Oral and Maxillofacial Surgery
https://doi.org/10.1007/s10006-018-0719-5
ORIGINAL ARTICLE
A novel noise filtered and occlusion removal: navigational accuracy
in augmented reality-based constructive jaw surgery
Bijaya Raj Basnet 1 & Abeer Alsadoon 1 & Chandana Withana 1 & Anand Deva 2 & Manoranjan Paul 1
Received: 16 March 2018 / Accepted: 28 August 2018
# Springer-Verlag GmbH Germany, part of Springer Nature 2018
Abstract
Purpose Augmented reality-based constructive jaw surgery has been facing various limitations such as noise in real-time images,
the navigational error of implants and jaw, image overlay error, and occlusion handling which have limited the implementation of
augmented reality (AR) in corrective jaw surgery. This research aimed to improve the navigational accuracy, through noise and
occlusion removal, during positioning of an implant in relation to the jaw bone to be cut or drilled.
Method The proposed system consists of a weighting-based de-noising filter and depth mapping-based occlusion removal for
removing any occluded object such as surgical tools, the surgeon’s body parts, and blood.
Results The maxillary (upper jaw) and mandibular (lower jaw) jaw bone sample results show that the proposed method can
achieve the image overlay error (video accuracy) of 0.23~0.35 mm and processing time of 8–12 frames per second compared to
0.35~0.45 mm and 6–11 frames per second by the existing best system.
Conclusion The proposed system concentrates on removing the noise from the real-time video frame and the occlusion. Thus, the
acceptable range of accuracy and the processing time are provided by this study for surgeons for carrying out a smooth surgical
flow.
Keywords Augmented reality navigation . 3D-2D matching . Image registration . Occlusion handling . Noise removal
Introduction
Corrective jaw surgery can be defined as a surgical procedure
that is performed on jaw bones. Corrective jaw surgery is
performed to correct the dental misalignment. The corrective
jaw surgery could contain various surgical procedures such as
drilling, cutting, resection, and implantation. The main problem of this surgery is the limited viewing space in the mouth of
the patient [1]. There is always a high risk of surgeons damaging the nerve channels or tooth root during dental surgery
[1, 2]. The traditional method of performing the jaw surgery
used the CT scan to report to plan the surgical procedure
manually by the surgeons [3]. Surgeons were required to identify the nerve channel and root canals manually with the use of
* Chandana Withana
cwithana@studygroup.com
1
School of Computing and Mathematics, Charles Sturt University,
Sydney Campus, Sydney, Australia
2
Faculty of Medicine and Health Sciences, Macquarie University,
Sydney, Australia
the CT scan report [3]. Due to the limitations such as difficulty
in identifying the nerves and accurate position of drilling in
the surgical procedure using the traditional method, a 2D virtual video-guided system was developed which helped the
surgeon by displaying virtual video on the monitor and then
augmented reality has emerged as the latest technology in
medical surgery [3]. Figure 1 shows the traditional, videoguided, and augmented reality-based surgery.
Augmented reality (AR)-based surgery uses both virtual
images from the pre-surgery and the real-time image during
surgery to create the augmented view for the user [4].
Augmented reality-based surgeries superimpose the virtual
jaw onto the real jaw during surgery which provides the surgeon with the 3D view in real time. Augmented reality provides the surgeons in the surgical environment with more
realistic and intuitive information during surgery which can
guide the surgeons during the surgical procedure [4]. It provides the surgeon with the information about the cutting lines
and drilling position in the jaw bone and also helps to find the
nerve channel and location of disease [3].
The 3D view is provided by augmented reality by
superimposing the various virtual images onto the real-time
Oral Maxillofac Surg
Fig. 1 a Traditional surgery. b Video guided. c AR guided. These images are downloaded using Google search engine; the image is free to use, share or
modify, even commercially
images [4, 5]. Augmented reality has been providing a huge
benefit in the medical field. Augmented reality is generally
used in the surgery of the complicated and sensitive areas like
the heart, kidney, brain, pelvis, breast, arteries, and jaw, but its
implementation has been limited in the jaw surgery.
Numerous research has been conducted in the past and present
in the field of corrective jaw surgery. AR in jaw surgery has
been facing various limitations such as image registration,
occlusion, noise in real-time images, high processing time,
and poor occlusion handling which has limited the implementation of AR in corrective jaw surgery [4]. Hence, 3D view
accuracy and processing time plays a vital role in augmented
reality-based surgery. The best system should be able to provide better accuracy, low processing time, and better occlusion
handling capacity. It is necessary to provide the surgeons with
accurate real-time navigational guidance for higher precision
and accuracy in surgery through accurate object tracking, navigation, and real-time registration process [6].
In the current context of augmented reality technologies in
the medical sector, video-based display, see-through display,
and projection-based display are the main categories of augmented reality technologies [4].
This paper aims to improve the accuracy of the real-time
video accuracy by removing the noise in the real-time video
caused by a range of factors such as machine vibration, camera movement, and image sensors and also by removing the
occlusion caused by surgical instruments, the surgeon’s hands,
etc. The noise removal is necessary for augmented realitybased constructive jaw surgery because the noise deteriorates
the image edges which impacts negatively on image registration, navigation, and image overlay. The features of the modified kernel non-local means (MKNLM) filter are used to denoise the real-time video images. This feature is used in the
de-noising process because this filter is less sensitive to outliers and produces the constant regular results while the other
filters are outlier sensitive and tend to produce incorrect
results.
The tracking-learning-detection (TLD) cannot handle the
occlusion and eventually fails if the occlusion is present [7].
This will require re-initialization of the TLD. Failure of the
TLD eventually results in image registration failure. This research proposes a new TLD system with depth mappingbased occlusion removal to improve the image tracking and
image registration (overlay).
A significant body of research exists that focuses on increasing accuracy and lowering processing time in augmented
reality-based surgery. [8] proposed a portable surgical navigation device and technique to reduce the bone resection error.
This solution proposes a resection plane that automatically
computes the resection margin with an error of 1.02 mm.
This solution has used markers but failed to consider the deformities caused by the patient’s movement during surgery.
[9] proposed a projection-based augmented reality solution
to eliminate the necessity of monitoring several display monitors and co-ordinates during surgery. They proposed a technique for projecting the pre-surgical images (virtual image)
onto the body of the patient but did not improve the accuracy
(range 1.4–7.4 mm) and failed to consider the patient’s movements and the occlusion present during surgery. Therefore,
these solutions do not provide possibilities for further
improvement.
[10] proposed the concept of a differential map to determine the shape changes during the bone tumor resection surgery to allow the surgeons to visualize the remains of a tumor
to be resected (cut) and provide depth information through a
graphical overlay. However, the researcher has used fiducial
markers which could change their position with the movement
of the patient. Furthermore, the author has not considered the
occlusion that is present due to surgical tools and blood. [11]
presented an optical see-through head-mounted display-based
augmented reality for navigation to improve accuracy and
reliability using an optical tracking system and surface-based
registration. The solution improved the accuracy but was not
able to address the processing time. A further limitation is that
the latency occurrence in an anatomical structural movement
which decreases the real-time performance of the system was
not addressed. In addition, the weight of the head-mounted
display could cause problems for the surgeon during long
surgeries. Thus, these solutions offer no major possibilities
for improvement, either in accuracy or in processing time.
Oral Maxillofac Surg
[12] conducted a study to evaluate the navigational accuracy of implants in an augmented reality-based navigational
system for zygomatic implant placement. The study concluded that the real-time navigation-based surgery demonstrated
higher accuracy but did not consider the presence of saliva and
blood that could cause occlusion. Furthermore, deviations in
accuracy analysis that could influence the implant failure,
generally caused by the invasion of other anatomical structure,
were not taken into consideration. [13] proposed a method to
track the patient-specific 3D-printed implant during the intraoperative placement process with the use of point-based [14]
and surface-based registration [15]. This solution was able to
increase the accuracy of implant placement but ignored the
deformities caused by soft tissues, patient movement, and
noise from the breathing of the patient while registering the
patient’s body position on the 3D image set. Thus, these solutions offer no major possibilities for improvement in accuracy
or processing time.
[4] proposed a marker-less registration solution with the
use of a stereo camera and a half-silvered mirror for depth
perception. Even though the burden of marker usage was
eliminated, this solution failed to improve the processing time
as integral videography has high processing time. The researchers also did not address the impact of blood and other
fluids which could cause inaccuracy in contouring and decrease the registration accuracy. [1] have also proposed a solution with a stereo camera and half-silvered mirror for tracking of the surgical instruments, patients’ movements, contours, and ICP (iterative closet point) for patient-image registration. However, the proposed framework still has issues in
the initial registration process where there are chances of errors which could lead to surgical inaccuracies and inconsistencies. Furthermore, the use of a stereo camera which has to
be re-calibrated and maintained for high accuracy causes difficulties in daily clinical use. Thus, these solutions offer no
major possibilities for improvement in accuracy or processing
time.
[16] proposed a marker-less registration system that enables AR visualization by projecting the CT image directly
onto the real patient’s body. The author has used kinetic surface segmentation, a two-phase registration process (initial
and fine registration), color image fusing, and CT data for
the AR view. However, repetitive initial registration (manual
registration) is required in case of movement of the object or
the camera. Furthermore, the solution was developed for the
forensic field which means that it works for non-deforming
objects. In addition, accuracy and the processing time of this
solution are relatively higher than that in any of the other
solutions presented. Thus, this method offers no major possibilities for improvement in accuracy or processing time.
[6] also conducted research into the use of stereo cameras
and a translucent mirror with the use of a 3D calibration model
in integral imaging to remove the initial registration error and
display undistorted 3D images. However, even though the
processing time has improved in this solution, accuracy remains unchanged, with an additional limitation arising from
a lack of consideration of occlusion. Thus, this solution offers
no major possibilities for improvement in accuracy or processing time. [17] proposed wafer-less maxillary positioning with
the help of interactive IGV (image-guided visualization) display complemented surgical navigation that can offer an alternative approach to the use of arbitrary splints and 2D
orthognathic planning. However, this model did not reduce
the surgical time which was high due to setting up the technical and recording process. Thus, this solution offers no major
possibilities for improvement in accuracy or processing time.
[18] introduced a video see-thorough system that uses a
hierarchy of images, TLD tracking (frame to frame) proposed
by [7], and iterative closest point (ICP) developed by [19] for
3D pose refinement. Ulrich’s method [20] is used for initial
registration. The bounding box is used for object tracking
which reduces object matching time through limiting of the
search area and iterative closest point (ICP) is used to refine
the 3D pose for higher registration accuracy [21, 22].
Limitations arise from the fact that the solution has failed to
address depth perception and occlusion present in surgical
procedures due to the presence of surgical tools and blood.
Thus, this solution offers no major possibilities for improvement in accuracy or processing time.
[3] proposed a rotational matrix and translation vector algorithm to improve the geometric accuracy in oral and maxillofacial surgery in the Wang model [18]. This solution has
addressed the depth perception using two stereo cameras.
Similar to the Wang model, this solution uses an aspect graph
to create multiple models to be matched in real time.
Tracking-learning-detection developed by [7] is used to track
the object in the video frame with the use of a bounding box
which decreases the search area. Ulrich’s method is used for
initial registration and enhanced ICP [3] is used for final pose
refinement with the use of a novel rotational matrix and translation vector algorithm that improves the geometric error.
This system reduced the overlay error to 0.30~0.40 mm and
processed10–13 frames per second. However, this system
failed to consider the time consumed due to the use of the
3D stereo camera, noise in real-time video due to machine
vibration, patient movements, image sensors, and also the
occlusion caused by the surgical tools and the surgeon’s body
parts, as well as blood, etc. The noise issue was addressed
with the use of the modified kernel non-local means
(MKNLM) filter by [23]. There are various other noise removal filters, but this filter is less sensitive to outliers and
provides more consistent results when compared to the other
filters [23]. Addition of this feature to the model mentioned
above can improve the image registration accuracy and hence
this feature is a significant addition to improve the quality of
the system.
Oral Maxillofac Surg
The model proposed by [3] addresses the accuracy and
processing time and has lower image overlay error and
processing time in comparison to other proposed systems.
This research is focused on this model to improve the
results produced for a better-augmented reality view.
This paper works on the model proposed by [3] and particularly focuses on the tracking-learning-detection (TLD)
algorithm called Tracking. This paper illustrates that better results can be achieved by removing the occlusion
caused by surgical tools, the surgeon’s body part, and
blood, etc. during surgery.
The paper is organized into three parts. The first part contains a “System Overview” that discusses the current best
model proposed by [3]. It also includes the description of the
proposed system, the associated flowchart, and pseudocode
for the proposed formula. The second part discusses
“Results” where the proposed system is tested with a range
of samples from maxillary and mandible jaw bones. This is
Fig. 2 State-of-the-art AR System
followed by a “Discussion” and comparison between the results of the state-of-art and the proposed system results and a
conclusion is provided.
System Overview
State of the art
This section describes the current state-of-art solution (Fig. 2)
with limitations (highlighted in red—Fig. 2). The model proposed by [3] provides a better image overlay with the use of a
rotational matrix and a translation vector (RMaTV) algorithm.
This system has higher accuracy through a lower image overlay error (0.35~0.45 mm) and the best processing speed of 10–
13 frames per second. The model is divided into pre-operative,
intra-operative, and pose refinement phases (Fig. 2).
Oral Maxillofac Surg
Pre-operative environment
The pre-operative planning of the surgery is done with the use
of the CT image of the patient which is segmented and an
aspect graph (hierarchy of the model) is created in the offline
phase as shown in Fig. 2. This permits matching of the different models of the segmented CT scan in the online phase
against the real-time video frame.
Intra-operative environment
Two 3D stereo cameras are used for capturing the real-time
surgical video with the translucent mirror for visualizing the
augmented reality view. Video frames are generated from the
real-time video and the hierarchy of the video frame is created
based on its resolution. The image with the lowest resolution
is used for tracking and detecting the region of interest (ROI).
However, this solution did not consider the need for regular
re-calibration and maintenance of the stereo camera to maintain performance at levels of high accuracy which is not possible in a real-time scenario. Furthermore, this solution failed
to consider the processing time required to convert the 3D
stereo video image frames to 2D image video frames for tracking of the object of interest using the tracking-learningdetection (TLD) algorithm [18]. Further limitations come
from the fact that the need for a strict viewing angle for the
stereo camera was not considered which may result in image
overlay inaccuracies, if the correct viewing angle is not
achieved [18]. In addition, the solution failed to consider the
noise present in the real-time video due to vibrations from the
machinery and optical sensors. This noise results in a deterioration of the image edges and may also lead to contour leakage
which would then negatively affect the image overlay and
registration accuracy.
The tracking-learning-detection algorithm (TLD) is used
for tracking the region of interest (surgical area). The TLD
uses a bounding box to match the object of interest with the
aspect graph created during the offline phase. The search for
the bounding box is carried out from the top level of the video
frame (lowest resolution) to the lowest level of the video frame
from the hierarchy (highest resolution). Once the match is
found, the 2D image is overlaid onto the real-time video creating an accurate 2D model. The initial registration is performed using a method proposed by [20] also known as
“Ulrich’s method” which uses shape similarity matching and
online matching [18]. After the initial registration, the ICP
(iterative closest point) is used for post-refinement for achieving an accurate 3D model. A rotational matrix and a translation vector (RMaTV) algorithm are used to remove the geometric error proposed by [3]. The refined 3D model along with
the real-time video is projected onto the translucent mirror
creating an augmented reality view for the surgeon.
Tracking an object can be defined as the estimation of displacement of the object between the two-image frames [7].
Tracking is necessary as failure to track the object of interest
results in an incorrect image overlay of pre-surgical and intrasurgical images. The quality of the image overlay depends on
how well the object of interest has been tracked. First of all, the
object needs to be tracked and then detected before the image
overlay and registration can take place. The TLD uses Lukas–
Kanade median flow (LKMF) for tracking an object of interest
(Kalal et al., 2012. [7]). It uses object feature point flow estimation for tracking (Fig. 3). The TLD uses LKMF tracker with
failure detection features that detect the TLD failure based on
the median displacement of the feature points being tracked. A
TLD failure is established if the median displacement of the
object feature point is greater than the threshold (Fig. 3). This
tracker is highly susceptible and prone to occlusion. Lukas–
Kanade Median flow tracker fails once the object gets occluded because it cannot track the feature point of an object of
interest and computes the median displacement as greater than
a threshold which results in failure of the TLD. With the use of
this tracker, the model achieved an accuracy of 0.35~0.45 mm
but fails when an occlusion occurs because it cannot compute
and track the feature point in the object once the object is
occluded. The Lucas-Kanade median flow tracker is presented
in Table 1 and the flowchart in Fig. 3.
The Lucas-Kanade median optical flow for tracking is calculated as Eq. 1:
V ¼uþd
ð1Þ
V = final location, u = image point in 2D in first image
frame, d= image velocity vector (optical flow) that reduces
residual function calculated as Eq. 2:
∈ d ; ¼ ∈ d x; d y
¼
ux þωx
uy þωy
∑
∑
x¼ux −ωx
y¼uy −ωy
I ðx; yÞ− J x þ d x ; y þ d y
2
ð2Þ
dx,dy = x and y point of optical flow matrix, ∈ = residual
function, ux, uy = image point at u, ωx, ωy = two integers, I, J =
two gray scaled image, I(x, y) = gray scaled image I at point
X(x, y).
RMaTV algorithm-based pose refinement
A rotational matrix and a translation vector algorithm
(RMaTV) proposed by [3] are used to eliminate the geometric
error with the help of rotational and translation vectors. ICP
used to register the images produces higher image overlay
accuracy with the use of the RMaTV algorithm. The
RMaTV algorithm helps to eliminate the estimation of the
wrong pose hence improves the image overlay accuracy.
Oral Maxillofac Surg
Fig. 3 Tracking in TLD using
Lukas–Kanade optical flow
tracker
Start
Select First Image frame(I) from hierarchy of images
Select Feature Points
Calculate the optical flow
and mean displacement
Display
Bounding
N
Is mean
displacement >
the threshold?
Calculate the position of the feature
point in next image frame (J)
Ye
Occlusion has
occurred
Tracking-Learning-Detection fails
Proposed solution
A range of techniques and models from existing augmented
reality-based surgery have been analyzed and reviewed in
depth to identify strengths and weaknesses. The main
Table 1
problems relating to the augmented reality-based surgery are
accuracy, processing time, noise, and occlusion handling.
Most models have primarily focused on accuracy and processing time and, to the lesser extent, on noise removal and occlusion handling. [3] model has been selected as the base model
Lukas–Kande optical flow tracker
Algorithm: Lucas- kanade method to track the object of interest
Input: Two images frames image1(I) and image2 (J) from the hierarchy of images created.
Image1 (I)=current image frame in TLD where feature point is tracked
Image2(J)=Next image frame the feature point is to be tracked
Output: Optical flow (d) which is the estimated displacement of feature point between two images i.e., from image
(I) to image (J)
BEGIN
Step 1: First of all, the image hierarchy are created
Step 2: Select the feature point to be tracked in the object of interest.
Step 3: Get the 2d image position (feature point) of the point in image I (u).
Step 4: calculate the optical flow (d) and mean displacement of each co-ordinates of the point.
Step 5: Check if the median displacement is greater than the threshold
Step 6: If the median displacement is greater than the threshold, discard the point else
Calculate the new position of the image point in the image J; = +
Step 7: Apply step 3-4 on each feature point selected.
END
Oral Maxillofac Surg
for the proposed solution which includes a range of features
from the base model. In addition, it proposes a noise filtered
video frame and occlusion removal based on an enhanced
TLD algorithm to overcome the noise and occlusion problem
in augmented reality-based corrective jaw surgery. This has
improved the tracking of the jaw which in turn improves the
image registration through better tracking and detection of the
region of interest (jaw).
Furthermore, features from the second-best solution were
adapted to improve the processing time through a highperformance optical camera as shown in Fig. 4 [18] eliminating the need for regular re-calibration and maintenance of the
stereo camera (Fig. 2). This also improves the processing time
by removing the image conversion process of the 3D stereo
video images frame to a 2D image video frame used in TLD
(Fig. 4). Using the optical camera will improve the viewing
angle for the surgeon by not limiting the view of the object of
interest (jaw) to only one defined angle. The optical camera
captures the real-time surgical video with a single highdefinition camera and follows the remaining state-of-the-art
solution.
Fig. 4 The proposed AR solution
We propose an enhanced video frame with noise removal
and an enhanced TLD with an occlusion removal system to
remove noise in real-time video frames and occlusion in the
tracking and detecting phase. This will improve the tracking
accuracy and also the augmented accuracy by reducing the
overlay error to 0.23~35 mm compared to 0.35~0.45 mm.
The processing time was improved from 6~11 frames per
second to 8~12 frames per second.
TLD algorithm history
TLD also known as tracking-learning-detection was developed by [22]. The recorded 2D video, in the intra-operative
phase, needs to be sent through a TLD algorithm to find and
segment the exact location of the surgical area with the help of
bounding box tracking to reduce the search area and speed up
the process. However, TLD cannot handle the full occlusion
and movement of the object of interest from a frame and
terminates when these two factors occur [7]. The TLD has in
recent years undergone significant re-development to improve
it, yet the TLD algorithm remains a major subject for research.
Oral Maxillofac Surg
The TLD algorithm has three parts: tracking, learning, and
detection.
Area of improvement
The proposed modification focuses on the first stage of
TLD—tracking, to improve tracking accuracy of the object of interest for image registration. So far, the TLD
algorithm has failed if the object of interest is occluded
or exits the video frame, unable to deal with the loss of
the object of interest, proposing the wrong estimate of the
position of the object of interest and finally failing. The
TLD algorithm requires re-initiation after failure because
feature points get lost once the tracker fails and needs to
be assigned again to track. This slows the algorithm process as a search for feature points takes longer when the
object is occluded.
The proposed system consists of three major parts as
shown in Fig. 4 also known as the pre-operative environment,
intra-operative environment with noise removal, and tracking
with occlusion removal and RMaTV algorithm-based pose
refinement.
Pre-operative environment
In this pre-operative environment, a CT scan of the patient is
taken and segmented. A hierarchy of the segmented image
model is created so that the real-time images can be matched
with these aspect graph images from different angles and perspectives. The CT scan is used because it is superior to other
medical images in providing details and information about
bones and nerves.
Intra-operative environment using the optical camera
An optical camera is used to capture real-time videos during
surgery. This reduces the processing time by eliminating the
need to convert the 3D video frames into 2D video frames that
are used in TLD tracking. It also eliminates the necessity of
regular re-calibration and maintenance of the stereo camera
(Fig. 4). Furthermore, it improves the processing time by removing the image conversion process of the 3D stereo video
images frame to 2D image video (Fig. 4). With the use of the
optical camera, the viewing angle of the object of interest
improves as it does not limit the viewing angle the way the
optical camera does.
When the video frames are generated from the video,
they contain noise from the vibration of surgical machinery, the image sensors used, the sensors used to monitor
the patient’s health condition, and also the movements of
the patient. This noise deteriorates the quality of the image, especially the object edges present in the video which
eventually affects the image registration accuracy. Thus, a
modified kernel non-local means (MKNLM) noise filter
proposed by [23] is used to eliminate the noise from the
real-time video frames. MKNLM filter is a robust filter
and removes the noise from the image frame because it is
insensitive to outliers and produces accurate results on a
consistent basis. The use of the MKNLM filter in the
system removes the noise from the image frame and preserves the image edges from deterioration which plays a
vital role in the image overlay process (Appendix).
A hierarchy model of the image is created in 5 level images
with respect to the resolution. The highest level of the image
(lowest resolution) is sent to the TLD for tracking and detecting of the object of interest for online matching. The lowest
resolution image is used in TLD to decrease the processing
time. High-resolution images take more time to process in
comparison with the low-resolution image. The TLD searches
for the bounding box and matches it with the aspect graph
from the offline phase. The bounding box is used to decrease
the search area for an object of interest within the image frame
and hence speed up the online matching. It also reduces the
possibility of matching with an object that is outside the
bounding box. This matching process continues until the lowest level image is found. However, if the bounding box cannot
be found, the occlusion removal technique is applied which is
described below.
Occlusion removal using image reconstruction-based
technique
The TLD algorithm method of tracking and detecting an object of interest uses the feature point within the bounding box
from the initial frame to track the same feature point in the
next image frame with the help of a Lukas–Kanade median
flow tracker [22]. This tracker fails if the object of interest is
occluded.
With the proposed solution, the failure of the tracker can be
prevented by reconstructing the occluded object of interest as
shown in Eq. 3 below.
Proposed equation
Image reconstruction-based occlusion removal uses a technique based on image pixel classification by [21] Image pixel
classification output will be either the object of interest image
or not. It is the key to achieve the high-quality image reconstructed after the occlusion has occurred.
Input is the superposition of the shifted elemental image,
center position to the image sensor, the number of the pixel
from the object to reconstruct the number of the pixels that
have been occluded, the weight of the image pixel in the range
of [1,0] depending on the pixel belonging to the object of
interest or not.
Oral Maxillofac Surg
Table 2
Filtered video frame and enhanced TLD with occlusion removal
Algorithm: Proposed Filtered Video Frame and Enhanced TLD with Occlusion Removal
Input: Two Image Frame, image1(I) & image2 (J)
Output: Noise free and occlusion removed image frame
BEGIN
Step 1: Get the lowest resolution image frame from the hierarchy of image
Step 2: Apply MKLMN noise removal filter to the image frame.
Step 3: Get the 2d image position (feature point) of the point in the image I (u).
= ( , ), where =image point, x= image point position in x-axis
and y= image point position in y-axis
Step 4: calculate the optical flow (d) and mean displacement of each co-ordinates of the point.
Step 5: Check if the median displacement is greater than the threshold
|>10 pixel (threshold).
If | −
Where =displacement of single pint in optical flow,
= mean displacement
Step 6: If the median displacement is greater than the threshold, apply occlusion removal algorithm
1
+
1
, +
1
+
1
, +
1
Step 6: Calculate the new position of the image point in the image J
= +
where V= new position for the feature point;
u=original position of the feature point in Image I
d= optical flow of the feature point
Step 7: Repeat Step 3-7 for each feature point
The weight of the image pixel is classified into 1 and 0
to eliminate the use of any pixels from the occluding
object as the use of this pixel will result in wrong image
reconstruction of the occluded object of interest leading to
a higher image overlay error. The pixel is determined as
the object pixel if the statistical variance of the pixel is
greater than the threshold and hence the weight of the
pixel is determined as 1.
For each of the pixel position that has been occluded that
belongs to the object of interest, the following equation can be
used to reconstruct the occluded image (J) hence removing the
occlusion as Eq. 3.
J¼
1 N M
∑ ∑E
H i¼1 j¼1
1
1
1
xþ
Cx; y þ
C y *W x þ
M0
M0
M0
Cx; y þ
The following Eq. 4 is used to calculate the number of
pixels relating to the object using the technique proposed by
[21].
N
1
1
H ¼ ∑ W xþ
ð4Þ
Cx ; y þ
Cy
M0
M0
i¼1
where, W=weight in [1,0], 1 if the pixel belongs to object
else 0; (x, y) = point belonging to object; i= ith column of
the sensor; j= jth column of the sensor; N, M = sample
intensities of the point.
Equation 5 is used to calculate the weight of pixel in terms
of 0 and 1
w ¼ 1 if v < t else 0
ð3Þ
where v is the statistical variance of the pixel being calculated
which is given as Eq. 6 and t is the pre-defined threshold [21]:
1
Cy
M0
where, E = superposition of shifted elemental image; Cx =
center position of the image sensor in x; Cy = center position
of the image sensor in y; H = pixel numbers from object class
which is calculated as Eq. 4:
N
∑
v¼
ð5Þ
N
∑ E2
i¼1 j¼1
NM
ð6Þ
Oral Maxillofac Surg
where E = statistical mean of the pixel being calculated
which is calculated as Eq. 7 [21].
E¼
N
N
∑
∑ I
i¼1 j¼1
NM
ð7Þ
where I is the intensity of the pixel being calculated.
Why image reconstruction-based occlusion removal?
Image reconstruction-based occlusion removal is a simple
technique of reconstructing the occluded object of interest
by removing the occlusion through classifying the image pixel
as the object of interest (image pixel) or the pixel that does not
belong to the object of interest. With the aid of this algorithm,
the occluded object will be reconstructed so that the TLD does
not fail.
The current Lucas–Kanade median optical flow tracker
(Eq. 2) cannot handle the occlusion of the object of interest.
The current TLD employs Lukas–Kanade median flow
tracker for tracking the object of interest. The tracker fails
when it cannot track the displacement of the feature point that
it is tracking in the object of interest. This results in failure to
re-initialize and, thus, in a failure of the entire trackinglearning-detecting process, slowing down the whole process.
The additional time required for the tracker to search for the
feature point after the occlusion occurs and the time required
for re-initialization time after failure results in an increase of
the processing time. The proposed system detects the tracker
failure with the help of calculating the median displacement of
the feature point. If the median displacement is greater than
the threshold (generally 10), the tracker detects an occlusion
and fails.
After the detection of the occlusion, the proposed system
(Table 2) based on imaged reconstruction-based occlusion removal is executed and removes the occlusion before the
tracker fails. This not only helps to improve the tracking and
detecting accuracy, it will finally help the registration accuracy. It further reduces the time processing time by removing the
prolonged time required for the tracker to search for the feature point after the occlusion occurs and the re-initialization
time after its failure. The processing time of the system will
also increase with the use of the optical camera as it removes
the system overload of converting 3D images into 2D image
frames for tracking and detection. Also, the image overlay
accuracy increases with the help of noise removal by the modified kernel non-local means (MKLMN) filter proposed by
[23].
With the assistance of the proposed method (Table 2
explains the steps involved), the MKLMN filter removes the
noise from the live image frame and removes the occlusion
through image reconstruction-based occlusion removal after
the tracker detects the occlusion whereas the state-of-the-art
solution has no noise removal technique in live video frames
and the TLD method has no ability to remove the occlusion
and, thus, fails when the occlusion occurs. The proposed system can produce an image overlay error of 0.23~35 mm compared to 0.35~0.45 mm produced by state of the art in a jaw
image overlay. Furthermore, the proposed system is able to
achieve a processing speed of 8–12 frames per second compared to 6–11 frames per second achieved by the state of the
art.
The proposed filtered video frame and enhanced TLD with
occlusion removal system are presented in Table 2 and the
flowchart is illustrated in Fig. 5.
RMaTV algorithm-based pose refinement
Rotational matrix and Translation vector algorithm (RMaTV)
proposed by [3] is used to eliminate the geometric error with
the help of rotational and translation vector. ICP used to register the images produces a higher image overlay accuracy
with the use of the RMaTV algorithm. RMaTV algorithm
helps to eliminate the estimation of the wrong pose hence
improve the image registration accuracy.
Start
Select First Image frame(I)
Remove Noise using
MKLMN Algorithm
Select Feature Points
Calculate the optical flow
and mean displacement
Display Bounding
Is
mean
displacement >
threshold
N
Calculate the position of the
feature point in next image frame
Y
Occlusion
occurred
Remove occlusion using
image reconstruction technique
Fig. 5 Flowchart for the proposed algorithm
Oral Maxillofac Surg
Table 3
Results for mandibular and maxillary jaw bone
S. No
1.
Sample
details
Lower Left
mandible
(Age-27)
(Male)
Origi
nal video
Process
ed sample
Current solution
Processing
Accurac
time
y by
(Frames per
overlay
second)
error
0.42
mm
0.45
mm
3.
Lower
Right Mandible
(Age-37)
(Male)
Lower
Frontal
Mandible
(Age-42)
(Male)
Image overlay
9 fps
If patient moves?
7 fps
If surgical tools move?
6 fps
0.7 mm
7 fps
0.25mm
10 fps
0.33
mm
8 fps
0.32mm
7 fps
7 fps
0.7 mm
7 fps
0.36mm
Image overlay
9 fps
0.25mm
10 fps
0.42mm
If patient moves?
8 fps
0.30mm
8fps
0.44mm
If surgical tools move?
7 fps
0.32mm
8 fps
0.7 mm
7 fps
0.22mm
10 fps
Image registration
0.7 mm
0.38
mm
7 fps
Image overlay
8 fps
If patient moves?
9 fps
0.28
mm
9 fps
If surgical tools move?
7 fps
0.30mm
8 fps
0.7 mm
Image registration
7 fps
0.7 mm
7 fps
0.36mm
Image overlay
9 fps
0.26mm
10 fps
0.41
mm
Lower
posterior
Mandible
(Age-17)
(Male)
7 fps
Image registration
0.7 mm
0.35
mm
4.
Processin
g time
(Frames
per second)
Image registration
0.7 mm
0.35mm
2.
Proposed solution
Process
Accurac
ed sample
y
by
overlay
error
0.38
mm
0.43
mm
If patient moves?
6 fps
If surgical tools move?
7fps
0.27
mm
8fps
0.31
mm
8 fps
Oral Maxillofac Surg
Table 3
(Continued)
5.
Upper
Frontal Maxilla
(Age-7)
(Male)
Image registration
0.7 mm
7 fps
0.7 mm
7 fps
0.22mm
10 fps
Image overlay
0.34mm
0.38
mm
0.42
mm
6.
Upper Right
Maxilla
(Age-18)
(Male)
0.7 mm
0.41
mm
0.43
mm
Upper
Frontal Maxilla
(Age-15)
(Male)
0.7 mm
0.40
mm
0.44
mm
Upper left
Maxilla
(Age-32)
(Female)
If surgical tools move?
7 fps
0.27
mm
9fps
0.31
mm
8 fps
7 fps
Image overlay
9 fps
If patient moves?
7 fps
If surgical tools move?
6 fps
0.7 mm
7 fps
0.27mm
10 fps
0.30
mm
8fps
0.31mm
8 fps
0.7 mm
7 fps
0.25mm
12 fps
Image registration
0.36
mm
8.
If patient moves?
9 fps
Image registration
0.39mm
7.
10 fps
0.7 mm
0.37
mm
0.41
mm
0.45
mm
7 fps
Image overlay
10 fps
If patient moves?
9 fps
0.28
mm
9 fps
If surgical tools move?
8 fps
0.32mm
8 fps
Image registration
7 fps
0.7 mm
7 fps
Image overlay
9 fps
If patient moves?
6 fps
If surgical tools move?
7 fps
0.26
mm
10 fps
0.29
mm
8fps
0.35
mm
8 fps
Oral Maxillofac Surg
Table 3
(Continued)
9.
Upper
anterior
Maxilla
(Age-67)
(Male)
Image registration
0.7 mm
0.35mm
0.39
mm
0.43
mm
10.
Upper
posterior
Maxilla
(Age-42)
(Female)
7 fps
Image overlay
9 fps
If patient moves?
7 fps
If surgical tools move?
5 fps
0.7 mm
7 fps
0.25mm
10 fps
0.28
mm
8fps
0.32
mm
8 fps
Image registration
0.7 mm
0.32
mm
7 fps
Image overlay
11 fps
0.7 mm
7 fps
0.42mm
12 fps
If patient moves?
0.36
mm
0.43
mm
Results
The proposed model was implemented in the MATLAB
R2017b [24]. The model was implemented with the use of
10 video samples and 10 CT scan samples from a various
age group in maxillary and mandibular jaw bones
(Table 3). The videos and CT scan samples were gathered
from various sources that are available online for research
and study purposes. The image overlay accuracy and
Fig. 6 Grouping of the jaw bone
9 fps
If surgical tools move?
7 fps
0.25
mm
0.31mm
9 fps
8 fps
processing were calculated to measure the performance
of the system proposed. The jaw is divided into lower left
mandible, lower right mandible, lower front mandible,
upper right maxilla, upper left maxilla, and upper frontal
maxilla.
The proposed system works in three main stages: the preoperative stage, the intra-operative stage with noise removal
and tracking (with occlusion removal), and pose refinement
with RMaTV. In the intra-operative phase, the CT scan images
Oral Maxillofac Surg
Fig. 7 a Real-time image. b Image registration. c Image overlay
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
initialization after this process. Thus, the image
reconstruction-based occlusion removal that uses pixel
classification to identify the object of interest (or not) is
used to remove the occlusion and improve the quality of
the reconstructed image. A bounding box is detected
which defines the object of interest through the hierarchy
of images. The bounding box helps to decrease the search
area and processing time. Ulrich’s method is used for the
initial alignment. This method eliminates the need for initial manual registration and decreases the possibility of
human error. This helps to generate the 2D pose object
of interest. Once the best image is found, it is sent for 3D
pose refinement with the use of the ICP algorithm.
In pose refinement with RMaTV phase, a rotational matrix
and a translation vector are used to eliminate the geometric
error. This algorithm is used to eliminate the possibility of
wrong pose selection, hence improving the image overlay
and registration accuracy (Fig. 7).
Sample videos and images were used in Matlab 2017b [24]
to simulate the state-of-the-art system and the proposed system. A range of reports and graphs were generated in terms of
accuracy and speed to evaluate and compare the state-of-theart with the proposed system. The comparison graphs are
displayed below. Figure 8 compares features of the mandibular jaw in terms of image overlay accuracy and Fig. 9 draws a
comparison in terms of the processing time. Similarly, Fig. 10
Processing in terms of
frames per second
Accuracy in terms of
overlay error (mm)
of the patient are collected and segmented as per the object of
interest (jaw). These segmented images are used to create the
aspect graph that contains images with various models, used to
match with the online images from a different perspective (angles). The generation of the aspect graph depends on the type of
camera parameters used and the jaw model used (Fig. 6). In our
case, the aspect graph generation was less than 45 s.
In the intra-operative phase with noise removal and
tracking with noise removal, optical cameras are used to
capture the live surgical video. The video frames are generated from the real-time video and MKLMN filter is used
to remove the noise from the image, improving the edges
of the objects and preventing contour leakage through denoising. A pyramid of the video frame images is created
based on its resolution. The highest-level image (image
with the lowest resolution) is taken to the TLD for tracking and detecting the object of interest. The tracking
phase of the TLD algorithm uses Lucas–Kande median
flow tracker to track the object of interest and display
the bounding box. The Lucas–Kanade Median flow
tracker uses feature points from the object of interest
and calculates its displacement (optical flow) to track
the same feature points in the next image frame. But if
the object of interest is occluded, then the Lucas–Kande
tracker cannot track the feature points and will eventually
fail, resulting in failure of the TLD as well, requiring re-
Image
registraon
Image overlay
If paent
moves?
If instrument
moves?
12
10
8
6
4
2
0
Image
registraon
Image overlay If paent moves? If instrument
moves?
Different stages
Different stages
State-of-Art
Proposed soluon
Fig. 8 Accuracy results in mandibular jaw bones samples
State-of-Art
Proposed soluon
Fig. 9 Processing time result in mandibular jaw bone samples
Oral Maxillofac Surg
Discussion
Accuracy in terms of
overlay error (mm)
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Image
registraon
Image overlay
If paent
moves?
If instrument
moves?
Different stages
State-of-Art
Proposed soluon
Fig. 10 Accuracy results in maxillary jaw bones samples
and Fig. 11 compare the maxillary jaw bones in terms of
accuracy and processing time between the state-of-the-art
and proposed the system.
The main factor deciding the accuracy and speed of an
augmented reality-based system is the image overlay error
and the processing time. The image overlay error is the difference in the superimposition of the projected offline 2D image
onto the real-time video. The processing time of the augmented reality system is the total number of the image frames
processed by the system in a given timeframe. We calculated
the speed of the system in terms of seconds. The samples that
were collected were simulated in Matlab 2017b using both the
state-of-the-art and the proposed system.
This simulation was carried out for different age groups
ranging from 7 to 67 and a variety of jaw bones were used
(Fig. 6) to simulate and test the accuracy and processing time
of both systems. During the simulation, the proposed system
has been able to achieve a lower image overlay error by
approx. 0.12 mm and improved the processing time by 3–4
frames per second. Test data are presented as bar graph to
compare the accuracy and processing time of state-of-the-art
and the proposed the system.
Processing in terms of
frames per second
12
10
8
6
4
2
Table 3 represents the comparison between the accuracy and
processing time achieved by the state-of-the-art solution and
the proposed solution. The results for accuracy and processing
time are compared in terms of registration, image overlay,
patient movement, and surgical tool movement.
The result achieved by the implementation of the proposed
system, in terms of overlay error, was 0.23~0.35 mm in comparison to 0.35~0.45 mm achieved by the state of art. Furthermore,
the use of an optical camera reduces the system processing time
by eliminating the necessity of conversion of the 3D video frame
to a 2D video frame in the online phase, leading to an increase in
frames per second (8–12) in comparison to 6~11 frames per
second achieved by state of art solution.
An augmented reality system is the combination of a range
of techniques and methods that work simultaneously to provide better AR results and view. When improved techniques
are combined, they form a superior AR system. Our proposed
system uses an optical camera to reduce the time overhead and
a noise removal technique that improves the quality of the live
video frames. The aspect graphs help to match the object in
real-time videos from various perspectives (angles, rotation,
etc.). Initial registration through Ulrich’s method eliminates
human error, possible when the initial registration is done
manually. The use of the TLD helps with long-term tracking
and detection in real-time videos and the use of the bounding
box reduces the search area in the real-time video, reducing
the processing time. The use of RMaTV removes the geometric error and improves registration accuracy and reduces the
image overlay error.
A significant body of research exists in the field of augmented
reality-based surgery, especially in constructive jaw surgery but,
to date, accuracy and processing time remain an area of concern.
This study aimed to improve the current best solution which
produced image overlay accuracy of 0.35~0.45 mm and processing time of 6~11 frames per second.
The proposed method of noise removal and image
reconstruction-based occlusion removal in the TLD was simulated in Matlab to demonstrate that the proposed method can
reduce the image overlay error while positively affecting the
processing time achieved by the state-of-art solution. The current method removes the noise from the real-time video frame
and improves tracking and detection through occlusion removal that improves the image overlay to 0.23~0.35 mm
and achieves a processing speed of 8~12 frames per second.
0
Image
registraon
Image overlay
If paent
moves?
If instrument
moves?
Future research
Different stages
State-of-Art
Proposed soluon
Fig. 11 Processing time results in maxillary jaw bone samples
Future research may be able to improve the other stages of the
TLD including Learning and Detection. Furthermore, the image reconstructed through the image reconstruction occlusion
Oral Maxillofac Surg
removal process could be improved, further improving the
accuracy of the system.
Acknowledgements This work was supported in part by Study Support
Manager Angelika Maag from the Sydney Study Centre of Charles Sturt
University, Sydney, Australia.
7.
8.
9.
Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of
interest.
10.
Ethical approval Not Applicable.
Informed consent Not Applicable.
12.
Appendix
Table 4
11.
Abbreviations for the terms used in the paper
AR
Augmented Reality
ICP
TLD
Iterative closest point algorithm
Tracking Learning Detection algorithm
CT
Computed Tomography
LKMF
RMaTV
MKNLM
Lukas–Kanade Median Flow Tracker
Rotational Matrix and Translation Vector Algorithm
Modified Kernel Non-Local Means Filter
3D
2D
Three-Dimensional
Two-Dimensional
13.
14.
15.
16.
References
1.
2.
3.
4.
5.
6.
Wang J, Suenaga H, Hoshi K, Yang L, Kobayashi E, Sakuma I,
Liao H (2014) Augmented reality navigation with automatic
marker-free image registration using 3-D image overlay for dental
surgery. IEEE Trans Biomed Eng 61(4):1295–1304
Bruellmann D, Tjaden H, Schwanecke U, Barth P (2012) An optimized video system for augmented reality in endodontics: a feasibility study. Clin Oral Investig 17(2):441–448
Murugesan Y, Alsadoon A, Paul M, Prasad P (2018) A novel rotational matrix and translation vector (RMaTV) algorithms: geometric accuracy for augmented reality (AR) in oral and maxillofacial
surgeries. Int J Med Rob Comput Assisted Surg 14:e1889. https://
doi.org/10.1002/rcs.1889
Suenaga H, Tran H, Liao H, Masamune K, Dohi T, Hoshi K, Takato T
(2015) Vision-based markerless registration using stereo vision and an
augmented reality surgical navigation system: a pilot study. BMC Med
Imaging 15(1):51. https://doi.org/10.1186/s12880-015-0089-5
Sielhorst T, Feuerstein M, Navab N (2008) Advanced medical displays: a literature review of augmented reality. J Disp Technol 4(4):
451–467. https://doi.org/10.1109/jdt.2008.2001575
Wang J, Suenaga H, Liao H, Hoshi K, Yang L, Kobayashi E,
Sakuma I (2015) Real-time computer-generated integral imaging
and 3D image calibration for augmented reality surgical navigation.
Comput Med Imaging Graph 40:147–159. https://doi.org/10.1016/
j.compmedimag.2014.11.003
17.
18.
19.
20.
21.
22.
Kalal Z, Mikolajczyk K, Matas J (2012) Tracking-learning-detection. IEEE Trans Pattern Anal Mach Intell 34(7):1409–1422 https://
doi.org/10.1109/tpami.2011.239
Choi H, Park Y, Lee S, Ha H, Kim S, Cho H, Hong J (2017) A
portable surgical navigation device to display resection planes for
bone tumor surgery. Minim Invasive Ther Allied Technol 26(3):
144–150. https://doi.org/10.1080/13645706.2016.1274766
Wu J, Wang M, Liu K, Hu M, Lee P (2014) Real-time advanced
spinal surgery via visible patient model and augmented reality system. Comput Methods Prog Biomed 113(3):869–881. https://doi.
org/10.1016/j.cmpb.2013.12.021
Nakao M, Endo S, Nakao S, Yoshida M, Matsuda T (2016)
Augmented endoscopic images overlaying shape changes in bone
cutting procedures. PLoS One 11(9):e0161815. https://doi.org/10.
1371/journal.pone.0161815
Chen X, Xu L, Wang Y, Wang H, Wang F, Zeng X, Wang Q, Egger J
(2015) Development of a surgical navigation system based on augmented reality using an optical see-through head-mounted display. J
Biomed Inform 55:124–131. https://doi.org/10.1016/j.jbi.2015.04.003
Hung K, Wang F, Wang H, Zhou W, Huang W, Wu Y (2017)
Accuracy of a real-time surgical navigation system for the placement of quad zygomatic implants in the severe atrophic maxilla: a
pilot clinical study. Clin Implant Dent Relat Res 19(3):458–465.
https://doi.org/10.1111/cid.12475
Chen X, Xu L, Wang Y, Hao Y, Wang L (2016) Image-guided
installation of 3D-printed patient-specific implant and its application in pelvic tumor resection and reconstruction surgery. Comput
Methods Prog Biomed 125:66–78. https://doi.org/10.1016/j.cmpb.
2015.10.020
Fitzpatrick J, West J, Maurer C (1998) Predicting error in rigidbody point-based registration. IEEE Trans Med Imaging 17(5):
694–702. https://doi.org/10.1109/42.736021
Schicho K, Figl M, Seemann R, Donat M, Pretterklieber M,
Birkfellner W et al (2007) Comparison of laser surface scanning
and fiducial marker–based registration in frameless stereotaxy. J
Neurosurg 106(4):704–709. https://doi.org/10.3171/jns.2007.106.
4.704
Kilgus T, Heim E, Haase S, Prüfer S, Müller M, Seitel A, Fangerau
M, Wiebe T, Iszatt J, Schlemmer HP, Hornegger J, Yen K, MaierHein L (2014) Mobile markerless augmented reality and its application in forensic medicine. Int J Comput Assist Radiol Surg 10(5):
573–586. https://doi.org/10.1007/s11548-014-1106-9
Zinser M, Mischkowski R, Dreiseidler T, Thamm O, Rothamel D,
Zöller J (2013) Computer-assisted orthognathic surgery: waferless
maxillary positioning, versatility, and accuracy of an image-guided
visualisation display. Br J Oral Maxillofac Surg 51(8):827–833.
https://doi.org/10.1016/j.bjoms.2013.06.014
Wang J, Suenaga H, Yang L, Kobayashi E, Sakuma I (2016) Video
see-through augmented reality for oral and maxillofacial surgery.
Int J Med Rob Comput Assisted Surg 13(2):e1754 https://doi.org/
10.1002/rcs.1754
Gold S, Rangarajan A, Lu C, Pappu S, Mjolsness E (1998) New
algorithms for 2D and 3D point matching. Pattern Recogn 31(8):
1019–1031. https://doi.org/10.1016/s0031-3203(98)80010-1
Ulrich M, Wiedemann C, Steger C (2012) Combining scale-space
and similarity-based aspect graphs for fast 3D object recognition.
IEEE Trans Pattern Anal Mach Intell 34(10):1902–1914. https://
doi.org/10.1109/tpami.2011.266
Xiao J, Gerke M, Vosselman G (2012) Building extraction from
oblique airborne imagery based on robust façade detection. ISPRS
J Photogramm Remote Sens 68:65–68
Kalal Z, Mikolajczyk K, Matas J (2010) Forward-backward error:
automatic detection of tracking failures. In: 2010 20Th international
conference on pattern recognition. https://doi.org/10.1109/icpr.
2010.675
Oral Maxillofac Surg
23.
Kazemi, M., Mohammadi, E., sadeghi, P., & Menhaj, M. (2017). A
Non-local means approach for Gaussian noise removal from images
using a modified weighting kernel. In Iranian Conference on Electr
Eng Tehran: ICEE
24.
He C., Liu Y. and Wang Y. Sensor-fusion based augmented-reality
surgical navigation system. 2016. [Online]. Available: http://
ieeexplore.ieee.org/document/7520404. Accessed 7 Jan 2018