DR Kumar Rajamani Thesis Final

Aus dem MEM Forschungszentrum,
Institut für Chirurgische Technologien und Biomechanik

Direktor: Prof. Dr.-Ing. L.-P. Nolte
Arbeit unter der Leitung von Dr. Miguel Angel Gonzalez Ballester,
Prof. Dr.-Ing. L.-P. Nolte,
Prof. Martin Styner
Three Dimensional Surface Extrapolation from Sparse Data using

Deformable Bone Models
Inaugural-Dissertation zur Erlangung der Doktorwürde der Philosophie im Fach

Biomedizinische Technik der Medizinischen Fakultät der Universität Bern
Vorgelegt von:
Kumar T. Rajamani
von Indian
Von der Medizinischen Fakultät der Universität Bern auf Antrag der Dissertations-
kommission als Dissertation genehmigt.
Promotionsdatum:
Der Dekan der Medizinischen Fakultät:

Three Dimensional Surface Extrapolation from Sparse Data using
Deformable Bone Models
by
Kumar T. Rajamani
Dissertation submitted to the MEM Research Centre

Institute for Surgical Technology and Biomechanics
Faculty of Medicine
University of Bern
for the requirements of the degree of
Doctor of Philosophy in Biomedical Engineering
2006
Advisory Committee:
Dr. Miguel Angel Gonzalez Ballester,

Prof. Martin Styner , Advisor
Prof. Dr.-Ing. L.-P. Nolte, Co-Adviser
Prof. Martin Styner, Tutor
ABSTRACT
Title of Dissertation: STATISTICAL DEFORMABLE MODELS
FOR SURFACE EXTRAPOLATION
FROM SPARSE DATA
Kumar T. Rajamani, Doctor of Philosophy, 2006
Dissertation directed by: Dr. Miguel Angel Gonzalez Ballester,

Prof. Martin Styner
MEM Research Centre,
Institute for Surgical Technology and Biomechanics
A majority of pre-operative planning and navigational guidance during com-

puter assisted orthopedic surgery (CAOS) routinely uses three-dimensional (3D)
models of patient anatomy. These models enhance the surgeon’s capability to
decrease the invasiveness of surgical procedures and increase their accuracy and
safety. A common approach for this is to use computed tomography (CT) or mag-
netic resonance imaging (MRI). These have the disadvantages that they are expen-
sive and/or induce radiation to the patient. Additionally a number of orthopedic
surgeries such as total hip arthoplasty (THA) and total knee arthoplasty (TKA)
do not warrant a pre or intra-operative scan. Visualization in such image-free pro-
cedures entails set of landmarks, navigational primitives and surgical instruments.
The mental recognition of 3D structures from 2D views could be very demanding
for the surgeon.
The main goal of this thesis is the formulation and evaluation of techniques to
construct a patient-specific three-dimensional model that provides an appropriate
intra-operative visualization without the need for a pre or intra-operative imaging.
The 3D model is reconstructed by fitting a statistical deformable model to mini-
mal sparse 3D data consisting of digitized landmarks and surface points that are
obtained intra-operatively.
A statistical model was initially constructed from a collection of previous CT
data-sets. The primary application that we focus on is hip surgery such as total
hip replacement (THR) we began by concentrating on the proximal femur. Our
database comprised of 30 CT scans of patient hips. The proximal femurs were
separated and were manually segmented to get a population of 3D surface meshes.
Correspondence establishment is the first and vital step in constructing a statistical
model. A comparison of some of the existing correspondence establishing methods
was therefore studied. This helped to identify the strategy to construct a compact
model with fine model characteristics. Dense point correspondences are used to
construct the shape model.
The first shape reconstruction method that was realized was based on pro-
gressive elimination of variation. The shape is controlled by selecting locations
of certain landmarks and surface points. The positions of these landmarks intro-
duce boundary conditions for the shape of the organ. After one control vertex of
the surface is adjusted, the variability that causes the displacement of that vertex
is removed from the model and this process is repeated. The number of control
vertex that could be defined were limited.
An alternative shape reconstruction method was formulated as a least squares
error minimization with additional regularization terms that computes the Maha-
lanobis distance of the predicted model. We solve for the shape parameters that
minimize the residual errors between the reconstructed model and the cloud of
landmarks and surface points. This reconstruction method could seamlessly han-
dle both small and large sets of digitized points and provide real time interactivity.
This technique has been validated on plastic and dry cadaver bones.
We then explored using ultrasound imaging for non-invasive intra-operative
surface points digitization. We have evaluated the application of our deformable
bone models concurrently with automatic segmentation of 2D B-mode ultrasound
contours.
An optimal estimation framework was finally established by setting up a three-
stage estimation process. The first stage iteratively estimates the optimal scale
and coordinate transformation between the mean shape of the statistical model
and cloud of input points. This is then used as input to the Mahalanobis weighted
least square fit, which optimally and robustly reconstructs the shape. A kernel-
based shape deformation is used as a third stage to further improve the reconstruc-
tion accuracy. The framework has been extensively evaluated using leave-one-out
experiments and validation has been carried out on plastic and dry cadaver bones.
DEDICATION
To my MOTHER
v
vi
ACKNOWLEDGEMENTS
I would first like to express my sincere gratitude to Prof. Lutz-Peter Nolte

for giving me an opportunity to explore this fascinating field of computer assisted
surgery. I would like to acknowledge his permanent support and advice as well for
his helpful disposition.
I heartily thank Dr. Miguel Gonzalez Ballester for his support and guiding me
through the various stages of my project and also giving me valuable ideas as well
as enthusiasm, which largely contributed to the completion of this dissertation. I
gratefully recognize the numerous stimulating suggestions he contributed at each
stage of the work.
I would also like to warmly thank Dr. Martin Styner for his continuous support
in the course of the past four years. He introduced me to the concepts of medical
image analysis and statistical modeling. His inspiration right from the initial stages
of conception of ideas through the long and tedious experimental and validation
stages provided great motivation. I gratefully cherish the many fruitful discussions
we have had together.
It was a great experience working together with Haydar Talib. He provided a
number of interesting ideas and ways in which this work could be extended and
extensively validated. My sincere thanks also to him for standing by throughout
whenever I had needed help and also for his valuable input and expert comments
in proof reading this thesis.
Many thanks to my colleagues and friends from the MEM research center. Spe-
cial thanks to the Medical Image Analysis ”MIA” group, Jason Paulonis, Digvijay
Singh , Manon Blouin, Ravi Iyer, Rudolf Sidler, Jaime Garcia, Nina Kozic, Laura
Belenguer Querol, Thierry Zimmermann, for making the time spent together mem-
orable and enjoyable. I would also like to acknowledge the encouragement and
support from Dr. Ion Pappas, who was always behind my back providing lot of
confidence and also for all the administrative and organizational support from the
secretaries Esther Gnahore and Annelies Neuenschawander.
I also received an enormous amount of personal support from my family back
in India, whose constant encouragement helped in successfully completing this
dissertation. Accordingly, I wish to thank, deeply from the bottom of my heart,
my spiritual guides, Sri Sathya Sai Baba and Sri Mata Amritanandamayi, my
parents, T.M.Rajamani and K.E. Padmavathy, my brother T.R. Sridhar and my
sister T.R. Srividya.
vii
viii
TABLE OF CONTENTS
Acronyms xi
List of Tables xii
List of Figures xiv
1 Introduction 1
2 Anatomical Shape Reconstruction from Sparse Data 7

2.1 Statistical Model Generation . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Model Deformation by Progressive elimination of variation . . . . . 9
2.3 Model Deformation by Minimization of Mahalanobis distance . . . . 10
2.4 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Ultrasound-Initialized Deformable Bone Models . . . . . . . . . . . 14
2.6 Kernel based shape deformation . . . . . . . . . . . . . . . . . . . . 15
3 Conclusions and Outlook 19
4 Publications 21
4.1 The Complete List of Publications . . . . . . . . . . . . . . . . . . . 22
4.1.1 Journal Articles . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1.2 Conference Articles . . . . . . . . . . . . . . . . . . . . . . . 22
Bibliography 25
ix
x
ACRONYMS
xi
xii
LIST OF TABLES
2.1 Mean and Median surface errors from their actual surfaces in five
different trials for the two cast femur bones, using ultrasound to
acquire surface points . . . . . . . . . . . . . . . . . . . . . . . . . . 15
xiii
xiv
LIST OF FIGURES
1.1 Planning module for performing tumour biopsies in neurosurgery . . 2
1.2 Basic components of a computer assisted surgery system, including

representations of the associated local coordinate systems. Top left
is the positional sensor, in the middle is the surgical object with the
end-effector, and on the right is the virtual representation of the
surgical object on the screen of a navigation system. . . . . . . . . . 3
1.3 Figure illustrating the concept of correspondence establishment. Three

different proximal femurs are shown each represented as a dense col-
lection of points (surface models). The correspondence establishes
one-to-one relationship across identical locations, as shown by the
dotted lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Top row and bottom left: Graphs with error plots of compact-
ness (C(M)), generalization (G(M)) and specificity (S(M)) for the
femoral head study. Bottom row, right: Table with average, maxi-
mal and minimal mean absolute distances (MAD) between the man-
ual landmarks and the studied methods for the femoral head study.
There is little change between DetCov and MDL. SPHARM clearly
shows the worst performance of all studied methods. For compari-
son, the mean landmark selection error was 2.5mm. . . . . . . . . . 8
2.2 The first two eigen modes of variation of our model after the variabil-
ity associated with three landmark points has been removed from
the population. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 View of the predicted most probable surface overlaid on top of the
actual object in a leave-one-out analysis with six points being dig-
itized only on the femoral head. The average error of the bone
morphing in this case was at 4.2 mm . . . . . . . . . . . . . . . . . 11
xv
2.4 Left: A typical proximal femur of the population that was used in
the leave-one-out test. Middle: The average shape of the popula-
tion with color coded distance map to the actual shape. The mean
surface error is 3.37 mm and the median surface error is 2.65 mm.
Right: The shape based on only 6 digitized points with color coded
distance map to the actual shape. The mean surface error is 1.50
mm and the median surface error is 1.25 mm . . . . . . . . . . . . 12
2.5 Reconstruction errors of 10 different femurs using leave-one-out ex-
periments with 10 digitized points. The maximum, 95-percentile,
median and mean error with standard deviation are plotted for each
femur. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.6 Ultrasound-based prediction: Predicted models overlaid onto “gold”
references. Bone 1 (left): 3.08 mm mean error and Bone 2 (right):
2.94 mm mean error . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.7 One of the reconstruction examples using points only from the sil-
houettes of the surface; first column: the anterior-posterior silhou-
ette (top) and the lateral-medial silhouette (bottom); second col-
umn: the actual surface model rendered together with the silhou-
ettes; third column: the final reconstructed surface rendered to-
gether with the actual surface. . . . . . . . . . . . . . . . . . . . . 17
2.8 Reconstruction errors of 7 dry cadaver femurs using the estimation
framework when 90 points were used in each case. . . . . . . . . . 17
xvi
Chapter 1
Introduction
The domain of computer assisted surgery (CAS), and more specifically computer
assisted orthopaedic surgery (CAOS), is an emerging area of research which aims
at uniting the disciplines of computer science and engineering with classical surgi-
cal interventions. The goal of CAOS is to enhance the dexterity, visual feedback
and information integration for the surgeon and enable him to carry out surgical
interventions that are more precise and less invasive than conventional procedures.
There are several technologies that have been introduced to achieve this goal and
these include three-dimensional imaging modalities such as CT and MRI, powerful
computers with sophisticated graphics capabilities, robotics and surgical naviga-
tion.
The mark of successful computer assisted surgery is to unify the pre-operative
planning, intra-operative procedure, and post-operative followup that make up, in
a general sense, the surgical workflow. Typically, quantitative information obtained
from any of these phases is not easily transferred to another phase. For instance,
measurements made during a pre-operative planning step are difficult to correlate
to intra-operatively acquired images, e.g. fluoroscopy, which are used to visually
(and not quantitatively) assess the ongoing procedure. The aim of CAS is to link
these phases through the use of computers, and also to provide more information
to the surgeon at every phase. For the pre-operative and post-operative phases,
the introduction of computers may simply imply storing all information (images,
planning or assessment measurements) in a digital format that can be accessed
through a number of other computer stations. The greatest difference emerges
intra-operatively when the concept of navigation is introduced into the operating
room (OR). The purpose of navigation is to provide quantitative feedback about
the position of the anatomical structures and surgical tools, via tracking devices
such as optical cameras or magnetic trackers. Thus, intra-operative information
can be linked to pre-operative information. Pre-operative information comes from
planning on acquired images. Figure 1.1 shows an example of a planning module
used to plan a trajectory for performing a tumor biopsy in neurosurgery. Intra-
1
Figure 1.1: Planning module for performing tumour biopsies in neurosurgery
operative information comes from the position of tools at all times during the
surgery. Tools are tracked with high accuracy and they can also be used to acquire
3D locations from the anatomy of interest. This is called point digitization.
Navigation systems can be characterized by three major components: the sur-
gical object, the virtual object, and the navigator as portrayed in Figure 1.2. In
CAOS, the surgical object is the set of bones and accompanying tissues in the
surgical field. The virtual object is the virtual representation of the surgical object
obtained through the various imaging means. It enables the surgeon to plan the
intended intervention virtually. Finally, the navigator establishes a coordinate sys-
tem in which the location and orientation of the target as well as “end-effectors” are
expressed. The end-effectors can be surgical instruments or active devices (Nolte
and Beutler, 2004). In the context of CAOS surgery, navigation systems differ in
the way information on the surgical object is acquired. Current navigation systems
can be classified into CT-based, fluoroscopy-based, and image-free categories.
Certain surgical interventions, such as total hip and knee replacements are
typically performed without the use of any pre- or intra-operative radiological
images. This motivated the development of image-free approaches. For some
procedures, in which pre-operative images are routinely acquired, it may also be
desirable to obtain image-free solutions in order to reduce costs and radiation
exposure to the patient. Some procedures also use intra-operative imaging, such
as fluoroscopy, and in such cases both patient and surgical staff can benefit from
reduced radiation exposure with the use of 3D anatomical models. Typically,
the use of intra-operative imaging cannot be fully circumvented, but if additional
visualization is possible, then it may not be necessary to perform as many scans
2
Figure 1.2: Basic components of a computer assisted surgery system, including
representations of the associated local coordinate systems. Top left is the positional
sensor, in the middle is the surgical object with the end-effector, and on the right
is the virtual representation of the surgical object on the screen of a navigation
system.
as is normally done.
The basic concept of image-free surgical navigation is to create a virtual repre-
sentation of the surgical object using a tracking system and a tracked pointer to de-
fine locations on the anatomical structure of interest. Since pre- or intra-operative
radiological images are unavailable, the virtual representation is “surgeon-defined”.
This concept is also termed computer assistance based on surgeon defined anatomies.
A major challenge to the surgeon in image-free surgical navigation is that the points
digitized give only a partial representation of the shape of the object, and thus re-
quire mental recognition of 3D structures. This is very demanding and needs
extensive training even for well-defined simple structures. Visualization in such
procedures usually entails a set of geometric landmarks, navigational primitives
and surgical instruments.
In order to overcome this problem, statistical shape models (Cootes et al.,
1994) were introduced in CAS. The basic idea of statistical shape models is to
build a model representing the “average” shape of an object, as well as the main
patterns of shape variability with respect to this average shape. A first phase
consists of learning the shape characteristics from a training set (a group of data
that represents a population) of examples (e.g. bone surfaces segmented from CT
images). Once this model is built, we obtain a compact representation of the
shape variability in the training population, and it is possible to generate new
3
Figure 1.3: Figure illustrating the concept of correspondence establishment. Three
different proximal femurs are shown each represented as a dense collection of points
(surface models). The correspondence establishes one-to-one relationship across
identical locations, as shown by the dotted lines.
valid shapes from this representation. Typically, this model is used to constrain
segmentation approaches, as to only generate shapes that are contained in the
model. Statistical shape models, in the context of image-free surgery, can be
used to provide a complete 3D representation of the anatomy of interest from
the original sparse set of digitized points, i.e. it provides a means to extrapolate
surface information. As a result of this process, it is possible to obtain a 3D shape
representation of the target anatomy without the need of any images, and still
remain compliant with minimally invasive surgical approaches.
One key step in the construction of a statistical shape model is correspondence
establishment. Each shape in the training set is represented as a dense collection
of 3D points. In order to build a valid statistical model, it is of key importance
that identical locations across the different shapes are consistently identified. One
way to reach this goal is to manually identify similar landmark locations across the
different shapes. This approach can be quite tedious and makes it very difficult
to establish dense correspondence across the whole surface. Automatic methods
have been proposed to build a parametrization of each surface that complies with
the correspondence constraint, as described by Brechbühler et al. (1995). This
approach yields a dense correspondence establishment across the different shapes.
The concept of correspondence is illustrated in Figure 1.3.
Once correspondence has been established, statistical multivariate factor anal-
ysis techniques are used to decompose the shape variation of the training shape.
The most popular statistical method employed to this end is principal component
analysis(PCA) (Jolliffe, 1986). The purpose of principal component analysis is to
capture the variability across the different shapes. First, the mean and covariance
matrix of the set of shapes is computed. Then an eigendecomposition of the co-
variance matrix is computed. This results in an ordered set of eigenvectors with
decreasing corresponding eigenvalues. Each of these eigenvectors represents a prin-
cipal mode of shape variation, and the associated eigenvalue gives an idea of the
4
amount of shape variance explained by each mode of variation. New objects can
be computed as a linear combination of the mean shape and eigenvectors of the
covariance matrix (ref. another chapter).
The variability captured by the PCA can be used in the construction of fitting
schemes to construct patient-specific 3D models. The idea of the fitting scheme is
to find the scalar (notation that you use in the formula) parameters that best fit
the intra-operatively acquired input data. The patient specific information that
is provided as input is the set of bone surface points. This error minimization
problem is typically formulated as the minimization of Euclidean distance of the
estimated shape (drawn from the statistical model) and the set of input points.
In CAOS, the use of statistical shape models for shape recovery from digitized
point data has been explored by Fleute and Lavallée (1998), who fit the deformable
model surface to intra-operatively digitized point data via jointly optimizing de-
formation and pose (orientation of the anatomy). Chan et al. (2003) optimize
deformation and pose separately using an iterative method. Current shape pre-
diction techniques demand a large and dense collection of surface data as input.
This can be very time-consuming for the surgeon and also error prone. Addition-
ally, there is often lack of accessibility to the entire surface, especially in the case
of minimally invasive procedures. The motivation of this work is to address the
problem of robustly reconstructing anatomical shapes given very sparse surface
information. The work presented in this thesis describes our proposed methods
for extrapolating the three dimensional surface of a given anatomy using minimal
surface information, based on the use of statistical shape models. In the following
chapter we present the various steps for the realization of such a framework and
the different methods that were developed and explored in the course of this work.
5
6
Chapter 2
Anatomical Shape Reconstruction

from Sparse Data
This chapter briefly describes the evolution through various stages of our proposed
anatomical shape reconstruction from sparse data, which include (1) statistical
model generation; (2) model deformation by progressive elimination of variation;
(3) model deformation by minimization of Mahalanobis distance; (4) ultrasound-
initialized deformable bone models and (5) kernel based shape deformation. Details
regarding the methodological formulation and validation studies are outlined in the
respective publications, which will be cited in this chapter, and are provided as
annexes to the thesis.
2.1 Statistical Model Generation

A statistical shape model is constructed from a training database.
The primary application that we focus on targets hip interventions, such as
total hip replacement (THR). Our database is therefore comprised of CT scans
of patient hips, and we obtained a total of 30 scans. Several different geometric
representations have been used to model anatomy (Bookstein (1986), Cootes et al.
(1995), Kelemen et al. (1999)). For our model building we have employed the
representation of shapes using point distribution models (PDM) (Cootes et al.,
1995). The proximal femurs were segmented semi-automatically (ref. Szemely),
and surface models of the bones were extracted.
A key step in this model building involves establishing a dense correspondence
between shape boundaries over a reasonably large set of training data. We com-
pared the methods introduced by Brechbühler et al. (1995) (SPHARM), Kotcheff
and Taylor (1998) (DetCov), Davies et al. (2002) (MDL) and a fourth method
based on manually initialized subdivision surfaces similar to Wang et al. (2000)
(MSS). We analyzed both the direct correspondence via manually selected land-
marks as well as the properties of the model implied by the correspondences, in
7
Compactness C(M) Generalization G(M)
Pelvis Generalisation Ability
5 Pelvis Compactness
x 10
3 8
MSS
SPHARM
2.5
DetCov
7
MDL
2
6
G (M)
C (M)
1.5
5
1
4
MSS
SPHARM
0.5
DetCov
MDL 3
0
0 2 4 6 8 10 12 14 2 4 6 8 10 12
M M
Specificity S(M)
Pelvis Specificity Landmark error table
5.8
Method Femoral head

5.6
Mean Max Min
5.4
MSS 3.13 mm 6.48 mm 1.73 mm
5.2
SPHARM 7.24 mm 12.2 mm 6.07 mm
S (M)
5 DetCov 3.24 mm 7.36 mm 1.71 mm

4.8 MDL 3.40 mm 6.41 mm 1.71 mm
MSS
4.6 SPHARM
DetCov
MDL
2 4 6 8 10 12
M
Figure 2.1: Top row and bottom left: Graphs with error plots of compactness
(C(M)), generalization (G(M)) and specificity (S(M)) for the femoral head study.
Bottom row, right: Table with average, maximal and minimal mean absolute dis-
tances (MAD) between the manual landmarks and the studied methods for the
femoral head study. There is little change between DetCov and MDL. SPHARM
clearly shows the worst performance of all studied methods. For comparison, the
mean landmark selection error was 2.5mm.
regard to compactness, generalization and specificity. The results of our study are
shown in Figure 2.1.
Our comparison study of these common correspondence establishing methods
revealed that for modeling purposes the best among the correspondence methods
was Minimum Description Length (MDL) (Davies et al. (2002)). Based on the
study, for our model building, correspondence was initialized using MSS and then
optimized based on the MDL criteria. The details of this study and the results are
elaborated in our publication (Styner et al. (2003)).
The statistical shape model is constructed based on the established point corre-
spondences. This is achieved using Principal Component Analysis (PCA) (Jolliffe
(1986)). Each member of the training population is described by individual vectors
8
containing all 3D point coordinates. For the computation of PCA, the mean shape
and the covariance matrix are computed from the set of object vectors. The sorted
eigenvectors of the covariance matrix are the principal directions spanning a shape
space with mean shape at its origin. Objects in that shape space can be described
as linear combinations of eigenvectors.
2.2 Model Deformation by Progressive elimina-

tion of variation
Given the statistical shape model that enabled us to capture the variability of the
different shapes and the sparse set of patient specific surface points we started
determining ways in which the model could be deformed to fit this set of points.
The first idea was based on similar work for initializing a statistical model for
segmentation explored by Hug et al. (2000).
The statistical model is first initialized to the point space (this is the patient
information, which can be a cloud of points spanning the anatomy). Initialization
can be, for instance, performed manually as the model is aligned to the data, or
semi-automatically, as we propose by using three user-defined anatomical land-
marks. The purpose of this step is to establish a rough correspondence between
the mean model and the data. After the first step, each point is considered inde-
pendently, and a predicted shape is given iteratively as the points are treated one
by one.
For each point from the point cloud, a corresponding point on the model’s
surface is obtained. This again, can be done manually to ensure more or less
accurate correspondence, but we propose a nearest-point search, using Euclidean
distance. A three dimensional unit vector is then extracted from the corresponding
point on the surface, which controls the deformation of the model.
The vector contains information on the possibilities of deformation, which are
constrained in the sense that the new shape should remain ”near” the mean shape.
This constraint is achieved in the extraction of the vector, which is unique and
minimizes the Mahalanobis distance of the new shape with respect to the mean
shape of the deformable model.
This vector is then appropriately scaled and then added to the mean shape
to estimate a new shape that passes through the chosen point. To ensure that
subsequent shape modifications do not alter an already fixed point, we remove
those components from the shape space that cause a displacement of this point.
This is done by subtracting the variation coded by the point from each instance
and the statistical model is reconstructed using these new instances.
This procedure of point selection and variability removal is repeated until a
close approximation to the patient anatomy is achieved. The final extrapolated
surface represents the most probable surface in the shape space given the digitized
9
√ √ √ √
-2 λk -1 λk +1 λk + 2 λk
λ1
λ2
Figure 2.2: The first two eigen modes of variation of our model after the variability
associated with three landmark points has been removed from the population.
landmarks. Figure 2.2 shows the first two eigen modes of variation of our model
after the variability associated with three landmark points has been removed from
the population. Figure 2.3 shows one of the estimation results in a leave-one-out
analysis with six points being digitized only on the femoral head. The probable
surface is overlaid on top of the actual object. The average error of this deformation
scheme computed across the whole surface was 4.2mm. The major limitation was
that the number of control vertices that could be defined were limited by the
number of instances used for constructing the model. The details of the methods
and the results are elaborated in our publication (Rajamani et al. (2004b)).
2.3 Model Deformation by Minimization of Ma-

halanobis distance
The core of the algorithmic development is the formalization and refinement of an
elegant yet simple shape reconstruction technique. The key factor is the observa-
tion that objects in our shape space, and by our hypothesis the 3D shape that we
try to estimate, can be described as the mean shape plus a weighted linear com-
bination of eigenvectors. The problem is therefore formulated as estimating the
weights for this unknown shape, such that the errors between the reconstructed
model and the cloud of digitized surface points is minimized. This can be formu-
lated as a least-squares problem.
A standard least-square solution would be adequate, if the shape to be esti-
mated is sufficiently constrained. This is the case if a large collection of digitized
surface points are available. In the scenario that we consider of sparse set of surface
10
Figure 2.3: View of the predicted most probable surface overlaid on top of the
actual object in a leave-one-out analysis with six points being digitized only on
the femoral head. The average error of the bone morphing in this case was at 4.2
mm
points, it turns out that such a straightforward solution tend to produce contorted
shapes. Hence additional regularization is needed to stabilize the estimation. We
have therefore formulated the problem as a least squares error minimization with
an additional regularization term that computes the Mahalanobis distance of the
predicted model to the shape distribution (Rajamani et al., 2004a). The Maha-
lanobis distance term enables stable prediction with minimal number of known
surface points.
The objective function that we minimize is defined as follows
 
N m
( m )
X α2

 

~ +
X X
f =ρ wk kY~k − (X̄ j αi p~i (j))k2 + (1 − ρ) i
(2.1)

 k=1 i=1

 i=1
λi
j=Index(k)
The first term in the function is the Euclidean distance between the N digitized
points Y~ and the estimated shape comprising the mean X̄ ~ plus a weighted sum
of the eigenvectors p~i . The corresponding point for each of the digitized points
Y~k is computed using closest point correspondence from the current estimated
shape. This is denoted as X̄ ~ , where j = Index(k) is the index of the closest point
j
corresponding to the k th digitized point. The second term is the Mahalanobis
distance of the predicted shape from the mean and controls the probability of the
11
Figure 2.4: Left: A typical proximal femur of the population that was used in
the leave-one-out test. Middle: The average shape of the population with color
coded distance map to the actual shape. The mean surface error is 3.37 mm and
the median surface error is 2.65 mm. Right: The shape based on only 6 digitized
points with color coded distance map to the actual shape. The mean surface error
is 1.50 mm and the median surface error is 1.25 mm
predicted shape. This term ensures that the predicted shapes are valid by favoring
those that are closer to the mean.
The parameter ρ relaxes the effect of the Mahalanobis distance term as ad-
ditional points are digitized. This makes the surface less constrained to remain
close to the mean and allows it to more freely deform (Rajamani et al. (2004c)).
Hence the error between the predicted surface and the set of digitized points is bet-
ter minimized. The parameter wk are M-estimators based weights (Styner et al.
(2000)), and enable robust rejection of outliers.
We present here some of the illustrative results. Figure 2.4 shows an exam-
ple, with mean surface error of 1.44 mm obtained with 20 digitized points. The
color-coded 3D rendering is calculated using Hausdorff’s Distance to measure the
distance between discrete 3D surfaces (Aspert et al., 2002). The reconstruction er-
rors of 10 different femurs using leave-one-out experiments with 10 digitized points
is potrayed in Figure 2.5. The maximum, 95-percentile, median and mean error
with standard deviation are plotted for each femur. The detailed mathematical
formalization and extended validation results including experiments on dry cadaver
bones are presented in our paper (Rajamani et al., 2006).
Our formalization can handle both small and large sets of digitized points with
little computational burden, and is performed in real time as shape parameters are
determined by solving a single linear system. The formalization also enables the
incorporation of the complete set of eigenvectors for the shape estimation.
2.4 Validation
At each step of our statistical deformable bone model formulation, and in the de-
velopment of applications, it was necessary to study different aspects of our system
in order to determine the potential of using our approach in clinical scenarios. We
12
Shape Estimation Using 10 Points
14
Maximum
95−percentile
Median
12
Mean
Reconstruction Error [mm]

10
0
Femur 1 Femur 2 Femur 3 Femur 4 Femur 5 Femur 6 Femur 7 Femur 8 Femur 9 Femur 10
Figure 2.5: Reconstruction errors of 10 different femurs using leave-one-out exper-

iments with 10 digitized points. The maximum, 95-percentile, median and mean
error with standard deviation are plotted for each femur.
list these briefly in this subsection, though they are described in more detail in
other subsections as well as in the adjoining publications.
• Leave-one-out tests were performed when we were developing the ideal de-
formation algorithm for our intended application, as described in section 2.3
and section 2.2. The aim of these tests was to establish a proof-of-concept
of the methods we compared, and each method was performed several times.
Each test consisted of excluding one data set from the training population,
and using the remaining data to construct the model that would be used to
estimate the shape of the removed data. All possibilities were tested, and
in the case of a 30-data-set population, 30 tests were run per deformation
method.
• Once our method was deemed suitable (high accuracy), we began using it
in the framework of a navigation system. A tracked pointer was used to
acquire bone surface points from cast femurs, and the predicted shape was
compared to a “gold” reference obtained from a CT scan. The aim was to
begin approaching a more realistic scenario in evaluating our method (Talib
et al., 2005).
• In conjunction with the pointer-based study, we also evaluated the use of

navigated ultrasound imaging for bone surface point acquisition. More de-
tails on this study will be given in section 2.5 as well as in (Talib et al.,
13
2005).
• We observed that the different bones we were using for the pointer-based
and US-based approaches were behaving differently in terms of accuracy of
predictions, and so we studied the performance of our deformable model
with respect to varying population sizes, as well as using cadaveric bones
(Rajamani et al., 2005).
• The latest formulation of our deformable model incorporated the use of non-
linear deformation, and this was studied with fluoroscopy-based point ac-
quisition, as will be described in section 2.6 and detailed in (Zheng et al.,
2006a).
Each of these studies provided insight to the various aspects of our statistical
model deformation method, and provided useful validation in determining the
feasibility of our approach.
2.5 Ultrasound-Initialized Deformable Bone Mod-

els
In Chapter 1, we briefly discussed how digitizing pointers are typically used to
acquire bone surface information for deformable models. Due to limited surgical
access especially in minimal invasive surgery, it may be difficult to acquire a set
of points that sufficiently spans the patient’s anatomy to ensure accurate shape
prediction of a given statistical model. When considering the standard approach
to digitizing bone surface points in CAOS, the main constraint is the bone region
that is exposed during the procedure. The exposed region may be large enough for
the operation, but does not necessarily provide enough information to represent
the shape of the complete anatomy. As such, it became interesting to investigate
a method that would enable the surgeon to non-invasively acquire points from
non-surgically accessible areas of the anatomy, such as through medical imaging.
For this study we chose ultrasound (US) imaging technology in conjunction
with CAOS with the aim of capturing bone surface points from the inaccessible
areas. The choice for US imaging stems from its benefits, namely that: it can be
used intra-operatively due to its mobility; it is relatively safe and radiation-free; it
is a common and inexpensive imaging device; and its properties make it suitable
for imaging bone, even though the typical US system is used to view soft tissues
(US waves are fully reflected at tissue-bone interfaces).
In order to use US imaging in the scope of CAOS, the US probe is rigidly fixed
with a dynamic reference base (DRB). This enables the probe to be tracked like
any other tool. The coordinates of the US imaging plane are determined with
respect to this DRB, and this is achieved through a calibration step (Kowal et al.,
14
Bone 1 Bone 2
error [mm] error [mm]
Trial Num. Mean Median Mean Median
1 3.88 3.56 2.94 2.12
2 4.37 4.28 5.30 4.98
3 6.88 6.34 3.79 3.48
4 4.75 4.51 4.57 4.54
5 3.08 2.53 3.12 2.84
Average 4.59 4.35 3.95 3.59
Table 2.1: Mean and Median surface errors from their actual surfaces in five differ-
ent trials for the two cast femur bones, using ultrasound to acquire surface points
2003). Because of this, each US image that is acquired is localized with respect
to the anatomy of interest (which is also tracked). A series of ultrasound images
are acquired and the bone contours from the US images are obtained using an
automatic segmentation procedure(Kowal et al., 2001). The segmentation pro-
cess yields a cloud of bone surface points that can be provided as input to the
deformation algorithm.
We tested the technique on two different cast proximal femurs that were im-
mersed in a water-bath. A series of 5 trials per bone was carried out, the results
of which are tabulated in Table 2.1. The results in this table show the mean and
median surface errors for the predicted shapes with respect to the “gold” refer-
ences for each bone, using 24 - 26 digitized surface points. Fig. 2.6 shows for each
bone one case of ultrasound based shape prediction, with predicted shape overlaid
to its respective “gold” reference. Further experimental results with comparison
to pointer based experiments are detailed in our paper (Talib et al., 2005). From
our study, we were able to observe the limitations and potential quality of shape
prediction using navigated ultrasound.
2.6 Kernel based shape deformation

The methods that we have explored so far allow us to estimate shape instances that
fall within the shape space. The shape space could be limited by the information
provided by the training samples that were used initially in constructing the model.
The shape to be estimated might have abnormal local shape deviations due to
pathology. It was realized that using our method in conjunction with local non-
linear deformation schemes would help us to better estimate pathological cases.
Non-linear deformation also enables the reduction of reconstruction error, and
yields sub-millimeter accuracy in the general case.
The Mahalanobis distance regularized least squares shape reconstruction scheme
15
Figure 2.6: Ultrasound-based prediction: Predicted models overlaid onto “gold”
references. Bone 1 (left): 3.08 mm mean error and Bone 2 (right): 2.94 mm mean
error
has therefore been incorporated into a three stage optimal estimation framework
(Zheng et al. (2006b) and Zheng et al. (2006a)).
A dense point distribution model is first constructed. This is achieved by re-
cursively tessellating the mesh. A simple subdivision scheme called Loop scheme,
invented by C.T.Loop (1987), was employed for this. Three new vertices are in-
serted to divide a triangle in coarse resolution to four smaller triangles in fine
resolution.
In the first stage registration parameters between the mean shape of the model
and the set of input points is iteratively estimated. The Iterative Closest Point
(ICP) algorithm is used to optimally estimate the registration parameters. Three
anatomical landmarks are used to initialize the registration procedure. k-D trees
are used to speed up search for the closest paired points. The registration results
from ICP are used to establish point correspondence for the second stage, which
optimally and robustly generates the initial estimate using the statistical method
detailed above.
The dense surface estimated in the second stage is taken as the template sur-
face for the third stage. A smooth nonlinear deformation transformation is used for
deforming the template surface to further reduce the reconstruction error. Thin
plate spline kernels are used as they incorporate smoothness constraints on the
deformation. The spline function smoothly maps corresponding landmarks. The
points in the neighborhood of a landmark are moved similar to the way the land-
mark moves towards its corresponding landmark. A linear equation system is set
up to estimate the affine transformation and kernel interpolation coefficients. One
of the reconstruction examples using points only from the silhouettes of the surface
is shown in Figure 2.7. The final reconstructed surface is color coded rendered to-
16
Figure 2.7: One of the reconstruction examples using points only from the silhou-
ettes of the surface; first column: the anterior-posterior silhouette (top) and the
lateral-medial silhouette (bottom); second column: the actual surface model ren-
dered together with the silhouettes; third column: the final reconstructed surface
rendered together with the actual surface.
Figure 2.8: Reconstruction errors of 7 dry cadaver femurs using the estimation
framework when 90 points were used in each case.
17
gether along with the actual surface. Reconstruction errors using this framework of
7 dry cadaver femurs when 90 points were digitized are presented as an illustrative
result in Figure 2.8. Detailed mathematical formalization and extended results are
elaborated in our paper Zheng et al. (2006a).
18
Chapter 3
Conclusions and Outlook
In this thesis we studied shape reconstruction techniques to extrapolate the three-

dimensional model of a given anatomy using statistical shape models from sparse
data. A systematic methodology has been adopted starting from building a sta-
tistical model to the final realization of an optimal estimation framework.
Correspondence establishment was identified as a key aspect in the construction
of statistical models. A comparison study was carried out using some common
dense 3D correspondence establishing methods and Minimum Description Length
(MDL) was chosen for building a compact statistical model.
The method of progressive elimination of variation was explored as a potential
means for shape reconstruction. The variability associated with landmarks and
identified surface points were iteratively removed from the shape space. Shape
vectors were computed that could displace a given landmark with minimum defor-
mation. The major constraint was that the number of surface points that could
be constrained was limited.
A simple least-squares formalization with Mahalanobis distance based shape
regularization was proposed as a shape reconstruction alternative. The formulation
of the problem as a sum of two distance terms enabled an efficient solution scheme
by linear system solving. The proposed scheme could seamlessly handle small and
large sets of digitized points. The technique was refined by identifying and fine
tuning the parameters for enabling better convergence and achieving resistance to
outliers.
An evaluation study was carried out for using 2D ultrasound in conjunction
with our statistical deformable bone model, and a digitizing pointer was also used
to acquire points, in order to compare the methods. The experiments made use
of two different cast proximal femurs immersed in a water bath. Bone contours in
the ultrasound images were automatically segmented, thereby yielding a cloud of
points to be used for the deformable bone model. CT surface models were used
as “gold” references for error measurements of the predicted bone models (Talib
et al., 2005).
19
The shape reconstruction scheme was then incorporated into a multi-stage op-
timal estimation framework. Kernel based shape deformation based on using Thin-
Plate splines was used to smoothly deform and further improve the reconstruction
accuracy (Zheng et al., 2006a).
Though we obtained promising results from the various studies, there are some
technical challenges that need to be addressed in order to enhance the estimation
scheme:
• The order and the location of the surface points that are provided to the
shape deformation algorithm might have an influence on the overall estima-
tion result. Evaluation and quantification of this dependence could help in
identifying the best ordered set of landmarks and surface points for a given
application.
• Some arising errors are due to inaccurate localization of surface points. These
could be due to improper calibration, tracking error or error due to segmen-
tation in the case of ultrasound. These could perhaps be modeled and incor-
porated into the framework for a robust estimation scheme that can account
for these localization errors. The method presented here is quite general, and
it could also be extended to segmentation applications.
• It would also be interesting to study whether the attributes of the patient,

such as age, sex, height, weight, ethnicity, etc. influence the shape charac-
teristics.
• The training population of bone shapes should be expanded to build more

comprehensive shape models.
• We will also target other anatomies like the distal femur, the complete femur
and the spine. It would be interesting to study the constraints if any, imposed
by the shape of proximal to the distal shape and vice versa.
• Regarding the use of ultrasound for shape estimation, we are also working
on improving the methods for bone detection to improve the accuracy.
The results from our different experiments show potential for our method to be
applicable in clinical settings. We are confident that our method can already be
used for clinical visualization applications, where we can provide 3D models from
very limited sets of digitized points. The proposed technology brings a variety
of advantages to orthopaedic and other surgical procedures, such as improved
accuracy and safety, often reduced radiation exposure, as well as 3D visualization.
In particular, navigation based on shape deformation opens the door to more
minimally invasive approaches.
20
Chapter 4
Publications
The following papers are included in this thesis
Publicaion 1, Medical Image Analysis 2006 (accepted), Statistical Deformable
Bone Models for Robust 3D Surface Extrapolation from Sparse Data
Publication 2, published in Computer Aided Surgery 2005, A Comparison
Study Assessing the Feasibility of Ultrasound-Initialized Deformable Bone Model
Publication 3, IEEE Transactions on Biomedical Engineering (submitted) Ac-
curate and Robust Reconstruction of Proximal Femur from Sparse Intraoperative
Data and Dense Point Distribution Model for Surgical Navigation
Publication 4, published as a chapter in MICCAI 2004, (Springer-Verlag,
LNCS 3217, pp. 478-485) A Novel Approach to Anatomical Structure Morphing
for Intra-operative Visualization
Publication 5, published in SPIE Medical Imaging 2004, Bone Morphing with
statistical shape models for enhanced visualization
Publication 6, published in IPMI 2003, (Springer-Verlag, LNCS 2732, pp.
63-75) Evaluation of 3D Correspondence Methods for Model Building
21
4.1 The Complete List of Publications
4.1.1 Journal Articles
1. Statistical Deformable Bone Models for Robust 3D Surface Extrapolation
from Sparse Data
Kumar T. Rajamani, M. Styner, H. Talib, L.P. Nolte, M. A. Gonzalez
Ballester
Medical Image Analysis 2006 (to appear).
2. A Comparison Study Assessing the Feasibility of Ultrasound-Initialized De-

formable Bone Model
H. Talib, Kumar T. Rajamani, J. Kowal, M. Styner, and M. A. Gonzalez
Ballester
Computer Aided Surgery, 10(5), p.293-299, 2005.
3. Accurate and Robust Reconstruction of Proximal Femur from Sparse Intra-

operative Data and Dense Point Distribution Model for Surgical Navigation
G. Zheng, Kumar T. Rajamani, X. Dong, X. Zhang, M.A. Gonzalez Ballester,
M. Styner
IEEE Transactions on Biomedical Engineering 2006 (under review revisions).
4.1.2 Conference Articles

1. Accuracy of reconstructing patient specific 3D bone model from clinically
relevant sparse points obtained by magnetic tracking: Initial Results
R. Thoranaghatte, G. Zheng, L.-P. Nolte, and K. T. Rajamani
Computer Assisted Orthopaedic Surgery (CAOS) 2006, Montreal, Canada
2. A robust and accurate apporach for reconstructing patient specific 3D bone

model from sparse point sets
G. Zheng, X. Dong, K. T. Rajamani, L.-P. Nolte
3. Feasibility of 3D Ultrsound initialized deformable bone modeling

H. Talib, G. Zheng, K. T. Rajamani, X. Zhang, M. Styner, and M.A. Gon-
zalez Ballester
22
4. A Comparison Study Assessing the Feasibility of Ultrasound-Initialized De-
formable Bone Models
H. Talib, K. T. Rajamani, J. Kowal, M. Styner, M.A. Gonzalez Ballester
SPIE Medical Imaging 2006, San Diego, USA.
5. An optimal three-stage method for anatomical shape reconstruction from

sparse information using a dense surface point distribution model
Guoyan Zheng, K. T. Rajamani
6. Use of a Dense Surface Point Distribution Model in a Three-Stage Anatom-

ical Shape Reconstruction from Sparse Information for Computer Assisted
Orthopaedic Surgery: A Preliminary Study
Guoyan Zheng, K. T. Rajamani, Lutz-Peter Nolte
ACCV 2006, Hyderabad, India.
7. Evaluation and Initial Validation Studies of Anatomical Structure Morphing

K. T. Rajamani, H. Talib, M. Styner, and M. A. Gonzalez Ballester
IEEE Engineering in Medicine and Biology (EMBC), Shanghai, China, 2005.
8. Kernel Regularized Bone Surface Reconstruction from Partial Data Using

Statistical Shape Model
G. Zheng, K. T. Rajamani, X. Zhang, X. Dong, M. Styner, L.P. Nolte
IEEE Engineering in Medicine and Biology (EMBC), Shanghai, China, 2005.
9. Validation studies of anatomical structure morphing

K. T. Rajamani, H. Talib, M. Styner, and M. A. Gonzalez Ballester.
Computer Assisted Orthopaedic Surgery (CAOS), Helsinki, Finland, 2005.
10. Feasibility of ultrasound-initialized bone morphing: early experiences and

evaluation of a computer-assisted surgical technique
H. Talib, K. T. Rajamani, J. Kowal, M. Styner, and M. A. Gonzalez Ballester
Computer Assisted Orthopaedic Surgery (CAOS), Helsinki, Finland, 2005.
11. A Novel and Stable Approach to Anatomical Structure Morphing for En-
hanced Intra-operative 3D Visualization
K. T. Rajamani, M. A. Gonzlez Ballester, L.P. Nolte, M. Styner
23
12. A Novel Approach to Anatomical Structure Morphing for Intra-operative
Visualization
K. T. Rajamani, L.P. Nolte, M. Styner
Medical Image Computing and Computer-Assisted Intervention (MICCAI)
2004, Rennes, France.
13. A Novel approach to bone morphing using statistical shape models

K. T. Rajamani, M. Styner
Computer Assisted Orthopaedic Surgery (CAOS) 2004 Chicago, USA.
14. Bone Model Morphing for enhanced surgical visualization

K. T. Rajamani, S. C. Joshi, M. A. Styner
IEEE International Symposium on Biomedical Imaging: From Nano to Macro
(ISBI) 2004, Arlington, USA.
15. Bone Morphing with statistical shape models for enhanced visualization
K. T. Rajamani, L. P. Nolte, M. Styner
16. Evaluation of 3D Correspondence Methods for Model Building

M. Styner, K .T. Rajamani, G. Zsemlye, G. Szekely, C. J. Taylor, R. H.Davies
Information Processing in Medical Imaging (IPMI), 2003, Ambleside, UK.
24
BIBLIOGRAPHY
Aspert, N., Santa-Cruz, D., Ebrahimi, T., 2002. Mesh - measuring errors between
surfaces using the hausdorff distance. Proceedings of the IEEE ICME. pp. 705–
708.
Bookstein, F. L., 1986. Size and shape spaces for landmark data in two dimensions
(with discussion). Statistical Science 1, 181–242.
Brechbühler, C., Gerig, G., Kübler, O., 1995. Parametrization of closed surfaces
for 3-D shape description. Computer Vision, Graphics, Image Processing: Image
Understanding 61, 154–170.
Chan, C. S., Edwards, P. J., Hawkes, D. J., 2003. Integration of ultrasound-based

registration with statistical shape models for computer-assisted orthopaedic
surgery. SPIE, Medical Imaging. pp. 414–424.
Cootes, T., Taylor, C., Cooper, D., Graham, J., 1995. Active shape models - their
training and application. Computer Vision and Image Understanding 61, 38–59.
Cootes, T. F., Hill, A., Taylor, C. J., Haslam, J., 1994. The Use of Active Shape
Models for Locating Structures in Medical Images. Image and Vision Computing
12 (6), 355–366.
C.T.Loop, 1987. Smooth subdivision surfaces based on triangles. M.S. Thesis, De-
partment of Mathematics, University of Utah.
Davies, R. H., Twining, C. J., Cootes, T. F., Waterton, J. C., Taylor, C. J., 2002.
A minimum description length approach to statistical shape modeling. IEEE
Transaction on Medical Imaging 21 (5), 525–537.
Fleute, M., Lavallée, S., 1998. Building a complete surface model from sparse data
using statistical shape models. MICCAI. pp. 879–887.
Hug, J., Brechbühler, C., Székely, G., June 2000. Model-based initialisation for
segmentation. In: Vernon, D. (Ed.), Proceedings 6’th European Conference on
Computer Vision - ECCV 2000, Part II. Lecture Notes in Computer Science.
Springer, pp. 290–306.
25
Jolliffe, I., 1986. Principal Component Analysis. Springer-Verlag.
Kelemen, A., Székely, G., Gerig, G., October 1999. Elastic model-based segmen-
tation of 3d neuroradiological data sets. IEEE Transactions on Medical Imaging
18, 828–839.
Kotcheff, A., Taylor, C., 1998. Automatic construction of eigenshape models by
direct optimization. Medical Image Analysis. pp. 303–314.
Kowal, J., Amstutz, C., Nolte, L., 2001. On b-mode ultrasound based registration
for computer assisted orthopaedic surgery. 6th Int. Symposium on CAOS.
Kowal, J., Amstutz, C., Nolte, L., 2003. On the development and comparative
evaluation of an ultrasound b-mode probe calibration unit. Computer Aided
Surgery, 107–119.
Nolte, L. P., Beutler, T., 2004. Basic principles of caos. Injury 35, 6–16.
Rajamani, K., Styner, M., Talib, H., Nolte, L., Gonzalez Ballester, M., 2006.
Statistical deformable bone models for robust 3d surface extrapolation from
sparse data. Medical Image Analysis, to appear.
Rajamani, K., Talib, H., Styner, M., Gonzalez Ballester, M., 2005. Evaluation and
initial validation studies of anatomical structure morphing. IEEE Engineering
in Medicine and Biology (EMBC).
Rajamani, K. T., Joshi, S., Styner, M., 2004a. Bone model morphing for enhanced
surgical visualization. IEEE International Symposium on Biomedical Imaging:
From Nano to Macro ISBI. pp. 1255–1258.
Rajamani, K. T., Nolte, L., Styner, M., 2004b. Bone morphing with statistical
shape models for enhanced visualization. SPIE Medical Imaging. pp. 122–130.
Rajamani, K. T., Nolte, L., Styner, M., 2004c. A novel approach to anatomical
structure morphing for intraoperative visualization. MICCAI. pp. 478–485.
Styner, M., Brechbühler, C., Székely, G., Gerig, G., 2000. Parametric estimate of
intensity inhomogeneities applied to mri. IEEE Transactions on Medical Imaging
19 (3), 153–165.
Styner, M. A., Rajamani, K. T., Nolte, L. P., Zsemlye, G., Szekely, G., Taylor,
C. J., Davies, R. H., 2003. Evaluation of 3d correspondence methods for model
building. IPMI. pp. 63–75.
Talib, H., Rajamani, K., Kowal, J., Nolte, L., Styner, M., Gonzalez Ballester,
M., 2005. A comparison study assessing the feasibility of ultrasound-initialized
deformable bone models. Comput Aided Surgery, 293–299.
26
Wang, Y., S., P. B., Staib, L. H., 2000. Shape-based 3d surface correspondence
using geodesics and local geometry. CVPR. pp. 644–651.
Zheng, G., Rajamani, K., Dong, X., Zhang, X., Gonzalez Ballester, M., Styner,
M., Nolte, L.-P., 2006a. Accurate and robust reconstruction of proximal femur
from sparse intraoperative data and dense pdm for surgical navigation. IEEE
Transactions on Biomedical Engineering, submitted.
Zheng, G., Rajamani, K., Nolte, L.-P., 2006b. Use of a dense surface point dis-
tribution model in a three-stage anatomical shape reconstruction from sparse
information for computer assisted orthopaedic surgery: A preliminary study.
ACCV. pp. 52–60.
27
Statistical Deformable Bone Models for
Robust 3D Surface Extrapolation from Sparse
Data
Kumar Rajamani a,∗ Martin Styner b Haydar Talib a
Lutz Nolte a Miguel A. González Ballester a

a MEM Research Center, Institute for Surgical Technology and Biomechanics
University of Bern, Switzerland
b Departments of Computer Science and Psychiatry, Neurodevelopmental Disorders
Research Center, University of North Carolina at Chapel Hill, USA
Abstract
A majority of pre-operative planning and navigational guidance during computer
assisted orthopaedic surgery routinely uses three-dimensional models of patient
anatomy. These models enhance the surgeon’s capability to decrease the invasiveness
of surgical procedures and increase their accuracy and safety. A common approach
for this is to use computed tomography (CT) or magnetic resonance imaging (MRI).
These have the disadvantages that they are expensive and/or induce radiation to
the patient. In this paper we propose a novel method to construct a patient-specific
three-dimensional model that provides an appropriate intra-operative visualization
without the need for a pre or intra-operative imaging. The 3D model is reconstructed
by fitting a statistical deformable model to minimal sparse 3D data consisting of
digitized landmarks and surface points that are obtained intra-operatively. The
statistical model is constructed using Principal Component Analysis from training
Preprint submitted to Elsevier Science 14 March 2006

objects. Our deformation scheme efficiently and accurately computes a Mahalanobis
distance weighted least square fit of the deformable model to the 3D data. Relaxing
the Mahalanobis distance term as additional points are incorporated enables our
method to handle small and large sets of digitized points efficiently. Formalizing
the problem as a linear equation system helps us to provide real-time updates to
the surgeons. Incorporation of M-estimator based weighting of the digitized points
enables us to effectively reject outliers and compute stable models. We present here
our evaluation results using leave-one-out experiments and extended validation of
our method on nine dry cadaver bones.
Key words: Medical Image Analysis, Shape Analysis, Statistical Shape Model,
Principal Component Analysis, Deformable Models
1 Introduction
In surgery, the computerized visualization of three-dimensional patient data
models has both pre- and intra-operative purposes. Pre-operatively, simulators
may be used to train practitioners in basic surgical tasks as well as complete
interventions. Patient specific models allow the practice of complex proce-
dures prior to working with the patient directly. This could be used for effec-
tive diagnosis and procedural planning methods. Intra-operatively, it presents
opportunities in navigation by augmenting the surgeon’s view of the operat-
ing field with computer-generated data. The common approach to obtain 3D
models is to use imaging techniques such as CT or MRI scans. These have
the disadvantage that they are expensive and/or induce high radiation doses
∗ Phone: ++41-31-631-5952, FAX: ++41-31-631-5960.

Email address: Kumar.Rajamani@MEMcenter.unibe.ch (Kumar Rajamani).
2
to the patient. Additionally a number of orthopaedic surgeries such as total
hip arthoplasty (THA) and total knee arthoplasty (TKA) do not warrant a
pre or intra-operative scan. The alternative is to build a statistical deformable
model and adapt it to the patient anatomy.
Statistical shape analysis (Dryden and Mardia (1998), Kendall (1989), Small
(1996)) is emerging as an important tool for understanding anatomical struc-
tures from medical images. Statistical models give an efficient parameteri-
zation of the geometric variability of the anatomy. Model based approaches
are popular (Kelemen et al. (1999), Turk and Pentland (1991), Cootes et al.
(1994)) due to their ability to robustly represent objects. Intra-operative 3D
anatomical visualization can be potentially achieved through the use of sta-
tistical shape models. Statistical model building consists of establishing legal
variations of shape from a training population. The statistical model is then
adapted, or fitted to the patient anatomy using intra-operatively digitized bone
surface points. Thus the aim of statistical shape model fitting is to extrapolate
from an extremely sparse set of 3D points a complete and accurate anatomical
surface representation. This is particularly interesting for Minimally Invasive
Surgery (MIS), largely due to the operating theater setup. Statistical modeling
technologies allow minimal intrusion on the surgical environment as 3D com-
puterized models may be directly injected into the scene, enabling enhanced
visualization.
Extrapolation via principal component analysis (PCA) based statistical shape
models has been explored by several scientists. Fleute and Lavallée (1998)
fit the deformable model surface to intra-operatively digitized point data via
jointly optimizing deformation and pose. This technology developed by Fleute
et al has been clinically evaluated and these results have been published (Stin-
3
del et al., 2002). Chan et al. (2003) optimize deformation and pose separately
using an iterative closest point (ICP) method. In our prior work (Rajamani
et al., 2004b) we proposed to iteratively remove shape information coded by
digitized points from the PCA model. The extrapolated surface is then com-
puted as the most probable surface in the shape space given the data. Unlike
earlier approaches, this approach was also able to include non-spatial data,
such as patient height and weight. It is applicable for very small set of known
points.
We present here a novel bone deformation method that can seamlessly handle
both small and large sets of digitized points and provide real time interactivity.
We have formulated the problem as a least squares error minimization with
additional regularization terms that computes the Mahalanobis distance of the
predicted model (Rajamani et al., 2004a). We solve for the shape parameters
that minimize the residual errors between the reconstructed model and the
cloud of random points. The novelty is that the Mahalanobis distance term
enables stable prediction with minimal number of known surface points. In
addition, the computation is performed in real time as shape parameters are
determined by solving a single linear system. The formalization also enables
the incorporation of the complete set of eigenvectors for the shape estima-
tion. This scheme was then improved to have better convergence behaviour
by having an additional parameter in the objective function that relaxes the
Mahalanobis distance term as additional points are digitized (Rajamani et al.,
2004c). As more information in terms of additional digitized points is received
we relax the constraint on the surface to remain close to the mean and allow it
to deform so that the error between the predicted surface and the set of digi-
tized points is minimized as far as possible. Finally, the usage of M-estimators
4
based weighting enables for a smart estimation mechanism that is robust to
outliers.
Surface points are typically acquired by use of a tracked digitizing pointer. It
can, due to limited surgical access, be difficult to acquire a set of points that
sufficiently spans the patient’s anatomy to ensure accurate shape prediction
of a given statistical model. Hence we explored using ultrasound imaging for
non-invasive intra-operative surface points digitization. We briefly illustrate
the application of our deformable bone models concurrently with automatic
segmentation of 2D B-mode ultrasound contours (Kowal et al., 2003), to pro-
vide for a rapid, automatic intra-operative visualization for navigation and
planning especially in minimally invasive orthopaedic surgery.
This paper is structured as follows. Section 2 briefly describes model con-
struction using principal components and outlines the method we chose for
building our model. In Section 3 we describe in detail the evolution of our
bone deformation algorithm. In Section 4 we present our evaluation results
using leave-one-out experiments, extended validation of our method on plas-
tic and dry cadaver bones and finally application of our deformation algorithm
in conjunction with ultrasound contours resulting in rapid, automatic intra-
operative visualization. We conclude by discussing the results and limitations
and possible extensions of our work.
2 Statistical model construction
The first step is to build a statistical shape model from a training database.
Several different geometric representations have been used to model anatomy.
5
Bookstein (Bookstein, 1986) uses landmarks to capture the important geo-
metric features. The active shape model (ASM) of Cootes and Taylor (Cootes
et al., 1995) represents an object’s geometry as a dense collection of boundary
points. Cootes et al. (1998) have augmented their statistical models to in-
clude the variability of the image information as well as shape. Kelemen et al.
(1999) use a spherical harmonic (SPHARM) decomposition of the object ge-
ometry. Recent researchers are also exploring methods towards constructing
a Statistical Shape Model Using Nonrigid Deformation of a Template Mesh
(Heitz et al., 2005), thereby making point correspondance no longer an issue
for model computation.
For our model building we have employed the representation of shapes us-
ing point distribution models (PDM). The basic idea is to compute the mean
shape and to establish from the training set the pattern of legal variations in
the shapes for a given class of images. This is achieved using Principal Com-
ponent Analysis (PCA) (Jolliffe, 1986). PCA defines a linear transformation
that decorrelates the parameter signals of the original shape population by
projecting the objects into a linear shape space spanned by a complete set
of orthogonal basis vectors. The axes of the shape space are oriented along
directions in which the data has its highest variance. If the parameter signals
are highly correlated, then the major variations of shape are described by the
first few basis vectors. Furthermore, if the joint distribution of the parameters
describing the surface is Gaussian, then a reasonably weighted linear combi-
nation of the basis vectors results in a shape that is similar to the existing
ones.
A key step in this model building involves establishing a dense correspondence
between shape boundaries over a reasonably large set of training data. In 2D,
6
correspondence is often established using manually determined landmarks, but
this is a time-consuming, error-prone and subjective process. In principle, the
method extends to 3D, but in practice, due to very small sets of reliably identi-
fiable landmarks, manual landmarking becomes impractical. Most automated
approaches posed the correspondence problem as that of defining a parame-
terization for each of the objects in the training set, assuming correspondence
between equivalently parameterized points. We compared the methods intro-
duced by Brechbühler et al. (1995) (SPHARM), Kotcheff and Taylor (1998)
(DetCov), Davies et al. (2002b) (MDL) and a fourth method based on man-
ually initialized subdivision surfaces similar to Wang et al. (2000) (MSS). We
analyzed both the direct correspondence via manually selected landmarks as
well as the properties of the model implied by the correspondences, in regard
to compactness, generalization and specificity. Our comparison study (Styner
et al., 2003) of these popular correspondence establishing methods revealed
that for modeling purposes the best among the correspondence methods was
Minimum Description Length (MDL) (Davies et al., 2002b). Based on the
study, for our model building, correspondence was initialized using MSS and
then optimized based on the MDL criteria
The statistical shape model is constructed based on the established point
correspondences. Each member of the training population is described by in-
dividual vectors x~i containing all 3D point coordinates. The aim of building
this model is to use several training datasets to compute the principal com-
ponents of shape variation. PCA is used to describe the different modes of
variation with a small number of parameters. For the computation of PCA,
the mean vector ~x̄ and the covariance matrix D are computed from the set
of object vectors(1). The sorted eigenvalues λi and corresponding p~i of the
7
√ √ √ √
-2 λ -1 λ +1 λ +2 λ
Fig. 1. The first two eigenmodes of variation of our model built from 13 segmented
proximal femoral surface data. Each individual surface in the model and the shapes
generated are described by a sparse triangle mesh list containing 4098 vertices. The
√
shape instances were generated by evaluating ~x̄ + ω λi p~i with ω ∈ {−2, .., 2}
covariance matrix are the principal directions spanning a shape space with
~x̄ representing its origin(2). Objects x~i in that shape space can be described
as linear combinations with weights b~i calculated by projecting the difference
vectors x~i − ~x̄ into the eigenspace(3).
n
1 X
D= (x~i − ~x̄) · (x~i − ~x̄)T (1)
n−1 1
P = {~ pi }; D · p~i = λi · p~i ; (2)
b~i = D T (x~i − ~x̄); x~i = ~x̄ + P · b~i (3)
Figure 1 shows the variability captured by the first two modes of variation of
our proximal femur model varied by ±2 standard deviations.
8
3 Model deformation algorithm
The aim of this step is to recover the patient-specific 3D shape of the anatomy
from the few available digitized landmarks and surface points. Our approach
uses the shape model built earlier to infer the anatomical information in a
robust way and provides the best statistical shape that corresponds to the
patient. The key factor is the observation that objects in our shape space, and
by our hypothesis the patient’s 3D shape, can be described as the mean shape
plus a weighted linear combination of eigenvectors. The problem is therefore
formulated as estimating the weights for this unknown shape, such that the
errors between the reconstructed model and the cloud of digitized surface
points is minimized.
Our model fitting algorithm is formulated as a least squares error minimization
with additional regularization terms that computes the Mahalanobis distance
of the predicted model (Rajamani et al., 2004a). The Mahalanobis distance
term enables stable prediction with minimal number of known surface points.
Where Fleute (Fleute and Lavallée, 1998) and Chan (Chan et al., 2003) con-
sider a truncated set, we include the complete set of eigenvectors, or shape
variations, without exorbitant increase in the computation time. The method
consists of two steps
• Initially a small point-set of anatomical landmarks with known correspon-
dence to the model is digitized. This is used to register the patient anatomy
to the model. This also provides an initial estimation of the 3D shape with
only a few digitized points.
• To improve the prediction additional points can be interactively incorpo-
9
rated via closest distance correspondence.
The objective function that we minimize is defined as follows
 

 
 ( )
 N
X m
X  m
Xαi2
f =ρ ~ +
wk kY~k − (X̄ 2
αi p~i (j))k + (1 − ρ) (4)
j
 

 k=1 i=1 
 i=1 λi
j=Index(k)
The first term in the function is the Euclidean distance between the N digitized
~ plus a weighted sum
~ and the estimated shape comprising the mean X̄
points Y
of the eigenvectors p~i . The corresponding point for each of the digitized points
Y~k is computed using closest point correspondence from the current estimated
~ , where j = index is the index of the closest point
shape. This is denoted as X̄ j k
corresponding to the k th digitized point. The second term is the Mahalanobis
distance of the predicted shape from the mean and controls the probability of
the predicted shape. This term ensures that the predicted shapes are valid by
favoring those that are closer to the mean.
The parameter ρ in the objective function enables the deformation scheme to
have better convergence behaviour (Rajamani et al., 2004c). This is enabled by
relaxing the effect of the Mahalanobis distance term as additional points are
digitized. This makes the surface less constrained to remain close to the mean
and allows it to more freely deform. Hence the error between the predicted
surface and the set of digitized points is better minimized. Since the error
typically decreases exponentially, we chose ρ to increase logarithmically with
the number of digitized points, and was therefore defined according to the
following equation
10





 N ≤6
 0.5

ρ= (5)



 log { M axN
N
(ge−1)+1}

 + 0.5 N >6
 2log(ge)
where N is the number of digitized points, MaxN is the total number of points
in the model, g is a factor which determines the rate of growth of ρ. To achieve
faster growth rate for ρ, g was empirically set to be the number of members
in the population.
The parameter wk enables the realization of stable predictions and the robust
rejection of outliers (Rajamani et al., 2005). Instead of using a box-filter based
rigid threshold to reject the outliers we decided to employ a smarter and ro-
bust M-estimator based weighting mechanism that creates an inverted valley
function as shown in Figure 2. This helps us to effectively reject outliers and
compute stable models. Incorporating the M-estimator based weight-function
analytically in the objective function, would make the system to be solved
non-linear as there is no differentiable, linear weight function that is an M-
estimator. Instead we decided to include the computed weights in the objective
function as single independent constants for each digitized point (computed
based on its distance from the closest point in the current estimated shape).
These weights could be updated iteratively in an ICP like mechanism until
there are no more significant changes in the weights. This feature was incorpo-
rated by having the weighting parameter wk in the objective function, adapted
from Styner et al. (2000) and defined according to the following equation
11
wk = [(1 − dist2 )/(dist2 + sp)] + 1, where (6)
m
X
~ +
dist = kY~k − (X̄ αi p~i (j))k2 (7)
j
i=1
i.e. dist is the euclidean distance between the digitized surface point and its
closest point in the current estimated shape and sp is a parameter that defines
the sharpness or steepness of the inverted valley function. The greater the value
of sp, the sharper the valley is and the harder the outliers are disregarded. To
have a gentle handling of outliers that lie within a radius of 10mm the value
of sp was empirically fixed in our application at 27.
We briefly explain here our solution strategy, where we formulate the problem
as a linear equation system. To determine the shape parameters α that best
describe the unknown shape, the function f is differentiated with respect to
the shape parameters and equated to zero. Differentiating f with respect to
αn yields
! 2
XN m
∂f
=ρ wk
∂ ~
Y k ~ + X α p~ (j) + (1 − ρ) 2αi
− X̄ (8)
j i i
∂αn ∂αn λn
k=1 i=1
f is differentiated with respect to each of the α and for each of the resulting
equations collating the different α terms, and dividing throughout by 2ρwk
yields a linear equation system of the form Aα = b with A being
12
 
 1−ρ 1 
 p~1 (j).p~1 (j) + ρw λ1 ... p~n (j).p~1 (j) ... p~m (j).p~1 (j) 
 k 
 
 
 .. 
 


. 

 
XN  
 
 1−ρ 1 
 p~1 (j).p~n (j) . . . p~n (j).p~n (j) + ... p~m (j).p~n (j) 
k=1 

ρwk λn 

 
 


.. 

 . 
 
 
 
 1−ρ 1 
p~1 (j).p~m (j) ... p~n (j).p~m (j) . . . p~m (j).p~m (j) + ρwk λm
(9)
The unknowns in our system are (α1 . . . αn . . . αm ). Collating the constant
terms yields b, the right hand side of our system as follows
 


~ ).p~ (j) 
(Y~k − X̄ 
j 1
 
 
 
 .. 
 


. 

 
N  
X 
 ~ ).p~ (j)  (10)
 (Y~k − X̄ j n 
k=1 



 
 


.. 

 . 
 
 
 
 ~ ).p~ (j) 
(Y~k − X̄ j m
This results is a m×m linear equation system over α. This is solved using
standard linear equations system solvers using QR decomposition. This novel
solution strategy enables real time estimation of shape parameters hence facil-
itating incorporation of the complete set of eigenvectors for shape estimation.
13
1.4
1.2
Weighting Factor Value
0.8
0.6
0.4
0.2
0
−50 −40 −30 −20 −10 0 10 20 30 40 50
Distance of Digitized Point From Current Estimate Shape
Fig. 2. The inverted valley function that was generated using the M-estimator based
weights and used in the objective function to achieve outlier resistance
4 Results
The primary application that we focus on is hip surgery such as total hip
replacement (THR) and knee surgery such as total knee arthroplasty (TKA)
and anterior cruciate ligament surgery (ACL). Hence we began by concentrat-
ing on the proximal femur. Our initial database comprised of 14 segmented
proximal femoral surface data. Each individual surface is described by a sparse
triangle mesh list containing 4098 vertices. The triangle mesh was the direct
result of a octahedron based subdivision scheme after parametrically mapping
the surface models onto the sphere. The proof of concept using leave-one-out
experiments were carried out using this database. We continue to expand our
training population, and for the study of effects of the various parameters on
shape estimation and for the cadaver experiments, the database consisted of
14
a set 30 CT scans of the proximal femur with a similar sparse triangle mesh
description as our initial training set.
The CT data sets in our database were segmented, and surface models of the
bones extracted, for the statistical model construction. Dense correspondence
between points on the surface of the bones in the training database was initial-
ized with a semiautomatic landmark-driven method and then optimized using
the Minimum Description Length criterion (Davies et al., 2002a) to construct
a compact optimal model. Three anatomical landmarks, the femoral notch and
the upper and the lower trochanter are used as the first set of digitized points.
This is used to initially register the model to the patient anatomy. This first
set of points is also used for computing the bounding box of the shapes in our
databse and the bounding box of the set of digitized points. This aids in scale
normalization of the shapes prior to modeling and normalizing the predicted
shape, making our method size invariant. The remaining points are added uni-
formly across the parameterization so that they occupy different locations on
the bone surface. We first demonstrate proof of principle of our method using
leave-one-out tests and then detail validation studies on cast and dry cadaver
bones.
4.1 Leave-one-out Experiments
A series of leave-one-out experiments was carried out to evaluate our method.
Surface points were chosen uniformly from the surface model of the left out
object so that they occupy different locations on the bone surface. We stud-
ied two different correspondence methods for incorporating additional points
into the estimation scheme. The first was to use the implicit correspondence
15
Fig. 3. Left: A typical proximal femur of the population that was used in the
leave-one-out test. Middle: The average shape of the population with color coded
distance map to the actual shape. The mean surface error is 3.37 mm and the me-
dian surface error is 2.65 mm. Right: The shape based on only 6 digitized points
with color coded distance map to the actual shape. The mean surface error is 1.50
mm and the median surface error is 1.25 mm
implied by the modeling (MDL). Figure 3 shows an example of a very good
estimate with mean error of 1.5mm obtained with as few as 6 digitized points
using MDL correspondence. The color-coded 3D rendering is calculated using
Hausdorff’s Distance to measure the distance between discrete 3D surfaces
(Aspert et al., 2002). The second correspondence scheme does not use the im-
plicit correspondence. This is the realistic case, as no correspondence would
be available in a real scenario. In this case the correspondence was established
via closest distance. The errors are therefore magnified due to mis-correspon-
dences and hence more points would be needed to be digitized to get a good
estimation. Figure 4 shows the cumulative statistics from the leave-one-out
experiments using MDL and closest point correspondence. The cumulative
mean/median surface error using MDL correspondence was 2.13mm/1.70mm
and using closest correspondence 3.25mm/2.66mm with 30 digitized surface
points. The leave-one-out experiments helped us evaluate the proof-of-concept
of our method and showed that we can extrapolate a three dimensional shape
from sparse data with a mean surface estimation error of 2-3mm.
16
Leave One Out Experiment (MDL Correspondence) Leave One Out Experiment (Closest Point Correspondence)
5 5
Mean
Mean Median
4.5 Median 4.5
4 4
3.5 3.5
Error (mm)
Error (mm)
3 3
2.5 2.5
2 2
1.5 1.5
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
N : Number of points digitized N : Number of points digitized
Fig. 4. Statistics cumulated from the different leave-one-out experiments with MDL
(left) and closest point correspondence (right). The average of the mean error and
the average of the median are plotted against the number of digitized points
4.2 Influence of ρ parameter
As explained earlier, the ρ parameter helps us to relax the probability term
to get a much better estimate. Our formulation makes ρ a factor of the num-
ber of digitized surface points and hence it is adapted automatically as more
points are digitized. To evaluate the influence of the ρ factor we studied the
performance of our deformation algorithm with and without the ρ parame-
ter. Figure 5 shows the cumulative statistics of all leave-one-out experiments
with and without ρ factor using the closest point correspondence. There is a
significant improvement using the ρ factor as can be deduced from the plots.
Incorporating ρ ensured better convergence and the error factor gain is about
6% in this correspondence scenario The influence of the ρ factor is more promi-
nent when a larger population is used to build the model. This is evident from
our previous study (Rajamani et al., 2004c) of the role of the ρ parameter
in Hippocampus model generated from 172 hippocampus instances where the
error factor gain was about 10%.
17
Leave One Out Experiment (Closest Correspondence)
2.5
Mean Without p
Median Without p
Mean With p
Median With p
Error (mm)
1.5
1
0 10 20 30 40 50 60 70 80 90 100
N : Number of points digitized
Fig. 5. Statistics cumulated from the different leave-one-out experiments of the
proximal femur study with and without the ρ factor. The average of the mean error
and the average of the median are plotted against the number of digitized points
4.3 Effect of wk parameter
A series of experiments was carried out to evaluate our method with regards to
robustness to outliers. A proximal cast femur with attached reference base was
used for this experiment. Tracking was done by using an in-house navigation
environment and maintaining an optimal distance of around 2 m to the optical
tracking camera (Optotrack , NDI, Waterloo, Canada). The accuracy for such
tracking systems when used ex vitro with exposed fiducial markers is lesser
than 1 mm. Figure 6 shows screen shots of our method, when the plastic
femur was estimated using surface points digitized using a calibrated navigated
pointer. In the first run, 12 surface points were digitized which comprised four
outliers. In spite of the large set of outliers a stable prediction was realized.
The second run had fewer outliers among digitized 32 surface points. Our
experiments verified that the M-estimator based weighting function was very
successful in robustly eliminating outliers and enabling stable predictions.
18
Fig. 6. Figure showing the screen shots of our method, when the shape of a plastic
femur was estimated using surface points digitized using a pointer in our navigation
environment. As can be seen the outliers are well eliminated and a stable prediction
is realized
4.4 Cadaver Validation Study
Nine different dry cadaver femur bones were chosen for this validation study.
High-resolution CT scans of these bones were segmented and fine 3D surface
models were generated. The experiment trials were carried out in the CT
coordinate system. The three anatomical landmarks and additional 51 bone
surface points were digitized on the surface model of each of the cadaver bones.
The deformation procedure was then employed to estimate the 3D model that
best approximates the digitized set of points.
We carried out the experiments on two models, built from different initial
training population. The first model was constructed from the entire 30 proxi-
mal femurs and the second model was constructed from a subset of 14 proximal
femurs, with correspondence optimized across the respective training sets. This
helped us to evaluate the effect of training size on the deformation algorithm.
Table 1 shows the error results for each of the cadaver bones with different
number of digitized surface points using the larger and smaller population.
19
The mean surface error with 3, 27 and 54 selected surface points are tabu-
lated. Figure 7 (left) shows the surface model of the estimated 3D shape for
one of the cadaver bones with color coded distance map to the actual shape.
The mean error here with 54 digitized points was 0.85 mm and a median of
0.66 mm using the larger population of 30 proximal femurs. Figure 7 (right)
shows the cumulated statistics across all the cadaver bones. The average of
the mean and median errors across the entire set of 9 cadaver bones is plotted
against the number of digitized points for both the models generated from
the smaller and larger population. The average mean surface error with 10
digitized points lies between 2.1-2.6mm and with 54 digitized points the error
is 1.7-1.9mm. The results for predicting the cadaver bones are in the same
error range as the leave-one-out experiments. The cadaver experiments helps
us conclude that we can indeed estimate quite accurately the 3D shape of an
anatomy even with very sparse information.
4.5 Ultrasound-Initialized deformable bone models
Two different cast proximal femurs were chosen for this study. Their CT sur-
face models were registered into the anatomy’s co-ordinate space using paired
point matching and refined using surface matching (Gong et al., 1997) inte-
grated into our in-house optical-tracking navigation system, yielding a reg-
istration error of 0.2 mm for this experiment. The registered surface models
were considered as “gold” references, used for error measurements (computed
with Mesh (Aspert et al., 2002)) of the predicted bone models.
To initialize as well as provide surface points for our bone deformation method
a classical ultrasound system (Kontron Sigma 330 , Basel, Switzerland) with
20
Cadaver Mean error [mm] w.r.t. # of points
bone # Large population Small population
3 27 54 3 27 54
1 2.08 1.90 1.72 2.57 2.02 1.85
2 0.96 0.91 0.85 2.03 1.49 1.23
3 2.44 2.28 2 3.02 2.69 2.50
4 2.55 2.45 2.03 2.92 2.63 2.12
5 2.18 1.99 1.85 1.98 1.87 1.72
6 3.49 3.1 2.54 4.44 3.79 2.65
7 1.73 1.59 1.39 3.15 2.61 2.23
8 2.01 1.87 1.67 1.91 1.75 1.58
9 2.06 2.04 1.83 2.22 2.14 1.64
Average 2.17 2.01 1.76 2.69 2.33 1.95

Table 1
Mean surface errors for nine dry cadaver bones with 3, 27 and 54 selected sur-
face points in the CT-based error scheme. The errors are tabulated for both the
experiments with the larger and smaller population.

a 7.5 MHz linear array transducer, was used for the experiments. The system
provides conventional B-Mode imaging. Calibration was achievable in less than
five minutes, using a minimum of ultrasound images, and has a high reported
accuracy (Kowal et al., 2003). The bone contours used in our experiment were
automatically segmented from the image planes of the calibrated tracked 2D
21
Statistics Cumulated From 9 Cadaver Bones
Mean (Large Population)

Median (Large Population)
Mean (Small Population)
2.6 Median (Small Population)
2.4
Error Across Whole Surface (mm)

2.2
1.8
1.6
5 10 15 20 25 30 35 40 45 50
N : Number of Points Digitized
Fig. 7. Left: The surface model that was estimated for one of the cadaver bones with
color coded distance map to the actual shape. The mean error with 54 digitized
points was 0.85 mm and a median of 0.66 mm. Right: Error statistics cumulated
across all the cadaver bones. The average mean and median errors are plotted
against the number of digitized points for both the models generated from smaller
and larger population.
B-mode probe, thereby yielding a cloud of points in the co-ordinate space of
the anatomy. Our automatic segmentation approach requires an average of
0.8 seconds of computation for each ultrasound image frame and has a mean
accuracy of 0.42 mm (Kowal et al., 2001).
The cloud of segmented ultrasound points were provided as input to the de-
formation algorithm. The result was a predicted model in the anatomy’s co-
ordinate space. To obtain stable estimates of the errors we performed a series
of 5 trials per bone, the results of which are tabulated in Table 2. The results
in this table show the mean and median surface errors for the predicted shapes
with respect to the “gold” references for each bone, using 24 - 26 digitized sur-
face points. They also include averaged mean and median surface errors for
each scenario, to help gauge the repeatability of each experiment. Fig. 8 shows
for each bone one case of ultrasound based shape prediction, with predicted
22
Bone 1 Bone 2
1 3.88 3.56 2.94 2.12
2 4.37 4.28 5.30 4.98
3 6.88 6.34 3.79 3.48
4 4.75 4.51 4.57 4.54
5 3.08 2.53 3.12 2.84
Average 4.59 4.35 3.95 3.59

Table 2
Mean and Median surface errors from their actual surfaces in five different trials for
the two cast femur bones, using ultrasound to acquire surface points
shape overlaid to its respective “gold” reference. From the results we can see
that ultrasound imaging could be used along with our deformation algorithm
to estimate an appropriate 3D model of the anatomy.
5 Discussion
In this paper we have presented a novel anatomical shape deformation tech-
nique to predict the three-dimensional model of a given anatomy using statis-
tical shape models. The proposed shape deformation is especially attractive
in the scenario of sparse set of surface points, and can also seamlessly han-
dle small and large set of digitized points which is an innovative concept. We
23
Fig. 8. Ultrasound-based prediction: Predicted models overlaid onto “gold” refer-
ences. Bone 1 (left): 3.08 mm mean error and Bone 2 (right): 2.94 mm mean error
have shown that we can robustly estimate a realistic patient-specific three-
dimensional model of a given anatomy. Our formulation of the problem as
a Mahanalobis weighted least squares error minimization, and the novel so-
lution scheme by linear system solving,enables us to generate 3D models for
visualization in real time. The running time of the algorithm on a Petium 4
machine with 512 MB of RAM, is in the order of seconds. Our formulation
also enables incorporating the complete set of eigenvectors.
The ρ parameter helps us to relax the probability term to get a better estimate
as more points are digitized. The effect of the ρ parameter is not significantly
noticed in the case when the population size is small. This is because the error
gets stabilized and uniform after the first few points are digitized and there is
not much information that could be extracted by adding additional points in
this case. Hence the ρ factor seems not to contribute much as was observed
in the proximal femur model with a population size of only 14 members. On
the contrary in the study using the hippocampus population (Rajamani et al.,
2004c) the effect of the ρ parameter was significantly visible and it contributes
in a significant way to decrease the error and achieve better convergence.
24
Rectangular or box filters are quite easy to incorporate for outlier rejection,
but they do not smartly handle the outliers and are quite rigid. Outlier resis-
tance based on Gaussian function seems to be an option but they do not fall off
sharply enough. The novel usage of M-estimators based weighting enables for
a smart estimation mechanism that is robust to outliers. Direct incorporation
of the M-estimator into the function definition makes the minimization prob-
lem non-linear and not easy to solve. Hence we incorporated the M-estimating
weight function into the fitting function. This gives us all the advantages of
the M-estimator and also we do not lose the linearity of the problem. Nor-
malization of the shapes prior to modeling and normalizing the digitized data
makes our method size invariant. Our experiments on the cast proximal femur
data show that our method can robustly reject outliers.
We also see that there is indeed a dependency of the deformation algorithm
on the size of the training population. The larger population as expected
estimates the shape better in most of the cases. In 7 out of the 9 bones the
larger population has better estimation properties. In two of the cases where
the smaller population seems to outperform, the difference between the two
is in the order of 0.1mm. It has to be emphasized that our population is still
not large enough to capture all the possible variability of the shape. It could
very well be the case that for a much larger population the shape variability
is better captured and estimates even these two bones better than the smaller
population. It is to be expected, therefore, that an increase in population size
would increase the accuracy of the method.
As a natural extension we have explored the use of ultrasound imaging for
non-invasive intra-operative surface points digitization. Ultrasound effectively
solves the problems posed by limited surgical access and is an ideal way to
25
acquire points from otherwise inaccesible regions. We have seen above that
ultrasound initialized deformable bone models in our experimental conditions
can provide a stable and repeatable prediction for bone visualization.
The technique of statistical shape deformation introduces novel navigation
and augmented reality concepts wherein reconstructed 3D models are overlaid
on top of 2D views. The proposed technology brings a variety of advantages
to orthopaedic and other surgical procedures, such as improved accuracy and
safety, often reduced radiation exposure, and improved surgical reality through
3D visualization and image overlay techniques. In particular navigation based
on shape deformation opens the door to more minimally invasive approaches.
There are several scopes for improving the technique that we have exlaborated.
It should be emphasised here that the series of experiments that were presented
are our initial efforts and experiments in realizing a working module for shape
estimation from sparse data. The estimation results that we currently generate
are still not satisfying the precision required in surgical guidance. Surgical
guidance using surface reconstructed models demand a targer error of less
than 1.5mm on average (Livyatan et al., 2003). We have therefore identified
the various possibilities for improvement like better initial registration, using
ICP for registration refinement, using densely sampled initial models etc, and
these are explored and finer precision issues are now addressed.
6 Acknowledgment
We thank Sarang Joshi from University of North Carolina at Chapel Hill,
USA for insightful discussions about modeling and shape prediction. The CT
26
datasets were provided by Frank Langlotz from MEM Research Center,Bern
Switzerland and the segmentation and MSS preprocessing of the CT data
were done by Gabor Zsemlye from ETH Zurich, Switzerland. The MDL tools
were provided by Rhodri H. Davies and Chris Taylor from the University
of Manchester, UK. This work was supported by the Swiss National Science
Foundation project Computer Aided and Image Guided Medical Interventions
(NCCR CO-ME).
References
Aspert, N., Santa-Cruz, D., Ebrahimi, T., 2002. Mesh - measuring errors be-
tween surfaces using the hausdorff distance. Proceedings of the IEEE ICME.
pp. 705–708.
Bookstein, F. L., 1986. Size and shape spaces for landmark data in two di-
mensions (with discussion). Statistical Science 1, 181–242.
Brechbühler, C., Gerig, G., Kübler, O., 1995. Parametrization of closed sur-
faces for 3-D shape description. Computer Vision, Graphics, Image Process-
ing: Image Understanding 61, 154–170.
Chan, C. S., Edwards, P. J., Hawkes, D. J., 2003. Integration of ultrasound-
based registration with statistical shape models for computer-assisted or-
thopaedic surgery. SPIE, Medical Imaging. pp. 414–424.
Cootes, T., Taylor, C., Cooper, D., Graham, J., 1995. Active shape models -
their training and application. Computer Vision and Image Understanding
61, 38–59.
Cootes, T. F., Edwards, G. J., Taylor, C. J., 1998. Active appearance models.
In: ECCV. pp. 484–498.
Cootes, T. F., Hill, A., Taylor, C. J., Haslam, J., 1994. The Use of Active
27
Shape Models for Locating Structures in Medical Images. Image and Vision
Computing 12 (6), 355–366.
Davies, R. H., Twining, C. J., Cootes, T. F., Waterton, J. C., Taylor, C. J.,
2002a. 3d statistical shape models using direct optimisation of description
length. In: ECCV (3). pp. 3–20.
Davies, R. H., Twining, C. J., Cootes, T. F., Waterton, J. C., Taylor, C. J.,
2002b. A minimum description length approach to statistical shape model-
ing. IEEE Transaction on Medical Imaging 21 (5), 525–537.
Dryden, I., Mardia, K., 1998. Statistical Shape Analysis. John Wiley and Sons.
Fleute, M., Lavallée, S., 1998. Building a complete surface model from sparse
data using statistical shape models. MICCAI. pp. 879–887.
Gong, J., Bächler, R., Sati, M., Nolte, L.-P., 1997. Restricted surface matching:
a new approach to registration in computer assisted surgery. In: CVRMed-
MRCAS ’97. Springer-Verlag, pp. 597–605.
Heitz, G., Rohlfing, T., Maurer, C., 2005. Statistical shape model generation
using nonrigid deformation of a template mesh. Vol. 5747 of SPIE. pp.
1411–1421.
Jolliffe, I., 1986. Principal Component Analysis. Springer-Verlag.
Kelemen, A., Székely, G., Gerig, G., October 1999. Elastic model-based seg-
mentation of 3d neuroradiological data sets. IEEE Transactions on Medical
Imaging 18, 828–839.
Kendall, D., 1989. A survey of the statistical theory of shape. Statistical Sci-
ence 4 (2), 87–120.
Kotcheff, A., Taylor, C., 1998. Automatic construction of eigenshape models
by direct optimization. Medical Image Analysis. pp. 303–314.
Kowal, J., Amstutz, C., Nolte, L., 2001. On b-mode ultrasound based regis-
tration for computer assisted orthopaedic surgery. 6th Int. Symposium on
28
CAOS.
Kowal, J., Amstutz, C., Nolte, L., 2003. On the development and comparative
evaluation of an ultrasound b-mode probe calibration unit. Computer Aided
Surgery, 107–119.
Livyatan, H., Yaniv, Z., Joskowicz, L., 2003. Gradient-based 2-d/3-d rigid reg-
istration of fluoroscopic x-ray to ct. IEEE Transactions on Medical Imaging
22 (11), 1395–1406.
Rajamani, K., Gonzalez Ballester, M., Nolte, L., Styner, M., 2005. A novel
and stable approach to anatomical structure morphing for intraoperative
3d visualization. Vol. 5744 of SPIE. pp. 718–725.
Rajamani, K. T., Joshi, S., Styner, M., 2004a. Bone model morphing for en-
hanced surgical visualization. IEEE International Symposium on Biomedi-
cal Imaging: From Nano to Macro ISBI. pp. 1255–1258.
Rajamani, K. T., Nolte, L., Styner, M., 2004b. Bone morphing with statistical
shape models for enhanced visualization. SPIE Medical Imaging. pp. 122–
130.
Rajamani, K. T., Nolte, L., Styner, M., 2004c. A novel approach to anatomical
structure morphing for intraoperative visualization. MICCAI. pp. 478–485.
Small, C., 1996. The Statistical Theory of Shape. Springer.
Stindel, E., Briard, J., Merloz, P., Plaweski, S., Dubrana, F., Lefevre, C.,
Troccaz, J., 2002. Bone morphing: 3d morphological data for total knee
arthroplasty. Computer Aided Surgery 7 (3), 156–168.
Styner, M., Brechbühler, C., Székely, G., Gerig, G., 2000. Parametric estimate
of intensity inhomogeneities applied to mri. IEEE Transactions on Medical
Imaging 19 (3), 153–165.
Styner, M. A., Rajamani, K. T., Nolte, L. P., Zsemlye, G., Szekely, G., Taylor,
C. J., Davies, R. H., 2003. Evaluation of 3d correspondence methods for
29
model building. IPMI. pp. 63–75.
Turk, M., Pentland, A., 1991. Eigenfaces for recognition. Journal of Cognitive
Neuroscience 3 (1), 71–86.
Wang, Y., S., P. B., Staib, L. H., 2000. Shape-based 3d surface correspondence
using geodesics and local geometry. CVPR. pp. 644–651.
30
A Comparison Study Assessing the Feasibility of
Ultrasound-Initialized Deformable Bone Models
Haydar Talib1 , Kumar Rajamani1 , Jens Kowal1 , Lutz-Peter Nolte1 , Martin

Styner2 and Miguel A. González Ballester1
1
MEM Research Center, Institute for Surgical Technology and Biomechanics
University of Bern, Switzerland
2
Departments of Computer Science and Psychiatry, Neurodevelopmental Disorders
Research Center, University of North Carolina at Chapel Hill, USA
{Haydar.Talib, Kumar.Rajamani}@MEMcenter.unibe.ch
Abstract. This paper presents a feasibility and evaluation study for us-
ing 2D ultrasound in conjunction with our statistical deformable bone
model in the scope of computer-assisted surgery (CAS). The final aim
is to provide the surgeon with an enhanced 3D visualization for surgi-
cal navigation in orthopaedic surgery without the need for pre-operative
CT or MRI scans. We unified our earlier work to combine several au-
tomatic methods for statistical bone shape prediction and ultrasound
segmentation and calibration to provide the intended rapid and accurate
visualization. We compared the use of a tracked digitizing pointer to ul-
trasound to acquire landmarks and bone surface points for the estimation
of two cast proximal femurs.
1 Introduction
The use of 3D anatomical models in computer-assisted surgery (CAS) provides

the surgeon with image guidance and enhanced visualization to assist navigation
and planning. Such models are typically obtained from preoperatively acquired
CT or MRI scans, which may not always be available, or may not be a necessity
if cheaper, radiation-free and/or intra-operative solutions can be provided.
Consequently, intra-operative 3D anatomical visualization can be potentially
achieved using an image-free or sparse information approach through the use of
statistical shape models. Building a patient-specific anatomical model is a non-
trivial challenge given sparse a priori patient anatomical data. Statistical model
building consists of establishing legal variations of shape from a training popu-
lation. The statistical model is then adapted, or fitted to the patient anatomy
using intra-operatively digitized bone surface points. Thus the aim of statistical
shape model fitting is to extrapolate from an extremely sparse set of 3D points
a complete and accurate anatomical surface representation.
Surface points are typically acquired by use of a tracked digitizing pointer.
It can, due to limited surgical access, be difficult to acquire a set of points that
sufficiently spans the patient’s anatomy to ensure accurate shape prediction of a
given statistical model. As such, a natural extension is to use ultrasound imag-
ing for non-invasive intra-operative surface points digitization. The use of ultra-
sound in CAS is a subject that has been broached by several scientists. Chan
[1] and Lavallée [2] have explored methods using ultrasound to instantiate 3D
deformable bone models without the need of preoperative CT or MRI scans. We
present here our first experiences using our method for automatic segmentation
of 2D B-mode ultrasound contours [3], concurrently with our 3D bone deforma-
tion method [4] to provide for a rapid, automatic intra-operative visualization
for navigation and planning in minimally invasive orthopaedic surgery.
2 Methods
2.1 Statistical model construction
The first step is to build a statistical shape model from a training database.
The basic idea is to compute the mean shape and to establish from the training
set, the pattern of legal variations in the shapes for a given class of images.
This is achieved using Principal Component Analysis (PCA). PCA finds a new
orthonormal basis for the training set, such that the axes are oriented along
directions in which the data has its highest variance. PCA based statistical
shape models were introduced by Cootes et al. [5] to establish point distribution
models (PDM).
Since potential clinical applications could be for hip and knee surgeries we
chose to begin by concentrating on the proximal femur. Prior work involving our
method used a database of 14 bones [4]. We continue to expand our training
population, and for the experiments described here, the database consisted of a
set 30 CT scans of the proximal femur. The CT data sets were segmented, and
surface models of the bones extracted, for the statistical model construction.
Dense correspondence between points on the surface of the bones in the training
database was initialized with a semiautomatic landmark-driven method and then
optimized using the Minimum Description Length criterion [6] to construct a
compact optimal model.
The Statistical shape model is constructed based on the established point
correspondences. Each member of the training population is described as an
individual vector containing all 3D point co-ordinates. The mean vector and
the covariance matrix are next computed from the set of object vectors. The
sorted eigenvalues and corresponding eigenvectors of the covariance matrix are
the principal directions spanning a shape space with the mean shape representing
its origin [5]. Fig. 1 shows the variability captured by the first two modes of
variation of our proximal femur model varied by ±2 standard deviation.
2.2 Model deformation algorithm
The aim of this step is to recover the patient-specific 3D shape of the anatomy
from the few available digitized landmarks and surface points. The key factor
√ √ √ √
-2 λ -1 λ +1 λ +2 λ
Fig. 1. The first two eigen modes of variation √

of our proximal femur model. The shape
instances were generated by evaluating x̄ + ω λk uk with ω ∈ {−2, .., 2}
is the observation that objects in our shape space, and by our hypothesis the
patient’s 3D shape, can be described as the mean shape plus a weighted linear
combination of eigenvectors. The problem is therefore formulated as estimating
the weights for this unknown shape, such that the errors between the recon-
structed model and the cloud of digitized surface points is minimized.
Our model fitting algorithm is formulated as a linear equation system with
additional regularisation terms, and computes a Mahalanobis distance weighted
least square fit of the model to the 3D data [7]. The Mahalanobis distance term
enables stable prediction with minimal number of known surface points. Where
Fleute [8] and Chan [9] consider a truncated set, we include the complete set of
eigenvectors, or shape variations, without exorbitant increase in the computation
time. The objective function that we minimize is defined as follows
 

 


 
 (m )

 N
X m
X 
 X α2
2 i
f =ρ∗ wk ∗ kYk − (Xj + αi pi (j))k + (1 − ρ)

i=1

i=1
λi
k=1

 


 

j = indexk
 
(1)
The first term in the function is the euclidean distance between the N digi-
tized points Y and the estimated shape comprising the mean X plus a weighted
sum of the eigenvectors pi . The second term is the Mahalanobis distance of
the predicted shape from the mean and controls the probability of the pre-
dicted shape. This term ensures that the predicted shapes are valid by favoring
those that are closer to the mean. ρ is dynamically adapted as additional points
are digitized, thereby relaxing the Mahalanobis distance term, enabling better
(a) (b)
Fig. 2. (a) DRB attached to ultrasound probe and (b) automatically segmented bone
contours in anatomical space.
convergence behaviour. M-estimator based weights wk help to effectively reject

outliers [10].
The shape parameters αi are estimated such that both terms are simulta-
neously minimized. The function f is differentiated with respect to the shape
parameters and equated to zero. This results in a linear system of m unknowns,
which is solved using QR decomposition.
We evaluated our statistical model deformation through leave-one-out tests,
and obtained acceptable predictions (mean surface error < 2.5 mm) for the
database of 14 proximal femur scans [4].
2.3 Calibration and automatic segmentation of ultrasound
To initialize as well as provide surface points for our bone deformation method,
we used bone surface contours extracted from the image planes of a tracked 2D
B-mode probe of a Kontron Sigma 330 c standard diagnostic ultrasound system.
The ultrasound probe was fitted to a dynamic reference base (DRB), allow-
ing for accurate tracking (fig. 2(a)). The necessary ultrasound calibration step
ensures that the co-ordinates of the imaging plane are known with respect to
the anatomy in question, which also bears a DRB. Calibration was achievable in
less than five minutes, using a minimum of ultrasound images, and has a high
reported accuracy [3]. Where Chan et al. manually segment bone contours from
ultrasound images [1], the bone contours used in our experiment were automat-
ically segmented, thereby yielding a cloud of points in the co-ordinate space of
the anatomy (fig. 2(b)). This automatic segmentation approach requires an av-
erage of 0.8 seconds of computation for each ultrasound image frame and has a
mean accuracy of 0.42 mm [11].
Average Mean Errors for Bone 1 Average Mean Errors for Bone 2
2.8 Small population 3.8 Small population

Large population Large population
2.7 3.6
Average Mean error
Average Mean error

2.6 3.4
2.5 3.2
2.4
3
2.3
2.8
2.2
2.6
2.1
5 10 15 20 25 5 10 15 20 25
# of points used for bone deformation # of points used for bone morphing
(a) (b)
Fig. 3. Statistics cumulated from the trials carried out by the two users on (a) Bone
1 and (b) Bone 2 using a tracked pointer to obtain surface points. The average mean
error is plotted against the number of surface points digitized.
2.4 Experimental set-up

Two statistical models were built: one using all 30 CT scans (large) and another
one using a subset of 14 scans (small ). This was done to study the effect of
population size on shape prediction, and will be further addressed in sec. 3.1.
The shape prediction algorithm requires initialization via three landmark points
(lesser trochanter, femoral notch, and greater trochanter), and the predicted
shape is consequently refined from subsequent surface points.
Using two different cast proximal femurs, their CT surface models were reg-
istered into the anatomy’s co-ordinate space using an in-house optical-tracking
navigation system, yielding a registration error of 0.2 mm for this experiment.
The registered surface models were considered as “gold” references, used for
error measurements (computed with Mesh [12]) of the predicted bone models.
Ultrasound bone contours provided the three anatomical landmarks and sur-
face points for bone model deformation. The result was a predicted model in the
anatomy’s co-ordinate space. To have a comparison of performance, a tracked
calibrated pointer was used in parallel to digitize points (24 to 26 points) for the
bone deformation method.
To obtain stable estimates of the errors and to have an initial impression of
the repeatability of our proposed method, two different research scientists each
performed a series of 5 - 6 trials per bone.
2.5 CT-based error

Let the CT-based error be defined as a lower-bound measure of accuracy, given
a training population for the bone shape prediction method used here. The
CT-based error for each cast bone was obtained by first initializing the bone
Table 1. Error statistics of predicted shapes for Bone 1
Bone 1 Pointer-based Ultrasound-based

User 1 1 2.20 1.71 14.34 14.34
2 1.99 1.46 9.52 8.91
3 1.98 1.25 7.01 6.69
4 2.06 1.63 7.97 7.18
5 1.79 1.56 11.94 10.68
6 1.91 1.65 N/A N/A
Average 1.99 1.54 10.15 9.56
User 2 1 2.24 1.81 3.88 3.56
2 2.00 1.74 4.37 4.28
3 1.76 1.59 6.88 6.34
4 2.11 1.85 4.75 4.51
5 2.21 1.90 3.08 2.53
Average 2.06 1.78 4.59 4.35
deformation algorithm with ideal localization of the three landmarks from the
3D bone surface model of the CT. Following this step, additional surface points
that sufficiently span the object were input to our shape prediction method. The
result is a bone shape prediction in the CT co-ordinate space, which provides a
prediction error that is isolated, in the sense that there is minimum human error
and no tracking, segmentation, registration or calibration errors. In this scenario,
if all surface points from the CT surface models were used for the deformation,
then the error would converge to the minimum prediction error for a given shape
with respect to the model training population. Therefore the CT-based error for
each bone was obtained for an equivalent number of surface points as in the
experimental trials to establish a “fair” error reference.
3 Results
The results in tables 1 and 2 show the mean and median surface errors for the
predicted shapes with respect to the “gold” references for each bone, using 24 -
26 digitized surface points for each of the two methods described in this paper.
They also include averaged mean and median surface errors for each scenario,
to help gauge the repeatability of each experiment. We will use mean error and
median error to refer to mean surface error and median surface error respectively.
3.1 Effects due to training population size
The results for all the pointer-based trials can be seen in fig. 3, which shows the
averaged mean error plots for each user and for the two training population sizes
with respect to the number of digitized surface points. It can be immediately
seen from this result that the first bone (Bone 1) was better estimated with
the large training population whereas the converse was true for the second bone
(Bone 2). The trend was observed for all the experimental set-ups, including
the CT-based error scenario. It should be noted that the error differences with
Table 2. Error statistics of predicted shapes for Bone 2
Bone 2 Pointer-based Ultrasound-based

User 1 1 2.95 2.78 5.63 5.72
2 2.78 2.62 4.19 4.33
3 2.77 2.59 2.90 2.58
4 2.91 2.53 6.54 6.35
5 2.56 2.39 3.48 3.51
Average 2.80 2.58 4.55 4.50
User 2 1 2.83 2.68 2.94 2.12
2 3.66 3.45 5.30 4.98
3 3.06 2.96 3.79 3.48
4 3.30 3.17 4.57 4.54
5 3.59 3.43 3.12 2.84
Average 3.29 3.14 3.95 3.59
respect to different training population sizes can be expressed as linear shifts;

typically 0.5 mm difference. The reported results in tables 1 and 2, as well as
those for CT-based error, are taken from the large, 30 bones training population.
Statistically speaking, the large population should yield better estimations
for a given set of surface points of an unknown shape. The latter statement holds
true, on average, for most unknown shapes, but not all. So it appears that Bone
2 would be an outlier. We can corroborate this assertion by briefly considering
the aforementioned CT-based error set-up. Nine dry cadaver femur bones were
gathered from an archive to test our shape prediction method with respect to
training population size, using surface points acquired from CT-generated sur-
face models. Shape prediction error was taken with respect to the number of
points digitized, and the two training populations’ effects on the outcome were
compared. It can be seen in table 3 that for seven of the nine bones, the large
training population produced better predictions than the small population. For
the two cases where the small population produced better results, the difference
in mean error was in the order of 0.1-0.2 mm. Fig. 4 shows the mean errors
obtained for each population, averaged over all nine bones, and it can be seen
that globally, the large population performed better than the small one.
3.2 CT-based error
For Bone 1, the CT-based deformation error consisted of a mean surface error
of 1.72 mm and a median surface error of 1.49 mm for 30 surface points. In the
case of Bone 2, the CT-based error comprised a mean error of 2.30 mm and a
2.07 mm median error for 30 surface points.
3.3 Evaluation of pointer-based anatomical prediction
For Bone 1, we did not notice a large observable discrepancy (0.07 mm average
difference in mean error) between the two users, as can be inferred from table 1.
On average, the predicted errors using this technique were within 0.3 mm of the
Mean surface error averaged across all nine bones
3
Cadaver Mean error [mm] w.r.t. # of points Large population
bone # Large population Small population 2.8
Small population
3 27 54 3 27 54
1 2.08 1.90 1.72 2.57 2.02 1.85 2.6
2 0.96 0.91 0.85 2.03 1.49 1.23

2.4
3 2.44 2.28 2 3.02 2.69 2.50
4 2.55 2.45 2.03 2.92 2.63 2.12 2.2
5 2.18 1.99 1.85 1.98 1.87 1.72
6 3.49 3.1 2.54 4.44 3.79 2.65 2
7 1.73 1.59 1.39 3.15 2.61 2.23
1.8
8 2.01 1.87 1.67 1.91 1.75 1.58
9 2.06 2.04 1.83 2.22 2.14 1.64
1.6
Average 2.17 2.01 1.76 2.69 2.33 1.95 5 10 15 20 25 30 35 40 45 50
# of points used for bone deformation
Table 3. Mean surface errors for Fig. 4. For each population, the av-
nine dry cadaver bones in the CT- erage mean error is plotted against
based error scheme. the number of digitized points.
CT-based error. All the results here were deemed acceptable in terms of accuracy
(< 2.5 mm mean error for large bones) for potential surgical applications.
For Bone 2, we noticed a larger (0.49 mm average difference) discrepancy
between the two users’ trials (table 2). The averaged mean errors for both users
were greater than 2.5 mm, and it can be noted that for the small population
they would have been acceptable (fig. 3(b)). The discrepancy between these error
results and the CT-based error is larger than for Bone 1. The first user had on
average 0.5 mm greater mean error, and the second user had on average 0.99
mm greater mean error than the CT-based error.
Fig. 5 shows for each bone one case of shape prediction using the pointer-
based approach, with predicted shape overlaid to its respective “gold” reference.
3.4 Evaluation of ultrasound-based anatomical prediction
Considering Bone 1, the first user obtained the worst results of the entire experi-
ment (table 1). These predicted models were quite erroneous with respect to the
pointer-based reference and we found that this was largely due to inadequate
localization of the initial three landmark points. The second user fared better
with Bone 1, and in the best trial (3.08 mm mean error), the result is comparable
to the pointer-based approach.
For Bone 2, there was not a large discrepancy in the results of the users’
ultrasound trials, as seen in table 2. The second user’s results were slightly better
than the first user’s, with a 0.6 mm average difference in mean error between
users. In this scenario, there were four trials (mean errors of 2.90, 3.48, 2.94
and 3.12 mm) that produced results comparable, if not better, than some of the
pointer-based trials.
Fig. 6 shows for each bone one case of shape prediction using the ultrasound-
based approach, with predicted shape overlaid to its respective “gold” reference.
Fig. 5. Pointer-based prediction: Predicted models overlaid onto “gold” references.
Bone 1 (left): 1.76 mm mean error and Bone 2 (right): 2.78 mm mean error
4 Discussion
Considering the size difference of the training populations, we can see that the
modes of shape variation in the large population better represented Bone 1,
whereas the converse was true for Bone 2 (fig. 3). Considering Bone 2, it can be
supposed that the large population became biased against accurate prediction
of this shape. This is the effect of the insufficient size of the training population,
and highlights the need for large databases for the statistical model construction.
The pointer-based trials produced results that were very close to those of the
CT-based error, although they represent less of a realistic, clinical situation in
which the access to the bone surface is limited.
We have seen above that ultrasound-based prediction in our experimental
conditions can provide a stable and repeatable prediction for bone visualization,
though the accuracy is still not to the level needed for clinical applications. We
identified that a severe cause for error was the inaccurate localization of the
initial 3 landmarks. With 2D ultrasound as a visual guide, it is quite an arduous
task to accurately identify a defined landmark. The users who performed these
experiments were not medical experts nor sonographers. As such, the greatest
source of error became a human one, and a different approach needs to be taken
for localization of the initial landmarks using tracked 2D ultrasound. A study
conducted by Cannon et al. [13] highlights the limitations of using tracked 2D
ultrasound for surgical guidance, by arguing that real-time 3D systems provide
improved accuracy when precise visualization is required. In our scenario, one
further step for future work could be to initialize the shape prediction algorithm
using a digitizing pointer, and obtaining further surface points with the use of
2D ultrasound. The location of the initial landmarks can be defined according
to a given procedure, to lie in areas where there would be surgical access. In a
Fig. 6. Ultrasound-based prediction: Predicted models overlaid onto “gold” references.
Bone 1 (left): 3.08 mm mean error and Bone 2 (right): 2.90 mm mean error
non-invasive manner, ultrasound can therein be used to obtain surface points of

unexposed regions of bone that make up the anatomy of interest.
In a clinical situation, although the landmarks may be defined with flexi-
bility (they should span the object space), there exists a challenge in localizing
them due to limited surgical access and possibly high patient-to-patient vari-
ability. We so far avoided using a registration algorithm in conjunction with our
bone deformation technique, since a geometric matching of points to a surface
does not necessarily reflect an anatomical correspondence. We do nevertheless
intend on reducing the method’s dependency on the initial landmarks by imple-
menting a recursive registration algorithm, such as ICP; which could improve
prediction and has been shown to yield good results in similar work [1]. From
our experience, it is feasible to obtain landmark points within 2-3 mm of their
actual location with ultrasound-based techniques, which would then provide a
reasonable initialization of the ICP registration. Finally, increasing the size of the
model’s training population would also improve the shape prediction accuracy
for a larger set of “unknown” bones.
Acknowledgements
We would like to thank the NCCR CO-ME for funding this project, and Kontron
Medical for the 2D diagnostic ultrasound system. We also thank our colleagues
Rudolf Sidler, Ségolène Tarte, Tobias Rudolph, Marc Puls, Christoph Anderegg
and Thibaut Bardyn for their support. As well, thanks go to Kati Haenssgen of
the Anatomy Department of the University of Bern for providing the cadaver
bones, as well as Mrs. Spielvogel and Dr. Sven Sarfert for all the CT scans.
References
1. Chan, C., Barratt, D., Edwards, P., Penny, G., Slomczkkowski, M., Carter, T.,
Hawkes, D.J.: Cadaver validation of the use of ultrasound for 3d model instan-
tiation of bony anatomy in image guided orthopaedic surgery. MICCAI (2004)
397–404
2. Lavallée, S., Merloz, P., Stindel, E., Kilian, P., Troccaz, J., Cinquin, P., Langlotz,
F., Nolte, L.: Echomorphing introducing an intra-operative imaging modality to
reconstruct 3d bone surfaces or minimally invasive surgery. CAOS (2004) 38–39
3. Kowal, J., Amstutz, C., Nolte, L.: On the development and comparative evaluation
of an ultrasound b-mode probe calibration unit. Computer Aided Surgery (2003)
107–119
4. Rajamani, K.T., Nolte, L., Styner, M.: A novel approach to anatomical structure
morphing for intraoperative visualization. MICCAI (2004) 478–485
5. Cootes, T.F., Hill, A., Taylor, C.J., Haslam, J.: The use of active shape models
for locating structures in medical images. In: IPMI. (1993) 33–47
6. Davies, R.H., Twining, C.J., Cootes, T.F., Waterton, J.C., Taylor, C.J.: 3d statis-
tical shape models using direct optimisation of description length. In: ECCV (3).
(2002) 3–20
7. Rajamani, K.T., Joshi, S., Styner, M.: Bone model morphing for enhanced surgical
visualization. IEEE International Symposium on Biomedical Imaging: From Nano
to Macro ISBI (2004) 1255–1258
8. Fleute, M., Lavallée, S.: Building a complete surface model from sparse data using
statistical shape models. MICCAI (1998) 879–887
9. Chan, C.S., Edwards, P.J., Hawkes, D.J.: Integration of ultrasound-based regis-
tration with statistical shape models for computer-assisted orthopaedic surgery.
SPIE, Medical Imaging (2003) 414–424
10. Rajamani, K., Ballester, M., Nolte, L., Styner, M.: A novel and stable approach
to anatomical structure morphing for intraoperative 3d visualization. SPIE (2005)
11. Kowal, J., Amstutz, C., Nolte, L.: On b-mode ultrasound based registration for
computer assisted orthopaedic surgery. 6th Int. Symposium on CAOS (2001)
12. Aspert, N., Santa-Cruz, D., Ebrahimi, T.: Mesh - measuring errors between sur-
faces using the hausdorff distance. Proceedings of the IEEE ICME (2002) 705–708
13. Cannon, J., Stoll, J., Salgo, I., Knowles, H., Howe, R., Dupont, P., Marx, G., del
Nido, P.: Real time 3-dimensional ultrasound for guiding surgical tasks. Computer
Aided Surgery 8 (2003) 82–90
Accurate and Robust Reconstruction of Proximal Femur from
Sparse Intraoperative Data and Dense Point Distribution Model
for Surgical Navigation
*Guoyan Zhenga, Kumar T. Rajamania, Xiao Donga, Xuan Zhanga,

Miguel A. Gonzalez Ballestera, Martin Stynerb, Lutz-Peter Noltea
a
MEM Research Center, University of Bern, CH-3014, Switzerland
b
Department of Computer Science and Psychiatry, University of North Carolina at Chapel Hill,
NC 27599-3175, USA
*Correspondence:
Dr. Guoyan Zheng,
MEM Research Center, University of Bern, Stauffacherstrasse 78, CH-3014, Switzerland
E-mail: Guoyan.Zheng@MEMcenter.unibe.ch
Phone: (41) 31 631 5956
Fax: (41) 31 631 5960
Abstract
Constructing a three-dimensional (3D) model from sparse data is a nontrivial

task. Here, we report an accurate and robust approach for reconstruction of
proximal femur from sparse intraoperative data and dense point distribution
model (DPDM). The problem is formulated as a three-stage optimal estimation
process. The first stage, registration, is to iteratively estimate the scale and the
rigid registration transformation between the mean shape of the DPDM and the
input points. The estimation results of the first stage are used to establish point
correspondences for the second stage, morphing, which optimally and robustly
estimates a template surface from the DPDM using a statistical approach. The
estimated template surface is then fed to the third stage, deformation, which uses
a kernel-based shape deformation to further improve the reconstruction accuracy.
3D surface models of seven dry cadaveric proximal femurs, obtained from CT
volume data, were used in validation experiments. Two different point acquisition
strategies were investigated identifying the applicable clinical settings for both
open and minimally invasive surgeries. Experimental results show a root mean
square error of 0.5-1.2 mm (95% percentile error of 1.1-2.6 mm), which
demonstrates that the proposed approach is accurate enough for surgical
navigation.
1. Introduction
With the recent introduction of navigation techniques in orthopedic surgery, three dimensional (3D)
models of the patient are routinely used to provide image guidance and enhanced visualization to a
surgeon to assist in planning and navigation. 3D models are typically derived from tomographic data
acquired from Computed Tomography (CT) or Magnetic Resonance Imaging (MRI). To avoid the high
costs and possible health hazards (CT-imaging) associated with these technologies, an alternative way is
to reconstruct surface using sparse data consisting of dozens of landmarks and surface points (e.g., 50
points) which are intraoperatively digitized by the surgeon using pointing device [1] or via ultrasound
[2], or of several calibrated fluoroscopic images (e.g. 2 images) [1]. However, constructing an accurate
3D model from sparse data is a challenging task. Moreover, inherent to the navigation application is the
high accuracy and robustness requirement. Similar to registration, when surface reconstruction is used
for the purpose of surgical guidance, a target error of less than 1.5 mm on average (2 to 3 mm worst
case) with a 95% successful rate is normally required [3]. In the present paper, we try to solve the
problem in an accurate and robust way. At the heart of our approach lies the combination of
sophisticated surface reconstruction techniques and a dense point distribution model (DPDM) of the
target anatomical structure.
The first part concerns the DPDM for inferring 3D shape. The distributions are learned from
segmented anatomical surfaces of real patients. A two-level approach is proposed to construct the
DPDM in fine resolution by subdividing optimally aligned surfaces in coarse resolution. The
motivations for introducing such a model are several. It is treated as one important way to incorporate
the a priori information about the topology of the target anatomical surface. Otherwise, it is a hard
problem to robustly reconstruct the correct topology of the complete surface from sparse data without
any a priori model. Moreover, it facilitates the setup of point correspondences for all stages of surface
reconstruction due to its dense description.

The second part of this work deals with fitting surfaces to the input sparse data. The fitting problem
is formulated as a three-stage optimal estimation process. The first stage, registration, is to iteratively
estimate the scale and the 6 degree-of-freedom rigid registration transformation between the mean shape
of the DPDM and the input sparse data using the iterative closest point (ICP) algorithm [4, 5, 6]. The
estimation results of the first stage are used to establish point correspondences for the second stage,
morphing, which optimally and robustly estimates a dense template surface from the DPDM using a
Mahalanobis distance based stabilizer [7]. The estimated dense template surface is then fed to the third
stage, deformation, where a general framework using a newly formulated kernel-based shape
deformation, derived from a reproducing kernel Hilbert space (RKHS) theory [8], have been proposed
to further reduce the reconstruction error. We consider it general framework because it does not depend
on particular kernel. Any kernel function derived from a RKHS can be used. The prior knowledge of
preservation of topology results in a smoothness constraint on the shape deformation, which leads us to
use thin plate spline (TPS) as the deformation kernel.
1.1. Related Work
In the area of statistical shape model, our work is inspired by Point Distribution Model (PDM) [9],
which can learn shape variations from a set of training data, containing a set of landmarks to define the
shape. When applied to surface reconstruction, this approach employs Principal Component Analysis
(PCA) to reduce the dimensionality of the shape parameter space and then performs shape prediction in
the reduced, low dimensional space. In Benameur et al. [10, 11], a statistical shape model of scoliotic
vertebrae was fitted to two conventional radiographic views by simultaneously optimizing both shape
and pose parameters. Chan et al [12] used a similar algorithm, but optimize the shape and pose
parameters separately. Following the seminal work of Blanz and Veter for the synthesis of 3D faces
using a morphable model [7], our recent work [13, 14, 15] incorporates a Mahalanobis distance based
stabilizer into the estimation for robust and stable anatomical shape reconstruction. Although this
method is quite popular and has been successfully applied to different medical imaging fields such as
segmentation of two-dimensional (2D) anatomical structures [16], non-rigid 3D/2D registration of the
knee [17, 18], 3D/2D segmentation and registration of scoliotic vertebrae [10, 11], enhanced
visualization [13, 14, 15], and tooth surface reconstruction from sparse data [19], the resulting model
accuracy cannot reach the level required for surgical navigation application without further processing.
This might be explained by following facts: Using such an approach for shape prediction is essentially
equivalent to assuming that the shape variation of future instances falls within a Gaussian distribution
with a constant mean and covariance that is calculated from a training database [9]. In the generative
case, the Law of Large Numbers justifies using this method [20]. Nevertheless, it may well be that part
of the shape variation of the future instance can not be fully accounted for by any element generated
from this distribution because: (1) the input sparse data may be deteriorated by noise or other errors; (2)
there are abnormal local shape deviations due to pathology; (3) the limited number of training samples
in our training database does not cover the full range of shape variation of the target object. The most
similar work to ours was reported in Stindel et al. [21], where statistical model based shape estimation
was combined with Octree splines based local deformation. However, there are three main differences
between their work and ours: (1) their statistical model based shape estimation applied a least-squares fit
without regularization, which makes the solution very unstable; (2) their Octree splines based local
deformation could be treated as a specification of our general framework for shape deformation; (3) For
a successful reconstruction with sub-millimetric root mean square accuracy, their approach required to
acquire a random cloud of as many as 1000 points. In contrast, we only need dozens of points.
Regularization proposed in computer vision community has also been applied to surface
reconstruction from noisy measurements in a technique that defines a unique solution for these
otherwise ill-posed problems by minimizing an additional smoothness functional [22]. The solutions to
the minimization problems result in either implicit surface interpolations using radial basis functions
[23, 24], or variational models such as deformable ballon model [25] or level set representation [26].
The former require either large numbers of scattered surface points as no prior information is used [23],
or manual interventions to find correctly the corresponding homologous features [24]. Both cases are
not appropriate for our applications when taking a sterilized environment into consideration. The latter
are often solved by costly iterations. The stability of such iterations and the preservation of topology
still remain to be solved [27].
1.2. Contribution of the Paper
The main contribution of the paper is the approach itself – the realization of a unique and innovative
framework for robust and accurate anatomical shape reconstruction from sparse data by combining
state-of-art techniques to tackle the problems in different stages. Our formulation of the problem as a
three-stage optimal estimation process results in a sub-millimetric average accuracy. Iterations are only
used in the first stage; the other stages are computed in single steps by solving associated linear
equations systems. Preservation of topology and reduction of reconstruction error are sequentially
achieved through projections to different sub-spaces as shown in Fig. 1. The inputs are: (a) shape vector
v ' described by the input sparse points, (b) the mean shape vector x of the DPDM described by dense
points, and (c) the statistical space S , taking x as its origin and spanned by the eigenvectors of the
DPDM. S denotes the orthogonal complement sub-space of S in the complete shape space. After
registration, the input shape vector v' and the means shape vector x of the DPDM are aligned and point
correspondences between them are established. The distance between the input shape vector and the
statistical space S is decomposed into two orthogonal terms, i.e., distance from statistical space (DFSS)
and distance in statistical space (DISS). The DISS is reduced to zero by the second stage of our method,
morphing, where a template shape vector x is robustly estimated through projecting the input shape
represented by the aligned shape vector v ' into the statistical space S . The DFSS remains non-zero
without further processing. The third stage of our method, deformation, is specially designed to reduce
this error and to preserve the topology of the template shape vector estimated from the second stage as
well. Let’s denote the corresponding sub-shape vector of v ' on the complete template shape vector x as
v . Then (v - v' ) is the residual shape vector. The aims of the third stage are achieved by projecting this
residual shape vector (v - v' ) to a kernel space K , which is a sub-space of a potentially infinite-
dimensional Hilbert space Y induced by reproducing kernels Φ , to estimate a smooth deformation
transform. The whole template shape vector x is then deformed using the estimated transform.
2. Dense Point Distribution Model
We propose a two-level approach to construct the dense point distribution model. The input data set is
the training shape database described in our previous work [13], which consists of 13 segmented
proximal femoral surfaces. Each individual surface is described by a triangle mesh list containing 4098
vertices. A sequence of correspondence establishing methods was employed to optimally align these
training shapes. It starts with a SPHARM-based parametric surface description [28] and then is
optimized using Minimum Description Length (MDL) based principle as proposed by Davies et al. [29].
The vertices for constructing the dense point distribution model in fine resolution are then obtained
by subdividing these aligned surfaces in coarse resolution. The basic idea of subdivision is to provide a
smooth limit surface which approximates the input data. Starting from a mesh in low resolution, the
limit surface is approached by recursively tessellating the mesh. The positions of vertices created by
tessellation are computed using a weighted stencil of local vertices. The complexity of the subdivision
surface can be increased until it satisfies the user’s requirement.
For our purpose, we use a simple subdivision scheme called Loop scheme, invented by Loop [30],
which is based on a spline basis function, called the three-dimensional quartic box spline. The reasons
why we choose Loop scheme are that it is defined for triangle meshes, and that it guarantees that the
limit surface is smooth. Its subdivision principle is very simple. Three new vertices are inserted to
divide a triangle in coarse resolution to four smaller triangles in fine resolution.
As mentioned before, the levels of subdivision depend on the user’s requirement. In our case, we
require that the maximum edge length of all triangles should be less than 1.5 mm, which is a value
determined by the target reconstruction accuracy. A single-level subdivision is enough for our purpose,
which results in totally 16386 vertices per training surface. One of the examples is given by Fig. 2.
The Loop subdivision doses not change the positions of vertices on the input meshes. Furthermore,
positions of the inserted vertices in fine resolution are interpolated from the neighboring vertices in
coarse resolution. As the input surfaces have already been optimized for establishing correspondence, it
is reasonable to conclude that the dense surfaces obtained by single-level subdivision are also aligned.
Following subdivision, the DPDM is constructed as follows. Let x i = ( p0 , p1 ,......, pN −1 ) , i = 0, 1, …,
m-1 be m (here m = 13) members of the aligned training population. Each member is described by
individual vectors x i containing N (here N = 16386) aligned 3D point coordinates. A statistical shape
model is constructed using PCA [9]:
D = (( m − 1) −1 ) ⋅ ∑m−1( x i − x )( x i − x )T
i=0 (1)
P = (p 0 , p1 ,...); D ⋅ p i = σ i2 ⋅ p i
where x and D represents the mean vector and the covariance matrix respectively. Since D is
symmetrical, the columns p i of P form an orthogonal set of eigenvectors. σ 0 ≥ σ 1 ≥ ... ≥ σ m−2 are the
standard deviations of the data along each of the principal directions p i spanning the shape sub-space S
with x representing its origin. The space spanned by p i is at most m' = ( m − 1) dimensional, and the
rank of D and P is at most m' . If we use the eigenvectors p i as basis, any new member from this shape
sub-space can be expressed as
x = x + ∑ m − 2α ip i (2)
i=0
The shapes reconstructed by varying the first three eigenvectors are shown in Fig. 3. And the
estimated normal distribution of the coefficients α i is
1 m−1
− ∑ m− 2 (α i2 / σ i2 ) −
2 i =0
p(α 0 , α 1 ,..., α m −2 ) = vα ⋅ e , vα = ( 2π ) 2
(3)
where ∑im=−02 (α i2 / σ i2 ) is the Mahalanobis distance from the mean according to the normal distribution.
3. The Proposed Approach
Given the positions of a reduced number n << N of input points in Euclidean space,
v' = {v'i = (x'i ,y'i ,z'i ) ; i = 0,1,..., n − 1} , the reconstruction problem is solved in three stages:
1. Registration: This is the only stage solved by iteration. In this stage, the scale and the rigid
registration transformation between the mean shape of the DPDM and the input sparse points is
iteratively determined; the key problem to be solved is the initialization.
2. Morphing: Using the estimated scale and pose information from the first stage, point
correspondences between the input points and the mean shape are first established. Then, a dense
template shape for the third stage is optimally estimated; the key problem to be solved is the
robustness and stability of the estimation algorithm.
3. Deformation: First, point correspondences between the template shape estimated from the second
stage and the aligned sparse points are established. Then, the estimated template shape is further
deformed to reduce the reconstruction error; the key problem to be solved is the accuracy and the
preservation of shape topology.
3.1. Registration
This is a well-known problem and several efforts have been made to solve it. One of the most popular
methods is the Iterative Closest Point (ICP) algorithm developed by Besl and McKay [4], Chen and
Medioni [5], and Zhang [6]. The ICP is based on the search of pairs of closest points, and the
computation of a paired-point matching transformation. The resulting transformation is then applied to
one set of points, and the procedure is iterated until convergence. Normally, when trying to register a set
of points to a surface described by a triangle mesh, a computation-intensive point-to-surface distance
needs to be computed. However, as the mean shape in our case is described by a dense point surface, a
cost-effective point-to-point distance is enough. An additional advantage is that this point-to-point
matching allows us to use k-D tree to speed up searching for the closest paired points [6].
It is well-known that ICP algorithm will converge to a local minimum without a proper
initialization. In our case, three anatomical landmarks shown in Fig. 4, i.e., the center of the femoral
head H(H), a point on the axis of the femoral neck N(N), and the most cranial point of the greater
trochanter G(G), are used as follows to initialize the registration procedure, which guarantees the
convergence of ICP algorithm.
Let’s denote the three landmarks on the mean shape of the DPDM as H, N and G; and their
corresponding landmarks on the anatomy as H', N', and G', respectively. H', N', and G' can be obtained
intraoperatively, either by point digitization (for point G') followed by geometric fitting (sphere fitting
for point H' and circle fitting for point N') [31], or by biplanar landmark reconstruction [1] when two or
more calibrated X-ray images are used. P' (P) is the orthogonal projection point of G' (G) on the line
H'N' (HN). Let’s take the point H' (H) as the origin and line N'H' (NH) as the x' (x) axis to build a local
coordinate system. The reason why we choose H' (H) is because, intraoperatively, the point H' can be
obtained more easily and more accurately than point G' and point N'. The initial scale S0 and the initial
rigid registration transformation (R0, T0) are then computed as follows.
(P − H) PH × GH z×x
x= , z= , y= ; R1 = [x y z ] (4)
|| ( P − H ) || || PH × GH || || z × x ||
(P ' − H ' ) P 'H ' × G 'H ' z ' × x'

x' = , z '
= , y '
= , R2 = [x ' y' z' ]
|| ( P ' − H ' ) || || P ' H ' × G ' H ' || || z ' × x' ||
where × means cross product between two vectors, and || ⋅ || denotes the usual Euclidean distance.
S0 = area( ∆H P G )
' ' '
area( ∆HPG )
, R 0 = R2 ⋅ R1−1 , T0 = H ' − S0 ⋅ R 0 ⋅ H (5)
3.2. Morphing
After registration, it is easy to find the corresponding homologous points of the input sparse points v'
on the dense smooth mean shape of the DPDM. Let’s denote these homologous points as:
x' = Cx = {( x j ) i ; 0 ≤ j ≤ N − 1; i = 0,1,..., n − 1} (6)

where N is the number of points on the dense mean surface of the DPDM; n << N is the number of the
input sparse points. C represents the correspondence operation. ( x j ) i denotes that the jth point x j on
the dense smooth mean shape x of the DPDM is the closest point to the ith input sparse points v'i .
Given n << N number of correspondences as well as the positions of these matched point pairs, our
task is to estimate the positions of all N vertices of a complete surface, which should have the same
topology as the mean shape x of the DPDM, and at the same time to minimize the distances between
those matched point pairs. Taking these two factors into consideration, we formulate the morphing
problem as the minimization of the following joint cost function:
Eα ( x' , v' , x ) = E( x' , v' , x) + ρ * E( x) ;

(7)
x = x + ∑ m − 2α k p k
k=0
where α k is the m-1 shape parameters that describe the to-be estimated surface x , E( x' , v' , x) is the
likelihood energy term and E(x) is the prior energy term (or the stabilization term), used to constrain the
estimated shape to a realistic result. ρ is a factor that controls the relative weighting between these two
terms.
Likelihood energy term: This term measures the fitting quality of the digitized landmark sites.
The likelihood is expressed by a measure of the least-squares distance between the digitized points and
the predicted shape:
E( x' , v' , x) = (n −1 )∑n−1|| vi' − ( ( x j ) i + ∑m−2 α k ⋅ p k (j)) ||2 ; (8)

i=0 k=0
where (x j ) i is the jth point on the mean shape of the DPDM that is closest to the digitized point vi' as
described above, p k ( j ) is the jth tuple of the kth shape basis eigenvector.
Prior energy term: As described above, coefficients α k are independent and follow a normal law
with a null mean and variance σ k2 . The topology of the target shape is known a priori to be the same as
that of the mean shape. To penalize the deviation of the predicted shape from the mean shape x of the
DPDM, Mahalanobis distance from the mean according to the normal distribution is used as the energy
term of this prior model:
E( x) = (1/ 2) ⋅ ∑m−2 (α k2 / σ k2 ) (9)

k=0
To determine each α k , the cost function is differentiated with respect to the shape parameters and
equated to zero resulting in a linear system of m unknowns, which is solved with standard methods such
as LU decompositions. As soon as the coefficients α k are determined, the estimated surface is
calculated by:
x = x + ∑ m − 2α k p k (10)
k=0
The estimated surface has the same topology and the same number of vertices as the mean shape x of
the DPDM but different positions for each vertex. It is also a dense surface.
3.3. Deformation
The dense surface estimated in the second stage is taken as the template surface for this stage. The aims
of this stage are to further reduce the reconstruction error by deforming the template surface and at the
same time to preserve the topology of the template surface. Our task is to find a smooth nonlinear
deformation transformation.
General framework: Similar to the second stage, we also need to find the corresponding homologous
points of the input sparse points v ' on the template surface x . Let’s denote these homologous points as:
v = {vi = ( x l )i = (xi ,y i ,zi ) ; i = 0,1,..., n − 1} (11)
where the lth point x l on the template surface x is the closest point to the ith input point v'i .
To compensate the possible positional difference between the input sparse points and the template
3 3
surface, and to estimate a nonlinear mapping t = ( f , g, h)T : → that describe the shape
deformation (here f , g, h are the transforms of the nonlinear mapping t along x, y, and z direction,
respectively), we formulate the shape deformation problem as minimizing following cost function:
E(t ) = (n −1 )∑n−1||v'i − Avi − t(vi ) ||2 +(τ ⋅ ln( m) / ln( n)) ⋅ L[t ] (12)
i=0
where A is an affine transformation to compensate the possible positional difference between v ' and
v . The first term of above equation measures the fitting quality. The second term, L[t ] ≥ 0 , is a
regularization functional defined on the nonlinear mapping t . τ ≥ 0 is a weighting parameter between
the fitting quality and the regularization constraint. m is the number of samples in the training
population; n is the number of digitized points; v ' and v are the n input sparse points and their
corresponding points on the template shape, respectively. t(v ) = {t(vi ) ; i = 0,1,..., n − 1} is the results of
applying the mapping on these corresponding points.
The advantage of such a formulation is that it will adaptively adjust the weight for regularization
term according to the information contained in the input data. The deviation of the deformed surface
from the template shape is penalized more when larger training data (as m becomes bigger) are available
and less when more digitized points (as n becomes bigger) are added. In the extreme cases when n → ∞ ,
it reduced to a pure least-squares nonlinear fitting.
From regularization theory [4], L[t ] can be defined as a norm in a reproducing kernel Hilbert space
(RKHS) which can be uniquely induced by a positive definite (or conditionally positive definite) kernel
function Φ(vi , v j ) . In the support vector community, reproducing kernels are often referred to as Mercer
kernels [4]. They provide an elegant way of dealing with nonlinear deformation algorithms by mapping
them to linear ones in some feature space K nonlinearly related to the input space. Any kernel function
derived from such a space can be used. This is why we consider eq. (12) as a general framework.
Thin-plate spline based specification: To complete the specification of the general framework of
shape deformation, we now discuss a specific form of kernels. There are infinite numbers of
reproducing kernels that can be used for our purpose. One of our prior knowledge is the preservation of
topology, which results in the smoothness constraint on the deformation. The smoothness constraint is
necessary because it discourages mappings which are too arbitrary. One of the measures of the
smoothness of deformation is the space integral of the square of the second order derivatives of the
nonlinear mapping. This leads us to use the 3D thin-plate spline (TPS) kernel Φ(v, vi ) =|| v − vi || , which
is conditionally positive definite and its null space is the affine subspace [32]. Now L[t ] , the measure of
the smoothness of the nonlinear mapping, has the form:
L[t ] = ∫∫∫R3 (I ( f ) + I ( g ) + I ( h))dxdydz

∂2 2 ∂2 2 ∂2 2 ∂2 2 ∂2 2 ∂2 2 (13)
I (⋅) = ( ) + 2( ) + ( ) + 2( ) + ( ) + 2( )
∂x 2 ∂x∂y ∂y 2 ∂y∂z ∂z 2 ∂z∂x
Another advantage of using TPS is that the affine transformation A in eq. (12) is automatically
recovered, as the resulting TPS mapping t = ( f , g ,h ) has the following form:
⎧ f (v ) = a + a x + a y + a z + ∑n−1γ Φ(v, v )
⎪ 1 2 3 4 i=0 i i
⎪ n−1θ Φ(v, v )
⎨ g ( v ) = b + b x + b y + b z + ∑ (14)
1 2 3 4 i=0 i i
⎪ n −1
⎪⎩ h(v ) = c1 + c 2 x + c3 y + c 4 z + ∑i=0 ωi Φ(v, vi )
where a = ( a1 , a 2 , a3 , a 4 )T , b = (b1 , b2 , b3 , b4 )T , c = (c1 , c 2 , c3 , c 4 )T represent the affine coefficients and
γ = (γ 0 ,..., γ n−1)T , θ = (θ 0 ,...,θ n−1 )T , ω = (ω0 ,..., ωn−1)T are the kernel interpolation coefficients. And the
measure of the smoothness of the nonlinear mapping is given by:
L[t ] = γ T Kγ + θT Kθ + ω T Kω (15)
where kij = Φ(vi , v j ) ; i , j = 0,1,..., n − 1 are the elements of matrix K
To determine the affine transformation coefficients a , b , c , and the kernel interpolation coefficients
γ , θ , ω , the cost function is differentiated with respect to all these transform parameters and equated to
zero resulting in following linear system:
⎛ x' y' z ' ⎞ ⎛ (K + (τ ⋅ ln (m) / ln (n)) ⋅ I) P ⎞⎛ γ θ ω ⎞

⎜ ⎟=⎜ ⎟⎜ ⎟ (16)
⎜o
⎝ o o ⎟⎠ ⎜⎝ PT O ⎟⎠⎜⎝ a b c ⎟⎠
where O is a 4 x 4 matrix of zeros, o is a 4 x 1 column vectors of zeros, P = (1, x, y, z). Note that
1 = (1, ...,1)T ; and that x = ( x0 ,..., xn−1 )T , y = ( y 0 ,..., y n−1 )T , z = ( z0 ,..., zn−1 )T ; x ' = ( x0' ,..., xn' −1 )T ,
y ' = ( y 0' ,..., y n' −1 )T , z ' = ( z 0' ,..., z n' −1 )T represents coordinates of the input sparse points and their
corresponding points on the template surface, respectively.
4. Experimental Results
Seven 3D surface models of dry cadaveric proximal femurs as shown in Fig. 5, obtained from CT
volume data, were used for our experiments. Points directly picked from the surfaces of the bones were
used together with the DPDM as the input for reconstruction. A series of experiments were performed
in order to validate the proposed novel reconstruction approach. Two different point acquisition
strategies were investigated: (1) when points can be acquired from any region of the surface of the bone
without any restriction, a typical scenario of open surgery; and (2) when points can only be acquired
from silhouettes determined from multiple different viewing angles, a scenario of minimally invasive
surgery where two or more calibrated X-ray images are used as inputs. In all the studies, two
regularization parameters were experimentally set to ρ = 1.0 and τ = 0.1 , respectively. Each time the
reconstruction result was directly compared to the actual surface model of the corresponding dry femur,
which was taken as the ground truth.
4.1. Reconstruction Error Measurement
To quantify the reconstruction error, Target Reconstruction Error (TRE) was used. The TRE is defined
as the distance between the actual and the reconstructed position of selected target features, which can
be landmark points or bone surfaces themselves. The difficulty in estimating the TRE using landmark
points lies in the determination of the actual positions of these landmarks, which themselves are
subjective and prone to measurement errors. For this reason, we propose to use the distance between the
actual and the reconstructed surfaces as the measure of reconstruction error, which is defined as follows.
Let us define d( v ' , T ) as the distance between a point v ' on the reconstructed surface and the actual
surface T as:
d( v' , T ) = min
v∈T
|| v − v' || (17)
From this definition, we can define a root mean square (RMS) error drmse (T ' , T ) between the
reconstructed surface T ' and the actual surface T as:
1
drmse (T ' , T ) =
| T' | ∫∫
v'∈T '
d( v ' , T ) 2 dT ' (18)
where | T ' | denotes the area of T ' .
To calculate drmse (T ' , T ) , we use the open source tool MESH [33], which implemented an efficient
method to estimate drmse (T ' , T ) between two discrete surface represented by triangular 3D meshes
through surface sampling.
We also adapted this tool to calculate order statistics such as the 5% percentile error, the first
quartile error, the medial error, the third quartile error, and the 95% percentile error, after ordering all
the errors measured as defined by eq. (17) on all the sampled points. For example, the 95% percentile
error d95% Percentile (T ' , T ) is defined as a smallest value that is larger than the errors measured on 95% of all
sample points.
4.2. Experiment 1: using points acquired from the surface
In open surgery, it is possible for the surgeon to access the complete surface. The surgeon can acquire
any point from the surface of the bone. In this section, we first investigate the effect of the number of
points on the reconstruction quality. One of the seven dry cadaveric femurs was randomly chosen for
this experiment. Different numbers of points were picked from the surface of this dry femur. The results
are shown in Fig. 6. It was found that the more points that were input, the smaller the reconstruction
error. Even when only 20 points were used, the RMS error was smaller than 1.5 mm. But a lower
number of points also led to a reduction in the robustness of the registration and required a closer
initialization.
We also applied the new reconstruction approach to all the seven dry cadaveric femurs. Fig. 7.
presents the results when 90 points were used in each case. In all cases, the RMS errors were less than 1
mm and the 95% percentile errors were smaller than 1.5 mm, which demonstrates the stability and the
accuracy of the proposed approach. Fig. 8. shows one of the reconstruction examples.
4.3. Experiment 2: using points only from silhouettes of the surface
In minimally invasive surgery, it is normally very difficult, if not impossible, to access the complete
surface of the target anatomical structure using a pointing device without damaging the surrounding soft
tissues. Instead, intraoperative imaging such as ultrasound or fluoroscopy is used to percutaneously
acquire points. When ultrasound is used, the results are similar to those in Experiment 1. In this section,
we investigate the situation when two or more calibrated fluoroscopic images are used. In such a
situation, only points from the silhouettes of the surface, determined by the detected bony contours in
the images and the projection parameters of the input fluoroscopic images, can be used. Details on
finding these silhouettes are beyond the scope of this paper and the interested may refer to [17]. Fig. 9.
shows the experimental results on all seven dry cadaveric femurs, when only the points from the
silhouettes determined by two simulated perspective views are used. In each case, about 340 points
were acquired from two silhouettes. It was found that the reconstruction errors in all cases were bigger
than the corresponding cases in experiment 1 where only 90 points were used. It demonstrates that not
only the number of points but also the distribution of the acquired points has an effect on the
reconstruction quality. Nevertheless, in all cases the RMS errors were less than 1.5 mm and the 95%
percentile errors were smaller than 2.6 mm. It shows that the reconstruction accuracy of the proposed
approach is appropriate for surgical navigation application. One of the reconstruction examples using
only points from silhouettes is shown in Fig. 10.
5. Conclusion
We have presented a novel approach for robust and accurate anatomical shape reconstruction from
sparse data and dense point distribution model. The reconstruction problem is formulated as a three-
stage optimal estimation process taking the dense point distribution model as the a priori information of
the target anatomical surface. In each stage, the best result is optimally estimated under the assumption
for that stage, which guarantees a topologically preserved solution when only sparse data are available.
The proposed approach was tested using 7 dry cadaveric proximal femurs. Two different point
acquisition strategies were investigated identifying the applicable clinical settings for both open and
minimally invasive surgeries. The effects of both the number of points as well as the distribution of the
acquired points on the reconstruction accuracy were studied. Experimental results show a root mean
square error of 0.5-1.2 mm (95% percentile error of 1.1-2.6 mm).
Our experimental results demonstrate the efficacy of our proposed approach. Our carefully
simulated experiments mimicking clinical settings make our technique readily usable in various
orthopaedic applications. The reconstruction accuracy of our proposed approach demonstrates that it is
appropriate for surgical navigation application. The enormous gains in accurate visualization of patient
specific 3D model of the anatomy with minimal demands in the surgical settings requiring sparse data
makes our approach an attractive and useful tool for surgeons.
The proposed approach is generic and can be easily extended to other rigid anatomical structures,
though in this paper we only demonstrate its application for the reconstruction of proximal femur. Our
future work will focus on applying the proposed approach for reconstruction of spinal vertebrae.
6. Acknowledgments
We acknowledge support from the Swiss National Centers of Competence in Research CO-ME. We
thank Paul Thistlethwaite for his kind help during preparation of this manuscript.
References
[1] G. Zheng, A. Marx, U. Langlotz, K.-H. Widmer, M. Buttaro, L.-P. Nolte, “A hybrid CT-free navigation system
for total hip Arthroplasty”, Comp Aid Surg., vol. 7, pp. 129-145, 2002
[2] S. Lavalle, P. Merloz, E. Stindel, P. Kilian, J. T. P. Cinquin, F. Langlotz, and L.-P. Nolte, “Echomorphing
introducing an intra-operative imaging modality to reconstruct 3d bone surfaces for minimally invasive surgery”,
CAOS 2004, pp. 38–39, 2004.
[3] H. Livyatan, Z. Yaniv, L. Joskowicz, “Gradient-based 2-D/3-D rigid registration of fluoroscopic X-ray to CT”,
IEEE T Med Imaging, vol. 22, no. 11, pp. 1395-1406, 2004.
[4] P. J. Besl and N. D. McKay, “A method for registration of 3D shapes,” IEEE T Pattern Anal, Vol. 14, no. 2, pp.
239-256, 1992.
[5] Y. Chen and G. Medioni, “Object modeling by registration of multiple range images,” Image and Vision
Computing, Vol. 10, no. 3, pp. 145-155, 1992.
[6] Z. Zhang, “Iterative point matching for registration of free-form curves and surfaces,” International Journal of
Computer Vision, Vol. 13, no. 2, pp. 119-152, 1994.
[7] V. Blanz and T. Vetter, “A morphable model for the synthesis of 3d faces,” In SIGGRAPH 1999, pp. 187-194.
[8] T. Evgeniou, M. Pontil, T. Poggio, “Regularization networks and support vector machines”, Adv Comput Math,
vol. 13, pp. 1-50, 2000
[9] T. F. Cootes, C. J. Taylor, D. H. Cooper, J. Graham, “Active shape models – their training and application”,
Comput Vis Image Und, vol. 61, no.1, pp. 38-59, 1995
[10] S. Benameur, M. Mignotte, S. Parent, H. Labelle, W. Skalli, J.A. de Guise, “3D Biplanar Reconstruction of
Scoliotic Vertebrae using Statistical Models,” 20th IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, CVPR’01, Kauai Marriott, Hawaii, USA, vol. 2, pp. 577-582, December 2001.
[11] S. Benameur, M. Mignotte, S. Parent, H. Labelle, W. Skalli, J.A. de Guise, “3D/2D registration and
segmentation of scoliotic vertebrae using statistical models”, Comput Med Imag Grap, vol. 27, pp. 321-337, 2003.
[12] C. S. Chan, P. J. Edwards, D. J. Hawkes, “Integration of ultrasound-based registration with statistical shape
models for computer-assisted orthopedic surgery” SPIE, Medical Imaging, pp. 414-424, 2003.
[13] K. T. Rajamani, L.-P. Nolte, M. Styner, “Bone morphing with statistical shape models for enhanced
visualization”, SPIE Medical Imaging, Volume 5367, pp. 122-130, 2004
[14] K. T. Rajamani, S. Joshi, M. Styner, “Bone model morphing for enhanced surgical visualization”, IEEE
International Symposium on Biomedical Imaging: From Nano to Macro ISBI, pp. 1255-1258, 2004.
[15] K. T. Rajamani, M. A. Gonzalez Ballester, L.-P. Nolte, M. Styner, “A novel and stable approach to
anatomical structure morphing for enhanced intraoperative 3D visualization”, SPIE Medical Imaging, Volume
5744, pp. 718-725, 2005.
[16] T.F. Cootes, G.J. Taylor, J. Haslam, “The use of active shape models for locating structures in medical
images”. Image Vision Comput, vol. 12, no. 6, pp. 355-365, 1994
[17] M. Fleute, S. Lavallee, “Nonrigid 3D/2D registration of images using a statistical model”, Medical image
computing and computer-assisted intervention (MICCAI), Berlin: Springer; pp. 138-147, 1999.
[18] M. Fleute, S. Lavallee, “Building a complete surface model from sparse data using statistical shape models:
application to computer assisted knee surgery system”, Medical image computing and computer-assisted
intervention (MICCAI), Berlin: Springer; pp. 879-887, 1998.
[19] V. Blanz, A. Mehl, T. Vetter, and H.-P. Seidel “A statistical method for robust 3D surface reconstruction
from sparse data”, 3DPVT 2004, pp. 293-300, 2004
[20] P. Golland, W. E. L. Grimson, M. E. Shenton, R. Kikinis, “Small sample size learning for shape analysis of
anatomical structures”, MICCAI 2000, pp. 72-82, 2000
[21] E. Stindel, J.L. Birard, P. Merloz, S. Plaweski, F. Dubrana, C. Lefevre, J. Troccaz, “Bone morphing: 3D
morphological data for total knee Arthroplasty”, Comp Aid Surg., vol. 7, pp. 156-168, 2002
[22] D. Terzopoulos, “The computation of visible-surface representations”, IEEE T Pattern Anal, vol. 10, no. 4, pp.
417-438, 1988.
[23] D.S. Morse, T.S. Yoo, P. Rheingans, D.T. Chen, K.R. Subramanian, “Interpolating implicit surfaces from
scattered surface data using compactly supported radial basis functions”, Proceedings of the International
Conference on Shape Modeling & Applications, pp. 89-98, 2001.
[24] R. J. A. Lapeer, R. W. Prager, “3D shape recovery of a newborn skull using thin-plate splines”, Comput Med
Imag Grap, vol. 24, pp. 193–204, 2000.
[25] T. McInerney, D. Terzopoulos, “A finite element model for 3D shape reconstruction and nonrigid motion
tracking”, Proc. Fourth International Conf. on Computer Vision (ICCV'93) Berlin, Germany, pp. 518-523, May,
1993
[26] S. Osher and J.A. Sethian, “Fronts propagating with curvature-dependent speed: Algorithms based on
hamilton-jacobi formulations”, J. Comp. Phys., vol. 79, no. 1, pp. 12-49, 1988
[27] S. Wang, J. X. Ji, Z.-P. Liang, “Landmark-based shape deformation with topology-preserving constraints“,
Proceedings of the Ninth IEEE International Conference on Computer Vision (ICCV-2003), vol. 2, pp. 923-930,
2003.
[28] C. Brechbuehler., G. Gerig, O. Kuebler, “Parameterization of Closed Surfaces for 3D Shape Description, ”
Comp Vision and Image Under, Vol. 61, pp. 154-170, 1995
[29] R. H. Davies, C. J. Twining, T. F. Cootes, J. C. Waterton, and C. J. Taylor, “3D statistical shape models using
direct optimization of description length”, in Computer Vision – ECCV 2002 Pt Iii, vol. 2352, Lecture Notes in
Computer Science, pp. 3-20, 2002.
[30] C. T. Loop, “Smooth subdivision surfaces based on triangles,” M.S. Thesis, Department of Mathematics,
University of Utah, August 1987.
[31] J. Stoer, R. Bulirsch, “Introduction to Numerical Analysis”, third edition, New York: Springer-Verlag, 2002.
[32] F. Bookstein, “Principal warps: thin-plate splines and the decomposition of deformations”, IEEE T Pattern
Anal, vol. 11, no. 6, pp. 567-585, 1989.
[33] N. Aspert and D. Santa-Cruz, T. Ebrahimi, “MESH: Measuring Errors between Surfaces using the Hausdorff
Distance”, ICME, vol. 1, pp. 705-708, 2002.
Figure Captions
Fig. 1. Schematic view of the proposed reconstruction approach.
Fig. 2. Subdivision example for one of the surfaces in the training database. Left: original mesh
described with 4098 vertices; right: subdivided mesh described with 16386 vertices.
Fig. 3. Shapes reconstructed by varying the first three eigenvectors: in each row the middle one is the
mean shape; the others are obtained by varying corresponding eigenvector -3 σ i (left) or 3 σ i (right)
Fig. 4. Schematic view of the three anatomical landmarks used for initializing the ICP algorithm.
Fig. 5. Surface models of seven dry proximal femurs obtained from CT volume data (the first seven
images) and the mean model of the DPDM (bottom right image).
Fig. 6. Reconstruction errors of one randomly chosen dry femur using different number of points picked
from the surface of the bone
Fig. 7. Reconstruction errors of all 7 dry femurs when 90 points were used in each case.
Fig. 8. One of the reconstruction examples using 90 points; first column: the actual surface model (top)
rendered together with the picked points (bottom); second column: estimated surface after registration
and morphing (top) rendered together with the actual surface (bottom); third column: the final
reconstructed surface (top) rendered together with the actual surface (bottom);
Fig. 9. Reconstruction errors of all 7 dry femurs when using points from silhouettes of the surfaces
Fig. 10. One of the reconstruction examples using points only from the silhouettes of the surface; first
column: the anterior-posterior silhouette (top) and the lateral-medial silhouette (bottom); second column:
the actual surface model rendered together with the silhouettes; third column: the final reconstructed
surface rendered together with the actual surface.

Fig. 1.
Fig. 2.
Fig. 3.
Fig. 4.
Fig. 5.
Fig. 6.
Fig. 7.
Fig. 8.
Fig. 9.
Fig. 10.
A Novel Approach to Anatomical Structure
Morphing for Intraoperative Visualization
Kumar Rajamani, Lutz Nolte, and Martin Styner
M.E. Müller Institute for Surgical Technology and Biomechanics, University of Bern,
Switzerland,
kumar.rajamani@MEMcenter.unibe.ch,
Abstract. In computer assisted surgery 3D models are now routinely

used to plan and navigate a surgery. These models enhance the sur-
geon’s capability to decrease the invasiveness of surgical procedures and
increase their accuracy and safety. Models obtained from specifically ac-
quired CT scans have the disadvantage that they induce high radiation
dose to the patient. In this paper we propose a novel method to construct
a patient-specific model that provides an appropriate intra-operative 3D
visualization without the need for a pre or intra-operative imaging. The
3D model is reconstructed by fitting a statistical deformable model to
minimal sparse 3D data consisting of digitized landmarks and surface
points that are obtained intra-operatively. The statistical model is con-
structed using Principal Component Analysis from training objects. Our
morphing method then computes a Mahalanobis distance weighted least
square fit of the model by solving a linear equation system. The re-
fined morphing scheme has better convergence behaviour because of the
additional parameter that relaxes the Mahalanobis distance term as ad-
ditional points are incorporated. We present leave-one-out experiments
with model generated from proximal femors and hippocampi.
1 Introduction
Three dimensional (3D) models of the patient are routinely used to provide im-
age guidance and enhanced visualization to a surgeon to assist in navigation and
planning. These models are usually extracted from 3D imagery like CT or MRI.
To avoid the high radiation dose and costs associated with such scans, image
free approaches have been researched extensively and are becoming popular es-
pecially in orthopedic surgery. In an image free approach, building a 3D model
that is specific to the patient anatomy is quite challenging as only very sparse
patient data is available.
For this purpose, statistical models of shape have been extensively researched.
The basic idea in model building is to establish from a training set the pattern
of legal variations of shape. The model is adapted to the patient anatomy using
digitized landmarks and bone surface points obtained during surgery. The main
problem here is to extrapolate this extremely sparse three-dimensional set of
points to obtain a complete surface representation. The extrapolation or morph-
ing procedure is done via a statistical principal component analysis (PCA) based
2
shape model. Fleute et al fit the morphed model surface to sparse intra-operative
data via jointly optimizing morphing and pose [1]. Chan et al [5] optimize mor-
phing and pose separately using an iterative closest point (ICP) method. In our
prior work [6] we proposed to iteratively remove shape information coded by dig-
itized points from the PCA model. The extrapolated surface is then computed
as the most probable surface in the shape space given the data. Unlike earlier
approaches, this approach was also able to include non-spatial data, such as pa-
tient height and weight. It is only applicable though for a small set of known
points. In our earlier work [7] we presented a novel morphing scheme that com-
putes a Mahalanobis distance weighted least square fit of the model by solving
a linear equation system.
We propose a enhanced morphing scheme that has better convergence be-
haviour. This is achieved by having an additional parameter in the objective
function that relaxes the Mahalanobis distance term as additional points are
digitized. As more information in terms of additional digitized points is received
we relax the constraint on the surface to remain close to the mean and allow it to
deform so that the error between the predicted surface and the set of digitized
points is minimized as far as possible. In this paper we demonstrate proof of
principle of our method using a proximal femur model as well as hippocampus
model and evaluate these models using leave-one-out experiments.
2 Method
2.1 Model Construction
The first step is to build a deformable model from a training database. The
basic idea of building a statistical model based on PCA is to establish, from
the training set, the pattern of legal variations in the shapes for a given class
of images. Statistical PCA models were introduced by Cootes et al[2] based on
point distribution model (PDM).
A key step in this model building involves establishing a dense correspon-
dence between shape boundaries over a reasonably large set of training images.
Our previous comparison study [4] of some of the popular correspondence estab-
lishing methods revealed that for modeling purposes the best of the correspon-
dence method was Minimum Description Length (MDL) [3]. Correspondence was
initialized with a semi-automatic landmark driven method and then optimized
based on the MDL criterion.
We construct a deformable statistical shape model based on the correspond-
ing point positions. Each member of the training population is described by in-
dividual vectors x̄i containing all 3D point coordinates. The aim of building this
model is to use several training datasets to compute the principal components of
shape variation. PCA is used to describe the different modes of variations with
a small number of parameters. For the computation of PCA, the mean vector
x̄ and the covariance matrix D are computed from the set of object vectors(1).
The sorted eigenvalues λi and corresponding eigenvectors pi of the covariance
3
√ √ √ √
-2 λ -1 λ +1 λ +2 λ
Fig. 1. The first two eigen modes of variation √

of our proximal femur model. The shape
instances were generated by evaluating x̄ + ω λk uk with ω ∈ {−2, .., 2}
matrix are the principal directions spanning a shape space with x̄ representing
its origin(2). Objects xi in that shape space can be described as linear combina-
tion with weights bi calculated by projecting the difference vectors xi − x̄ into
the eigenspace(3).
n
1 X
D= (xi − x̄) · (xi − x̄)T (1)
n−1 1
P = {pi }; D · pi = λi · pi ; (2)
bi = DT (xi − x̄); xi = x̄ + P · bi (3)
Figure 1 shows the variability captured by the first two modes of variation
of our proximal femur model varied by ±2 standard deviation.
2.2 Morphing
Anatomical structure Morphing is the process of recovering the patient specific
3D shape of the anatomy from the few available digitized landmarks and surface
points. Our approach uses the statistical based shape model built earlier to infer
the anatomical information in a robust way. This is achieved by minimizing the
residual errors between the reconstructed model and the cloud of random points,
and provides the best statistical shape that corresponds to the patient.
Earlier morphing methods were based on fitting procedures in Euclidean
space and have the disadvantage that these are often computationally expensive
and only a small set of shape variations can be considered. The morphed model
also does not represent the most probable shape given the input data but rather a
constrained fit. Our novel morphing method operates directly in the PCA shape
space incorporating the full set of possible variations. The method consists of
two steps
4
– Initially a small point-set of anatomical landmarks with known correspon-

dence to the model is digitized. This is used to register the patient anatomy
to the model. This also provides an initial estimation of the 3D shape with
only a few digitized points.
– To improve the prediction additional points can be interactively incorporated
via closest distance correspondence. A color coded feedback is given to the
surgeon which shows regions where the prediction is accurate and regions
where the prediction could be improved. This assists the surgeon in deciding
the location where to digitize extra points.
The morphing computation is based on formulating the problem as a linear

equation system and then solving for the shape parameters that best describe
the unknown shape. An additional term in the objective function minimizes the
Mahalanobis shape distance. The objective function that we minimize is defined
as follows
 

 


 
 ( )

 N
X m
X 
 m
X
2 α2 i
f = ρ∗ γ ∗ kYk − (Xj + αi pi (j))k +(1 − ρ) (4)

 
 λi

 k=1 i=1 
 i=1

 

j = indexk
with N the number of points that are digitized, Yk is the kth digitized point,
Xj is the point in the mean model that is closest to Yk , pi (j) is the jth tuple
of the ith shape basis vector, λi the ith eigen value and α0i s are the m shape
parameters that describe the shape. The first term of the function minimizes
the distance between the predicted shape and the set of digitized points. This is
similar to the Euclidean distance term used by Fleute [1] . The second term con-
trols the probability of the predicted shape. This term ensures that the predicted
shape has minimal Mahalanobis shape distance. The factor γ is a parameter that
weights the two terms of the function and ensures that a valid shape is predicted
in the scenario when there are relatively few digitized points. A series of tests
with varying values of gamma was carried out to determine the optimal value
of gamma. The granularity of gamma was chosen using binary selection scheme
where the region containing the current best value of gamma was further divided
to find gamma to an acceptable level of accuracy. Our series of tests revealed
that for our current application the best results with the least prediction mean
and median errors were obtained when the value of gamma was fixed at one.
Hence based on our tests the optimal value of γ was empirically fixed at one.
We modified the morphing scheme to one that is enhanced and has better
convergence behaviour. This is achieved by having an additional parameter ρ in
the objective function that relaxes the Mahalanobis distance term as additional
points are digitized. As more information in terms of additional digitized points
is received we relax the constraint on the surface to remain close to the mean and
allow it to deform so that the error between the predicted surface and the set
5
Fig. 2. Left: A typical proximal femur of the population that was used in the leave-
one-out test. Middle: The average shape of the population with color coded distance
map to the actual shape. The mean error is 3.37 mm and the median is 2.65 mm.
Right:The shape based on only 6 digitized points with color coded distance map to the
actual shape. The mean error is 1.50 mm and the median error is 1.25 mm
of digitized points is minimized as far as possible. As the error ideally decreases

exponentially with the increase in the number of digitized points, we chose ρ to
increase logarithmically, and was defined according to the following equation
(
0.5 N ≤6
ρ = log{ M axN
N
(g∗e−1)+1} (5)
2∗log(g∗e) + 0.5 N >6
where N is the number of digitized points, MaxN is the total number of points
g is a factor which determines the rate of growth of ρ. To achieve faster growth
rate for ρ, g was empirically set to be the number of members in the population.
To determine the shape parameters αi that best describe the unknown shape,
the function f is differentiated with respect to the shape parameters and equated
to zero. This results in a linear system of m unknowns, which is solved with
standard linear equations system solvers using QR decomposition.
3 Results
In this paper we demonstrate proof of principle of our method using the proxi-
mal femur structure. 14 CT scans of the proximal femur were segmented and a
sequence of correspondence establishing methods was employed to compute the
optimal PCA model [4]. A series of leave-one-out experiments was carried out to
evaluate the new method. Three anatomical landmarks, the femoral notch and
the upper and the lower trochanter are used as the first set of digitized points.
This is used to initially register the model to the patient anatomy. The remaining
points are added uniformly across the spherical parameterization so that they
occupy different locations on the bone surface.
Our studies with the two different correspondence methods, MDL and closest
correspondence for incorporating additional points along with different error
plots are discussed in [7]. Figure 2 shows a example of a very good estimate
with mean error of 1.5mm obtained with as few as 6 digitized points using MDL
6
Fig. 3. Statistics cumulated from the different leave-one-out experiments of the prox-
imal femur with and without the ρ factor. The average of the mean error and the
average of the median is plotted against the number of digitized points Left: Shows
the error plot obtained using MDL correspondence. Right: Shows the error plot using
Closest Point Correspondence
correspondence. The color-coded 3D rendering is calculated using Hausdorff’s

Distance to measure the distance between discrete 3D surfaces[8].
Here we present results using our refined morphing scheme and also compare
it to our initial version. Figure 3 shows the cumulative statistics of all leave-one-
out experiments with and without ρ factor using the MDL and Closest point
correspondence. In both the cases there seemed to be no significant improvement
using the ρ factor, mainly due to low number of subjects in our proximal femur
study population.
To evaluate the influence of the ρ factor we studied the enhanced morphing
scheme in Hippocampus model generated from 172 hippocampus instances[9].
Here the larger population helps us to efficiently capture the shape variability
and also helps us to evaluate better the influence of the ρ factor. Figure 4 shows
the cumulative statistics from ten randomly chosen leave-one-out experiments
with and without ρ factor using the MDL and Closest point correspondence for
the Hippocampus population. Here we can clearly see the excellent influence of
the ρ factor. The better convergence and the error factor We gain is about 10%
in the MDL scenario and about 5% in the closest correspondence case.
4 Discussion
In this paper we have presented a refined novel anatomical structure morph-
ing technique to predict the three dimensional model of a given anatomy using
statistical shape models. Our scheme is novel in that it operates directly in
the PCA shape space and incorporates the full set of possible variations. It is
also fully interactive, as additional bone surface points can be incorporated in
real-time. The computation time is mainly independent of the number of points
7
Fig. 4. Statistics cumulated from ten randomly chosen hippocampus leave-one-out

experiments with and without the ρ factor. The average of the mean error and the
average of the median is plotted against the number of digitized point. Left: Shows
the error plot obtained using MDL correspondence. Right: Shows the error plot using
Closest Point Correspondence
intra-operatively digitized, and largely depends on the number of members in the

population. The enhancement of this scheme compared to our earlier approach
is that we achieve smaller errors and better convergence as additional points are
digitized.
The gamma parameter plays a vital role in balancing the predictive error
term and the probability term. We empirically fixed its value to adapt to the
case when small number of points are digitized. The ρ parameter helps us to
relax the probability term to get a much better estimate as more points are dig-
itized. The effect of the ρ parameter is not significantly noticed in the case when
the population size is small. This is because the error gets stabilized and uniform
after the first few points are digitized and there is not much information that
could be extracted by adding additional points in this case. Hence the ρ factor
seem not to contribute much as was observed in the proximal femur model with
a population size of only 14 members. On the contrary in the hippocampus pop-
ulation the effect of the ρ parameter was significantly visible and it contributes
in a significant way to decrease the error and achieve better convergence.
Another interesting observation that we can make is that the average mean
error in the hippocampus population is far less compared to the proximal femur
population. With 20 digitized points the average mean error in the proximal
femoral population is about 2.25mm whereas in the hippocampus population
it is only 0.37mm. The reason for this is because the hippocampus is a simple
shape and we had a large population for the hippocampus model. Interestingly
the error reduction that we achieve with 20 digitized points is about 35% for
both the models.
There are a number of extensions that we plan to incorporate to this idea.
We have a fully developed and validated technology at M.E. Müller Institute to
8
extract bone contours from Ultrasound (US) images. First we plan to this use this
large set of bone surface points from US images into the morphing scheme. Using
this technique we can non-invasively get a large set of bone surface points intra-
operatively. We also plan to incorporate fluoroscopic images into the process to
extract surface points.
The concept of anatomical structure morphing has many interesting medical
applications. The primary application that we focus is on hip surgery such as
total hip replacement (THR) and knee surgery such as total knee arthroplasty
(TKA) and anterior cruciate ligament surgery (ACL). Several current navigation
systems for TKA/THR do not require preoperative CT or planning. By moving
the joint, the center of motion is obtained. The hip, knee, and ankle motion
centers give the functional axes of the femur and tibia. The surgeon is usually
provided with a digital readout and a single display of the relative bone posi-
tions or angles. It is sometimes difficult for surgeons to intuitively understand
such displays. The technique of anatomical structure morphing introduces novel
navigation concepts wherein reconstructed 3D bony images are overlaid on top
of 2D views of the axes. The proposed technology brings a variety of advantages
to orthopaedic procedures, such as improved accuracy and safety, often reduced
radiation exposure, and improved surgical reality through 3D visualization and
image overlay techniques. In particular navigation based on anatomical structure
morphing opens the door to larger minimally invasive approaches.
References
1. Fleute, M., Lavallee, S.: Building a Complete Surface Model from Sparse Data Using
Statistical Shape Models, MICCAI (1998) 879-887
2. Cootes, T., Hill, A., Taylor, C.J., Haslam, J.: The Use of Active Shape Models for
Locating Structures in Medical Images. Img. Vis. Comp. (1994) 355-366
3. Davies, Rh.H, Twining, C.J., Cootes, T.F., Waterton, J. C., Taylor, C.J.: A Mini-
mum Description Length Approach to Statistical Shape Model. IEEE TMI (2002)
4. Styner, M.A., Kumar .T.R., Nolte L.P., Zsemlye G., Szekely, G., Taylor, C.J., Davies
Rh.H.,: Evaluation of 3D Correspondence Methods for Model Building, IPMI (2003)
63-75
5. Chan, C.S., Edwards, P.J., Hawkes, D.J., : Integration of ultrasound-based registra-
tion with statistical shape models for computer-assisted orthopaedic surgery, SPIE,
Medical Imaging (2003) 414-424
6. Kumar T.R., Nolte L.P., Styner M.A.,: Bone morphing with statistical shape models
for enhanced visualization, SPIE Medical Imaging (2004)
7. Kumar T.R., Joshi, S.C., Styner M.A., : Bone model morphing for enhanced surgical
visualization, IEEE International Symposium on Biomedical Imaging: From Nano
to Macro ISBI (2004)
8. Aspert,N., Santa-Cruz, D., Ebrahimi, T.,: MESH:-Measuring Errors between Sur-
faces using Hausdorff Distance, IEEE ICME (2002) 705-708
9. Styner, M.A., Lieberman, J., Gerig, G. Boundary and Medial Shape Analysis of the
Hippocampus in Schizophrenia, MICCAI (2003)
Bone Morphing with statistical shape models for enhanced
visualization
Kumar T. Rajamania , Johannes Hugb , Lutz-Peter Noltea , Martin Stynera
a M.E. Müller Research Center for Orthopaedic Surgery, Institute for Surgical Technology and
Biomechanics, University of Bern, P.O.Box 8354, 3001 Bern, Switzerland
b sd&m Schweiz AG, World Trade Center, Zürich, Switzerland
ABSTRACT
This paper addresses the problem of extrapolating extremely sparse three-dimensional set of digitized landmarks
and bone surface points to obtain a complete surface representation. The extrapolation is done using a statistical
principal component analysis (PCA) shape model similar to earlier approaches by Fleute et al. 1 This extrapolation
procedure called Bone-Morphing is highly useful for intra-operative visualization of bone structures in image-free
surgeries. We developed a novel morphing scheme operating directly in the PCA shape space incorporating the
full set of possible variations including additional information such as patient height, weight and age. Shape
information coded by digitized points is iteratively removed from the PCA model. The extrapolated surface is
computed as the most probable surface in the shape space given the data. Interactivity is enhanced, as additional
bone surface points can be incorporated in real-time. The expected accuracy can be visualized at any stage of
the procedure. In a feasibility study, we applied the proposed scheme to the proximal femur structure. 14
CT scans were segmented and a sequence of correspondence establishing methods was employed to compute the
optimal PCA model. Three anatomical landmarks, the femoral notch and the upper and the lower trochanter are
digitized to register the model to the patient anatomy. Our experiments show that the overall shape information
can be captured fairly accurately by a small number of control points. The added advantage is that it is fast,
highly interactive and needs only a small number of points to be digitized intra-operatively.
Keywords: Statistical Shape Model, Principal Component Analysis, Bone Morphing, Deformable Models,
Computer Assisted Visualization
1. INTRODUCTION
Computer Assisted imaging techniques such as CT, MRI have gained great acceptance for use in diagnosis
and therapy planning. To avoid the high radiation dose and costs associated with such scans, current trends
are towards non-ionized or minimal imaging. Image free techniques using model building is also being ac-
tively researched to provide surgical guidance. These techniques are applicable in surgeries such as Total Knee
Arthroplasty (TKA), Total Hip Arthroplasty and Anterior Cruciate Ligament reconstruction (ACL) where only
pre-operative X-Ray is available. Constructing a patient specific 3D surface in non-image based approach is
quite challenging. This is usually done by building a deformable model and adapting the model to the patient
anatomy.
The use of model based a-priori knowledge to simplify and stabilize problems has long been explored in the
computer vision community. It began with the introduction of deformable models in its various forms such as
snakes,3 deformable templates4 and active appearance models.5 The amount of prior knowledge included in
these models varies from simple smoothness assumptions to very detailed knowledge about the surface. In the
field of medical imaging, the usage of statistical shape models has found widespread use6, ,7 since the notion of
biological shape could be best defined by a statistical description of a large population.
In order to provide similar sophisticated visualization similar to image based CAS systems, Fleute et al 1
proposed a model based technique to extrapolate the 3D patient specific surface from digitized landmarks and
Further author information: (Send correspondence to Kumar T. Rajamani)
Kumar Rajamani: E-mail: kumar@MEMcenter.unibe.ch, Telephone: +41 31 632 9994
Figure 1. Four selected proximal femur structure from our collection that consists of 14 samples
bone surface points obtained during surgery. The extrapolation procedure also called bone morphing is done
via a statistical principal component analysis (PCA) based shape model. The model surface is fitted to the
sparse intra-operative data via jointly optimizing morphing and pose. Chan et al2 use a similar algorithm, but
optimize morphing and pose separately using an iterative closest point (ICP) method. 8 These approaches to
bone morphing based on fitting procedures in Euclidean space have the disadvantage that these are often com-
putationally expensive and only a small set of shape variations can be considered. Also non-spatial information
such as patient height, weight, sex information cannot be incorporated in these earlier techniques.
In this paper we propose a novel bone morphing scheme that generates patient-specific 3D knee surfaces
from sparse intra-operative digitized data. Our main goal is to provide an interactive 3D visualization tool that
takes into account the prior knowledge of shapes as far as possible. We also need to get a good estimate and
describe the ”unknown” shape of the object with minimal number of digitized points. In our novel bone morphing
scheme we progressively remove shape information represented by digitized points from the PCA model. 10 The
extrapolated surface is then computed as the most probable surface in the shape space given the data. The
3D model is hence obtained by deforming the statistical model to match the digitized data. We can also get
additionally a map displaying the expected remaining variability. In this paper we demonstrate proof of principle
of our method using a proximal femur model generated from 14 CT datasets and evaluated using leave-one-out
experiments.
In the following, Section 2 reviews shortly the statistical shape analysis using principal components and
explains the mathematical notation. In Section 3, we discuss in detail bone morphing via the progressive
subtraction of variation. Section 4 presents out first results of bone morphing using our technique. Finally,
Section 5 concludes this report and outlines the next steps.
2. MODEL CONSTRUCTION USING PRINCIPAL COMPONENTS

In order to have a efficient statistical shape description at our disposal, we employ a representation that is
based on a principal component analysis (PCA) of all object instances in our database. The basic idea of a
PCA model11 consists in separating and quantifying the main variations of shape that occur within a training
population of objects. PCA defines a linear transformation that decorrelates the parameter signals of the original
shape population by projecting the objects into a linear shape space spanned by a complete set of orthogonal
basis vectors. If the parameter signals are highly correlated, then the major variations of shape are described
by the first few basis vectors. Furthermore, if the joint distribution of the parameters describing the surface is
Gaussian, then a reasonably weighted linear combination of the basis vectors results in a shape that is similar
to the existing ones.
A key step in the construction of statistical model is establishing a dense correspondence between the surface
boundaries for the members of the training set. In 2D, correspondence is often established using manually de-
termined landmarks, but this is a time-consuming, error-prone and subjective process. In principle, the method
extends to 3D, but in practice, due to very small sets of reliably identifiable landmarks, manual landmarking
becomes impractical. Most automated approaches posed the correspondence problem as that of defining a param-
eterization for each of objects in the training set, assuming correspondence between equivalently parameterized
points.
In our earlier paper12 we performed a comparative study of some of the popular correspondence establishing
methods for model based applications. We analyzed both the direct correspondence via manually selected
landmarks as well as the properties of the model implied by the correspondences, in regard to compactness,
generalization and specificity. The studied methods include a manually initialized subdivision surface (MSS)
method and three automatic methods that optimize the object parameterization: SPHARM, 13 MDL14 and the
covariance determinant (DetCov)15 method. In all studies, DetCov and MDL showed very similar results. The
model properties of DetCov and MDL were better than SPHARM and MSS. The results suggested that for
modeling purposes the best of the studied correspondence method are MDL and DetCov.
[1] [1] [1]
Our population consisted of 14 proximal femor instances, given as point distribution models p i = [xi , yi , zi ,-
[M [M ] [M ]
..., xi , yi , zi ]T with M = 4096 points. Five selected examples of population are illustrated in Fig. 1.
Correspondence among the population members was initialized using the semi-automatic landmark driven
method(MSS) and then optimized based on the MDL criteria. The shapes are first centered by calculating
the average model p̄ and computing the instance specific difference vector ∆pi s .
N
1 X
p̄ = pi , ∆pi = pi − p̄, ∆P = [∆p1 · · · ∆pN ] (1)
N i=1
PNThe difference vectors span a N −1 dimensional space. The missing dimension is due to the linear dependence:
3M ×3M
i=1 ∆p i = 0. The corresponding covariance matrix Σ ∈ IR is hence rank deficient. This circumstance is
exploited to speed the calculation of the valid eigenvalues and eigenvectors. Instead of calculating the full eigen
system of the covariance matrix Σ , the multiplication of the eigenvectors of a smaller matrix Σ̆ with ∆P leads
to the correct principal components:
1 PCA
Σ̆ = ∆P T ∆P = ŬΛ0 Ŭ T , Λ0 = diag(λ1 , · · · , λN −1 , 0) (2)
N −1

U = [u1 · · · uN −1 , uN ] = ψ ∆P Ŭ ψ (A) = N ormalize columns of A (3)
(4)
The sorted eigenvalues λi and corresponding eigenvectors ui of the covariance matrix are the principal direc-
tions spanning a shape space with p̄ representing its origin. Objects pi in that shape space can be described as
linear combination with weights bi calculated by projecting the difference vectors (pi − p̄) into the eigenspace.(5)
bi = U T ∗ ∆pi ; pi = p̄ + U · bi (5)
Figure 2 shows the variability captured by the first two modes of variation of our model varied by ±2
standard deviation. The shapes representing the first eigenmode in the first row are calculated by adding the
weighted first eigenvector u1 to the average model p̄. The bottom row with the second eigenmodes are calculated
correspondingly.
3. BONE MORPHING USING PROGRESSIVE ELIMINATION OF VARIATION

We developed a novel morphing scheme operating directly in the PCA shape space incorporating a large set of
possible variations including parameters additionally to spatial information such as patient height, weight and
age. Our method is based on the iterative removal of shape information associated with the digitized points.
First we calculate the most probable shape subject to the boundary conditions that are related to the three
initially digitized landmarks. The resulting outline is used for registration and as an initial configuration for
computing the most probable shape for the next digitized point. Using statistical shape analysis, we examine
the remaining shape variability, after the surface information coded by the digitized points is progressively
subtracted. This procedure of point selection and variability removal is repeated until a close approximation
√ √ √ √
-2 λk -1 λk +1 λk + 2 λk
λ1
λ2
Figure√ 2. The first two eigen modes of variation of our model. The shape instances were generated by evaluating
x̄ + ω λk uk with ω ∈ {−2, .., 2}
to the patient anatomy is achieved. The final extrapolated surface represents the most probable surface in the
shape space given the digitized landmarks.
We detail below the method used to compute the most probable shape given the position of an arbitrary
point, and also the method to subtract the variation coded by digitized points.
3.1. Shape-based Basis Vectors for One Point

In the context of PCA the mean model is the most probable surface. Hence the ’most probable surface’ that
satisfies all the boundary conditions would be the surface that has minimal deviation from the mean. This means
that we must choose the model with minimal Mahalanobis-distance Dm .
To determine the most probable surface given the position of an arbitrary point j, we must thus find the
minimally Mahalanobis-distant surface ~b that contains the point j at the required position pj . In order to do
so for all possible displacements pj , we should seek for three shape vectors ∈ span(U ) that translate the point j
in either x- y- or z-direction (in the object space), thereby only inducing a minimal deformation of the overall
surface. More precisely, the three shape vectors shall effect a unit displacement of the vertex j in either x- y- or
z-direction, and, as there are probably many such vectors, they shall be of minimal Mahalanobis-length D m . If
we have found them, we can satisfy all possible boundary conditions pj with minimal deformation of the surface
by just adding the three appropriately weighted “basis” vectors to the mean. This problem gives rise to the
following constrained optimization:
Let rxj , ryj and rzj denote the three unknown basis vectors causing unit x- y- and z- translation of point j,
respectively. The Mahalanobis-length Dm of these three vectors then given by:
2
−1 [e]
N
X rk
Dm (rk ) = (U rk )T Σ−1 U rk = rTk Λ−1 rk = , k ∈ {xj , yj , zj } (6)
e=1
λe
Taking into account that xj , yj and zj depend only on three rows of U , we define the sub-matrix Uj according
to the following expression:
     [2j−2]   
xj xj U xj
 yj  = y j  + U [2j−1]  b =  yj  + Uj b, U [j] = j th row of U (7)
zj zj U [2j] zj
In order to minimize the Mahalanobis-distance Dm subject to the constraint of a decoupled x- y- or z-

translation by one unit, we establish — as is customary for constrained optimizations — the Lagrange function
L: 2      
N
X −1 r [e] 1 0 0
k
L(rk , lk ) = − lTk [Uj rk − ek ] , e{xj ,yj ,zj } = {0 , 1 , 0} (8)
λe
e=1 0 0 1
The vectors lxj , lyj and lzj contain as usual the required Lagrange multipliers. To find the minimum of
L(rk , lk ), k ∈ {xj , yj , zj }, we calculate the derivatives with respect to all elements of rxj , ryj , rzj , lxj , lyj and
lzj and set them equal to zero:
     
2
.. .. .. .. ..
. . . . .
   λ1 



 
  
δL(rk , lk )   .. .. .. .. .. .. 
=0,
 . . −Uj  T

r
 x j . r yj . r z j  

 0 . 0 . 0 

δrk  .   .. ..   .. .. 
 2 ..   =  (9)
δL(rk , lk ) 
  λN −1   . .   . . 
=0      
δlk  . . . . . . . . . . . . . . . . . . . . . . . . .  . . . . . . . . . . . . . . . . . . .  . . . . . . . . . . . . . . . . . . . 

    .. ..
.. .. ..
U j . 0 l xj . l yj . lzj ex . ey . ez
j j j
If the basis vectors and the Lagrange multipliers are combined according to Rj = [rxj ryj ] and Lj = [lxj lyj ],
Eq. (9) can be rewritten as two linear matrix equations:
2Λ−1 Rj = UjT Lj (10)
Uj R j = I (11)
The three basis vectors rxj , ryj and rzj then result from simple algebraic operations ( resolve (10) for Rj and
replace Rj in (11) by the result, use the resulting equation to find Lj = 2[Uj ΛUjT ]−1 . Substitute for Lj in (10)
) and RJ is given by: −1
Rj = rxj ryj rzj = Λ UjT Uj Λ UjT (12)
While rxj describes the translation of xj by one unit with constant yj , zj and minimal shape variation, ryj
and rzj alter yj and zj correspondingly. The most probable surface p̌ given the displacement [∆xj , ∆yj , ∆zj ]T
of an arbitrary control vertex j is then obviously determined by
 
∆xj
p̌ = p + U Rj  ∆yj  . (13)
∆zj
3.2. Point-wise Subtraction of Variation

In the previous section we have seen how to estimate the most probable shape given the position of one specific
control vertex j. Before we proceed to the next control point, we ensure that subsequent shape modifications
will not alter the previously adjusted vertex j. To do so, we remove those components from the statistic that
cause a displacement of this point. Therefore, we subtract the variation coded by the point j from each instance
√ √ √ √
-2 λk -1 λk +1 λ1 + 2 λk
λ1
λ2
Figure 3. The first two eigen modes of variation of our model after the variability associated with one landmark point
has been removed from the population.
i, and rebuild the statistic afterwards. For the first part of this operation, we must subtract the basis vectors R j
weighted by the specific displacement [∆xj , ∆yj , ∆zj ]Ti , from each object instance i:
 
∆xj
bĵi = bi − Rj  ∆yj  = bi − Rj Uj bi = (I − Rj Uj ) bi (14)
∆zj i
By doing so for all instances, we obtain a new shape description that is invariant with respect to point j.
On the assumption that the initial shape parameters are however correlated, in particular, we obtain a point-
normalised population whose total variability is smaller than the one of the original collection. In order to verify
this assumption and to rebuild the statistic, we apply anew a PCA to the normalised set of shape instances
{bĵi | i ∈ {1, . . . , N }}. Note that the dimensionality of the resultant eigenspace decreases by three (compared to
the original one), as we have just removed three degrees of freedom. The point-normalised principal components,
denoted by U ĵ , confirm the expected behaviour and validate the removal of the variation in respect of point j.
The first two dominant one-point invariant eigenmodes are illustrated in Fig.3.
3.3. Computing the Most Probable Surface

Having such a progressive shape analysis technique at our disposal, we can now find the most probable surface by
repeatedly applying the point selection and elimination procedure, until the overall shape variability is sufficiently
small. The morphed surface is hence extrapolated from few landmarks and digitized points. Analogous to
traditional parametric curve representations, each digitized point j is associated with the three principal basis
functions U Rj that are globally supported.
The final most probable surface p̌l is given by reversing the point elimination procedure, that is, by combining
the surface information that is coded by the various control points. Accordingly, we combine the mean surface
ŝ
p with all the weighted principal basis functions Rjkk−1 :
 
l
X ∆xjk
ŝ
p̌l = p + U ŝk−1 Rjkk−1  ∆yjk  (15)
k=1 ∆zjk
ŝ
The weights [∆xjk , ∆yjk , ∆zjk ]T for the basis vectors U ŝk−1 Rjkk−1 depend on the surface defined by the
previous principal landmarks ŝk−1 . To emphasise the hierarchical structure of our formalism and to simplify the
algorithmic implementation, we use the following recursive definition instead of Eq. (15):
 
∆xjk
ŝ
p̌0 = p , p̌k = p̌k−1 + U ŝk−1 Rjkk−1  ∆yjk  (16)
∆zjk
With this shape representation of p̌k , we have a novel bone morphing method that can estimate unknown
patient specific models using the simple adjustment of a small number of points, taking into account all the prior
knowledge of the surface.
4. RESULTS
In this paper we demonstrate proof of principle of our method using proximal femur structures. 14 CT scans
of the proximal femur were segmented and a sequence of correspondence establishing methods was employed to
compute the optimal PCA model. Then, leave-one-out experiments were carried out to evaluate the new method.
Three anatomical landmarks, the femoral notch and the upper and the lower trochanter are used as the first
set of digitized points. This simulates the surgical scenario where we need to initially register the model to the
patient anatomy. The remaining points are added uniformly across the spherical parameterization so that they
occupy different locations on the bone surface.
In most surgeries only a portion of the bone is accessible. We simulated this in our leave-one-out analysis
with six digitized points, by digitizing only on the femoral head. The average error of the bone morphing across
the whole surface was 4.2mm (Figure 5).
Our scheme is novel in that it operates directly in the PCA shape space and incorporates the full set of
possible variations including other patient parameters such as height, weight and age. It is fully interactive, as
additional bone surface points can be incorporated in real-time. The expected accuracy of the current model can
be visualized at any stage of the procedure. This could be used to give additional visual (color-coded) feedback
to the surgeon to highlight regions of high and poor expected accuracy.
5. CONCLUSIONS
In this paper we have demonstrated a novel technique to predict the three dimensional model of a given anatomy
using statistical shape models. Earlier bone morphing methods based on fitting procedures in Euclidean space
have the disadvantage that these are often computationally expensive and only a small set of shape variations
can be considered. Also non-spatial patient information cannot be incorporated in these approaches. Using
these methods, the morphed model does not represent the most probable shape given the input data but rather
a constrained fit. Our scheme is novel in that it operates directly in the PCA shape space and incorporates the
full set of possible variations. We can include parameters additionally to spatial information such as patient
height, weight and age. It is fully interactive, as additional bone surface points can be incorporated in real-time.
Our experiments on this proximal femur data show that our method can capture the overall shape information
fairly accurately by a small number of control points. This method has the advantage in that it is fast, highly
interactive and needs only a small number of points to be digitized intra-operatively.
A visual (color-coded) feedback to highlight regions of high and poor accuracy assists the surgeon to choose
regions to digitize additional points. The expected accuracy of the current model can be visualized at any stage
of the procedure.
√ √ √ √
-2 λk -1 λk +1 λk + 2 λk
λ1
λ2
Figure 4. The first two eigen modes of variation of our model after the variability associated with three landmark point
has been removed from the population.
Figure 5. View of the predicted most probable surface overlaid on top of the actual object in a leave-one-out analysis
with six points being digitized only on the femoral head. The average error of the bone morphing in this case was at 4.2
mm
The relation between shape and parameters such as sex, height, weight, pathology is an interesting area of
research. The formulation and implementation of our bone morphing is such that we can include these other
patient parameters into the estimation procedure. Alternatively we can also predict the parameters based on the
surface using the morphing method. In our future work we plan to study the importance and the predictive power
of these parameters. We also plan to incorporate surface points extracted from ultrasound into the morphing
scheme.
ACKNOWLEDGMENTS
We thank Guido Gerig, Gábor Székely and Gabor Zsemlye for insightful discussions about modeling and
shape prediction. The datasets were provided by Frank Langlotz via the Swiss CO-ME network. The MDL
tools were provided by Rhodri H. Davies. The error computations were carried out using MESH available at
http://mesh.epfl.ch. This work was funded by AO/ASIF foundation.
REFERENCES
1. M. Fleute and S. Lavallee, “Building a complete surface model from sparse data using statistical shape
models,” MICCAI, pp. 879–887, 1998.
2. C. S. Chan, P. J. Edwards, and D. J. Hawkes, “Integration of ultrasound-based registration with statistical
shape models for computer-assisted orthopaedic surgery,” SPIE, Medical Imaging, pp. 414–424, 2003.
3. M. Kass, A. Witkin, and D. Terzopoulos, “Snakes: Active contour models,” Int. Journal of Computer Vision
1(4), pp. 321–331, 1988.
4. A. Yuille, D. Cohen, and P. Hallinan., “Feature extraction from faces using deformable templates.,” Proc.
CVPR’89, pp. 104–109, 1989.
5. T. Cootes, C. Taylor, D. Cooper, and J. Graham, “Active shape models - their training and application,”
Computer Vision and Image Understanding 61, pp. 38–59, 1995.
6. A. Kelemen, G. Székely, and G. Gerig, “Three-dimensional Model-based Segmentation,” in Proc. of
MB3IA’98, 1998.
7. A. Kelemen, G. Székely, and G. Gerig, “Elastic model-based segmentation of 3d neuroradiological data
sets,” IEEE Transactions on Medical Imaging 18, pp. 828–839, October 1999.
8. Y. Ge, C. Maurer, and J. Fitzpatrick, “Image registration using the iterative closest point algorithm with a
closestpoint transform,” in SPIE, 2710, pp. 358–367, 1996.
9. D. Zorin, P. Schröder, and W. Sweldens, “Interactive multiresolution mesh editing,” Computer Graphics
31(Annual Conference Series), pp. 259–268, 1997.
10. J. Hug, C. Brechbühler, and G. Székely, “Model-based initialisation for segmentation,” in Proceedings 6’th
European Conference on Computer Vision - ECCV 2000, Part II, D. Vernon, ed., Lecture Notes in Computer
Science, pp. 290–306, Springer, June 2000.
11. T. Cootes and C. Taylor, “Active Shape Models - ‘Smart Snakes’,” in British Mach. Vision Conf., pp. 266–
275, Springer-Verlag, 1992.
12. M. A. Styner, K. T. Rajamani, L. P. Nolte, G. Zsemlye, G. Szekely, C. J. Taylor, and R. H. Davies,
“Evaluation of 3d correspondence methods for model building,” IPMI, pp. 63–75, 2003.
13. C. Brechbühler, G. Gerig, and O. Kübler, “Parametrization of closed surfaces for 3-D shape description,”
Computer Vision, Graphics, Image Processing: Image Understanding 61, pp. 154–170, 1995.
14. R. Davies, C. Twining, T. Cootes, J. Waterton, and C. Taylor, “A minimum description length approach
to statistical shape modeling,” IEEE Transactions on Medical Imaging 21, pp. 525–537, May 2002.
15. A. Kotcheff and C.J.Taylor, “Automatic construction of eigenshape models by direct optimization.,” Medical
Image Analysis, pp. 303–314, 1998.
Evaluation of 3D Correspondence Methods for
Model Building
Martin A. Styner1 , Kumar T. Rajamani1 , Lutz-Peter Nolte1 , Gabriel Zsemlye2

and, Gábor Székely2 , Chris J. Taylor3 , and Rhodri H. Davies3
1
M.E. Müller Institute for Surgical Technology and Biomechanics, University of
Bern, P.O.Box 8354, 3001 Bern, Switzerland Martin.Styner@MEMcenter.unibe.ch
2
Computer Vision Lab, Gloriastrasse 35, ETH-Zentrum, 8092 Zürich, Switzerland
3
Division of Imaging Science and Biomedical Engineering, Stopford Building,
Oxford Road, University of Manchester, Manchester, M13 9PT, UK ?
Abstract. The correspondence problem is of high relevance in the con-

struction and use of statistical models. Statistical models are used for a
variety of medical application, e.g. segmentation, registration and shape
analysis. In this paper, we present comparative studies in three anatom-
ical structures of four different correspondence establishing methods.
The goal in all of the presented studies is a model-based application.
We have analyzed both the direct correspondence via manually selected
landmarks as well as the properties of the model implied by the corre-
spondences, in regard to compactness, generalization and specificity. The
studied methods include a manually initialized subdivision surface (MSS)
method and three automatic methods that optimize the object param-
eterization: SPHARM, MDL and the covariance determinant (DetCov)
method. In all studies, DetCov and MDL showed very similar results.
The model properties of DetCov and MDL were better than SPHARM
and MSS. The results suggest that for modeling purposes the best of the
studied correspondence method are MDL and DetCov.
1 Introduction
Statistical models of shape show considerable promise as a basis for segmenting,

analyzing and interpreting anatomical objects from medical datasets [5, 14]. The
basic idea in model building is to establish, from a training set, the pattern of
legal variation in the shapes and spatial relationships of structures for a given
class of images. Statistical analysis is used to give a parameterization of this
variability, providing an appropriate representation of shape and allowing shape
?
We are thankful to C. Brechbühler for the SPHARM software and to G. Gerig for
support and insightful discussions. D. Jones and D. Weinberger at NIMH (Bethesda,
MD) provided the MRI ventricle data. J. Lieberman and the neuro-image analysis
lab at UNC Chapel Hill provided the ventricle segmentations. This research was
partially funded by the Swiss National Centers of Competence in Research CO-
ME (Computer assisted and image guided medical interventions). The femoral head
datasets were provided within CO-ME by F. Langlotz.
2
constraints to be applied. A key step in building a model involves establish-

ing a dense correspondence between shape boundaries over a reasonably large
set of training images. It is important to establish the correct correspondences,
otherwise an inefficient parameterization of shape will be determined. The im-
portance of the correct correspondence is even more evident in shape analysis, as
new knowledge and understanding related to diseases and normal development is
extracted based on the established correspondence [10, 21]. Unfortunately there
is no generally accepted definition for anatomically meaningful correspondence.
It is thus difficult to judge the correctness of an established correspondence.
In 2D, correspondence is often established using manually determined land-
marks [1], but this is a time-consuming, error-prone and subjective process. In
principle, the method extends to 3D, but in practice, due to very small sets of
reliably identifiable landmarks, manual landmarking becomes impractical. Most
automated approaches posed the correspondence problem as that of defining a
parameterization for each of objects in the training set, assuming correspondence
between equivalently parameterized points. In this paper we compare methods
introduced by Brechbühler[2], Kotcheff[16] and - [8]. A fourth method is based
on manually initialized subdivision surfaces similar to Wang[24]. These meth-
ods are presented in more detail in sections 2.1-2.4. Similar approaches have also
been proposed e.g. Hill[11] and Meier [18]. Christensen[4], Szeliski[22] and Rueck-
ert[20] describe conceptionally different methods for warping the space in which
the shapes are embedded. Models can then be built from the resulting defor-
mation field [13, 9, 20]. Brett[3], Rangarajan[19] and Tagare[23] proposed shape
features (e.g. regions of high curvature) to establish point correspondences.
In the remainder of the paper, we first present the studied correspondence
methods and the measures representing the goodness of correspondence in order
to compare the methods. In the result section we provide the qualitative and
quantitative results of the methods applied on three populations of anatomical
objects (left femoral head, left lateral ventricle and right lateral ventricle).
2 Methods
Alignment - As a prerequisite for any shape modeling, objects have to be nor-
malized with respect to a reference coordinate frame. A normalization is needed
to eliminate differences across objects that are due to rotation and translation.
This normalization is achieved in studies based on the SPHARM correspon-
dence (section 2.2) using the Procrustes alignment method without scaling. In
the study based on the MSS correspondence (section 2.1) the alignment was
achieved using manually selected anatomical landmarks. MDL and DetCov can
align the object via direct pose optimization, an option not used in this paper.
Principal Component Analysis (PCA) model computation - A training
population of n objects described by individual vectors xi can be modeled by
a multivariate Gaussian distribution. Principal Component Analysis (PCA) is
performed to define axes that are aligned with the principal directions. First the
mean vector x̄ and the covariance matrix D are computed from the set of object
3
Fig. 1. Left: Visualization of the left and right lateral ventricle in a transparent human
head. Right: The manually selected landmarks on the left ventricle template.
vectors(1). The sorted eigenvalues λi and eigenvectors pi of the covariance matrix

are the principal directions spanning a shape space with x̄ at its origin. Objects
xj in the shape space are described as a linear combination
√ of the
√ eigenvectors
on x̄ (2). The shape space here is defined within [−3 · λi . . . 3 · λi ].
n n
1X 1 X
x̄ = xi ; D = (xi − x̄) · (xi − x̄)T (1)
n 1 n−1 1
P = {pi }; D · pi = λi pi ; xj = x̄ + P · b (2)
2.1 MSS: Manually initialized subdivision surfaces

This method is the only semi-automatic one, all others are fully automatic.
The correspondence starts from a set of predefined anatomical landmarks and
anatomically meaningful curves determined on the segmented objects using a in-
teractive display (e.g. spline on the crista intertrochanterica). After a systematic
discretization of the higher dimensional landmarks, a sparsely sampled point set
results, which is triangulated in a standardized manner and further refined via
subdivision surfaces. The correspondence on the 0th level meshes is thus given
by the manually placed control curves, on the subsequent levels by the subdi-
vision rule: the triangles are split to four smaller ones, the new vertices are the
midpoints of the pseudo-shortest path between the parent vertices. This path is
the projection of the edges connecting in three-space the parent vertices to the
original surface. The direction of the projection is determined by the normals of
the neighboring triangles. This method was successfully applied on organs with
a small numbers of anatomical point-landmarks.
2.2 SPHARM: Uniform area parameterization aligned to first order

ellipsoid
The SPHARM description was introduced by Brechbühler[2] and is a paramet-
ric surface description that can only represent objects of spherical topology. The
spherical parameterization is computed via optimizing an equal area mapping of
4
the 3D quadrilateral voxel mesh onto the sphere and minimizing angular distor-
tions [2]. The basis functions of the parameterized surface are spherical harmon-
ics. SPHARM can be used to express shape deformations [15], and is a smooth,
fine-scale shape representation, given a sufficiently small approximation error.
Based on an uniform icosahedron-subdivision of the spherical parameterization,
we obtain a Point Distribution Model (PDM) directly from the coefficients via
linear mapping [15]. The correspondence of SPHARM is determined by aligning
the parameterization so that the ridges of the first order ellipsoid coincide. It is
evident that the correspondence of objects with rotational symmetry in the first
order ellipsoid is ambiguously defined.
2.3 DetCov: Determinant of the covariance matrix

Kotcheff et al [16] and later - [6] propose to use an optimization process that
assigns the best correspondence across all objects of the training population, in
contrast to MSS and SPHARM, which assign inherently a correspondence to
each individual object. This view is based on the assumption that the correct
correspondences are, by definition, those that build the optimal model given
the training population. For that purpose they proposed to use the determinant
of the covariance matrix as an objective function. The disadvantages of the
original implementation was the computationally expensive genetic optimization
algorithm, and the lack of a re-parameterization scheme. The implementation in
this paper is different and is based on the optimization method by - [6], which
efficiently optimizes the parameterization of the objects. This same optimization
scheme was also used for the MDL criterion described in the next section. DetCov
has the property to minimize the covariance matrix and so explicitly favors
compact models.
2.4 MDL: Minimum Description Length

- [6, 8] built on the idea of the DetCov method, but proposed a different objective
function for the optimization process using on the Minimum Description Length
(MDL) principle. The DetCov criterion can be viewed as a simplification of the
MDL criterion. The MDL principle is based on the idea of transmitting a dataset
as an encoded message, where the code originates from some pre-arranged set
of parametric statistical models. The full transmission then has to include not
only the encoded data values, but also the coded model parameters. Thus MDL
balances the model complexity, expressed in terms of the cost of transmitting
the model parameters, against the quality of fit between the model and the
data, expressed in terms of the coding length. The MDL objective function has
similarities to the one used by DetCov [6]. The MDL computations for all our
studies were initialized with the final position of the DetCov method.
2.5 Measures of Correspondence Quality

In this section we present the measures of the goodness of correspondence used
in this paper. Such measures are quite difficult to define, since there is no gen-
5
eral agreement on a mathematical definition of correspondence. All methods in

this paper produce correspondences that are fully continuous and have an inher-
ent description of connectivity without any self-crossings. Measures of goodness
evaluating a method’s completeness and continuity (e.g. as suggested in Meier
et al. [18]) are thus not applicable here.
We propose the use of four different measures, each biasing the analysis to its
viewpoint on what constitutes correct correspondence. The first goodness mea-
sure is directly computed on the corresponding points as differences to manually
selected anatomical landmarks. The three other ones are of indirect nature, since
they are computed using the PCA model based on the correspondence. Further
details not discussed in this paper about the following methods can be found in
[7]. These three model based methods are in brief:
– Generalization: The ability to describe instances outside of the training set.
– Compactness: The ability to use a minimal set of parameters.
– Specificity: The ability to represent only valid instances of the object.
Distance to manual landmarks as gold standard - In medical imaging

human expert knowledge is often used as a substitute for a gold standard, since
ground truth is only known for synthetic and phantom data, but not for the
actual images. In the evaluation of correspondence methods this becomes even
more evident, because the goal is not clearly defined, in contrast to other tasks
such as the segmentation of anatomical structures. We propose to use a small set
of anatomical landmarks selected manually by a human expert on each object
as a comparative evaluation basis. We computed the mean absolute distance
(MAD) between the manual landmarks and each method’s points corresponding
to the same landmarks in a template structure. For comparison, we report the
reproducibility error of the landmark selection.
Model compactness - A compact model is one that has as little variance as

possible and requires as few parameters as possible to define an instance. This
suggests that the the
PMcompactness ability can be determined as the cumulative
variance C(M ) = i=1 λi , where λi is the ith eigenvalue. C(M ) is measured as
a function of the number of shape parameters M . The p standard error of C(M )
PM
is determined from training set size ns : σC(M ) = i=1 2/ns λi
Model generalization - The generalization ability of a model measures its

capability to represent unseen instances of the object class. This is a fundamental
property as it allows a model to learn the characteristics of an object class
from a limited training set. If a model is overfitted to the training set, it will
be unable to generalize to unseen examples. The generalization ability of each
model is measured using leave-one-out reconstruction. A model is built using all
but one member of the training set and then fitted to the excluded example.
The accuracy to which the model can describe the unseen example is measured.
The generalization ability is then defined as the approximation error (MAD)
averaged over the complete set of trials. It is measured as a function G(M ) of
the number of shape parameters M used in the reconstruction. Its standard error
6
SPHARM DetCov MDL
Fig. 2. Visualization of the correspondences of a set of landmarks from the template

(see Figure 1) in three selected objects from the two ventricle populations using the
different methods. The manually determined landmarks are shown as star-symbols and
the SPHARM, DetCov and MDL corresponding locations are shown as spheres.
σG(M ) is derived from √

the sampling standard deviation σ and the training set
size ns as: σG(M ) = σ/ ns − 1
Model specificity - A specific model should only generate instances of the

object class that are similar to those in the training set. It is useful to assess
this qualitatively by generating a population of instances using the model and
comparing them to the members of the training set. We define the quantitative
measure of specificity S(M ) (again as a function of M ) as the average distance
of uniformly distributed, randomly generated objects in the model shape space
to their nearest √
member in the training set. The standard error of S(M ) is given
by: σS(M ) = σ/ N , where σ is the sample standard deviation of S(M ) and N is
the number of random samples (N was chosen 100 000 in our experiments). The
distance between two objects is computed using the MAD.
A minimal model specificity is important in cases when newly generated

objects need to be correct, e.g. for model based deformation or shape prediction.
Model specificity is of lesser importance in the case of shape analysis since no
new objects are generated.
7
Left ventricle Right ventricle

Compactness C(M ) Compactness C(M )
4 Left Ventricle Compactness 4 Right Ventricle Compactness
x 10 x 10
5.5 5.5
5 5
4.5 4.5
4 4
3.5 3.5
3 3
C (M)
C (M)
2.5 2.5
2 2
1.5 1.5
SPHARM SPHARM
1 1
DetCov DetCov
0.5 MDL 0.5 MDL
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
M M
Generalization G(M ) Generalization G(M )

Left Ventricle Generalisation Ability Right Ventricle Generalisation Ability
2.5
2.5
SPHARM
DetCov SPHARM
MDL DetCov
2 MDL
G (M)
G (M)
1.5
1.5
1
1
2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20
M M
Specificity S(M ) Specificity S(M )

Left Ventricle Specificity Right Ventricle Specificity
2
1.95 2.05
2
1.9
1.95
1.85
1.9
1.8
1.85
S (M)
S (M)
1.75
1.8
1.7
1.75
1.65
SPHARM 1.7
SPHARM
1.6 DetCov 1.65 DetCov
MDL MDL
1.55 1.6
1.55
1 2 3 4 5 6 7 8 9 10 11 2 4 6 8 10 12
M M
Fig. 3. Table with errors graphs of compactness (C(M )), generalization (G(M )) and
specificity (S(M )) for the two ventricle studies (left column: left lateral ventricle, right
column: right lateral ventricle). The plot view is zoomed to a M value below 30, since
for higher M the plot values did not change.
8
Method Left ventricle Right ventricle

Mean Max Min Mean Max Min
SPHARM 4.47 mm 6.57 mm 1.72 mm 4.32 mm 6.70 mm 1.11 mm
DetCov 4.00 mm 6.16 mm 1.50 mm 4.28 mm 6.69 mm 1.10 mm
MDL 4.00 mm 6.15 mm 1.48 mm 4.28 mm 6.68 mm 1.10 mm
Table 1. Table with mean, maximal and minimal MAD between the manual landmarks
and the studied methods for the ventricle studies. It is clearly visible that there is little
change between DetCov and MDL. On the left side, DetCov and MDL have better
results than SPHARM. For comparison, the mean landmark selection error was 1.9mm.
3 Results on 3D anatomical structures
In the following sections we present the results of the application of the studied
correspondence methods to 3 different population: a left femoral head popula-
tion of 36 subjects, and a left and a right lateral ventricle population of each 58
subjects. The application of the model constructed from the femoral head popu-
lations is the femoral model-based segmentation from CT for patients undergoing
total hip replacement. The application of the two ventricle populations is shape
analysis for finding population differences in schizophrenia. In this document,
we focus only on the correspondence issue. It is noteworthy that the studied
populations are comprised not only of healthy subjects, but also of patients with
pathologically shaped objects.
3.1 Lateral ventricles
This section describes the studies of the left and the right lateral ventricle struc-
ture (see Figure 1) in a population of 58 subjects. The segmentation was per-
formed with a single gradient-echo MRI automatic brain tissue classification [17].
Postprocessing with 3D connectivity, morphological closing and minor manual
editing provided simply connected 3D objects. The manual landmarks were se-
lected by an expert with an average error of 1.9mm per landmark.
In Figure 2 the results of the correspondence methods in three exemplary
cases are shown. The first row shows the good correspondence with the manual
landmarks, as it is seen in the majority of the objects in this study. The second
row shows the frequent case in the remaining objects, in which all three meth-
ods have a rather large difference to the manual landmarks. In most cases of
disagreement with the manual landmarks, all methods produced similar results.
The last row shows the rare case in which SPHARM is clearly further away from
the landmarks than DetCov and MDL. The opposite case was not observed.
Table 1 displays the landmark errors and Figure 3 displays the error plots,
which both suggest that DetCov and MDL produce very similar results. Both
show smaller errors than SPHARM.
9
Fig. 4. Visualization of bad alignment in the first femoral head study on 3 example
objects seen from the same viewpoint. A large rotational alignment error is clearly
visible for a rotation around the long axis of the first order ellipsoid, which is close to
the femoral neck-axis.
Fig. 5. Left: Display in MSS tool with a single femur head object and manually placed
anatomical curves (anterior viewpoint). Right: Visualization of the femoral head tem-
plate (posterior viewpoint) and four of its anatomical landmarks (Fovea, center lesser
trochanter, tip greater trochanter).
3.2 Femoral head
This section describes the results on a population of objects from the head region
of the femoral bone. The segmentations were performed from CT-images with a
semi-automated slice-by-slice explicit snake algorithm [12]. The correspondence
study was done in 2 steps. Initially we only computed SPHARM, DetCov and
MDL on all available cases (30 total). We realized that the SPHARM corre-
spondence was not appropriate, so the results of the following computation were
meaningless (discussed further down). In a second step, we selected only those
cases, which contained the lesser trochanter in the dataset. For these cases (16)
we then computed MSS, SPHARM, DetCov and MDL.
The first study was based on the full 30 cases including 14 datasets with
missing data below the calcar. The distal cut of the femoral bone was performed
through the calcar perpendicular to the bone axis. The alignment was performed
using the Procrustes alignment based on the SPHARM correspondence. MSS was
not computed in this case. We observed a bad alignment due to the SPHARM
correspondence. In Figure 4 we visualize this inappropriate alignment in three
cases. As a consequence the DetCov and MDL results were inappropriate. The
10
Compactness C(M ) Generalization G(M )

Pelvis Generalisation Ability
5 Pelvis Compactness
x 10
3 8
MSS
SPHARM
2.5
DetCov
7
MDL
2
6
G (M)
C (M)
1.5
5
1
4
MSS
SPHARM
0.5
DetCov
MDL 3
0
0 2 4 6 8 10 12 14 2 4 6 8 10 12
M M
Specificity S(M )
Pelvis Specificity
5.8
5.6 Landmark error table

Method Femoral head
5.4
Mean Max Min
5.2 MSS 3.13 mm 6.48 mm 1.73 mm
S (M)
5
SPHARM 7.24 mm 12.2 mm 6.07 mm
DetCov 3.24 mm 7.36 mm 1.71 mm
4.8
MDL 3.40 mm 6.41 mm 1.71 mm
MSS
4.6 SPHARM
DetCov
MDL
2 4 6 8 10 12
M
Fig. 6. Top row, Bottom row left: Table with Errors graphs of compactness (C(M )),
generalization (G(M )) and specificity (S(M )) for the femoral head study. Bottom row
right: Table with mean, maximal and minimal MAD between the manual landmarks
and the studied methods for the femoral head study. There is little change between
DetCov and MDL. SPHARM shows clearly the worst results of all studied methods.
For comparison, the mean landmark selection error was 2.5mm.
bad SPHARM correspondence resulted from a rotational symmetry along the

long first order ellipsoid axis, which is close to the neck-axis. Due of the bad
correspondence, we do not present here the error analysis of these cases.
The second study was based on a subset of the original population comprising
only of those datasets that include also the lesser trochanter. The distal cut of
the femoral bone was performed by a plane defined using the lesser trochanter
center, major trochanter center and the intertrochanteric crest. For the MSS
method the anatomical landmarks for the subdivision surfaces were chosen as
follows: the fovea, the half-sphere approximating the femoral head, the circle
approximating the orthogonal cross-section of the femoral neck at its thinnest
location, the intertrochanteric crest and the lower end of the major trochanter.
The landmarks for the MSS alignment were the lesser trochanter, the femoral
11
head center, and the center of the circle approximating the neck at its smallest
perimeter. Each landmark was selected on the respective 3D femur model either
directly on the reconstructed bone surface or using 3D spherical primitives. The
manual landmarks for the comparison were selected by a different expert with
an average error of 2.5mm per landmark. The landmarks sets for the MSS and
the comparison were not exclusive, due to the scarceness of good landmarks.
All correspondences were based on the MSS alignment. We observed that
the SPHARM correspondence was visually better behaved in this study due
to the inclusion of the lesser trochanter, which eliminated some problems with
the rotational symmetry. However, figure 6 shows that the landmark errors for
SPHARM alignment are clearly the worst of the studied methods. MSS shows
the best average agreement with the manual landmarks, which is not surprising
since the landmarks contained points used also to construct MSS. MDL was
surprisingly better than both DetCov and MSS in regard to the minimal and
maximal MAD, although the MAD differences are rather small. In figure 6 it is
clearly visible that MDL and DetCov have similar and better modeling properties
than SPHARM and MSS. Only for G(M ) MSS is better than SPHARM.
4 Conclusions
In this paper, we have presented a comparison of the SPHARM, DetCov, MDL

and MSS correspondence methods in three populations of anatomical objects.
The goal in all of the presented studies is a model-based application. We have
analyzed both the direct correspondence via manually selected landmarks as
well as the properties of the model implied by the correspondences, in regard to
compactness, generalization and specificity.
The results for SPHARM of the first femoral head study revealed that in case
of rotational symmetry in the first order ellipsoid, independent of the higher order
terms, the correspondence is inappropriate. Since correspondence and alignment
are dependent on each other, such a bad correspondence cannot be significantly
improved using methods like MDL and DetCov. In all studies, DetCov and MDL
showed very similar results. The model properties of DetCov and MDL were
better than both SPHARM and MSS. The findings suggest that for modeling
purposes the best of the studied correspondence method are MDL and DetCov.
The manual landmark errors are surprisingly large for all methods, even
for the MSS method, which is based on landmarks. This finding is due to the
high variability for the definition of anatomical landmarks definition by human
experts, which is usually in the range of a few millimeters.
In the lateral ventricle studies we plan to do the following shape analysis
with the model built on the MDL correspondence. Other current research in our
labs suggest that the shape analysis could gain statistical significance by using
MDL rather than SPHARM. In the femoral head study, we plan to use the MDL
model for shape prediction in the shape space. The results of the specificity error
is in this study very relevant, since it is desired to generate ’anatomically correct’
objects from the shape space.
12
References
1. Bookstein, F.L.: Morphometric Tools for Landmark Data: Geometry and Biology,
Cambridge University Press (1991)
2. Brechbühler, C., Gerig, G., Kübler, O.: Parameterization of Closed Surfaces for 3-D
Shape Description. Comp. Vision and Image Under. 61 (1995) 154–170
3. Brett, A.D., Taylor, C.J.: Construction of 3D Shape Models of Femoral Articular
Cartilage Using Harmonic Maps. MICCAI (2000) 1205–1214
4. Christensen, G., Joshi,S., Miller, M.: Volumetric Transformation of Brain Anatomy.
IEEE Trans. Med. Imag., 16 6 (1997) 864–877
5. Cootes, T, Hill, A., Taylor, C.J., Haslam, J.: The Use of Active Shape Models for
Locating Structures in Medical Images. Img. Vis. Comp. 12 (1994) 355-366
6. Davies, Rh.H, Twining, C.J., Cootes, T.F., Waterton, J. C., Taylor, C.J.: 3D Statis-
tical Shape Models Using Direct Optimization of Description Length. ECCV (2002)
I.
7. Davies, Rh.H: Learning Shape: Optimal Models for Analysing Natural Variability.
Dissertation University of Manchester, 2002.
8. Davies, Rh.H, Twining, C.J., Cootes, T.F., Waterton, J. C., Taylor, C.J.: A Mini-
mum Description Length Approach to Statistical Shape Model. IEEE TMI 21 (2002)
9. Fleute, M., Lavallee, S.: Building a Complete Surface Model from Sparse Data Using
Statistical Shape Models, MICCAI (1998) 879-887
10. Gerig, G., Styner, M.: Shape versus Size: Improved Understanding of the Morphol-
ogy of Brain Structures, MICCAI (2001) 24-32
11. Hill, A., Thornham ,A., Taylor ,C.J.: Model-Based Interpretation of 3D Medical
Images. Brit. Mach. Vision Conf. BMCV, (1993) 339-348
12. Hug, J., Brechbühler, C., Székely, G.: Tamed Snake: A Particle System for Robust
Semi-automatic Segmentation. MICCAI (1999) 106–115
13. Joshi, Banerjee, Christensen, Csernansky, Haller, Miller, Wang:Gaussian Random
Fields on Sub-Manifolds for Characterizing Brain Surfaces. IPMI (1997) 381–386
14. McInerney, T., Terzopoulos, D.: Deformable Models in Medical Image Analysis: A
Survey. Med. Image Analysis 1 2 (1996) 91–108
15. Kelemen, A., Székely, G., Gerig, G.: Elastic Model-Based Segmentation of 3D
Neuroradiological Data Sets. IEEE Trans. Med. Imag. 18 (1999) 828–839
16. Kotcheff, A.C.W., Taylor, C.J.: Automatic Construction of Eigenshape Models by
Direct Optimization. Med. Image Analysis 2 4 (1998) 303–314
17. Van Leemput, K., Maes, F., Vandermeulen, D., Suetens, P.: Automated Model-
based Tissue Classication of MR Images of the Brain, IEEE TMI 18 (1999) 897–908
18. Meier, D., Fisher, E.: Parameter Space Warping: Shape-Based Correspondence
Between Morphologically Different Objects. Trans. Med. Imag. 12 (2002) 31–47
19. Rangarajan, A., Chui, H., Bookstein, F.L.: The Softassign Procrustes Matching
Algorithm, IPMI (1997) 29–42
20. Rueckert, D., Frangi, A.F., Schnabel, J.A.: Automatic Construction of 3D Statis-
tical Deformation Models Using Non-rigid Registration. MICCAI (2001) 77–84
21. Styner, M., Gerig, G., Lieberman, J., Jones, D., Weinberger, D.: Statistical Shape
Analysis of Neuroanatomical Structures Based on Medial Models. Med. Image Anal.
22. Szeliski, R., Lavallee, S.: Matching 3-D anatomical surfaces with non rigid defor-
mations using octree-splines, Int. J. Computer Vision, 18 2 (1996) 290–200
23. Tagare, H.: Shape-Based Nonrigid Correspondence with Application to Heart Mo-
tion Analysis. IEEE Trans. Med. Imag. 18 7 (1999) 570–580
24. Wang, Y., Peterson B. S., Staib, L. H.,: Shape-based 3D Surface Correspondence
Using Geodesics and Local Geometry. CVPR 2 (2000) 644-651

DR Kumar Rajamani Thesis Final

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

DR Kumar Rajamani Thesis Final

Hochgeladen von

Copyright:

Verfügbare Formate

Aus dem MEM Forschungszentrum,

Institut für Chirurgische Technologien und Biomechanik

Three Dimensional Surface Extrapolation from Sparse Data using

Inaugural-Dissertation zur Erlangung der Doktorwürde der Philosophie im Fach

Der Dekan der Medizinischen Fakultät:

Dissertation submitted to the MEM Research Centre

Dr. Miguel Angel Gonzalez Ballester,

Title of Dissertation: STATISTICAL DEFORMABLE MODELS

FOR SURFACE EXTRAPOLATION

FROM SPARSE DATA

Kumar T. Rajamani, Doctor of Philosophy, 2006

Dissertation directed by: Dr. Miguel Angel Gonzalez Ballester,

A majority of pre-operative planning and navigational guidance during com-

I would first like to express my sincere gratitude to Prof. Lutz-Peter Nolte

List of Tables xii

List of Figures xiv

2 Anatomical Shape Reconstruction from Sparse Data 7

3 Conclusions and Outlook 19

1.1 Planning module for performing tumour biopsies in neurosurgery . . 2

1.2 Basic components of a computer assisted surgery system, including

1.3 Figure illustrating the concept of correspondence establishment. Three

Anatomical Shape Reconstruction

2.1 Statistical Model Generation

Method Femoral head

5 DetCov 3.24 mm 7.36 mm 1.71 mm

2.2 Model Deformation by Progressive elimina-

2.3 Model Deformation by Minimization of Ma-

Reconstruction Error [mm]

Figure 2.5: Reconstruction errors of 10 different femurs using leave-one-out exper-

• In conjunction with the pointer-based study, we also evaluated the use of

2.5 Ultrasound-Initialized Deformable Bone Mod-

2.6 Kernel based shape deformation

Conclusions and Outlook

In this thesis we studied shape reconstruction techniques to extrapolate the three-

• It would also be interesting to study whether the attributes of the patient,

• The training population of bone shapes should be expanded to build more

The following papers are included in this thesis

Publicaion 1, Medical Image Analysis 2006 (accepted), Statistical Deformable

Bone Models for Robust 3D Surface Extrapolation from Sparse Data

Publication 2, published in Computer Aided Surgery 2005, A Comparison

Study Assessing the Feasibility of Ultrasound-Initialized Deformable Bone Model

Publication 3, IEEE Transactions on Biomedical Engineering (submitted) Ac-

curate and Robust Reconstruction of Proximal Femur from Sparse Intraoperative

Data and Dense Point Distribution Model for Surgical Navigation

Publication 4, published as a chapter in MICCAI 2004, (Springer-Verlag,

LNCS 3217, pp. 478-485) A Novel Approach to Anatomical Structure Morphing

for Intra-operative Visualization

Publication 5, published in SPIE Medical Imaging 2004, Bone Morphing with

statistical shape models for enhanced visualization

Publication 6, published in IPMI 2003, (Springer-Verlag, LNCS 2732, pp.

63-75) Evaluation of 3D Correspondence Methods for Model Building

2. A Comparison Study Assessing the Feasibility of Ultrasound-Initialized De-

3. Accurate and Robust Reconstruction of Proximal Femur from Sparse Intra-

4.1.2 Conference Articles

2. A robust and accurate apporach for reconstructing patient specific 3D bone

3. Feasibility of 3D Ultrsound initialized deformable bone modeling

5. An optimal three-stage method for anatomical shape reconstruction from

6. Use of a Dense Surface Point Distribution Model in a Three-Stage Anatom-

7. Evaluation and Initial Validation Studies of Anatomical Structure Morphing

8. Kernel Regularized Bone Surface Reconstruction from Partial Data Using

9. Validation studies of anatomical structure morphing

10. Feasibility of ultrasound-initialized bone morphing: early experiences and

13. A Novel approach to bone morphing using statistical shape models