc local-
rotational alignment: EOS system versus com- ization and segmentation from multi-modality MR
puted tomography. Orthop Traumatol Surg Res 99: images. Med Image Anal 45:41–54
509–516 108. Janssens R, Zeng G, Zheng G (2017) Fully auto-
99. Zheng G, Schumann S, Alcoltekin A, Jaramaz B, matic segmentation of lumbar vertebrae from CT
Nolte L-P (2016) Patient-specific 3D reconstruction images using cascaded 3D fully convolutional net-
of a complete lower extremity from 2D X-rays.In: works. arXiv:1712.01509
Computer-Aided Orthopedic Surgery:
Incremental Shift or Paradigm 2
Change?