Page 1
Planning safe trajectories in image-guided
keyhole neurosurgery
A thesis submitted in fulfillment of the requirements for the degree of
Master of Science
By
Miri Trope
Supervised by
Prof. Leo Joskowicz Dr. Ruby Shamir
Dr. Yigal Shoshan
The Selim and Rachel Benin
School of Engineering and Computer Science The Hebrew University of Jerusalem
Jerusalem, Israel
July 2012
Page 2
2
Acknowledgements
I would like to thank the important people who helped me with this thesis. First, I deeply thank
my thesis supervisor Prof. Leo Joskowicz for the opportunity to work on such an interesting
innovative topic and for the guidance, the trust and confidence that you have shown in me during
our work together. I thank Dr. Ruby Shamir for his generosity in donating his helpful ideas and
sharing his knowledge in algorithms and neurosurgery. I also thank Dr. Yigal Shoshan for his
noble attitude in developing a method that significantly improves the skills of neurosurgeons, his
unlimited time for our meetings and for motivating his team to our experiments: Dr. Zvi Israel,
Dr. Idit Tamir, Dr. Frenando Ramirez and Samuel Moscovici. I also thank my colleagues Achia
Kronman and Refael Vivanti at the computer aided surgery and medical image processing
laboratory, and all the other members of the laboratory, who were very informative, hands on
helpers and supportive during the whole way.
I would like to thank my parents, Hana and Moshe Trope, for their faith in me and allowing me to
be as ambitious as I wanted. It was under their watchful eye that I gained so much drive and an
ability to tackle challenges head on. This work is as much yours as it is mine.
Page 3
3
Abstract
Many image-guided keyhole neurosurgery procedures require the precise targeting of tumors and
anatomical structures with a surgical tool inside the brain based on pre-operative CT/MRI images.
A misplacement of the surgical tool from the planned trajectory may result severe neurological
complications. Consequently, it is desired to select a trajectory that is located at a safe distance
from critical structures such as blood vessels, ventricles and some of the major functional areas of
the cerebral cortex which are correlated to the motor, sensory, vision and speech activities and are
represented by bundle of fibers and functional MRI.
We present a novel preoperative straight trajectory planning method that calculates the safest
trajectory of the region of interest in the head surface for image-guided keyhole neurosurgery.
Our method quantifies the risks of multiple candidate trajectories and presents them due to their
associated risk on a color coded head surface to assist the neurosurgeon in selecting the safest
path.
This thesis presents a new software platform including an automatic algorithm for the safest
trajectory planning and brain structures segmentation. For visualization, a color coded head
surface and safest trajectories are presented to assist the neurosurgeon based on a defined target
and a candidate entry point area on the outer head surface on preoperative MRI scans. The
software provides a friendly user interface suitable for the clinical applications including plugins
for each brain structure segmentation, trajectories calculation and evaluation methods.
A retrospective comparative study for a selected target on MRI head scans for five patients
showed a significant reduction in insertion trajectory risk. The safety of our method was
compared with three trainees and a senior neurosurgeon yielding excellent improvement of the
trajectory risk value of less than 25% of the manual trajectories risk values and a minimum
distances difference of 1.6 mm.
The suggested method may result in safer trajectories for complex cases with internal targets,
shorter preoperative planning time and accurate placement of the surgical tool with the farthest
distance from the important brain structures while avoiding possible complications in keyhole
neurosurgery.
Page 4
4
Contents
1. Introduction ………………………………………………………………………..……….…..6
1.1. Background ………………………………………………………………………..….…….6
1.2. Clinical procedures ………………………………………………..…………………..……7
1.3. Keyhole neurosurgery workflow …………………………………………………………...8
1.4. Support system.………..……………………………………………………….…….....…..9
1.5. Thesis overview.……………………………………………………………………....…...13
1.6. Novel aspects.…………………………………………………………………….………..14
1.7. Thesis organization………………………………………………………………...………14
2. Literature review…………………………………………………………………….…...........15
2.1. Planning a safe trajectory in keyhole neurosurgery………………………………...….....15
2.2. Blood vessels segmentation………………………………………………………..…..…..16
3. Method……………………………………………………………………………….………… 9
3.1. Overview…………………………………………………………………………..….…… 9
3.2. Structures segmentation……………………………………………………..…..………... 21
3.3. Calculating a risk map…………………………………………………………..…...……..29
3.4. Calculating and visualizing safe trajectories…………………………………..………..... 31
3.5. Graphical user interface…………………………………………………………..……...…32
4. Experimental results……………………………………………………………….…..…........35
4.1. Datasets description…………………………………………………………………......….35
4.2. Methodology……………………………………………………………….…..…….….….35
4.3. Results……………………………………………………………………….….…..…...…36
5. Conclusion……………………………………………………………………….…..……….…42
5.1. Summary……………………………………………………………………….…..…….…42
5.2. Discussion…………………………………………………………………….……..……...42
5.3. Contribution………………………………………………………………….…..………....45
5.4. Limitations………………………………………………………………….……..…….….46
5.5. Future work………………………………………………………………….……..….……46
6. Bibliography…………………………………………………………………….………..….….47
Page 5
5
List of Figures
1.1 Stereotactic frame 10
1.2 Navigation system 10
1.3 Robotic neurosurgery system 12
1.4 Interventional MRI 13
1.5 Probe eye view: manual versus automatic methods 14
3.1 Flow diagram of the trajectories planning system 20
3.2 Skull segmentation 22
3.3 Ventricles segmentation 23
3.4 Blood vessels segmentation 25
3.5 Segmented functional areas 26
3.6 Fibers bundle 27
3.7 High density fibers segmentation 28
3.8 Risk map 30
3.9 Color coded trajectories and head surface 32
3.10 Segmented structures and color coded head surface 32
3.11 Planning Trajectories - user interface 33
3.12 MITK platform diagram 34
4.1 Internal deviation of the brain, coronal view 36
4.2 Color coded head surface and trajectories for each patient 41
5.2 Comparison of trajectories risk values – manual versus automatic methods 44
Page 6
6
Chapter 1
Introduction
This thesis presents a method for planning safe insertion trajectories of a straight surgical tool in
keyhole neurosurgery. This chapter introduces the topics of the thesis. In section 1.1, we discuss
the background. In section 1.2, we describe the clinical procedures of the image-guided keyhole
neurosurgeries. In section 1.3, we briefly describe image-guided keyhole neurosurgery
procedures and characterize their commons. In section 1.4, we introduce support systems that
allow the implementation of these operations. Section 1.5 provides a brief overview of the
method and the research goals. Section 1.6 discusses the novel aspects of the method and Section
1.7 discusses the organization of the thesis.
1.1 Background
Neurosurgery is an invasive operation of the brain and spinal cord that its goal is to treat and
diagnose neurological disorders. The extent of the invasion is directly related to the operation
risk. Brain surgeries, in which a large part of the skull is removed, or which involve extensive
surgical intra-cranial intervention, are associated with higher complications and mortality rates
than those performed with a small amount of invasiveness. The development of better imaging
modalities and stereotactic surgery allows applying a minimal invasive approach in which a
neurosurgery is at first planned on the patient’s MRI/CT image to minimize the invasiveness
level of the intervention.
The minimal invasive operation is executed in the operating room using a support system.
Usually, the operation consists of a straight surgical tool inserted through a small hall on the
skull. These types of surgeries, called image-guided keyhole neurosurgeries, are the focus of this
thesis. The systems that support image-guided keyhole neurosurgery bridge between the
preoperative plan and the physical head, and allow the accurate execution of the preoperative
plan on the patient. Since instrument misplacement may result in ineffective treatment and/or
severe neurological complications, it is important to study and improve the systems accuracy and
safety.
Page 7
7
1.2 Clinical Procedures
We briefly present the aim and characteristics of various image-guided keyhole neurosurgeries.
They all have four important common properties:
1) They are minimally invasive surgeries (MIS), performed via a keyhole of 3-30mm diameter
opened on the skull.
2) They require precise targeting and mechanical guidance support.
3) The targets and entry points are determined preoperatively in a CT/MRI image.
4) It is assumed that little or no brain shift occurs due to the minimal invasive approach.
We describe the most common procedures next.
Brain biopsy
Brain biopsy is the gold standard for accurately determining tumor pathology. It consists of
harvesting, with a hollow needle a tissue sample, from a predefined target site within the brain so
it can be analyzed in the pathology laboratory. Biopsies are usually done making a small opening
(3-14mm burr hole) in the skull and carefully introducing a biopsy needle with a support
guidance system.
Hydrocephalus treatment
Treatment of congenital or acquired hydrocephalus is indicated for alleviating the abnormal
accumulation of cerebrospinal fluid (CSF) within the brain ventricles. It consists of diverting the
flow of fluid away from the ventricles by inserting a shunt. A valve in the shunt maintains the
CSF at a normal pressure and volume within the ventricles.
Hematoma evacuation
Evacuation of hematoma is a surgery used to reduce intracranial pressure caused by an
expanding bleeding resulting from head injury stroke, or bleeding into a tumor. The goal of the
surgery is to decrease morbidity and mortality and to relieve neurological symptoms. It consists
of placing an aspiration needle in the hematoma and aspirating the blood out.
Page 8
8
Ommaya catheter insertion
‘Ommaya’ catheter insertion surgery is indicated for the on-site repeated delivery of drugs (e.g.,
chemotherapy and antibiotics) to the ventricular system. It is also performed for repeated CSF
sampling and to evacuate cystic lesions in the brain when surgical excision is not appropriate.
The surgery consists of inserting a silicon catheter into the ventricles or cystic cavity. The
catheter is connected to a small silicon reservoir (the ‘Ommaya’ reservoir) implanted under the
scalp so that the reservoir can then be easily approached with a needle puncture through the
scalp.
Deep brain stimulation
Deep brain stimulation (DBS) is a procedure for effectively treating certain types of
Parkinsonism, primary tremor, dystonia, hemibalismus, and thalamic pain. It consists of
implanting, through one or more small skull openings, electrodes into specific targets of the brain
and providing electric stimuli to these locations.
Minimal access craniotomy
Minimal access craniotomy is performed to resect deep brain tumors, vascular malformations,
and brain abscesses. It consists of opening a small circular hole (20-30mm radius) in the skull
and introducing a surgical instrument reach the area of interest.
1.3 Keyhole neurosurgery workflow
All clinical procedures follow a similar surgical workflow that is described in Table 1 [1].
Preoperatively, markers for image-to-patient alignment may be attached to the patient (1a). Then,
the patient head is imaged (1b) and the neurosurgeon selects the preferred target and entry point
(1c). In the operating room, the patient is prepared for the operation and the support system is
installed (2a). Then, the head image is aligned (e.g. registered) with the actual patient head (2b).
Then, the neurosurgeon localizes the entry point and makes the burr hole (2c). A mechanical
guidance is installed and adjusted to support tool’s insertion (2d), and then, the tool is inserted
and operation is performed. If undesired complication has occurred, steps 2b-e can be repeated as
needed.
Page 9
9
Table 1.1: Typical keyhole image-guided neurosurgery protocol.
1.4 Support Systems
We use this protocol to compare current support systems. Four types of support systems for
minimally invasive keyhole neuro-surgery are currently available: 1. stereotactic frames; 2.
navigation systems; 3. robotic systems, and; 4. interventional imaging systems.
Stereotactic frames
Stereotactic frames (Fig. 1.1) provide precise positioning with a manually adjustable frame
rigidly attached to the patient skull. Prior to image acquisition, four frame position screws are
implanted in the patient’s skull. An imaging coordinate box, called indicator, is mounted on the
frame and the patient is scanned with it. The surgeon identifies the brain targets and entry points
on the images and computes the corresponding stereotactic frame coordinates based on the
imaged indicator. Intra-operatively, the stereotactic frame is adjusted according to the computed
coordinates and mounted on the immobilized patient skull at the implanted screws. Keyhole
surgery of the skull opening is then done. Optionally, a linear drive needle insertion guide is
mounted on the frame to automate needle insertion and retraction.
1. Pre-operatively
1a. Pre-imaging preparations – implant skull screws and/or attach skin markers.
1b. Image acquisition – acquire a CT/MRI image.
1c. Planning – elaborate the pre-operative plan and identify targets and entry points.
2. Intra-operatively
2a. Preparation – set up the support system and perform patient preparation.
2b. Registration – align the preoperative plan and image with the physical head.
2c. Localization – locate the entry point with a tracked tool and perform incision.
2d. Guidance – provide mechanical guidance for the needle/probe insertion.
2e. Insertion – insert the needle to a planned depth at the proper speed/force.
2f. Repeat steps 2b-e as necessary.
Page 10
10
(a) (b)
Figure 1.1: A stereotactic frame: preoperative (a) and intra-operative (b) setups.
(a) (b)
Figure 1.2: A navigation system allows (a) the manipulation of a tracked surgical tool with (b)
real-time localization feedback on the preoperative patient’s head image.
Page 11
11
Navigation systems
Navigation systems (e.g., Medtronic, USA and BrainLab, Germany) show in real time the
location of hand-held tools of the pre-operative image onto which targets have been defined
(Figure 1.2). The registration between the preoperative data and the patient is performed via skin
markers affixed to the patient’s skull before scanning, or by acquiring points on the patient’s face
with either a laser probe or direct contact. Augmented with a manually positioned tracked
passive arm (e.g., Phillips EasyTaxis™ or Image-Guided Neurologics Navigus™) they also
provide mechanical guidance for targeting. Since nearly all navigation systems use optical
tracking, careful camera positioning and the maintenance of a direct line of sight between the
camera and tracked instruments is required at all times. Their disadvantages are: 1) cost (at least
US$150,000); 2) require head immobilization; 3) require the maintenance of a line of sight
between the camera and the instruments, and; 4) require manual passive arm positioning, which
can be time-consuming and error-prone; The main advantages of navigation systems are:1)
provide continuous, real-time surgical tool location information with respect to the defined
target; 2) allow the selection of new target points during surgery; 3) are quickly gaining wide
clinical acceptance since their introduction in the early 1990s.
Robotic systems
Robotic systems (Figure 1.3) provide frameless stereotactic neurosurgery with a robotic arm that
automatically positions itself with respect to a target defined in the preoperative image. They
have the potential to address intra-operative localization, guidance, and insertion with a single
system. The registration between the pre-operative image and the intra-operative situation is
done by direct contact or with video images. Two floor standing commercial robots include the
NeuroMate™ (Integrated Surgical Systems, USA – now defunct) and the PathFinder™
(Armstrong HealthCare, UK). Their disadvantages are: 1) bulky and cumbersome; 2) pose a
potential safety risk due to their size and weight; 3) require head immobilization or real-time
tracking; 4) costly (US$300,000 -$500,000); 5) not commonly used, with only a dozen systems
currently deployed. Their advantages are: 1) provides a frameless integrated solution; 2) allow
intra-operative plan adjustment, and; 3) are rigid and accurate.
Page 12
12
Figure 1.3: Robotic neurosurgery system
Interventional imaging
Interventional imaging systems (Figure 1.4) produce images showing the actual needle/probe
position with respect to the brain anatomy and target. A few experimental systems also
incorporate real-time tracking (Stereotaxis, Inc) and robotic positioning devices. The main
advantage of these systems is that they provide up-to-date images that account for brain shift (a
secondary issue in the procedures we are considering), and needle bending. The main drawbacks
are: 1) limited availability; 2) cumbersome and time-consuming intra-operative image
acquisition; 3) high nominal and operational costs, and; 4) for intra-operative MRI, complete
expensive room shielding is required.
Page 13
13
Figure 1.4: Interventional MRI. (Image source: Univ. of California, Irvine, Dept. of
Neurosurgery website).
1.5 Thesis overview
We present a novel preoperative straight trajectory planning method for image-guided keyhole
neurosurgery. Our method quantifies the risks of multiple candidate trajectories and presents
them on the outer head surface to assist the neurosurgeon in selecting the safest path. For
visualization, we color-code all the trajectories according to their associated risk level and
present them all at once on the relevant parts of the outer head surface.
Uniquely, our approach is from the clinic to the method. That is, at first we characterize the
actual tool localization accuracy under the full clinical setup and in-vivo. Then, we develop and
test methods for improving safety. Figure 1.5 shows our goal in a visual aspect: avoiding
trajectories which intersect with blood vessels, ventricles and functional areas and fibers bundle.
We describe new methods to improve surgical tool placement accuracy and patient safety. The
methods were tested in-vivo and under the full clinical setup, and significant improvements in
accuracy and safety were observed.
Page 14
14
Figure 1.5: Probe eye view- manual versus automatic method
Trajectory which intersects motor (blue) and sensory (purple) areas (left) and the calculated
safest trajectory (right).
1.6 Novel aspects
This thesis is unique in its comprehensive approach combining clinical in-vivo experiments data
with novel technical and theoretical work into a set of methods showed to actually improve the
accuracy and safety in image-guided keyhole neurosurgery.
Our method for minimal-risk path planning was designed carefully with the neurosurgeon such
that safe trajectories can be conveniently selected and refined with a visual and quantitative
feedback. This method is unique in its visualization, risk formula, updated quantitative feedback
interface, and in incorporating tool localization uncertainty model. Moreover, a neurosurgeon
tested the method on clinical data and it was shown that trajectories planned with our method are
safer than those planed with the routine method.
1.7 Thesis organization
The thesis is organized as follows. Chapter 2 presents a survey of planning safe insertion
trajectories of a straight surgical tool in keyhole neurosurgery and brain’s blood vessels
segmentation techniques. Chapter 3 provides an overview of the new method. Chapter 4
describes the experimental validation and presents the results. Chapter 5 concludes with a
summary of contributions, limitations and possible future work.
All our methods were tested under the full clinical setup and showed to improve accuracy and
safety in image-guided keyhole neurosurgery.
Page 15
15
Chapter 2
Literature review
This chapter reviews methods for preoperative planning in keyhole neurosurgery and blood
vessels segmentation. Section 2.1 provides a brief summary of recent studies describing planning
a safe insertion trajectory of a straight surgical tool in keyhole neurosurgery (Table 2.1). Section
2.2 reviews methods for blood vessels segmentation.
2.1 Planning a safe trajectory in keyhole neurosurgery
Vaillant et al [5] quantify the risk of a candidate insertion trajectory based on the sum of the
intensity values of critical brain structure voxels that are intersected by the insertion trajectory,
weighted by their associated importance. The drawback of this approach is that it ignores the
distance of critical structures from the insertion trajectory, so the relative potential for incorrect
surgical tool placement is factored into the risk calculation. Lee et al [4] describe a method that
fuses MRI head scans with a registered atlas to support manual trajectory selection based on 3D
visualization of brain structures. Its main drawback is that the insertion trajectory is selected
manually without any quantitative information regarding nearby critical structures for a specific
patient. Tirelli et al [2] describe a method that assigns a risk value to each candidate insertion
trajectory based on a weighted sum of various factors. Its drawbacks are that it provides a single
risk measure for each trajectory and there is no risk visualization. Brunenberg et al [3] observe
that weighted sum techniques for computing trajectory risk can be misleading, and propose to
compute a maximum risk value for each voxel based on the Euclidean distance of the trajectory
from critical brain structures. Their method produces tens to hundreds of trajectories whose
distances are above a predefined safety threshold. Although this method significantly reduces the
number of candidate insertion trajectories, it still leaves a considerable amount of manual work
to the surgeon, without quantitative feedback for trajectory selection. Navkar et al [1] describe a
method that takes into account the minimum acceptable distance between an insertion trajectory
and the closest blood vessel, as well as the maximum allowed insertion trajectory length, and
shows safe entry point zones on the outer head surface with respect to these criteria. However, as
Page 16
16
in Brunenberg et al [3] the method requires a considerable amount of manual work from the
neurosurgeon and does not provide quantitative feedback. Note that none of the above studies
quantify trajectory risk reduction or time that may be saved with a more automated selection
process in comparison with manual trajectory selection. Ruby et al [6] describe a method which
quantifies the risk of candidate insertion trajectories on the surgeon defined target and candidate
entry point areas on the outer head surface extracted from preoperative MRI scans. For
visualization and ease of selection, the candidate entry points in the outer head surface areas are
color coded due to their risk. Their method presents as an output just the safest trajectory,
although further calculations can present the safest trajectory in each region on the surface head.
Our method planes safest trajectories by segmenting structures such as blood vessels, ventricles,
functional MRI and high density fibers bundle, calculating a risk map which is the weighted sum
of all those structures and computing for each optional entry point the summarized risk value.
The safest entry points are selected from each region on the surface head and for visualization a
color coded trajectories and head surface are presented.
2.2 Blood vessels segmentation
Vessels segmentation is of great importance, since every organ in the human body has blood
supply. We can divide blood vessels segmentation algorithms into three main categories:
(1) Pattern recognition methods: In the vessel extraction domain, pattern recognition
techniques are concerned with the automatic detection of vessel structures and the vessel
features. Chwialkowski et al [7] preformed contour detection approach, using multi-
resolution analysis based on wavelet transform. Tozaki et al [8] used a skeleton-based
approach which extracts the bronchus and blood vessels from CT scans of the lung. As a
first step, a threshold was used to segment the scans. Then, blood vessels and bronchus
were differentiated by using their anatomical characteristics. Finally, a 3D thinning
algorithm was applied to extract the centerline, which gave the skeleton of the blood
vessels. The resulting skeleton was used to analyze and classify the blood vessels.
Region growing methods can also be considered as pattern recognition techniques.
Schmitt et al [9] added to the growing regions techniques a cavity filling process to add
the cavities missed during seeded region growing process.
Page 17
17
The most popular pattern recognition approach is matching filters. Matching filters
approach convolves the image with multiple matched filters for the extraction of objects
of interest. Sato et al [10] introduces a 3D multi scale line enhancement filter for the
segmentation of curvilinear structures in medical scans. The 3D line filter was based on
the directional second derivatives of smoothed scans using Gaussian kernel using multi
scales with adaptive orientation selection using the Hessian matrix.
(2) Model based methods: In model based methods, explicit vessel models are used to
extract the vasculature. This can include wide spectrum of models, from deformable
models, or active contour to template matching. Kass et al [11] were the first to use active
contour models to extract vessel boundary. The local model, with variable stiffness
parameters, locates the smooth at the location where edge are missing. Edge segments are
extracted using directional gradient information. Level sets methods can also be classified
as model based approach. Caselles et al [12] use propagating interfaces under a curvature
dependent speed function to model anatomical shapes using level sets method. Krissian
et al [13] developed a multi scale model: first, they created a skeleton from local
maximum image, and then extended the skeleton to a full segmentation by using a new
response function which measures the contours of the vessels around the centerlines.
(3) Tracking based methods: Tracking based methods apply local operators on a region
containing a vessel and track it. Starting from an initial point, they detect vessel
centerlines or boundaries by analyzing the pixels orthogonal to the tracking direction. A
sophisticated approach on vessel tracking is the use of graph representation as described
by Freiman et al [14]. The segmentation process is reduced to finding the optimum path
in a graph representation of the aortic arch and carotids arteries from CTA scans. The
method starts with morphological based segmentation of the aorta and the construction of
a prior intensity probability distribution function for arteries. The carotid arteries are then
segmented with graph min-cut method based on a new edge weights function that
adaptively couples the voxel intensity, the intensity prior and geometric vesselness shape
prior.
Page 18
18
METHOD
Risk
computation
method
Multiple
trajectories risk
visualization
Multiple risk
parameters
computation
Safety zone
sleeve
computation
Experimental
validation
methodology
Vaillant
et al. 5
weighted sum yes no no qualitative
Lee et al. 4
No no no no qualitative
Tirelli et
al. 2
weighted sum no no no partially
quantitative
Brunenberg
et al. 3
Maximum
yes
partial
no
qualitative
user evaluation
partially
quantitative
Navkar et
al. 1
maximum +
trajectory
length
yes no no qualitative
user evaluation
partially
quantitative
Ruby et al. 6 maximum +
weighted sum
yes yes yes qualitative
user evaluation
quantitative
Our Method summation +
weighted sum
yes yes yes qualitative
user evaluation
quantitative
Table 2.1: A comparison of straight insertion trajectory preoperative planning methods for
image-guided keyhole surgery. The methods are compared with respect to five categories: (a) the
insertion trajectory risk computation method; (b) multiple trajectory risk visualization; (c)
multiple risk parameter computation, including insertion trajectory length, distance to closest
blood vessel, etc.; (d) insertion trajectory safety zone and its visualization, and; (e) experimental
validation, where the term qualitative indicates visual inspection to evaluate insertion trajectory
safety, user evaluation indicates that the user opinion was reported, and quantitative indicates
that quantitative comparative insertion trajectory risk parameters are reported. Partially
quantitative indicates that only one or two measures were obtained on one image, or that the
surgeon was not involved in the experiment.
Page 19
19
Chapter 3
Methods
This section describes the method for automatic planning of safe trajectories in neurosurgery.
Section 3.1 provides a brief overview of the workflow. Section 3.2 describes the segmentation
method of each structure in the brain. Section 3.3 discusses the risk map calculation and
weighing procedure. Section 3.4 analyses the algorithm for calculating and planning safe
trajectories. Section 3.5 describes the graphical user interface and the interactive medical image
processing software platform.
3.1 Overview
Our preoperative planning system consists of brain internal structures segmentation, calculating
and visualizing trajectories and preoperative planning - a graphical user interface for calculating
automatic trajectories. Our planning method from MRI T1 with gadolinium, functional MRI and
diffusion MRI consists of three steps (Fig. 3.1): The first step is brain internal structures
segmentation such as the ventricles, blood vessels, fMRI and high density fibers. The second
step is calculating risk map due to ranking and weighing process. The third step is calculating
and visualizing color coded trajectories and head surface. The user interface enables the
neurosurgeon to define a target point and number of required trajectories and as an output to
visualize the color coded head surface and trajectories.
Workflow
We propose the following five step preoperative workflow for planning an image-guided keyhole
neurosurgery (Fig. 3.1). (1) MRI datasets are uploaded to the planning station, and the
neurosurgeon defines the target location. (2) Brain structures, e.g. external head skin, blood
vessels, ventricles, functional areas and fibers bundles that the neurosurgeon considers relevant
to the surgery are identified, and are automatically or semi-automatically segmented. (3) Risk
volume is computed by automatically combining the weighted risk values of relevant voxels in
the MRI datasets. (4) Candidate trajectories between the possible entry points and the target are
automatically computed. (5) Candidate entry points on the predefined head outer surface area
and calculated trajectories are color-coded by risk value and displayed to the surgeon.
Page 20
20
Figure 3.1: Flow diagram of the trajectories planning system
Page 21
21
3.2 Structures segmentation
Skull
The skull was made by an automatic segmentation method: all the external non-zero pixels of the
brain were extracted in each axis of a three dimensional MRI image. When the first non zero
pixel was discovered, the last encountered pixel was added to the segmentation.
Delaunay method was calculated in order to generate optimal triangulation. That is, the
maximum interior angle of any triangle is less than or equal to that of any possible triangulation.
One way for detecting if a point D is within a triangle’s circumcircle of A, B, C is to evaluate the
determinant:
> 0
Smoothing was achieved by determining normals for each polygon and averaging them at shared
points. When sharp edges are presented, the edges are splitting and new points are generated to
prevent blurry edges. Shading was achieved by Gouraud’s shading method: the intensities at the
edge of each mesh’s line are calculated from the vertex intensities and the intensities along it.
The interpolation equation is as follows:
The intensity of one pixel can be calculated from the previous pixel according to the increment
of intensity.
ROI extraction
The region of interest had to be extracted from the entire head surface in order to optimize the
planning trajectories algorithm and minimize the run time. Thus the right or left hemisphere was
expelled due to the tumor location and the cerebellar tentorium was used as a lower edge.
Page 22
22
Figure 3.2: Skull segmentation
The red line on the MRI orthogonal slices illustrates the mesh representation of the head surface.
Ventricles
The segmentation method for the ventricles is semi automatic by the seeded region growing
which defined as:
While S = is the set of initial seeds, T is the set of un-allocated pixels which border at
least one of the regions and N(x) is a set of immediate neighbors of pixel x.
We chose that method because as the histogram of the ventricles image shows (Fig. 3.3), there
are two main picks: (1) ventricles and (2) Substantia grisea which is known as the gray matter
and is represented by the medium gray value.
Page 23
23
(a)
(b)
(c)
Figure 3.3: Ventricles segmentation
MRI slice (a) with the segmented ventricles (green) and tumor (purple) showing small frame
of ventricles (right), its histogram (b) and a three dimensional ventricles model (c).
Page 24
24
Blood vessels
The blood vessels segmentation was complicated due to the partial volume affect and over
segmentation as a result of other pixels with high intensity values. As a solution for those
problems we decided to use the Vesselness algorithm which considers the longitude structure
and the high contrast of the dark background and the bright vessels (thanks to the gadolinium).
The majority of the vesselness algorithm is based on the analysis of eigenvalues of the Hessian
matrix of image intensity. The mutual magnitude of eigenvalues is indicative of the shape of the
underlying object. Isotropic structures are associated to eigenvalues having a similar non-zero
magnitude, while vessels present one negligible and two similar non-zero eigenvalues.
Recalling Frangi’s formulation and indicating the eigenvalues of a Hessian matrix as , and
, with , vesselness is defined in 3D as
And α, β, γ are user defined parameters. The constraints on the sign of eigenvalues assume that
vessels to enhance are bright on a dark background. Thus we observed in the output probability
image the compensation for a non cylindrical structure in areas with a high contrast and a manual
correction was required.
Page 25
25
Figure 3.4: Blood vessels segmentation
Functional MRI
The functional MRI matrices were achieved from the functional imaging lab in Hadasa Ein
Karem. The functional areas that we used are: speech, motor, sensory and vision. The extraction
process includes of two main stages: the first stage is merging functional areas from different
activities but same functionality (e.g.: functionality: speech, different activities: visual verb
generation and auditory) and the second stage is assigning a threshold value for a probability
image which represents the probability for a functionality in this observed area.
A threshed image can be well produced from the best fitting threshold value definition which
would yield a non spreading spots in the segmented functional areas image.
Page 26
26
Figure 3.5: Segmented functional areas
Four main functional areas: motor (red), sensory (blue), speech (green) and vision (yellow).
High density fibers
A fibers bundle represents the neuronal tract and was processed by the diffusion MRI imaging.
This section includes of three processes: the first process is converting three dimensional points
in a matrix into polygonal lines that represent a fibers bundle. The second process is generating
an intensity image which represents the fibers density in each voxel. The third process is
thresholding the tract density image in order to achieve a segmented high density fibers image.
Fibers as polygonal lines
Each matrix represents a different functionality and includes of three dimensional points.
The following table presents the fibers bundle composition (from bottom to top):
Structure Type
3D points point
Fiber Line
Bundle of fibers Bundle of lines
In order to achieve a mesh representation each point has to be inserted correspondingly to its
fiber in the corresponding bundle.
Page 27
27
Figure 3.6: Fibers bundle
Representation of the Superior Longitudinal Fasciculus (SLF), a language-related fibers bundle.
Each fiber is colored due to its gradient. The Broka (green) and Wernicke (blue) functional areas
(fMRI) are also language-related sending neuronal massages by the SLF fiber tracts.
The fibers density image
The fibers density image is a normalized probability image that was calculated due to the
intersected number of fibers in a voxel and normalized by the highest number of fibers in a
voxel. The density value at the midpoint is determined by trilinear interpolation
from the eight nearest nodes on the mesh as follows:
Where are the coordinates of the eight nearest grid points.
With that probability image (Fig. 3.7b, left) we were able to estimate the appropriate threshold
which extracts the high density fibers. That threshold factor is extracted from the manipulated
histograms of the tract density image by the following four steps: the first stage is ejecting the
zero values pixels (the background) and calculating the initial histogram (Fig 3.7a, upper left).
The second stage is converting that histogram into a convenient logarithmic scale. The third
stage is smoothing by Gaussian and the last stage is thresholding by the median value twice. In
average, the number of segmented pixels was reduced by six times the initial number of fibers
and their presence in the lowest density value was reduced by ten times.
Page 28
28
(a)
(b)
Figure 3.7: High density fibers segmentation
(a) The Histograms illustrate the process of the high density fibers extraction
(b) To the left an axial slice of the tract density image in JET color map: the red represents a
high density value and the purple a low density value. To the right the fibers before
(white) and after (blue) the high density extraction process.
Page 29
29
3.3 Calculating a risk map
Weighing
Calculating the risk map is a necessary step for planning trajectories. The risk map set a risk
factor due to the influence of the four major structures that were mentioned in the previous
section (structures segmentation): ventricles, blood vessels, fMRI and high density fibers. As the
presence of the segmented structures grows, the risk value grows in each pixel. In a clinical
aspect it means that alerted areas which involve a higher number of significance factors would be
set with a higher risk value.
The following equation represents the risk value for each voxel :
α =
While is the normalized value due to the following decreasing order:
Structure
Blood vessels
fMRI
Fibers
Ventricles
Weighing and ranking functional areas
We offer the neurosurgeon a novel tool in which functional activities could be ranked due to
their significance. The user interface includes of four main functionalities: Motor, Sensory,
vision and speech while the neurosurgeon can choose the order of significance. The moral
human concept behind it is a neurosurgeon can rank functionality due to the patient’s skills and
specialty.
Page 30
30
Figure 3.8: Risk map
The graphical user interface (left) and the risk map image (right). The edges marks represent the
sensory (blue) and the motor (orange) functional areas. That GUI enables to drag and drop
functional areas due to their significance.
The risk map as a vector image
In order to avoid a case in which the trajectories are intersected with blood vessels or ventricles,
a vector image has been selected as the suitable representation.
In this case three components were added to each pixel: two Booleans represented the presence
of blood vessels or ventricles and one float value represented the risk factor as was calculated in
the risk map. This representation enables to retrieval data more efficiently and to decrease the
number of input images in the next stage of calculating trajectories.
Page 31
31
3.4 Calculating and visualizing safe trajectories
The role of this section is calculating a required number of trajectories while each trajectory is
picked from a different region in the head surface. The trajectories and each entry point in the
head surface are then colored due to their risk value.
In the first stage some entry points were rejected due to a geometrical angle constraint: if the
angle between the trajectory and the surface’s normal is higher or lower than 30 – that entry
point is rejected.
The equation that describes the trajectory’s risk value is as follows:
With that calculated trajectory’s risk values each entry point gets its value due to the summarized
risk. The entries points with the minimal risk values for each corresponding trajectory from each
region were colored due to their risk values after normalization. The result is color coded
trajectories and a color coded head surface (Fig 3.9, 3.10).
Note: as mentioned in the calculating risk map section, each pixel in the risk map image
represents three components: the calculated risk value and two Boolean parameters which
indicate whether or not that pixel is crossed by blood vessels or ventricles. Thus, in that
summarized process, an additional check was manipulated: testing the presence of blood vessels
and ventricles in order to reject trajectories that may cause to neuronal damages.
The selected trajectories were picked in each region due to the following equation:
In order to keep fluency, the selected trajectories from each region were sorted due to their risk
value and were visualized in increasing order to the user (Fig. 3.11).
Page 32
32
Figure 3.9: Color coded trajectories and head surface
The colored trajectories are the automatic trajectories while the white trajectories are the doctors’
trajectories and will be discussed in greater detail in the experimental result section.
Figure 3.10: Segmented structures and color coded head surface
The structures segmentation: red- blood vessels, dark purple- motor fMRI, blue- sensory fMRI,
brown- vision fMRI, yellow- vision fibers and light blue- motor fibers.
The color coded head surface: green- low risk, yellow- medium risk and red- high risk.
3.5 Graphical user interface
In order to be the most compatible to the neurosurgeon time limitation, we minimize the number
of input images to one (instead of three, thanks to the risk map vector image) and a head surface
for the optional entry points. The interaction with the user requires one click to determine the
target point and the required number of trajectories. The output is the color coded head surface
and trajectories in three orthogonal views and 3D models.
Page 33
33
Fig. 3.11: Planning Trajectories - user interface
A data manager (left), a coronal slice (second to the left), a 3D model of the color coded risk
map (third to the left) and the required controllers: the definition of the target point and the
number of trajectories (right).
Software platform
The software platform was based on the Medical Imaging Interaction Toolkit (MITK) which is a
free open-source software system for development of interactive medical image processing
software. MITK combines the Insight Toolkit (ITK) and the Visualization Toolkit (VTK) with an
application framework. As a toolkit, MITK offers those features that are relevant for the
development of interactive medical imaging software covered neither by ITK nor VTK.
Mitk platform includes of three main components: modules, plugins and apps and the correlation
between them is presented in Fig. 3.12. The modules component includes of algorithms which
may call to ITK/VTK libraries. The plugins component consists of two types of files: cpp for
managing the inputs that Qt gets such as nodes of images, values etc. For each of those
components CMake defines the source files for the compilation process. The startup project
which is the executable file of the program is generated by the solution file that CMake
generates. As Mitk is an open source program, the correlation with developers is made by Git, a
distribution revision control and code management system. Thus, in each software update – Git
has to be involved in order to keep the web updated.
Our executable version consists of four main plugins: the first for calculating trajectory, the
second for calculating risk map, with the weighing procedure, the third for structures
segmentation (separately for each structure) and the fourth for experimental results.
Page 34
34
Figure 3.12: MITK platform diagram
Page 35
35
Chapter 4
Experimental results
This chapter describes the results of the algorithm evaluation. Section 4.1 describes the data sets
used in the experiments. Section 4.2 describes the methodology and section 4.3 presents the
results numeric and visually.
4.1 Datasets description
We obtained MRI datasets of five patients with different types of neuronal diagnosis (Table 1).
Of these, three targets are located at the left hemisphere of the brain and two in the other
hemisphere. Each of the five datasets consists of two types of MRI image: T1 with gadolinium,
T1 without gadolinium and two types of data matrices that were acquired from diffusion MRI,
for visualizing fibers bundle, and functional MRI, for visualizing functional areas. The MRI
images were acquired at the Hadassa Ein Karem hospital. Slice resolution is 512 x 512 x 256.
The voxel size in the datasets is 0.47 x 0.47 x 1 . The scans show targets in various internal
locations: frontal, parietal, temporal and occipital lobes. For each patient different functional
areas were observed such as the motor, sensory, vision and speech. Table 1 shows the medical
data regarding each patient. The diagnosis was determined by the doctor, the various locations of
the target were mapped into regions due to an internal deviation (Figure 4.1) which separated the
head to hemispheres and lobes. The alignment between functional data and T1 scan was
automatically registered by rigid registration with minimum three points in curved structures.
4.2 Methodology
To evaluate our planning trajectories method to the doctors’ method four surgeons, one expert
and three trainers, defined their entry points selection, one trajectory for each patient while the
target point was defined once for each patient by the surgeon. The following protocol describes x
stages in the experiment. In the first stage each dataset, including the T1 with gadolinium MRI
image and the structures segmentation, were presented to the surgeon. The surgeon defined his
best choice of entry point due to his knowledge and experience while watching the structures
segmentation. With the predefined target point, a surface trajectory was generated and the
surgeon could observe his trajectory selection by the three orthogonal slices in probe eye view
mode. Each of the trajectories per every patient was compared to the automatic trajectories that
our software calculated. The minimum distance was calculated from each trajectory to the
structures segmentation as well as the trajectory’s risk value from the risk map image and the
trajectory length.
Page 36
36
Figure 4.1: Internal deviation of the brain, coronal view
Patient Diagnosis Hemisphere Region fMRI Fibers Alignment
EN LLG Left 4,6 Vision
Speech
Vision
SLF
Yes
SO GLIOBLASTOMA Left 1,3 Vision
Speech
Vision
SLF
Yes
HD OLIGO Right 6 Motor
Sensory
Motor
Vision
Sensory
No
AO KAVANORMA Right 3,4 Sensory
Motor
Vision
Wernicke
Motor
Sensory
Vision
SLF
No
ST LLG left 3,4 Motor
Speech
Motor No
Table 4.1: Datasets of patients
*fMRI of speech is divided into three components: Broka, Wernicke and peripheral areas.
4.3 Results
Structures segmentation
The structures segmentation is an essential stage before the calculating trajectories stage. Thus,
for each patient four main structures were segmented: blood vessels, ventricles, fMRI and fibers
bundles of different functional activities (Table 4.1). The structure segmentation algorithms are
manipulated and can be easily used by the plugins and scripts that we developed. Table 4.2
shows a time analysis of that process in order to achieve one dataset.
Page 37
37
Structure Procedure Tool Time Images
Blood
vessels
Vesselness algorithm and
corrections
Mitk- our developed module
Itk snap
2-3 h 1
Ventricles
Region growing Mitk- our developed plug in 2 m 1
Head
surface
Extraction algorithm
Triangulation, smoothing and
shading
ROI surface extraction
Mitk- our developed plug in
Para view and VTK
Para view
3m
5m
4m
1
High
density
Fibers
Turning points into lines
Selecting areas due to ROI
Tract density image
Thresholding
Matlab
Mitk
Mitk- our developed module
Matlab
1m
2m
1m
1m
2-4
fMRI
Turning points into a binary
image
Selecting areas due to ROI
Matlab
Mitk/Itk snap
1m
1m
2-4
Total Structures segmentation Mitk, Itk and Vtk, Matlab and
Paraview
0.5-2h(bv)+28-
42m
Table 4.2: Segmentation Process of a patient – time analysis
Minimum distance
In order to examine both of the methods, the automatic method and the doctors method, the
minimum distance of each trajectory to each structure was calculated. In order to keep the results
coherent, we calculated dummy values such as the average structure which is the average of all
the structures (blood vessels, ventricles fibers bundle and fMRI) and the average doctor which is
the average of all the doctors (three trainees and an expert). The trainees are Dr1, Dr2 and Dr3
and the senior doctor is the expert (Table 4.3).
Page 38
38
Color coded head surface and trajectories
To assist the neurosurgeon in selecting the safest insertion trajectory, candidate points in the
predefined entry point area on the outer head surface are color coded due to their risk value and
displayed on a 3D image. Each candidate entry point is colored with respect to the risk value that
was computed on its trajectory. Fig. 4.2 shows the color coded head surface for each patient and
the manual (white cylinders, black head) and automatic (color coded) trajectories. The head
surface was subdivided into small polygons such that the resolution would be higher (thus the
sharps colored polygons on the surface). To the left the region of interest (the optional entries
points) with the MRI T1gad image, to the right a view from inside the brain (the target point).
Structure Dr 1 Dr 2 Dr 3 Expert Doctors Auto
BV 1.7 (2.6) 1.9 (1.0) 1.9 (1.2) 2.7 (1.0) 2.0 (0.5) 3.3 (2.4)
ventricles 19.3 (13.4) 19.5 (13.4) 18.5 (12.9) 17.9 (17.6) 18.8 (0.7) 15.9 (12.7)
fMRI 20.4 (10.6) 19.1 (11.3) 18.3 (11.3) 28.0 (9.1) 21.4 (4.5) 22.7 (9.7)
Fibers 12.4 (7.6) 12.0 (7.5) 10.2 (7.1) 12.9 (1.8) 11.9 (1.2) 12.1 (5.8)
Average 13.4 (8.5) 13.1 (5.4) 12.2 (8.1) 15.4 (7.4) 13.5 (1.3) 13.5 (7.7)
Table 4.3: Minimum distance – Each structure for all the patients
Patient Dr 1 Dr 2 Dr 3 Expert Doctors Auto
HD 14.6 (11.0) 12.8 (11.3) 11.7 (10.2)
13.0 (1.4) 12.5 (7.0)
ST 14.3 (9.9) 14.7 (10.8) 13.3 (9.8) 13.3 (8.5) 14.1 (0.7) 11.9 (5.7)
AO 7.1 (6.3) 8.4 (6.2) 7.7 (6.7)
7.7 (0.6) 11.1 (11.0)
EN 14.0 (16.2) 10.4 (15.5) 10.1 (15.1) 14.3 (17.2) 11.5 (2.2) 13.8 (15.8)
SO 17.1 (15.7) 19.4 (14.5) 18.3 (14.7) 18.6 (14.5) 18.3 (0.9) 18.1 (15.1)
ALL 13.4 (11.8) 13.1 (11.7) 12.2 (11.3) 15.4 (13.4) 12.9 (1.2) 13.5 (10.9)
Table 4.4: Minimum distance – Each patient for all the structures
Distance Dr 1 Dr 2 Dr 3 Expert Doctors Auto DIFF
min dist 0.5 (0.9) 1.7 (1.3) 1.2 (1.0) 1.5 (1.2) 1.2 (1.1) 2.8 (2.5) +1.6
max dist 27.3 (9.5) 26.7 (8.4) 25.5 (8.5) 31.6 (9.2) 27.8 (8.9) 27.4 (8.2) -0.4
avg dist 13.4 (3.8) 13.1 (4.2) 12.2 (4.0) 15.4 (2.8) 13.5 (3.7) 13.5 (2.8) 0.0
Table 4.5: Minimum distance – distances of all patients and all structures
Page 39
39
Patient SO: GLIOBLASTOMA, left hemisphere
Patient AO: KAVANORMA, right hemisphere
Page 40
40
Patient ST: LLG, left hemisphere
Sagittal view
Coronal view
Page 41
41
Figure 4.2: Color coded head surface and trajectories for each patient
The white trajectories are the doctors’ trajectories and the colored are the automatic trajectories.
Red represents high risk, yellow medium risk and green low risk.
Sagittal view
Patient HD: OLIGO, right hemisphere
Patient EN: LLG, left hemisphere
Page 42
42
Chapter 5
Conclusions
This chapter concludes the thesis. Section 5.1 summarizes the thesis’s background and goals.
Section 5.2 discusses contributes. Section 5.3 discusses the method’s limitation. Section 5.4
discusses possible improvements and future work.
5.1 Summary
Many image-guided keyhole neurosurgery procedures require the precise targeting of tumors and
anatomical structures with a surgical tool inside the brain based on pre-operative CT/MRI
images. A misplacement of the surgical tool from the planned trajectory may result severe
neurological complications. Consequently, it is desired to select a trajectory that is located at a
safe distance from critical structures such as blood vessels, ventricles and functional areas such
as motor, sensory, vision and speech which are represented by fibers bundle and fMRI.
We have developed software, a fast automatic method for planning safe trajectories based on the
conventional 3D orthogonal planes. A graphical user interface (GUI) for the method was
developed and used as an experimental system by trainers and experienced neurosurgeons.
5.2 Discussion
The large amount of numbers which represent the minimum distance of each trajectory from
each structure, for each patient to each doctor, can be very confusing. In this section we will
explain our method results in order to conclude the properties of our method.
Complexity: The routine method requires the neurosurgeon to summarize in their minds the risk
value of any trajectory. Thus, a large amount of data such as a multiple number of functional
MRI and fibers bundles and internal targets might cause this difficult rise. Thus in these
representing cases the automatic method might be in advantage relatively to the routine method:
Page 43
43
1. Patient AO, in addition to his internal target (Figure 4.1), a large amount of functional
dataset such as fMRI and fibers bundles (Table 4.1), cause to trajectories selection with
the lowest standard deviation relatively to all the other patients. In this case, out method
has an advantage of both a minimum distance to each structure and a standard deviation
relatively to the routine method (Table 4.4).
2. The target of patient EN is located in the most internal relatively to all the other patients.
Thus we consider this case as the most complicated in term of trajectory selection. As the
results show (Table 4.4), this case has the most significance improvement in terms of
both the minimum distance and the standard deviation. Our notion is that this standard
deviation was achieved thanks to the unlimited ability of complex calculations of the
automatic method relatively to the human thinking limitation. Thus, even if the number
of neurosurgeons would be greater the same results would be achieved.
Trajectory risk value
In order to estimate the quality of each trajectory, the risk value was calculated from the risk
map. Each doctor’s trajectory calculated for each patient including the average trajectory of the
automatic method (Fig. 5.2). The tendency of those risk values shows that as the years of
experience grow, the trajectory quality is improved. In addition, the automatic trajectory risk
value achieved the lowest risk value for each one of the patient. An additional observation is the
improvement of the same doctor as the number of tries grows. Thus we can conclude that the
presentation of the structures segmentation in the experiment has a significance impact for the
planning trajectory process. Table 5.1 shows the normalized risk values for each patient which
were achieved by the trainees, the expert and the automatic method. The normalization was made
by the minimal and maximal risk values that calculated for each patient. As the last row presents,
the automatic method achieves a risk value of 25% of the expert’s risk value and 19% of the
average trainee’s risk values.
Page 44
44
Patient Average trainee Expert Auto
HD 0.22
0.06
ST 0.20 0.24 0.04
AO 0.31
0.06
EN 0.14 0.10 0.02
SO 0.17 0.15 0.02
Average patient 0.21 0.16 0.04
Table 5.1: Normalized trajectories risk values
Figure 5.2: Comparison of trajectories risk values - manual versus automatic methods
0.00
0.05
0.10
0.15
0.20
0.25
0.30
0.35
HD ST AO EN SO
Trajectories risk values per patient
Average trainee
Expert
Auto
0.00
0.05
0.10
0.15
0.20
0.25
Average trainee Expert Auto
Trajectories risk values - all patients
Page 45
45
5.3 Contributions
This thesis makes five main contributions to the state of the art:
1. Planning with functional MRI and DTI data: In addition to ventricles and blood
vessels segmentation, we added segmentations of fibers bundle and fMRI for planning
trajectories. The segmentation of fibers bundle was built as poly lines which can yield (a)
a density image as we calculated and extracted the high density fibers (b) a further
experiment such as tracking by the fiber’s direction and finding specific locations in the
brain such as in DBS procedure (described in section 1.2).
2. User interface for segmenting and planning: We designed and implemented GUI with
most of the basic image processing tools that a neurosurgeon need such as three
orthogonal axes, clicking ability for target definition, plugins for structures segmentation
and planning trajectories including calculations for examining trajectories.
3. Weighing and ranking functional activities: We developed a ranking and weighing
ability for risk map generation. This method can be considered as the first step towards a
personalize medicine by weighing major and minor functional activities due to the patient
information such as occupation and will be discussed later (the significance of the speech
ability of a politician patient should be higher than other functional activity).
4. Observing patients with internal targets: As we discussed earlier, the most significant
results were achieved for patients with internal targets due to the human thinking
limitations. Our automatic method has no calculations limitations, thus for any amount of
dataset for each patient in any location, our method plan the safest trajectories from each
region on the head.
5. Color coded head surface and trajectories: The automatic method preserves the risk
values of any optional trajectory in the region of interest. Thus the visualization of color
coded trajectories and head surface can be presented to the neurosurgeon as an indicator
of high a low risks areas due to the weighted risk map.
6. Simplicity and speed of the planning process: The simplicity of the user interface and
the output visualization makes our approach unique among others for planning
trajectories. While other methods require additional input or a lot of manual work to the
surgeon, our approach requires just one click for a target definition.
Page 46
46
5.4 Limitations
The following limitations were studied during the software development:
1. ROI of the output trajectories
Only the surgeon has the overall knowledge of the medical decisions that he has to make
when planning trajectories. The software, on the other hand, selects trajectories due the
region of interest that was inserted. Thus, long trajectories were selected as safest
trajectories even if they don’t acceptable to the surgeon. An idea of how to improve this
topic was to remove trajectories with a longer length of 40% the shortest length. But the
specific number is not sure yet and have to be considered in a further research.
2. Sulcus segmentation
The sulcus considered as a structure in the brain which we can’t cross by a trajectory (like
the blood vessels). Because of this structure wasn’t segmented, we manually checked that
the safest trajectory don’t pass thorough this structure. If we had achieved this
segmentation, we could have expended the vector risk map with this Boolean data.
5.5 Future work
Personalized medicine
Personalized medicine is a young but rapidly advancing field of healthcare that is
informed by each person's unique clinical, genomic, and environmental information. We
suggest a ranking process that optimizes the trajectory selection correspondingly. With
the GUI of the ranking functional activities that we offer, the neurosurgeon can prefer
activities due the patient preferences. Thus, in a further experiment an additional data
about each patient can be collected such as occupation.
Sulci segmentation
As we discussed in previous section, the Sulci wasn’t segmented. A vector risk map
which contains data about Sulci presence in each pixel would reject trajectories at the
early stage.
Page 47
47
Chapter 6
Bibliography
[1] N. V. Navkar, N. V. Tsekos, J. R. Stafford, J. S. Weinberg and Z. Deng. Visualization and
Planning of Neurosurgical Interventions with Straight Access. Proc. 1st Int. Conf. on Information
Processing in Computer-Assisted Interventions. Lecture Notes in Computer Science 6135, N.
Navab and P. Jannin Eds, Springer, pp. 1-11 (2010).
[2] P. Tirelli, E. D. Momi, N. A. Borghese and G. Ferrigno. An intelligent atlas-based planning
system for keyhole neurosurgery. Int. Journal of Computer Assisted Radiology and Surgery, Vol
4 (Suppl. 1), pp. S85-S86 (2009).
[3] E. J. Brunenberg, A. Vilanova, V. Visser-Vandewalle, Y. Temel, L. Ackermans, B. Platel and
B. M. ter Haar Romeny. Automatic trajectory planning for deep brain stimulation: a feasibility
study. Proc. 10th Int. Conf. on Medical Image Computation and Computer Assisted Intervention,
Lecture Notes in Computer Science 4791, N. Ayache and S. Ourselin, Springer, pp. 584-592
(2007).
[4] J. D. Lee, C. H. Huang and S. T. Lee. Improving stereotactic surgery using 3-D
reconstruction. IEEE Engineering in Medicine and Biology Magazine, Vol. 21, pp. 109-116
(2002).
[5] M. Vaillant, C. Davatzikos, R. H. Taylor and R. N. Bryan. A path-planning algorithm for
image-guided neurosurgery. Proc. 1st Joint Conference on Computer Vision, Virtual Reality, and
Robotics in Medicine and Medical Robotics and Computer- Assisted Surgery, Lecture Notes in
Computer Science 1205, J. Troccaz and R. Mosges Eds, Springer, pp. 467-476 (1997).
[6] R. Shamir, L. Joskowicz, I. Tamir, E. Dabool, L. Pertman, A. Ben-Ami and Y. Shoshan.
Reduced risk trajectory planning in image-guided keyhole neurosurgery. Med. Phys. 39, 2885
(2012)
[7] M.P. Chwialkowski, Y.M. Ibrahim, F.L. Hong, and R.M. Peshock, “A method for fully
automated quantitative analysis of arterial flow using flow-sensitized MR images”,Comp. Med.
Imaging and Graphics, vol. 20, pp. 365–378, 1996.
[8] T. Tozaki, Y. Kawata, N. Niki, H. Ohmatsu, and N. Moriyama, “3-d visualization of blood
vessels and tumor using thin slice CT”, in IEEE Nuclear Science Symposium and Medical
Imaging Conference, vol. 3, pp. 1470–1474, 1995.
Page 48
48
[9] H. Schmitt, M. Grass, V. Rasche, O. Schramm, S. Haehnel, and K. Sartor: “An x-ray based
method for the determination of the contrast agent propagation in 3-d vessel structures”, IEEE
Trans. on Med. Img., vol. 21, pp. 251–262, 2002.
[10] Y. Sato, S. Nakajima, N. Shiraga, H. Atsumi, S. Yoshida, T. Koller, G. Gerig, and R.
Kikinis: “3d multi-scale line filter for segmentation and visialization of curvilinear structures in
medical images”, IEEE Medical Image Analysis, vol. 2, pp. 143–168, 1998.
[11] M. Kass, A. Witkin, and D. Terzoopoulos: “Snakes: Active contour models”, Int. J. of
Comp. Vision, vol. 1, pp. 321–331, 1988.
[12] V. Caselles, F. Catte, T. Coll, and F. Dibos: “A geometric model for active contours in
image processing”, Numerische Mathematik, vol. 66, pp. 1–32, 1993.
[13] K. Krissian, G. Malandain, and N. Ayache: “Model-based multiscale detection and
reconstruction of 3d vessels”, In: Institut National de Recherche en Informatique et en
Automatique, Technical Report 3442, 1998.
[14] M. Freiman, N. Broide, M. Natanzon, E. Nammer, O. Shilon, L. Weizman, L. Joskowicz
and J. Sosna: “Vessels-Cut: A Graph Based Approach to Patient-Specific Carotid Arteries
Modeling” In: Modelling the Physiological Human LNCS 5903, pp. 1-12, 2009.
[15] The Insight Segmentation and Registration Toolkit (ITK), http://www.itk.org/.41
[16] The Visualization Toolkit (VTK), http://www.vtk.org/.
[17] R. R. Shamir, L. Joskowicz, S. Spektor and Y. Shoshan. Target and trajectory clinical
application accuracy in neuronavigation. Neurosurgery, Vol. 68, pp. 95-101 (2011).
[18] M. H. Moghari and P. Abolmaesumi. Understanding the effect of bias in fiducial
localization error on point-based rigid-body registration. IEEE Transactions on Medical Imaging,
Vol. 29, pp. 1730-1738 (2010).
[19] J. M. Fitzpatrick, J. B. West and C. R. Maurer, Jr. Predicting error in rigid-body pointbased
registration. IEEE Transactions on Medical Imaging, Vol. 17, pp. 694-702 (1998).
[20] R. R. Shamir and L. Joskowicz. Geometrical analysis of registration errors in pointbased
rigid-body registration using invariants. Medical Image Analysis, Vol. 15, pp. 85-95 (2010).
Page 49
49
[21] R. R. Shamir, L. Joskowicz, S. Spektor and Y. Shoshan. Localization and registration
accuracy in image guided neurosurgery: a clinical study. Int. Journal of Computer Assisted
Radiology and Surgery, Vol. 4, pp. 45-52 (2009).
[22] M. de Berg, M. van Kreveld, M. Overmars and O. Schwarzkopf. Computational Geometry:
Algorithms and Applications, 3rd Edition, (Springer, Berlin, 2009).
[23] R. R. Shamir, L. Joskowicz, L. Antiga, R. I. Foroni and Y. Shoshan. Trajectory
planning method for reduced patient risk in image-guided neurosurgery: concept and
preliminary results. Proc. SPIE Medical Imaging: Visualization, Image-Guided Procedures, and
Modeling, Vol. 7625, pp. 762520I:1-8 (2010).
[24] M. Sindou, J. Auque and E. Jouanneau, Neurosurgery and the intracranial venous
system. Acta Neurochirgica Supplementum, Vol. 94, pp. 167-175 (2005).
[25] M. Sindou. Practical Handbook of Neurosurgery: From Leading Neurosurgeons, Vol.
1, (Springer-Verlag/Wien, 2009).
[26] M. Freiman, L. Joskowicz, N. Broide, M. Natanzon, E. Nammer, O. Shilon, L.
Weizman and J. Sosna. Patient-specific carotid arteries modeling: a model-based
graph-cut approach. Medical Image Analysis, Vol. 15, pp. 125-132 (2011).
[27] L. Joskowicz, R. Shamir, M. Freiman, M. Shoham, E. Zehavi, F. Umansky and Y.
Shoshan. Image-guided system with miniature robot for precise positioning and
targeting in keyhole neurosurgery. Computer Aided Surgery, Vol. 11, pp. 181-193 (2006).
[28] A. L. Simpson, B. Ma, E. C. Chen, R. E. Ellis and A. J. Stewart. Using registration
785 uncertainty visualization in a user study of a simple surgical task. Proc. 9th Int. Conf.
Medical Image Computation and Computer Assisted Intervention, Lecture Notes in
Computer Science 4191, R. Larsen, N. Niessen, D. Hawkes Eds, Springer, pp. 397-404 (2006).
[29] Bériault S, Al Subaie F, Mok K, Sadikot AF, and Pike GB. Automatic trajectory
790 planning of DBS neurosurgery from multi-modal MRI datasets. Med Image Computer
Assist Interv.Vol. 14 (Pt 1), pp. 259-66 (2011).