A Synopsis on

STUDY OF SEGMENTATION TECHNIQUES USING

VERTEBRAL CT IMAGES

NAME OF CANDIDATE
Mr. Gundavade Vikas Bhausaheb

NAME OF SUPERVISOR
Dr. Patil S.A.

ABSTRACT
With the growing research on image segmentation of spinal vertebrae, it has become important to
categorise the research outcomes and provide an overview of the existing segmentation
techniques in it. In this project, different image segmentation techniques applied on 3D
segmentation of spinal vertebrae are reviewed. The selection includes sources from image
processing journals, conferences, books, IEEE transactions. The detail survey includes different
approach on segmentation. The state of art research on each category is provided with emphasis
on developed technologies and image properties used by them. The categories defined are not
always mutually independent. Hence, their interrelationships are also stated. Finally, conclusions
are drawn summarizing commonly used techniques and their complexities in application.

1. INTRODUCTION
Accurate vertebra segmentation in computed tomography (CT) images is important for numerous
medical applications, e.g., diagnosis of osteolytic or osteoblastic cancer metastases within the
spinal column [41], diagnosis of spine trauma, and detection of osteoporosis [32]. Accurate
knowledge of the shape of the individual vertebrae is also important for spinal biopsies, implants,
or the insertion of pedicle screws in spinal surgery [50]. However, manually delineating and
annotating vertebrae is a subjective, tedious, and error prone task. Preparing an automatic
vertebra segmentation system would greatly improve the process, thereby easing the workload
on radiologists while also removing operator variability.

Fig.1 Fractures of vertebrae are indicated by arrows

Fig.2 Incomplete, missing information of vertebrae in CT images.

Vertebra segmentation in CT images is a challenging task due to the presence of image artifacts,
contrast variations, presence of neighboring structures, and shape variation [50]. Recently, a
considerable amount of work has been done toward preparing automatic systems for detection
and segmentation of vertebrae. In this work, we mainly focus on comparison of the level set
segmentation with willmore flow and statistical shape modeling to compare the segmentation
results. In the following, we start by preparing a brief review on state-of-the-art regarding the
segmentation step. In Sec. 2, we explain in detail review of segmentation methods. In Sec. 3, we
prepare qualitative and quantitative results of these methods and compare the results achieved by
them.
2. LITERATURE SURVEY

Various attempts have been made at spine segmentation in recent years, but majority of them use
2-D images and/or require user intervention in the process. For example, Naegel [2] combined
the watershed method and morphological approaches to segment vertebrae. Although the
proposed method is promising in segmenting healthy bones from high-resolution images, manual
refinement is necessary to obtain accurate segmentations, and the level of refinement is patient
and resolution dependent. Ghebreab and Smeulders [3] constructed a deformable integral spine
model to segment vertebrae. The method learns the appearance of vertebrae boundaries a priori
from a set of training images. This model is then used to generate landmark points, in order to
reduce the complexity of the segmentation process through point-based shape representation.
However, it remains unclear if the landmark points correspond to the actual anatomical locations
and whether they capture the biologically meaningful variations across different subjects. The
method is also not fully automated and needs step-by-step inputs from the user, which makes the
whole process tedious and time consuming. Ma et al. [4] presented an automatic vertebra
segmentation and identification method on thoracic vertebra CT images. A learning-based bone
structure edge detection algorithm was used and a hierarchical, coarse-to-fine deformable
surface-based segmentation method was proposed based on the response maps from the learned
edge detector. Though satisfactory results were obtained, the segmented vertebrae were only in
2-D and reproducibility of results in 3-D was not known. Another limitation is the complexity
and/or inaccuracy of current segmentation methods. For example, Lorenze and Krahnstoever [5]
proposed a statistical shape model whereby the mean shape was constructed from a set of
training samples. The initialization of the shape model for segmentation was done manually and
is highly sensitive to dislocation. If the model is not located in the proximity of vertebrae,
segmentation may fail. More recently Klinder et al. [6] used a mesh-based method to extract

spine curves, and then generalized Hough transform and curved planar reformation to detect the
vertebrae. The proposed approach has a further identification step to the detected vertebrae via
rigid registration of appearance model. Although they achieved very competitive identification
rates for vertebrae, their algorithm depends heavily on spatial registration of the model, which is
computationally very expensive. In a paper by Mastmeyer et al. [7], a hierarchical 3-D technique
was developed to segment the vertebral bodies in order to measure bone mineral density. The
proposed framework needs excessive user intervention to precisely locate seed points to facilitate
region growing segmentation. This process is time consuming and impractical for unhealthy
bone segmentation. A similar approach integrating region growing segmentation with local shape
and intensity refinement for delineating vertebrae was proposed by Kang et al. [8]. First, locally
adaptive thresholds were used to facilitate region growing segmentations globally, followed by
3-D morphological operations to refine the segmented surfaces. This method still required a site
specific separation of individual bones, which remains a challenge for vertebrae segmentation.
Due to the aforementioned drawbacks of the existing spinal vertebrae segmentation methods, we
have developed a new method capable of segmenting spinal accurately from noisy images with
missing information. The method is developed by introducing an edge-mounted Willmore flow,
as well as a prior shape kernel density estimator, to the level set segmentation framework. While
the prior shape model provides much needed prior knowledge when information is missing from
the image, the edge-mounted Willmore flow helps to capture the local geometry and smoothes
the evolving level set surface.

2.1 Level Set

The level set method [10] has been widely used for image segmentation [11]. The level set
segmentation obtained good results for highly challenging segmentation tasks, such as medical
images, level set methods have achieved good results when coupled with prior knowledge or
prior shape models [12][15]. The level set method embeds an interface in a higher dimensional
function (the signed distance function) as a level set = 0. The equation that governs the
evolution of the level set function (t) is /t + F|| = 0, where F represents the speed
function. In more recent development, the variational framework is often considered. Under the
variational framework, an energy E() is defined in relation to the speed function, and
minimization of the energy generates the EulerLagrange equation and, hence, providing the
evolution equation through the calculus of variation

E ()
=
t

To improve the results of level set segmentation we consider the fusion of energy, i.e., using a
shape prior distribution estimator Es with an edge-mounted Willmore energy Ew0
E ( )=Es+ E 0

where (0 < 1) is the weight parameter. Details on Es and Ew0 will be described in the
following sections. In order to incorporate a prior dataset {1, 2, . . . , N }into the level set
segmentation framework, we adopt a shape dissimilarity measure based on the Kernel density
estimation (KDE) discussed by Cremers et al. [13]. This nonparametric distribution estimator
overcomes the two shortcomings of existing algorithms: 1) the assumption that the shapes are
Gaussian distributed, which is generally inappropriate when the number of training set is small,

and not practical for modeling shapes with high complexity and structure; 2) the shapes are
represented by signed distance functions, which constitute a nonlinear space that does not
include the mean.
2.1.1 Kernel Density Estimation
KDE is a nonparametric approach in statistics for estimating the probability density function of a
random variable. The underlying theory of KDE states that data with unknown statistical
distribution converge to its actual distribution as the number of samples approaches infinity. In
practice, KDE provides a fundamental smoothing estimator even with a small number of data
samples. In application with N samples of shape models, the density estimation is formulated as
a sum of Gaussian of shape dissimilarity measures d2 (H(),H(i)), i = 1, 2, . . .,N

1
P() e
N i=1

d (H ( ) , H (i ))
2
2

where H() is the Heaviside function, the shape dissimilarity measure [16][18] is

1
d 2 ( H ( ) , H ( i )) = (H ( )H ( i ) )2 dx
2

and 2 is the mean squared nearest neighbor distance

min d H ( i ) , H ( j)
N
1
2
=
2 i , j=1
Note that the L2-norm is invariant under translation and scaling with respect to the principal axis
of the shape. Hence, the shape dissimilarity measure d2 is also invariant under these
transformations when the prior shapes are normalized with respect to translation and scaling
accordingly [13].
The segmentation is obtained by maximizing the conditional probability of given image I
P ( /I )=

P ( I / ) P ()
p(I )

Since the negative logarithmic scale of the probability distribution P(|I) nicely defines an
energy that associates the evolution of with the minimization problem, the shape energy is
formulated as
Es () = log P(|I).
Hence, the variational with respect to becomes

i d2 ( H ( ) , H ( i))

E s i=1
=

2 2 i
i=1

2.1.2 Willmore Flow

Willmore energy is a function of mean curvature, which isa quantitative measure of how much a
given surface deviates from a round sphere. It has been applied to image inpainting, restoration
of implicit surfaces [19], [20], and to studies of the bending energy of biological cell membranes

as these cell membranes tend to position themselves to minimize Willmore energy [21].
Willmore flow is the gradient flow of Willmore energy.Willmore flow of a surface is the
evolution of the surface in time to follow variations of the Willmore energy. Willmore energy
was defined after the British Geometer T.Willmore [22] and is formulated as

E =

1
h2 dA
2M

where M is a d-dimensional surface embedded in Rd+1 and h the mean curvature on M.

here we integrate Willmore flow into the level set segmentation framework as a geometric
functional. Willmore energy is defined on the collection of level sets, and Willmore flow is
enabled by defining a suitable metric, the Frobenius norm, on the space of the level sets. The
Frobenius norm of an arbitrary matrix A = (aij )mn , coincides with the calculation for the
gradient decent. It is equivalent to the l2-norm (the Euclidean norm) of a matrix, More
importantly, it is computationally attainable comparing to l2-norm. As Frobenius norm is an
inner-product norm, the optimization in the variational method comes naturally. Based on the
formulation by Droske and Rumpf [23], Willmore flow or the variational form for the Willmore
energy with respect to .
In order to ensure that the smoothing effect of Willmore energy acts around the constructed
surface and does not affect adversely the edge of vertebrae, we propose to multiply the edge
indicator function
g ( I )=

1
2

1+| GI|

to the level set evolution

2.2.MODEL BASED SEGMENTATION

We classify the existing vertebra segmentation approaches to two main groups: i) the ones, which
do not consider shape prior information, and ii) the ones, which do. Regarding the first group, we
can point to the following works: Ghosh et al. [36] extract the vertebra border as high gradient
edges. Peng et al. [45] apply the Canny edge detector on 2D slices for vertebra segmentation.
Aslan et al. [33] utilize a level set algorithm for vertebra segmentation. However, these methods
[36,45,33] do not make use of shape prior knowledge. Therefore, they are vulnerable to leakage
and thus lead to less accurate segmentation results. Considering the second group, there exist
several vertebra segmentation methods which make use of shape prior information. Aslan et al.
[32] consider shape prior information in a graph cut-based framework. Ma et al. [33] propose a
template-based segmentation method. However, these methods only rely on mean shape
information and do not benefit from the principal modes of variation. Herring et al. [39] compute
a coarse segmentation by simple thresholding and then register it to a pre-computed vertebra
shape model. However, their method requires a manual initialization; similar to the works in
[44,49]. Klinder et al. [42] propose a modelbased segmentation approach using a region-based
appearance model, which includes variance information.

2.2.1 METHOD

The general pipeline of our method is shown in Fig. 1. The input to our system is a 3D CT image
of the spinal cord accompanied with vertebra-bounding box information. These bounding boxes,
which are represented by their center, orientation, and scale, can be estimated by applying a
vertebra body detection as proposed by Kelm et al. [41]. Combining our method with this
method would lead to a fully automatic vertebra detection and segmentation system which does
not require any user interaction. Our method combines statistical shape modeling (SSM) to
capture global vertebra shape information and machine learning (ML) to capture local
appearance-related prior information. We break down our method into two main steps: the
training step and the testing step. In the training step, we compute the SSM and the boundary
detector model. In the testing step, we make use of the trained models resulting from the

previous step to segment vertebrae in an unseen image accompanied with its vertebra bounding
box information.

2.2.2 Training Step

The training step of our framework consists of four main steps: i) finding the mesh point
correspondences, ii) normalizing the meshes and volumes, iii) extracting the SSM, and iv)
learning the boundary detector. Note that we do the learning step on cervical (V1 to V5), thoracic
(V6 to V17), and lumbar (V18 to V22) parts separately, where Vi represents the vertebra number.
In Fig. 2(b) the vertebrae inside the cervical, thoracic, and lumbar parts are represented by pink,
green, and orange, respectively.

2.2.3 Finding Mesh Point Correspondences

Since extracting a SSM requires a set of training shapes with well-defined correspondences [38],
we apply a spectral-based algorithm to compute the point correspondences between the vertebra
meshes [40]. In the respective block in Fig. 1, corresponding points between a pair of meshes are
represented with the same color. Note that in our implementation of finding mesh-point
correspondences, we use vertebrae V3, V12, and V20 of one patient in the training set as the
reference meshes for cervical, thoracic, and lumbar parts and register all other meshes of each
group to them.3.1.2. Normalizing Meshes and Volumes The next step as depicted in Fig. 1 is
performing spatial normalization on the vertebra volumes and meshes. Regions within the
bounding boxes are spatially normalized to image volumes with equal size, resolution, and

orientation. The spatial normalization step is important in our machine learning-based approach.
Extracting 3D steerable features from these normalized volumes simplifies learning due to more
stable appearance patterns of the vertebra edges. A similar normalization step has beed proposed
by Wels et al.[47] to extract local features from vertebral bodies for spinal bone lesion detection.
We apply the same normalization step (normalizing w.r.t the box information) to the meshes. The
normalized meshes are used for extracting the SSM, as explained in the following.

2.2.4. Extracting the SSM

After finding correspondences and normalizing the meshes, we apply Generalized Procrustes
Analysis (GPA) [37] to align the meshes rigidly. Let us represent aligned meshes by x1; x2; :::;
xN, N N+, where xi consists of the spatial coordinates of the surface points of the meshes.

Then, the mean shap

N

Xi/ N ,
i=1

x i x

x i x

S=
i =1

e and the corresponding covariance matrix S is given by:

By applying eigen decomposition on S, we can extract principal modes of variation _m

(eigenvectors) and their respective variance _m (eigenvalue). Based on the main concept of SSM
theory, each shape in the training dataset can be approximated by a linear combination of the first

mth modes, i.e. given by:

x x Pb

(1,...., m)
, where P =

is the

matrix of the m selected eigenvectors, and b = (b1; :::bm)T

are the shape parameters [4].
2.2.5. Learning the Boundary Detector
To learn the boundary detector, given the normalized meshes and volumes, the image voxels on
the mesh surface are interpreted as positive training samples. Then, a set of 3D steerable features
is extracted from the points on the surface [50]. The same set of features is extracted from
several neighboring sampling points along the normal line of the mesh surface points providing
negative training samples [50,48]. These feature vectors are used to train the boundary detector
using a Probabilistic Boosting-Tree Classifier [50,48].
2.2.6 Testing Step
Given an unseen image with its vertebra bounding box information, we first spatially normalize
the volumes inside the box. On the normalized volumes, an initial estimation of the shape of the

x
xx
vertebra x is estimated using the computed mean shape ,i.e.,
. Then, a set of steerable

features is extracted from the mesh points x and several neighbouring sampling points along the
normal line of the mesh-surface points. After applying the boundary detector to the extracted

x (x1, y1, z1,.....xn, yn, zn) T

feature vectors, x is updated by a displacement vector

. To

apply shape constraints on the updated mesh

x x x

, it is registered to the SSM model space

and projected such that it can be approximated by the mean shape and a linear combination of
eigenvectors [4] (see Sec. 2.1.3).
As shown in Fig. 1, the final estimation of all the vertebrae in the original image space is made
by projecting back the detected meshes in the normalized space to the original image space.
3. OBJECTIVE
Proposed research objectives are as follows,
a) Implementation of willmore flow level set segmentation and statistical shape modeling
on vertebral CT images.
b) Selecting the most suitable one for further analysis using following parameters,
1. Prior shape energy.
2. Average measurement
3. Sensitivity, Specificity and Dice Similarity Coefficient.
c) Testing of the better algorithm on vertebral MRI images.

4. METHODOLOGY

CT IMAGES OF VERTEBRAE

SEGMENTATION OF VERTEBRAE
USING WILLMORE FLOW LEVEL
SET SEGMENTATION

SEGMENTATION USING
STSTISTICAL SHAPE MODELING

COMPARE BOTH RESULTS

IMPLEMENTATION OF
BETTER ONE ON MRI IMAGES
OF VERTERBRAE

Image segmentation, is widely used in content based image retrieval, Machine vision, Medical
Imaging ,Object detection, Pedestrian detection, Face detection, Brake light detection, Locate
objects in satellite images, Recognition Tasks, Iris recognition, Traffic control systems. The main
applications of medical imaging are Locate tumors and other pathologies, Measure tissue
volumes, Diagnosis & study of anatomical structure. Medical imaging which consists mainly
combination of sensors recording the anatomical body structure like magnetic resonance image
(MRI), ultrasound or CT with sensors monitoring functional and metabolic body activities like
positron emission tomography (PET), single photon emission computed tomography (SPECT) or
magnetic resonance spectroscopy (MRS). Results can be applied, for instance, in radiotherapy
and nuclear medicine. This project mainly deals with the application on CT image on spinal
vertebrae.
Segmentation is often the key step in interpreting the image. Image segmentation is a process in
which regions or features sharing similar characteristics are identified and grouped together.
Image segmentation may use statistical classification, thresholding, edge detection, region
detection, or any combination of these techniques. The output of the segmentation step is usually
a set of classified elements. Most segmentation techniques are either region-based or edge based-

Region-based techniques rely on common patterns in intensity values within a cluster of

neighboring pixels. The cluster is referred to as the region, and the goal of the segmentation
algorithm is to group regions according to their anatomical or functional roles. Edge-based
techniques rely on discontinuities in image values between distinct regions, and the goal of the
segmentation algorithm is to accurately demarcate the boundary separating these regions.
Segmentation is a process of extracting and representing information from an image is to group
pixels together
In order to compare willmore flow level set segmentation with statistical shape model we create
a dataset of 20 ct images of normal spinal vertebrae images of patients. These images are
carefully selected by radiologists to form a representative group. The ground truths are obtained
by using TURTLESEG, 3D image segmentation software and verified by radiologists.
The willmore flow level set segmentation and statistical shape model are applied on these set of
images differently. a qualitative comparison between the segmented results are done between
these two models and analyze the better suited model.
For a quantitative evaluation we measure the errors of the segmented meshes from the model
which is better than other and try to implement the same on similar MRI images.

5.POSSIBLE OUTCOMES

Comparative results of willmore flow level set segmentation and statistical shape model
Better segmentation results for diagnosis
A method of segmentation for MRI images.

REFRENCES

[1] S. Looby and A. Flanders. (2011, Jan.). Spine trauma. Radiol. Clin. North. Am. [Online].
49(1), pp. 129163, Available: http://linkinghub.elsevier. com/retrieve/pii/S0033838910001454
[2] B. Naegel, Using mathematical morphology for the anatomical labeling of vertebrae from 3D CT-scan images, Comput. Med. Imaging Grap., vol. 31, no. 3, pp. 141156, 2007.
[3] S. Ghebreab and A. Smeulders. (2004, Oct.). Combining strings and necklaces for interactive
three-dimensional segmentation of spinal images using an integral deformable spine model.
IEEE

Trans.

Biomed.

Eng.

[Online].

51(10),

pp.

18211829,

Available:

http://ieeexplore.ieee.org/ lpdocs/epic03/wrapper.htm?arnumber=1337150
[4] J.Ma, L. Lu, Y. Zhan, X. Zhou, M. Salganicoff, and A. Krishnan, Hierarchical segmentation
and identification of thoracic vertebra using learningbased edge detection and coarse-to-fine
deformable model, in Proc. Int. Conf. Med. Imag. Comput. Comput. Aided Intervention, 2010,
pp. 1927.
[5] C. Lorenz and N. Krahnstoever, 3D statistical shape models for medical image
segmentation, in Proc. 2nd Int. Conf. 3-D Dig. Imag.Model., 1999, pp. 48.
[6] T. Klinder, J. Ostermann, M. Ehm, A. Franz, R. Kneser, and C. Lorenz. (2009, Jun.).
Automated model-based vertebra detection, identification, and segmentation in ct images. Med.
Imag. Anal. [Online]. 13(3), pp. 471482, Available: http://linkinghub.elsevier.com/retrieve/pii/
S1361841509000085
[7] A. Mastmeyer, K. Engelke, C. Fuchs, andW. A. Kalender, A hierarchical 3-d segmentation
method and the definition of vertebral body coordinate systems for qct of the lumbar spine,
Med. Image Anal., vol. 10, pp. 560577, 2006.

[8] Y. Kang, K. Engelke, andW. A. Kalender, A new accurate and precise 3d segmentation
method for skeletal structures in volumetric ct data, IEEE Trans. Med. Imag., vol. 22, no. 5, pp.
586598, May 2003.
[9] P. H. Lim, U. Bagci, O. Aras, Y. Wang, and L. Bai, A novel spinal vertebrae segmentation
framework combining geometric flow and shape prior with level set method, in Proc. IEEE Int.
Symp. Biomed. Imag., Barcelona, Spain, May 2012, pp. 17031706.
[10] S. Osher and J. Sethian, Fronts propagating with curvature-dependent speed: Algorithms
based on Hamilton-Jacobi formulations, J. Comput. Phys., vol. 79, pp. 1249, 1988.
[11] D. Feltell and L. Bai, Level set image segmentation refined by intelligent agent swarms,
presented at theWorld Congr. Computational Intelligence, Barcelona, Spain, 2010.
[12] A. Tsai, A. Yezzy, W. Wells, C. Tempany, D. Tucker, A. Fan, W. E. Grimson, and A. Willsky,
A shape-based approach to the segmentation of medical imagery using level sets, IEEE Trans.
Med. Imag., vol. 22, no. 2, pp. 137154, Feb. 2003.
[13] D. Cremers, S. J. Osher, and S. Soatto, Kernel density estimation and intrinsic alignment
for shape priors in level set segmentation, Int. J. Comput. Vis., vol. 69, no. 3, pp. 335351,
2006.
[14] M. Rousson and N. Paragios. (2008, Mar.). Prior knowledge, level set representations &
visual

grouping.

Int.

J.

Comput.

Vis.

[Online].

76(3),

pp.

231243,

Available:

http://www.springerlink.com/index/10. 1007/s11263-007-0054-z
[15] P. H. Lim, U. Bagci, and L. Bai, A new prior shape model for level set segmentation, in
Proc. Iberoamer. Congr. Conf. Progress Pattern Recog., Image Anal., Comput. Vis., and Appl.,
2011, ch. 14, pp. 125132.

[16] T. Chan andW. Zhu, Level set based shape prior segmentation, Comput. Appl. Math.,
Univ. California, Los Angeles, Tech. Rep. 0366, 2003.
[17] T. Riklin-Raviv, N. Kiryati, and N. Sochen, Unlevel sets: Geometry and prior-based
segmentation, in Proc. Eur. Conf. Comput. Vis., 2004, pp. 5061.
[18] G. Charpiat, O. Faugeras, and R. Keriven, Approximations shape metrics and application
to shape warping and empirical shape statistics, Found. Comput. Math., vol. 5, no. 1, pp. 158,
Feb. 2005.
[19] R. Schneider and L. Kobbelt, Generating fair meshes with G1 boundary conditions, in
Proc. Geom. Model. Process. Conf., 2000, pp. 251261.
[20] S. Yoshizawa and A. G. Belyaev, Fair triangle mesh generation with discrete elastica, in
Proc. Geom. Model. Process., 2002, pp. 119123.
[21] J. W. Barrett, H. Garcke, and R. Nrnberg. (2008, Jan.). Parametric approximation of
willmore flow and related geometric evolution equations. SIAM J. Sci. Comput. [Online]. 31(1),
pp. 225253, Available:http://epubs.siam.org/doi/abs/10.1137/070700231
[22] T. J. Willmore, Note on embedded surfaces, Analele Stiintifice ale Universit atii Al. I.
Cuza din Iasi. Serie Noua, vol. Ia 11B, pp. 493496, 1965.
[23] M. Droske and M. Rumpf. (2004). A level set formulation for Willmore flow. Interfac. Free
Boundar. [Online]. 6(3), pp. 361378,Available: http:// www.ems-ph.org/doi/10.4171/IFB/105
[24] A. Top, G. Hamarneh, and R. Abugharbieh, Spotlight: Automated confidence-based user
guidance for increasing efficiency in interactive 3D image segmentation, in Proc. Med. Imag.
Comput. Comput.-Assist.Interv. Workshop Med. Comput. Vis., 2010, pp. 204213.
[25] D. Adalsteinsson and J. A. Sethian, A fast level set method for propagating interfaces, J.
Comput. Phys., vol. 118, pp. 269277, 1995.

[26] M. Sussman, P. Smereka, and S. Osher, A level set approach for computing solutions to
incompressible 2-phase flow, J. Comput. Phys., vol. 114, no. 1, pp. 146159, 1994.
[27] V. Caselles, R. Kimmel, and G. Sapiro. (1997). Geodesic active contours. Int. J. Comput.
Vis.[Online].22(1),pp.6179,Available:

http://www.springerlink.com/openurl.asp?

id=doi:10.1023/A:1007979827043
[28] T. Chan and L. Vese, Active contour without edges, IEEE Trans. Imag. Process., vol. 10,
no. 2, pp. 266277, Feb. 2001.
[29] L. Dice, Measures of the amount of ecologic association between species, Ecology, vol.
26, pp. 297302, 1945.
[30] R. T. Rockafellar and R. J.-B. Wets, Variational Analysis. NewYork:Springer-Verlag, 2005.
[31] S. H. Zhou, I. D. McCarthy, A. H. McGregor, R. R. H. Coombs, and S. P. F. Hughes. (2000,
Jun.). Geometrical dimensions of the lower lumbar vertebraeanalysis of data from digitised CT
images. Eur. Spine J. [Online]. 9(3), pp. 242248, Available: http://www.springerlink.com/
openurl.asp?genre=article&id=doi:10.1007/s0 05860000140
[32] M. Aslan, A. Ali, A. Farag, H. Rara, B. Arnold, and P. Xiang. 3D vertebrae body
segmentation using shape based graph cuts. IEEE ICPR, pages 39513954, 2010.
[33] M. Aslan, A. Farag, B. Arnold, and X. Ping. Segmentation of vertebrae using level sets with
expectation maximization algorithm. IEEE ISBI, pages 20102013, 2011.
[34] Y. Boykov and G. Funka-Lea. Graph Cuts and Efficient N-D Image Segmentation. IJCV,
70:109131, 2006.
[35] T. Cootes, C. Taylor, D. Cooper, and J. Graham. Active shape models-their training and
application. Computer Vision and Image Understanding, pages 3859, 1995.

[36] S. Ghosh, R. Alomari, V. Chaudhary, and G. Dhillon. Automatic lumbar vertebrae

segmentation from clinical CT for wedge compression fracture diagnosis. SPIE,pages 19, 2011.
[37] J. Gower. Generalized Procrustes Analysis. Psychometrika, pages 3350, 1975.
[38] T. Heimann and H. Meinzer. Statistical shape models for 3D medical image segmentation: A
review. MIA, 2009.
[39] J. Herring and B. Dawant. Automatic lumbar vertebral identification using surface-based
registration. Computers and Biomedical Research, pages 7484, 2001.
[40] V. Jain and H. Zhang. Robust 3D shape correspondence in the spectral domain. IEEE Shape
Modeling International,pages 118129, 2006.
[41] M. Kelm, K. Zhou, M. Suehling, Y. Zheng, M.Wels, and D. Comaniciu. Detection of 3D
spinal geometry using iterated marginal space learning. MICCAI:Workshop on
Medical Computer Vision, pages 96105, 2009.
[42] J. Ma, L. Lu, Y. Zhan, X. Zhou, M. Salganicoff, and A. Krishnan. Hierarchical
segmentation and identification of thoracic vertebra using learning-based edge detection
and coarse-to-fine deformable model. MICCAI, pages 1927, 2010.
[43] A. Mastmeyer, K. Engelke, C. Fuchs, and W. Kalender. A hierarchical 3D segmentation
method and the definition of vertebral body coordinate systems for QCT of
the lumbar spine. MIA, pages 560577, 2006.
[44] Z. Peng, J. Zhong, W. Wee, and J. Lee. Automated vertebra detection and segmentation from
the whole spine MR images. Engineering in Medicine and Biology,pages 25272530, 2005.
[45] M. Sofka, K. Ralovich, N. Birkbeck, J. Zhang, and K. Zhou. Integrated Detection Network
(IDN) for pose and boundary estimation in medical images. IEEE Biomedical Imaging, pages
294 299, 2011.

[46] M. Wels, M. Kelm, A. Tsymbal, M. Hammon, G. Soza, M. Suehling, A. Cavallaro, and D.

Comaniciu. Multistage osteolytic spinal bone lesion detection from CT data with internal
sensitivity control. SPIE, pages 83151318, 2012.
[47] M. Wels, Y. Zheng, G. Carneiro, M. Huber, J. Hornegger, and D. Comaniciu. Fast and robust
3D MRI brain structure segmentation. MICCAI, pages 575839, 2009.
[48] T. Whitmarsh, L. Barquero, S. Gregorio, J. Sierra, L. Humbert, and A. Frangi. Age-related
changes in vertebral morphometry by statistical shape analysis. Pages 3039. 2012.
[49] Y. Zheng, A. Barbu, B. Georgescu, M. Scheuering, and D. Comaniciu. Four-chamber heart
modeling and automatic segmentation for 3D cardiac CT volumes using marginal space learning
and steerable features. IEEE TMI, pages 16681681, 2008.
[50] T. Klinder, J. Ostermann, M. Ehm, A. Franz, R. Kneser, and C. Lorenz. Automated modelbased vertebra detection identification and segmentation in CT images. MIA, pages 471482,
2009.