AUTOMATIC SEGMENTATION OF THE LOWER LIMB ANATOMY Lena Paelinck Student number: 01307508 Promotor: Prof. Dr. Emmanuel Audenaert Copromotor: Prof. Dr. Christophe Pattyn A dissertation submitted to Ghent University in partial fulfilment of the requirements for the degree of Master of Medicine in Medicine Academic year: 2016 – 2018
58
Embed
AUTOMATIC SEGMENTATION OF THE LOWER LIMB ANATOMY · AUTOMATIC SEGMENTATION OF THE LOWER LIMB ANATOMY Lena Paelinck Student number: 01307508 Promotor: Prof. Dr. Emmanuel Audenaert
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
AUTOMATIC SEGMENTATION OF THE LOWER LIMB ANATOMY
Lena Paelinck Student number: 01307508 Promotor: Prof. Dr. Emmanuel Audenaert Copromotor: Prof. Dr. Christophe Pattyn A dissertation submitted to Ghent University in partial fulfilment of the requirements for the degree of Master of Medicine in Medicine Academic year: 2016 – 2018
AUTOMATIC SEGMENTATION OF THE LOWER LIMB ANATOMY
Lena Paelinck Student number: 01307508 Promotor: Prof. Dr. Emmanuel Audenaert Copromotor: Prof. Dr. Christophe Pattyn A dissertation submitted to Ghent University in partial fulfilment of the requirements for the degree of Master of Medicine in Medicine Academic year: 2016 – 2018
III
ACKNOWLEDGEMENTS
After one and a half years of hard work accompanied with laughter and tears, the process of
writing this dissertation comes to an end. Many people have played a crucial role in the
accomplishment of this work and to them I want to express my gratitude.
First of all, I would like to thank the University of Ghent and the department of Orthopaedics
and Traumatology of the UZ Ghent for giving me the chance to become part of a group of
professionals that has been investigating automatic segmentation for years and for giving me
the permission to help them in this research and to write this dissertation.
Secondly, I would like to express my gratitude to my supervisors Prof. Dr. Emmanuel
Audenaert and Prof. Dr. Christophe Pattyn for their guidance. Their insightful advices and
support through these eighteen months played an essential part in accomplishing this work.
Thank you for finding the time to help me in this process of researching and writing.
A special word of thanks to my mentor Dr. Jan Van Houcke is in order. His comments and
directives were essential to establish this work. He encouraged me to put my own ideas into
this dissertation but guided me when I was heading in the wrong direction. I appreciate the
time given into answering my questions, the reading over and making improvements to my
work.
Finally, a profound expression of gratitude towards my parents, sister, family and friends. Their
everlasting support and encouragement not only in this dissertation, but through all the years
of studying medicine, mean a lot to me. Thank you for finding the time to read through this
work and improving it. Without, I would not have been able to accomplish this work. Also, I
would like to say thanks, to my boyfriend, Brent Van Riet, for giving me the strength to complete
this work, for helping me through difficult times and for always being there whenever I needed
it.
A heartfelt thank you to all!
IV
TABLE OF CONTENTS
ACKNOWLEDGEMENTS ......................................................................................................... III
TABLE OF CONTENTS ...........................................................................................................IV
LIST OF FIGURES ....................................................................................................................VI
LIST OF TABLES .....................................................................................................................VI
LIST OF ABBREVIATIONS.....................................................................................................VII
Figure 1: Anatomy of the pelvis ............................................................................................................................... 7 Figure 2: A) Anatomy of the Femur B) Anatomy of the Tibia C) Anatomy of the Fibula ......................................... 8 Figure 3: Anatomy of the Patella .............................................................................................................................. 9 Figure 4: A) Anatomy of the Talus B) Anatomy of the Calcaneus ............................................................................ 9 Figure 5: A) Anatomy of the Lumbar Vertebra B) Anatomy of the Sacrum ........................................................... 10 Figure 6: The Hounsfield scale................................................................................................................................ 12 Figure 7: Illustration of pixels ................................................................................................................................. 13 Figure 8: Common drawbacks on manual and automatic segmentation .............................................................. 14 Figure 9: Illustration of thresholding ...................................................................................................................... 15 Figure 10: Histogram based threshold ................................................................................................................... 16 Figure 11: Histogram of CT values .......................................................................................................................... 17 Figure 12: Illustration of gradient based method .................................................................................................. 18 Figure 13: Illustration of the SSM and the Eigenmodes ......................................................................................... 21 Figure 14: Illustration of the free deformation phase ........................................................................................... 22 Figure 15: Illustration of the Dice Similarity Index (DSI) ........................................................................................ 25 Figure 16: Average shape and appearance of the low- and high-resolution SSM ................................................. 28 Figure 17: Overview of the automatic segmentation pipeline .............................................................................. 31 Figure 18: Illustration of the correspondence feature ........................................................................................... 32 Figure 19: Illustration of solid cortical, cartilage and marrow volumes ................................................................ 32 Figure 20: Illustration of the Hausdorff Distance ................................................................................................... 34
LIST OF TABLES
Table 1: Accuracy and observer variation in segmentation of the respective segments ....... 36
Table 2: Comparison with previously developed automatic segmentation techniques .......... 40
VII
LIST OF ABBREVIATIONS
2D Two-dimensional
3D Three-dimensional
AD Average Distance
ASD Average Surface Distance
ASM Active Shape Model
CAOS Computer Aided Orthopaedic Surgeries
CAT Computed Axial Tomography
CT Computed Tomography
CTA Computed Tomography Angiography
DSI Dice Similarity Index
HD Hausdorff Distance
HT Higher Threshold
HU Hounsfield Unit
LAM Local Appearance Model
LT Lower Threshold
MR(I) Magnetic Resonance (Imaging)
OE Overlap Error
PCA Principal Component Analysis
RMS Root Mean Square
ROI Region of Interest
SD Standard Deviation
SSM Statistical Shape Model
1
ABSTRACT
Background: Image segmentation allows transforming a 2D image, obtained from CT or MRI,
into a 3D image. Hence, image segmentation has a growing interest, especially in diagnosis
and planning of surgery. It gains traction in orthopaedics as well as other specialities like
cardiology and neurology. However, image segmentation is mostly done manual - a laborious
and time-consuming task to be performed by trained professionals. The increase in the number
of segmentations results in a need for automatic segmentation techniques. The aim of this
study is to validate the accuracy of our proposed developed segmentation protocol compared
with the golden standard of manual segmentations. Secondly, the inter- and intra-observer
variability of the manual segmentation will be evaluated.
Materials and methodology: The pipeline of the proposed method for automatic
segmentation is presented. The fully automatic segmentation is divided into three phases.
First, there is an initialization and segmentation phase, separated in four steps. Afterwards,
the training of the shape model is supervised and the segmentation pipeline is continuously
updated. At last the appearance and solids phase occurs. Additionally, the method of validation
as well as the validation metrics and the inter- and intra-observer differences are explained.
Average surface distance (ASD) and Hausdorff distance (HD) are used as parameters for
validation. Inter- and intra-observer differences are reported.
Results: The mean error or the ASD, measured in 10 cases, ranges from 0,53 mm to 0,76
mm, with an average of 0,65 mm and the maximum error or the Hausdorff distance ranges
from 2,02 mm to 7,84 mm, with an average of 4,05 mm. The inter-observer variability,
measured in three cases, has the following results: an average error ranging from 0,39 mm to
0,61 mm, with an average of 0,44 mm and a maximum error ranging from 1,67 mm up to 3,74
mm, with an average of 2,29 mm. The intra-observer variability in one case was also measured
and the following results are found: an average error ranging from 0,17 mm to 0,32 mm and a
maximum error ranging from 0,78 mm till 2,29 mm.
Discussion: Several previously developed automatic segmentation techniques are compared
to ours. The results of these techniques range from 0,5 till 5,4 mm ASD. Automatic
segmentation techniques which only implement the segmentation of a part of a bone or a joint
are not included in the comparison. The limitations of our study are mentioned, such as the
small number of CT data sets included in the beginning of our statistical shape models, the
small number of manually segmented CT data sets as limitation for the validation of our
2
program and the small number of studies found to compare our results to. Finally, suggestions
for further research are described: a possibility to make an extension of our developed
automatic segmentation program to MRI, the full skeleton and other disciplines like cardiology
and neurosurgery.
Conclusion: It is important to note that this study is the first of its kind to develop an automatic
segmentation technique for the full lower limb. After comparing with previously developed
techniques, our automatic segmentation technique seems to generate highly competitive
results and, by consequence, was found to be the most accurate technique.
3
NEDERLANDSE SAMENVATTING
Achtergrond: Segmentatie laat toe om een 2D beeld, verkregen via CT of MRI, te
transformeren tot een 3D beeld. Bijgevolg is er een groeiende interesse in segmentatie, vooral
bij het stellen van diagnoses en plannen van chirurgie. Het is van toepassing in orthopedie,
maar ook in andere specialisaties zoals cardiologie en neurologie. Echter, heden ten dage
wordt segmentatie meestal manueel gedaan, wat een intensief en tijdrovend werk is, dat alleen
is weggelegd voor experten. Voorgenoemde zaken duiden op de nood aan automatische
segmentatie technieken. Het doel van deze studie is om ons nieuw ontwikkeld automatische
segmentatie programma te valideren door dit te vergelijken met de gouden standaard, in casu
de manuele segmentatie. Bijkomend worden ook de inter- en intra-observator variabiliteit van
de manuele segmentatie bestudeerd.
Materialen en methodologie: Het werkingsmechanisme van de voorgestelde methode om
automatisch te segmenteren wordt voorgesteld. Het voorgestelde automatische segmentatie
programma is onderverdeeld in drie fases. Eerst vindt de initialisatie en segmentatie plaats.
Deze eerste fase is dan nog eens onderverdeeld in 4 stappen. Daarna wordt de training van
het statistische vormmodel gesuperviseerd en de werking ervan voortdurend geüpdatet. De
laatste fase is diegene waarbij uiterlijke kenmerken aan het 3D model gegeven worden. Verder
wordt zowel de methode van valideren als de parameters die gebruikt worden voor de validatie
en de inter- en intra-observator verschillen bekeken. The gemiddelde oppervlakte afstand en
de maximale afstand worden gebruikt als parameters voor validatie. De inter- en intra-
observator verschillen worden beschreven.
Resultaten: De gemiddelde fout of de gemiddelde oppervlakte afstand, bestudeerd bij 10
patiënten, varieert van 0,53 mm tot 0,76 mm met een gemiddelde van 0,65 mm. De maximum
fout of de HD varieert van 2,02 mm tot 7,84 mm met een gemiddelde van 4,05 mm. De inter-
observator variabiliteit werd berekend bij 3 patiënten en gaf volgende resultaten: een
gemiddelde fout die varieert van 0,39 mm tot 0,61 mm, met een gemiddelde van 0,44 mm en
een maximum fout die varieert van 1,67 mm tot 3,74 mm met een gemiddelde van 2,29 mm.
De intra-observator variabiliteit in 1 CT data set werd berekend en volgende resultaten werden
bekomen: een gemiddelde fout variërend van 0,17 mm tot 0,32 mm en een maximum fout die
varieert van 0,78 mm tot 2,29 mm.
Discussie: Verschillende automatische segmentatie technieken, ontwikkeld door andere
onderzoeksgroepen, werden vergeleken met de onze. Het resultaat van deze technieken
4
varieerde van 0,5 tot 5,4 mm ASD. Studies waarbij de automatische segmentatie techniek zich
toelegde op een deel van een bot of op een gewricht werden uitgesloten uit deze vergelijking.
Verder worden ook de beperkingen van onze studie besproken, zoals o.a. het kleine aantal
CT data sets die opgenomen zijn in het begin van ons statistisch vormmodel, het kleine aantal
manuele segmentaties, die een limitatie vormen voor de validatie en het kleine aantal
gevonden studies waarvan de resultaten vergeleken kunnen worden met de onze. Als laatste
werden de mogelijkheden voor verder onderzoek beschreven. Zo is er onder andere een
mogelijkheid tot uitbreiding van deze techniek tot MRI, het volledige skelet en andere
specialisaties zoals cardiologie en neurochirurgie.
Conclusie: Het is belangrijk te vermelden dat deze studie de eerste is die een automatische
segmentatie techniek ontwikkelt voor het volledige onderste lidmaat. Na de vergelijking met
eerder ontwikkelde technieken, bleek deze studie sterk concurrerende resultaten te vertonen
en daarbij dus ook over de meest accurate automatische segmentatie techniek te beschikken.
5
1 INTRODUCTION
Ever since the Ancient World, several physicians studied the human body in multiple ways.
Hippocrates and Vesalius are just a few examples of famous physicians who had an enormous
impact on medicine from then and now (1,2). The skeleton, the organs and the details of the
human body have been long known, studied and illustrated before modern medicine existed.
An important difference between medicine from then and now is the way we study the human
body. Before, the human body was studied on corpses or during an operation. Nowadays with
the help of medical imaging, the human body can be studied without the need for invasive
actions.
Even though modern technology made it much easier to study the human body, Computed
Tomography (CT) and Magnetic Resonance (MR) scanners still depict it in two-dimensional
(2D) images. In order to allow for three-dimensional (3D) evaluation of the anatomical
structures, segmentation is required.
Image segmentation is the process of segmenting or partitioning a digital image into multiple
segments or in a series of pixels with the aim to change the representation of the image into
something that gives more information and is easier to scrutinize (3). It converts a complex
image to a simple image that is easy to interpret (4). In specific, image segmentation is the
process of labelling every pixel in the image so that pixels with the same label share the same
characteristics (5). In plain words, segmentation as described in this paper is the extraction of
the part one wants to see in 3D, e.g. the skin, the heart, bones, etc., from a set of 2D slices. It
is a mainly manual technique that involves selecting the region of interest in each 2D image
followed by computing the 3D shape.
However, manual segmentation is a laborious and time-consuming task that must be
performed by trained professionals with the right anatomical knowledge. It is estimated that
the segmentation of a full skeletal lower limb takes around 100 man hours (6). If 3D evaluation
of anatomy wants to reach more physicians, then automatic segmentation techniques need to
be developed.
Segmentation of individual anatomy is used to study the anatomic variation of structures.
Furthermore, it allows for running personalized computational models in mechanical simulation
studies.
6
Up till now, various semi- to automatic segmentation routines have been described in literature.
The average surface distance (ASD) ranges from 0,5 mm up to 5,4 mm. Hence, there is room
for improvement (7–13).
The aim of this thesis is to validate the accuracy of our newly developed segmentation protocol
and to compare this to the golden standard, manual segmentations, and previously developed
automatic segmentation techniques. Additionally, the inter- and intra-observer variability of the
manual segmentation will be evaluated. The current protocol is hypothesized to be superior to
previous techniques. Furthermore, the average of the inter- and intra-observer manual
segmentations are expected to be in the same range.
In light of this study, I performed the majority of the manual segmentations and performed the
statistical analysis comparing manual to automatic segmentation as well as the intra and inter
observer variability.
1.1 Anatomy of the lower limb
The bones of the human skeleton included in this study are the pelvic bones, the femur, the
tibia, the fibula, the patella, the calcaneus, the talus, lumbar vertebrae and the sacrum. In this
section, the anatomy of these bones is described extensively and the joint they take part in is
mentioned (14–18).
1.1.1 The Pelvis or the os coxae (Figure 1)
The pelvis consists of three parts: the os pubis or the pubic bone, the os ilium and the os ischii
or the ischial bone. In childhood, these three bones are connected by cartilage and by the age
of 15-17 the fusion begins. From the age of 20 to 25 the fusion is completed and together they
form an acetabulum, which is located at the lateral side of the hip bone. This is the cavity where
the femur (the thighbone) will fit in the pelvis. The acetabulum and the femur head form the hip
joint. The biggest bone in the pelvis is the os ilium. It is located at the top and the backside of
the pelvis and consists of two parts: the ala ossis ilii and the corpus ossis illii. The latter is the
part of the ilium that fusions with the other two bones. The second bone is the os pubis,
consisting of the corpus ossis pubis, the ramus superior ossis pubis and the ramus inferior
ossis pubis. The two latter form the foramen obturatum and again the corpus part of the os
pubis is the one that fuses with the os ilium and the os ischii. At last, we have the os ischii,
which consists of the corpus ossis ischii (the part that fusions) and the ramus ossis ischii. The
7
ramus forms the lower part of the foramen obturatum. Formed by the os pubis and the os ischii,
the foramen obturatum is the space where the obturator nerve, vein and artery pass.
As mentioned above, these three bones form the hip bone and each hip consists of two hip
bones which are connected by fibrocartilage at the symphysis pubica.
1.1.2 The Femur or the thighbone (Figure 2)
The femur consists of a proximal and a distal extremity, called the proximal and distal
epiphysis, and in between the corpus femoris or the diaphysis is located. At the upper or
proximal epiphysis, we find the caput femoris, which is the part that will rotate in the acetabulum
of the hip bone. The caput contains a depression which is called the fovea capitis femoris. This
fovea does not take part in the joint but serves as an attachment place for the ligamentum
teres. Between the caput and corpus femoris we find the collum femoris, which is also called
the neck of the femur. At the sides, we find two bulgings, the trochanter major at the lateral
side, and the trochanter minor at the medial-back side. These trochanters major and minor are
the insertion point of multiple muscles. At the lower or distal epiphysis is a surface, the facies
patellaris, where the patella touches the femur. At the lower end, we find 2 surfaces, the
epicondylus medialis and the epicondylus lateralis, where the tibia touches the femur. The
femur, the tibia and the patella form the knee joint.
Figure 1: Anatomy of the pelvis (18) (from Sobotta Atlas of Human Anatomy)
8
1.1.3 The Tibia or the shinbone (Figure 2)
The tibia has the same structure as the femur, it consists of a proximal and distal epiphysis
and a diaphysis or corpus tibiae which has a triangular structure. The proximal epiphysis has
a condylus medialis and a condylus lateralis, where the tibia touches the femur. It also has a
tuberositas tibia, which is a bulging for the insertion of the ligamentum patellae. The distal
epiphysis has an incisura fibularis, where the fibula touches the tibia, and a malleolus medialis,
which is also called the inner ankle.
1.1.4 The Fibula or the splint bone (Figure 2)
The fibula has the same structure as the other long bones like the femur and the tibia and
consists of a diaphysis and a proximal and distal epiphysis. The proximal epiphysis is the caput
fibulae, which has a protrusion called the apex capitis fibulae. This caput fibulae is connected
to the diaphysis by the collum fibulae. Like the tibia, the corpus of the fibula has a triangular
structure. At the distal epiphysis, there is again a bulging, called the malleolus lateralis or the
outer ankle.
Figure 2: A) Anatomy of the Femur B) Anatomy of the Tibia C) Anatomy of the Fibula (18)
(from Sobotta Atlas of Human Anatomy)
9
1.1.5 The Patella or the knee cap (Figure 3)
The patella is a little bone that has a role in the knee
joint. The other bones that form the knee joint are the
femur and the tibia. The patella has a heart shaped
structure and is located with the point (the apex
patella) downwards or distal. The basis patella is
located upwards or proximal. The dorsal side
articulates with the femur and the anterior side (the
facies anterior) is touchable through the skin.
1.1.6 The Talus or the tarsal bone (Figure 4)
The talus is the most important bone of the foot, since it carries the weight of the whole body.
It has a caput tali (the head) and a corpus tali (the body) are connected by a collum tali ( the
neck). On the corpus tali, there is a trochlea tali with a facies superior, a facies malleolaris
lateralis and a facies malleolaris medialis. The facies superior is where the talus touches the
tibia, the facies malleolaris lateralis touches the fibula and the facies malleolaris medialis
touches the tibia. Underneath, there is a facies articularis calcanea anterior and posterior,
where the talus touches the calcaneus.
1.1.7 The Calcaneus or the heel bone (Figure 4)
The calcaneus is the biggest bone of the foot. At the back, there is a tuber calcanei, also called
the heel and is the place where the Achilles tendon inserts. At the top (the cranial side), there
are three facies articularis, namely the facies articularis talaris anterior, media and posterior.
These three are the connection between the talus and the calcaneus.
Figure 3: Anatomy of the Patella (18) (from
Sobotta Atlas of Human Anatomy)
Figure 4: A) Anatomy of the Talus B) Anatomy of the Calcaneus (18) (from Sobotta Atlas of
Human Anatomy)
10
1.1.8 The Lumbar Vertebrae (Figure 5)
The lumbar vertebrae are a part of the vertebral column which consists of seven cervical
vertebrae, twelve thoracic vertebrae, five lumbar vertebrae, five sacral vertebrae (the sacrum)
and four or five coccyx vertebrae (the coccyx). The corpus of the lumbar vertebra is much
bigger than the corpora of the other vertebrae. The processus spinosus points to caudal. The
processi costales are the protrusions on the side and are typical for the lumbar vertebrae. After
the processus costales, the processus accessorius is located and together with the processus
mamillaris they are a remnant of the processus transversus. The pediculus and lamina arcus
vertebrae form the arcus of the lumbar vertebrae. In the central of this arcus there is a foramen
vertebrale for the spinal cord.
1.1.9 The sacrum (Figure 5)
The sacrum is the consequence of the fusion between the five sacral vertebrae and it consists
of 2 sides, the frontal concave side or the facies pelvina, and the dorsal convex side or the
facies dorsalis. The fifth lumbar vertebra leans on the only visible corpus vertebrae of the upper
sacral vertebra, the basis ossis sacri and on the ventral side also called the promontorium.
The apex ossis sacri touches the os coccyx. In the facies pelvina there are 4 geminated
foramina sacralia anterior (foramina sacralia posterior on the dorsal side). These foramina
sacralia are the apertures for the nervi spinales. Between the pairs of foramina sacralia anterior
there are lineae transversae, originated from the fusion of the sacral vertebrae.
Figure 5: A) Anatomy of the Lumbar Vertebra B) Anatomy of the Sacrum (18) (from Sobotta Atlas of Human Anatomy)
11
1.2 Imaging technology
1.2.1 Introduction to Computed Tomography
To study these bones without using an invasive procedure, there is a need for imaging
technology of which the purpose is to make a scan or copy of the human body in 2D. On this
scan, dependent on the type of scanning machine, one can see different structures of the body
like bones, muscles and/or organs. The two well-known scanning machines are the Computed
Tomography (CT) and the Magnetic Resonance Imaging (MRI). The CT scan is the only one
included in this study and will be explained shortly.
The CT (Computed Tomography) or CAT (Computed Axial Tomography) uses, just like the X-
ray radiography, X-radiation for making a copy of the human body. The X ray tube emits the X
rays and the detector, located on the other side of the patient and the X ray tube, receives the
radiation that goes through the body. Hence, the part of the body that is to be visualized, needs
to be located between the X ray tube and the detector. But instead of using one stationary X-
ray tube like X-ray radiography, the CT uses an x-ray tube that circles around the patient.
Equal to X-ray radiography, the CT scan makes a reconstruction of the density of structures in
the human body. The principle is that the radiation can go through some structures and while
others stop the radiation. It is comparable with visual light: visual light is also an
electromagnetic radiation, but contrary to X-rays it is visible and it cannot pass any tissue of
the human body. When you hold your hand towards the light, there is shadow on the floor. The
same applies for X-rays. Dense or radio-opaque structures like bone and metal (e.g.
pacemaker, artificial knee, etc.) stop the radiation, so the detector does not receive any
radiation. Bones and metal create a shadow on the detector and that is why bones and metal
appear white on the 2D image. Other low density or radiolucent structures such as air let the
radiation pass through the body what makes it appear black on the 2D image. Structures like
muscles, blood and fat are a shade of grey as they let a small amount of radiation pass through.
In contrast to the X-ray radiography, the CT scan detector gets more information from the
reception of X-rays. When the X-ray tube circles around the patient, it gives 3D information
(information in the depth) to the detector. Then it is the task of the computer algorithm to
transform this 3D information into a 2D image.
12
The attenuation coefficient is the measurement of how much a tissue can be penetrated by an
X-ray beam, i.e. it measures how much strength the beam has lost while passing through the
body. The intensity of the X-ray beam that went through the patient, is lower in dense or radio-
opaque structures. The Hounsfield scale is an alternative for the attenuation coefficient and is
expressed in Hounsfield Units (HU), where water is given a HU of 0 and air a HU of -1000.
The HU of water and air serve as reference values to which all other tissues are compared.
Bone has a HU of + 400 up to +3000. All the other tissues have a HU between – 1000 HU and
+ 1000 (19) (Figure 6).
When the X-ray tube finishes the circle, a single 2D slice is made by the information sent from
the detector to the computer. For the next 2D slice, the bed where the patient lies on, is moved
a little bit in the direction in which one wants to scan the human body. Then, the CT scan can
make a new full circle around the patient to create the next 2D image. In plain words, the
principle of a CT scan is to slice the body into multiple 2D images (20–24).
Just like the X-ray radiography, the CT has the disadvantage of exposing patients to radiation,
which can cause mutations in the DNA when the patient is frequently exposed. Compared to
the X-ray radiography, the CT scan gives a lot more information about soft tissues and bones.
The advantage compared to the MRI is the shorter waiting time as well as the time that is
needed for the patient to be lying in the CT scan to create a full copy of his body (21).
In this paper, as mentioned before, we only include CT imaging. MRI and X-ray radiography
are excluded. However, in addition to the conventional CT scan, the CTA (Computed
Tomography Angiography) is included as well. The principle of CTA is equal to a conventional
CT scan, but in addition a contrast agent is administered intravenously and this enables the
Figure 6: The Hounsfield scale (19) (Available from:
The automatic segmentation of the pelvis has an average error of 0,75 mm and a Hausdorff
distance of 7,84 mm in comparison with the golden standard. The femur has an average
surface distance of 0,65 mm and a Hausdorff distance of 4,79 mm. The tibia has an average
surface distance of 0,63 mm and a maximum error of 4,07 mm. The patella has an average
error of 0,65 mm and a maximum error of 2,02 mm. The fibula has an average error of 0,76
mm and a Hausdorff distance of 3,76 mm. The calcaneus and the talus have an average
surface error of 0,53 mm and 0,57 mm respectively and a Hausdorff distance of 2,90 mm and
2,97 mm respectively (Table 1).
The inter-observer variability has the following results: an average error that ranges from 0,39
mm to 0,61 mm, with an average of 0,44 mm and a maximum error that ranges from 1,67 mm
till 3,74 mm, with an average of 2,29 mm. The intra-observer variability in one case was also
measured and the following results were found: an average error that ranges from 0,17 mm to
0,32 mm and a maximum error that ranges from 0,78 mm till 2,29 mm. These result show us
that even in the golden standard, degrees of error exist.
When we make a comparison between the errors made by the automatic segmentation and
the errors made by 2 experts whose manual segmentations of the same bone were compared
to each other, small differences are found. The maximum difference in average error between
the automatic segmentation and the inter observer variability research was found in the pelvis
since the average surface distance between the two manual segmentations, made by two
different experts, is 0,41 mm and the average error of the automatic segmentation is 0,75 mm.
37
Therefore, the difference in error is 0,34 mm. On the other hand, a minimum difference in
average surface distance of 0,13 mm is found in the calcaneus and talus since the average
surface distance between the two manual segmentations is 0,40 mm and 0,44 mm respectively
and the average surface distance of the automatic segmentation is 0,53 mm and 0,57 mm
respectively. The average difference between the automatic segmentation and the inter
observer variability research in average error is 0,21 mm and in maximum error is 1,76 mm.
The maximum difference in maximum error is 4,10 mm, found in the pelvis and the minimum
difference in maximum error is found in the patella with a difference of 0,06 mm.
Also in the intra-observer variability research errors are found. Here the maximum average
error difference between the intra-observer variability research and the automatic
segmentation is found in the fibula, with an average surface distance difference of 0,59 mm.
The minimal difference in average surface distance is 0,21 mm, found in the calcaneus. The
maximum difference in maximum error is 6,48 mm, found in the pelvis, and the minimum
difference in maximum error is found in the patella, with a difference of 1,24 mm. The average
difference between the automatic segmentation and the intra-observer variability research in
average error is 0,43 mm and in maximum error is 2,85 mm.
38
4 DISCUSSION
4.1 General
Image analysis is becoming an important tool in medicine, especially in diagnosis and medical
decision making. The progress in image analysis has caused an increase of diagnosis on the
basis of imaging technology. Image segmentation of Computed Tomography and Magnetic
Resonance Imaging plays an important role in the image analysis, but is often still done
manually, and by consequence is time-consuming (26). E.g. the time needed for extracting the
articular surfaces of the knee joint was estimated to approximately two days of work and a for
a full lower limb over 3 weeks of work were reported (6).
Because of this time-consuming task, several professionals (engineers and doctors in
medicine) have described different automatic segmentation techniques for image
segmentation. The segmentation technique that gives the most consistent results is the model
based technique, the technique based on using the SSM and ASM. But the disadvantage of
this technique is that there is a need of a large training data set to create an accurate shape
model. Other methods are less accurate in images with a lot of noise or artefacts (35,47,64).
Because of these disadvantages, there is a growing interest in automatic segmentation
programs which are able to process large data sets within minimal time and with a minimal
need of prior knowledge, and therefore a minimal need of human interaction (27).
An overall error of 0,65 mm was found using the current segmentation pipeline. The overall
error in the inter-observer variance was 0,44 mm, where the inter-observer variance are two
manual segmentations and therefore two golden standard segmentations compared to each
other. The absolute difference in overall error between the current segmentation pipeline and
the golden standard is therefore only 0,21 mm.
The current segmentation pipeline has an overall maximum error of 4,05 mm and the inter
observer variance has an overall maximum of 2,29 mm. The absolute difference in the overall
maximum error is therefore 1,76 mm.
On the basis of these results, we can conclude that a very good approximation of golden
standard segmentation can be guaranteed by using our proposed segmentation pipeline.
39
It is important to note that this study is the first of its kind to develop an automatic segmentation
program that segments the full lower limb. Multiple automatic segmentation techniques, which
were developed in the past, are able to segment only a few bones and organs or were
specifically developed to segment joints like the hip joint or knee joint (35,56,65).
4.2 Comparison with previously developed programs
As said above, multiple professionals have already developed semi-automatic or automatic
techniques for making segmentation a less time-consuming task. In Error! Reference source not found., the results of several studies will be presented and a comparison with our
proposed technique will be made.
The following studies are mentioned: Lamecker H. and colleagues (7), Seim H. and colleagues
(8), Almeida D.F. and colleagues (10), Krcah M. and colleagues (27), Uozumi Y. and
colleagues (26), Younes L.B. and colleagues (11), Wu D. and colleagues (9).
In the heading of the table, the following abbreviations are used: � patients stands for the
number of patients, ASD stands for average surface distance, RMS stands for root mean
square, HD stands for Hausdorff distance and OE stands for overlap error.
Lamecker H. et al. proposed a new statistical shape model of the pelvic bone, generated by
manually segmenting 23 CT data sets of male patients. They divided the pelvic bone into 11
regions. Just like our automatic segmentation pipeline, they performed a PCA (Principal
Component Analysis) and determined the main modes of variation. Also, they applied a gray-
value profile analysis where they use the thresholding technique. The results were described
for a “leave all in” segmentation as well as a “leave one out” segmentation. In the first case,
the SSM contains the CT data set that is going to be segmented. The ASD will be a lot smaller
if the SSM already knows the shape of the CT data set that needs to be segmented. Therefore,
these results are not comparable with ours, since our statistical shape model does not contain
any CT data sets that still need to be segmented. On the contrary, every new segmented CT
data set is added to the SSM. The “leave one out” segmentation on the other hand is fitted for
the comparison since the SSM does not know the shape in advance. The mean surface
distance in this study was 1,6 +/- 0,2 mm and the maximum distance was 14,6 +/- 3,8 mm. On
the basis of these results, we can conclude that our automatic segmentation is more accurate
(7).
40
Ta
ble
2: C
ompa
rison
with
pre
viou
sly
deve
lope
d au
tom
atic
seg
men
tatio
n te
chni
ques
41
Seim H. and colleagues presented an algorithm for an automatic segmentation of the pelvis.
Again, this study is based on a statistical shape model. The following three steps are used:
initial placement of the average shape (SSM) of the pelvis within the CT data, adaption of the
SSM to the image by a variation of the shape models and finally a free form deformation step,
which is used to overcome the restrictiveness of the SSM. We can conclude that this proposed
method has a lot of similarities with ours. For the validation, they used 50 manually segmented
CT data sets by performing the “leave one out” method. The results of this proposed automatic
segmentation pipeline were an average surface distance of 0,7 +/- 0,3 mm and a maximum
distance of 16,5 +/- 5 mm. In this study however, the pelvis contains the sacrum. We can
conclude that our research is an extension to their study since they suggested to include an
evaluation of the inter-user variability, which would help to compare the errors made by the
automatic segmentation to those made by professionals and therefore the golden standard.
For the results, their ASD is slightly better than ours with a minimal difference of 0,05 mm, but
the maximum distance is bigger, which makes this study equivalent to ours (8).
Almeida D.F. and colleagues described a new, fully, automatic segmentation pipeline for CT
images of the femur. They used an ASM that is based on a SSM and a LAM (Local Appearance
Model). Almeida D.F. et al. described the results for low-resolution CT scans as well as high-
resolution CT scans. Since we only used high-resolution CT scans in our study, we do not
mention the results of the low-resolution CT scans. 148 datasets were segmented, but only 10
were considered for the comparative analysis (due to the time consuming task of manual
segmentation) and resulted in an average surface distance of 1,014 +/- 0,474 mm and a
Hausdorff distance of 4,336 +/- 0,861 mm. Compared to our automatic segmentation pipeline
results, the ASD of our study is much lower but the HD of this study is with a difference of 0,36
mm slightly better than ours (10).
Krcah M. et al. proposed a fully automatic segmentation technique of the femur. Basically, they
worked with a threshold-like technique (graph-cuts and a bone boundary filter) and in the last
step they separated the individual bones, since working with thresholding, the risk of leakage
to adjacent bones exists. They performed this automatic segmentation on 197 femurs and had
as result an average Hausdorff distance of 5,4 mm. With a difference of 0,6 mm, we can
conclude that our automatic segmentation pipeline is more accurate (27).
Uozumi Y. et al. described an automatic bone segmentation method which was tested on six
patients. They worked in three different steps: the segmentation between the femur and the
tibia, the segmentation of the femur and the patella and the segmentation of the tibia and fibula.
Their main segmentation method is an automatisation of the thresholding technique and they
42
too, compare the segmentations, performed by the automatic segmentation program to the
manual segmentation or the golden standard segmentation. But due to their use of matching
rate (in %) as validation metric, the comparison with our proposed method is not useful.
However, it seemed to be an added value to mention their results, as their study showed good
results and was published recently (2013). The matching rates were 95,84 +/- 0,57 % for the
femur, 94,12 +/- 1,01 % for the tibia, 94,49 +/- 0,83 % for the patella and 86,37 +/- 4,28 % for
the fibula. Further extension of our results to the DSI (Dice similarity index) would be necessary
to compare these results (26).
Younes L.B. and colleagues presented a fully automatic segmentation of the femur using SSM.
The first step was a primitive shape recognition. The presented method is divided into three
steps: detection of the femoral head and the femoral shaft as a sphere and as a cylinder
respectively, registration between primitive shapes of the SSM and CT image to initialise the
SSM into the image and at last, the fitting of the SSM tot the CT image. 8 CT data sets were
segmented and had as result an average surface distance of 1,48 +/- 0,28 mm and a maximum
error of 10,53 +/- 3,19 mm. Both values are much higher than ours, which makes this technique
not the most accurate one (11).
At last, Wu D. et al. proposed an automatic segmentation method, based on a combination of
different techniques. First, they used a marginal space learning for bone detection and
deformed it with a SSM. Second, the deformed model was used as a shape prior in a graph
cut for refined segmentation. At last, a multi-layer graph cut is used because there was a need
for segmenting each bone separately since it is possible that their results overlap. They tested
their proposed method on 248 data sets and the following results were obtained: an average
surface distance of 0,82 +/- 0,33 mm for the femur, an ASD of 0,69 +/- 1,25 mm for the tibia,
an ASD of 0,96 +/- 4,29 mm for the fibula and an ASD of 0,68 +/- 2,06 mm for the patella.
When we compare these results to ours, we can conclude that our automatic segmentation
pipeline is more accurate (9).
Furthermore, many other studies were not included in this comparison since they only
segmented a part of the bone. Often, those studies focused on segmenting joints like the hip
joint or the knee joint. Therefore, their results are not comparable with ours. Still, a short
summary of their results is made since these studies had a big impact on other articles. The
segmentation methods will not be extensively mentioned as these studies will not be used for
comparison.
43
Chu C. et al proposed a fully automatic CT segmentation method for the hip joint. The main
method is the use of a SSM and for validation, the manual segmentation was taken as the
golden standard. An ASD of 0,52 +/- 0,10 mm for the pelvis, an ASD of 0,45 +/- 0,10 mm for
the left proximal femur and an ASD of 0,48 +/- 0,08 mm for the right proximal femur, were
obtained as results (56). Secondly, Cheng Y. et al described an automatic segmentation
technique for the acetabulum and the femoral head. Their method is a combination of
thresholding and an iterative process based on neighbourhood information with an average
ASD of 1,22 +/- 0,98 mm and a DSI of 91,55 +/- 4,82 % as result (35). Thirdly, Ramme A.J.
and colleagues presented a semi-automatic segmentation method for the knee joint. They
tested their method on the distal femur and the proximal tibia of 72 CT data sets. The following
results are obtained: A DSI of 95% for the femur as well as the tibia (66). Further, Zoroofi R.A.
et al. presented an automatic segmentation method for the hip joint (acetabulum and femoral
head), using 60 CT data sets. They used a combination of different techniques including
thresholding and obtained the following result: an ASD of 1,31 +/- 1,12 mm and a DSI of 90,36
+/- 5,31% (35,65). Moreover, Yokota F. and colleagues proposed an automated segmentation
method for a diseased hip. They made use of different SSM’s and had an ASD of 1,49 +/- 1,04
mm and a DSI of 90,14 +/- 1,95 % as results (35,64).
The only study found researching an automatic segmentation technique for the calcaneus was
the study of Görres J. and colleagues. They performed a segmentation of two calcaneal
surfaces on the basis of an ASM on 50 cases and have shown an ASD of 0,59 and 0,46 mm
respectively (67).
Unfortunately, there is no literature found on the automatic segmentation of the full calcaneus
and the talus. Thus, the results of our calcaneus and talus segmentation are not mentioned in
Error! Reference source not found. since comparison is not possible.
After comparing the results of the previously mentioned studies, which are the most important
studies comparable with ours, we can conclude that our proposed automatic segmentation
method is the most accurate one with the best results. Our segmentation method is similar to
the one used by Seim H. et al. (8). Their results were also close to ours, what makes this
technique the most promising one.
44
4.3 Limitations
This study was confronted with some limitations that have to be taken into account. Firstly, the
statistical shape models that are used (the low and high resolution statistical shape models)
contain in the beginning only a couple of CT data sets. As explained above, the SSM obtains
more accuracy by adding more CT data sets. After all, the ASM is a measurement of the mean
shape of the CT data sets that are included in the training data set. And it is logical that in case
of a large training data set, the average shape will have a more accurate shape, that ‘knows’
a lot of abnormalities and variations of the normal anatomy and therefore can be used for the
whole population. In case of a small training data set, the ASM will not be able to fit in a CT
data set with abnormalities, which is not included in the training data set. Since the program
adds every new segmented CT data set to the ASM, the ASM will become more accurate when
more CT data sets are segmented by the pipeline.
Secondly, the small number of manually segmented CT data sets is also a limitation for the
validation of our program. The manual segmentation is a very time-consuming task and for
validation there is a need for many manual segmentations, as the comparison between the
manual and the automatic segmentation are the basis of the validation. Because of the time-
consuming nature of this work, only a few full lower limb segmentations were performed
manually. Therefore, it was only possible to perform a global estimate of accuracy. Regional
analysis could not be performed, since the number of manual segmentations was too low for
this type of analysis. In the end, all other studies found, did not perform a regional analysis
either.
Finally, to compare our newly developed program to previously developed automatic
segmentations techniques, a literature study has been performed. Multiple studies that
included the automatic segmentation of the pelvis and femur were found. But on the other
hand, a very small number of studies were found that included the tibia, the fibula and the
patella. Not one study was found relating to the automatic segmentation of the calcaneus, talus
or the full lower limb. For this reason, a comparison with other programs could not be
performed optimally.
4.4 Suggestions for further research
The current automatic segmentation technique can be extended to the full human skeleton.
This study only includes the lower limb, but could be extended to the upper limb, thorax and
skull. Also, the automatization of segmentation is not only useful in orthopaedics, but can be
45
used in gastroenterology, cardiology, pneumology, gynaecology, neurology, oncology, etc.
This extension would imply that different statistical shape models must be created for each
organ, bone or tissue and for the validation, a lot of manual segmentations must be performed
in order to compare them to the automatic segmentation. As well as the creation of the
statistical shape model as the manual segmentation are time-consuming tasks and therefore
the extension of this program will cost a lot of time. But if the same results can be obtained,
segmentation and therefore analysis of the whole human body will be easier and quicker.
In this study, only the CT is included. It would be interesting to expand the segmentation
pipeline to other imaging techniques in particular Magnetic Resonance Imaging technology.
MRI is currently the most used imaging technology for the knee joint and the spine. However,
MRI will confront the researcher with some new challenges due to the different nature of
contrast intensity. The fact that one could start with elaborate SSM’s based on CT images
could facilitate the step towards to MRI segmentation.
46
5 CONCLUSION
The interest in automatic segmentation techniques is rising all over the world in various medical
specialties. Image segmentation makes it possible to see the region of interest in 3D, which
yields a lot more information than the 2D image obtained from CT or MRI. The need of
automatic segmentation techniques is high, since manual image segmentation is a time-
consuming task that can only be performed by trained professionals. With an average surface
distance ranging from 0,53 mm to 0,76 mm and a Hausdorff distance ranging from 2,01 mm
to 7,84 mm, the accuracy of our automatic segmentation technique has been illustrated.
Furthermore, the errors made by the automatic segmentation technique were compared to the
errors made by the manual segmentation or golden standard, resulting in a minimal average
difference in error of 0,21 mm.
Several scientists and engineers presented their own automatic segmentation techniques and
described an average surface distance ranging from 0,4 mm to 5,4 mm (10,12,13,27,35,56).
The ASD differs according to the different techniques, number of patients and quality of CT
scan, which makes comparison difficult. However, taking the differences into account, an
attempt to compare various studies with ours is made. We can conclude that our method has
highly competitive results and therefore is the most accurate one. Besides the study of Seim
H. and colleagues (8), neither one of the mentioned studies presented results similar to ours.
On one hand, it is important to note that this study is the first of its kind developing an automatic
segmentation technique for the entire skeleton of the lower limb. Previous studies focused on
one bone or on a joint only (56,64,65). On the other hand, further research can be performed.
For example, extension to MRI would be interesting. Also, the extension to the full skeleton
and to other specialties like cardiology and neurology would be an added value.
This study, with several advantages of automatic segmentation, can even be of use in other
specialties and even in non-medical professions like crime investigation (e.g. face and finger
print recognition) (32). The automatization of image segmentation has a lot of advantages for
professionals, since time is always a precious element. It makes transformation to 3D images
possible and therefore it can facilitate operations or it can allow for a diagnosis in an earlier
stage.
47
6 REFERENCES
1. Schiefsky MJ. Hippocrates On ancient medicine. Brill; 2005. 415 p. 2. O’malley CD. ANDREAS VESALIUS 1514-1564: IN MEMORIAM. Med Hist. 1964
Oct;8:299–308. 3. Shapiro L, Stockman G. Computer Vision. October. 2001. 608 p. 4. Saini S, Arora K. A Study Analysis on the Different Image Segmentation Techniques.
Int J Inf Comput Technol. 2014;4(14):1445–52. 5. Wikipedia, the free encyclopedia: Image segmentation [Internet]. [cited 2017 Oct 10].
Available from: https://en.wikipedia.org/wiki/Image_segmentation 6. Henak CR, Anderson AE, Weiss JA. Subject-Specific Analysis of Joint Contact
Mechanics: Application to the Study of Osteoarthritis and Surgical Planning. J Biomech Eng. 2013 Feb 11;135(2):21003.
7. Lamecker H, Seebass M, Hege H-C, Deuflhard P. A 3D statistical shape model of the pelvic bone for segmentation. In 2004. p. 1341.
8. Seim H, Kainmueller D, Heller M, Lamecker H, Zachow S, Hege H-C. Automatic Segmentation of the Pelvic Bones from CT Data Based on a Statistical Shape Model. In: Proceedings of the First Eurographics conference on Visual Computing for Biomedicine. 2008. p. 93–100.
9. Wu D, Sofka M, Birkbeck N, Zhou SK. Segmentation of multiple knee bones from CT for orthopedic knee surgery planning. Med Image Comput Comput Assist Interv. 2014;17(Pt 1):372–80.
10. Almeida DF, Ruben RB, Folgado J, Fernandes PR, Audenaert E, Verhegghe B, et al. Fully automatic segmentation of femurs with medullary canal definition in high and in low resolution CT scans. Med Eng Phys. 2016 Dec;38(12):1474–80.
11. Younes BL, Nakajima Y, Saito T. Fully automatic segmentation of the Femur from 3D-CT images using primitive shape recognition and statistical shape models. Int J Comput Assist Radiol Surg. 2014 Mar 8;9(2):189–96.
12. Almeida D, Folgado J, Fernandes PR, Ruben RB. 3D shape prior active contours for an automatic segmentation of a patient specific femur from a CT scan. In: Computational Vision and Medical Image Processing IV. 2014. p. 271–5.
13. Liu B, Zhang H, Hua S, Jiang Q, Huang R, Liu W, et al. An automatic segmentation system of acetabulum in sequential CT images for the personalized artificial femoral head design. Comput Methods Programs Biomed. 2016;127:318–35.
14. Kishner S, Chowdhry M. Hip Joint Anatomy: Overview, Gross Anatomy [Internet]. 2015 [cited 2017 Sep 30]. Available from: http://emedicine.medscape.com/article/1898964-overview#a2
16. Kerckaert I. Locomotorisch stelsel en huid, Deel III: Osteologie. Gent: Academia Press; 2013. 58 p.
17. D’Herde K, Schoemaker E, Veening J. Sesam Atlas van de anatomie deel 1: Bewegingsapparaat. 10th ed. Amersfoort: ThiemeMeulenhoff; 2012. 467 p.
18. Paulsen F, Waschke J. Sobotta Atlas of Human Anatomy: General anatomy and musculoskeletal system. 15th ed. Munich: Elsevier GmbH,Urban & Fischer; 2011. 400 p.
19. Chapter 1: Introduction to CT physics [Internet]. 2004 [cited 2017 Oct 19]. p. 16. Available from: https://web.archive.org/web/20070926231241/http://www.intl.elsevierhealth.com/e-books/pdf/940.pdf
20. Understanding How MRI, CT and X-ray Are Different | CDI, Insight Imaging & Affiliated Providers [Internet]. [cited 2017 Oct 2]. Available from: https://www.mycdi.com/knowledge_center/difference_between_mri_ct_and_x-ray/
48
21. Computed Tomography (CT) | National Institute of Biomedical Imaging and Bioengineering [Internet]. [cited 2017 Oct 2]. Available from: https://www.nibib.nih.gov/science-education/science-topics/computed-tomography-ct
22. Thurston M, Nadrljanski MM. Computed tomography | Radiology Reference Article | Radiopaedia.org [Internet]. [cited 2017 Oct 3]. Available from: https://radiopaedia.org/articles/computed-tomography
23. Seeram E. Computed tomography : physical principles, clinical applications, and quality control. 487 p.
24. X-rays | National Institute of Biomedical Imaging and Bioengineering [Internet]. [cited 2017 Oct 2]. Available from: https://www.nibib.nih.gov/science-education/science-topics/x-rays
25. Matej “Retro” J. Pixels and voxels, the long answer – Retronator Magazine – Medium [Internet]. 2014 [cited 2017 Oct 3]. Available from: https://medium.com/retronator-magazine/pixels-and-voxels-the-long-answer-5889ecc18190
26. Uozumi Y, Nagamune K. An automatic bone segmentation method based on anatomical structure for the knee joint in MDCT image. In: 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE; 2013. p. 7124–7.
27. Krcah M, Szekely G, Blanc R. Fully automatic and fast segmentation of the femur bone from 3D-CT images with no shape prior. In: 2011 IEEE International Symposium on Biomedical Imaging: From Nano to Macro. IEEE; 2011. p. 2087–90.
28. Hosseini M-P, Nazem-Zadeh MR, Pompili D, Soltanian-Zadeh H. Statistical validation of automatic methods for hippocampus segmentation in MR images of epileptic patients. Conf Proc . Annu Int Conf IEEE Eng Med Biol Soc IEEE Eng Med Biol Soc Annu Conf. 2014;2014:4707–10.
30. Zaitoun NM, Aqel MJ. Survey on Image Segmentation Techniques. Procedia Comput Sci. 2015;65:797–806.
31. Umaa Mageswari S, Sridevi M, Mala C. An Experimental Study and Analysis of Different Image Segmentation Techniques. Procedia Eng. 2013;64:36–45.
32. Dass R, Priyanka, Devi S. Images Segmentation Techniques. IJECT. 1993;3(1):66–70. 33. Kang W-X, Yang Q-Q, Liang R-P. The Comparative Research on Image Segmentation
Algorithms. In: 2009 First International Workshop on Education Technology and Computer Science. IEEE; 2009. p. 703–7.
34. Pham DL, Xu C, Prince JL. Current Methods in Medical Image Segmentation. Annu Rev Biomed Eng. 2000 Aug;2(1):315–37.
35. Cheng Y, Zhou S, Wang Y, Guo C, Bai J, Tamura S. Automatic segmentation technique for acetabulum and femoral head in CT images. Pattern Recognit. 2013 Nov;46(11):2969–84.
36. Saleh Al-Amri S, Kalyankar N V, D KS. Image Segmentation by Using Thershod Techniques. J Comput. 2010;2(5):2151–9617.
37. Otsu N. A Threshold Selection Method from Gray-Level Histograms. IEEE Trans Syst Man Cybern. 1979 Jan;9(1):62–6.
38. Daniel P, Raju R, Neelima G. Image Segmentation by using Histogram Thresholding. IJCSET. 2012;2(1):776–9.
39. Zhang J, Yan C-H, Chui C-K, Ong S-H. Fast segmentation of bone in CT images using 3D adaptive thresholding. Comput Biol Med. 2010 Feb;40(2):231–6.
40. Pal NR, Pal SK. A REVIEW ON IMAGE SEGMENTATION TECHNIQUES. Pattern Recoynition. 1993;26(9):1277–94.
41. Kang Y, Engelke K, Kalender WA. A new accurate and precise 3-D segmentation method for skeletal structures in volumetric CT data. IEEE Trans Med Imaging. 2003 May;22(5):586–98.
42. Jacobs D. Image Gradients: Class Notes for CMSC 426. 2005 p. 3.
49
43. Kaur N, Singh J, Sharma V. Analysis and Comprehensive Study: Image Segmentation Techniques. 2015;2321–9653.
44. Heena A, Biradar N, Maroof NM. A Novel Approach to Review Various Image Segmentation Techniques. Int J Innov Res Comput Commun Eng An ISO Certif Organ. 2007;32975(2).
45. Khan AM, Ravi S. Image Segmentation Methods: A Comparative Study. Int J Soft Comput Eng. 2013;3(34):2231–307.
46. Sarkalkan N, Weinans H, Zadpoor AA. Statistical shape and appearance models of bones. Bone. 2014 Mar;60:129–40.
47. Heimann T, Meinzer H-P. Statistical shape models for 3D medical image segmentation: A review. Med Image Anal. 2009 Aug;13(4):543–63.
48. Cootes TF, Taylor CJ. Statistical Models of Appearance for Computer Vision. 2004. 49. Cootes TF, Taylor CJ, Cooper DH, Graham J. Active Shape Models - Their Training and
Application. Comput Vis image Underst. 1995;61(1):35–59. 50. Morse BS. Lecture 21: Image Understanding. 2000 p. 6. 51. Pichumani R. Snakes: an active model [Internet]. 1997 [cited 2017 Oct 25]. Available
55. Xu C, Prince JL. Gradient Vector Flow: A New External Force for Snakes. IEEE Proc Conf Comp Vis Patt Recog. 1997;66–71.
56. Chu C, Chen C, Liu L, Zheng G. FACTS: Fully Automatic CT Segmentation of a Hip Joint. Ann Biomed Eng. 2015 May 4;43(5):1247–59.
57. Yeghiazaryan V, Voiculescu I. An Overview of Current Evaluation Methods Used in Medical Image Segmentation. 2015;21.
58. 3D Medical Image Processing Software | Materialise Mimics [Internet]. [cited 2017 Nov 1]. Available from: http://www.materialise.com/en/medical/software/mimics
59. Paz ICL A, Ldg B, Bruno L. Bone mineral density: review. Brazilian J Poult Sci Rev Bras. 2006;8(2):69–73.
60. McClung MR. The relationship between bone mineral density and fracture risk. Curr Osteoporos Rep. 2005 Jun;3(2):57–63.
61. Assessment of Fracture Risk [Internet]. 2005 [cited 2017 Nov 20]. Available from: https://www.medscape.org/viewarticle/506083_2
62. Weisstein EW. Root-Mean-Square [Internet]. Wolfram Research, Inc.; [cited 2017 Oct 20]. Available from: http://mathworld.wolfram.com/Root-Mean-Square.html
63. Huttenlocher DP, Klanderman GA, Rucklidge WJ. Comparing Images Using the Hausdorff Distance. IEEE Trans Pattern Anal Mach Intell. 1993;15(9):850–63.
64. Yokota F, Okada T, Takao M, Sugano N, Tada Y, Sato Y. Automated segmentation of the femur and pelvis from 3D CT data of diseased hip using hierarchical statistical shape model of joint structure. Med Image Comput Comput Assist Interv. 2009;12(Pt 2):811–8.
65. Zoroofi RA, Sato Y, Sasama T, Nishii T, Sugano N, Yonenobu K, et al. Automated segmentation of acetabulum and femoral head from 3-D CT images. IEEE Trans Inf Technol Biomed. 2003 Dec;7(4):329–43.
66. Ramme AJ, Criswell AJ, Wolf BR, Magnotta VA, Grosland NM. EM segmentation of the distal femur and proximal tibia: A high-throughput approach to anatomic surface generation. Ann Biomed Eng. 2011;39(5):1555–62.
67. Görres J, Brehler M, Franke J, Vetter SY, Grützner PA, Meinzer HP, et al. Articular surface segmentation using active shape models for intraoperative implant assessment. Int J Comput Assist Radiol Surg. 2016;11(9):1661–72.