Beruflich Dokumente
Kultur Dokumente
Arbeit unter der Leitung von Dr. Miguel Angel Gonzalez Ballester,
Prof. Dr.-Ing. L.-P. Nolte,
Prof. Martin Styner
Vorgelegt von:
Kumar T. Rajamani
von Indian
Von der Medizinischen Fakultät der Universität Bern auf Antrag der Dissertations-
kommission als Dissertation genehmigt.
Promotionsdatum:
by
Kumar T. Rajamani
Advisory Committee:
To my MOTHER
v
vi
ACKNOWLEDGEMENTS
vii
viii
TABLE OF CONTENTS
Acronyms xi
1 Introduction 1
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
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
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.
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
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.
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)).
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
0
Femur 1 Femur 2 Femur 3 Femur 4 Femur 5 Femur 6 Femur 7 Femur 8 Femur 9 Femur 10
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).
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.
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.
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
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.
• 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
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).
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.
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
SPIE Medical Imaging 2005, San Diego, USA.
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.
15. Bone Morphing with statistical shape models for enhanced visualization
K. T. Rajamani, L. P. Nolte, M. Styner
SPIE Medical Imaging 2004, San Diego, USA.
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.
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
Data
Abstract
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
without the need for a pre or intra-operative imaging. The 3D model is reconstructed
digitized landmarks and surface points that are obtained intra-operatively. The
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
enables us to effectively reject outliers and compute stable models. We present here
Key words: Medical Image Analysis, Shape Analysis, Statistical Shape Model,
1 Introduction
dures prior to working with the patient directly. This could be used for effec-
the disadvantage that they are expensive and/or induce high radiation doses
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
Statistical shape analysis (Dryden and Mardia (1998), Kendall (1989), Small
are popular (Kelemen et al. (1999), Turk and Pentland (1991), Cootes et al.
surface points. Thus the aim of statistical shape model fitting is to extrapolate
Surgery (MIS), largely due to the operating theater setup. Statistical modeling
puterized models may be directly injected into the scene, enabling enhanced
visualization.
models has been explored by several scientists. Fleute and Lavallée (1998)
fit the deformable model surface to intra-operatively digitized point data via
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
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.
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
the incorporation of the complete set of eigenvectors for the shape estima-
tion. This scheme was then improved to have better convergence behaviour
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-
4
based weighting enables for a smart estimation mechanism that is robust to
outliers.
can, due to limited surgical access, be difficult to acquire a set of points that
struction using principal components and outlines the method we chose for
tic and dry cadaver bones and finally application of our deformation algorithm
The first step is to build a statistical shape model from a training database.
5
Bookstein (Bookstein, 1986) uses landmarks to capture the important geo-
metric features. The active shape model (ASM) of Cootes and Taylor (Cootes
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.
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-
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
nation of the basis vectors results in a shape that is similar to the existing
ones.
between shape boundaries over a reasonably large set of training data. In 2D,
6
correspondence is often established using manually determined landmarks, but
method extends to 3D, but in practice, due to very small sets of reliably identi-
terization for each of the objects in the training set, assuming correspondence
(DetCov), Davies et al. (2002b) (MDL) and a fourth method based on man-
that for modeling purposes the best among the correspondence methods was
study, for our model building, correspondence was initialized using MSS and
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-
the mean vector ~x̄ and the covariance matrix D are computed from the set
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
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
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
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.
term enables stable prediction with minimal number of known surface points.
Where Fleute (Fleute and Lavallée, 1998) and Chan (Chan et al., 2003) con-
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
9
rated via closest distance correspondence.
( )
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
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
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
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
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
rigid threshold to reject the outliers we decided to employ a smarter and ro-
based on its distance from the closest point in the current estimated shape).
there are no more significant changes in the weights. This feature was incorpo-
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
We briefly explain here our solution strategy, where we formulate the problem
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
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)
~ ).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
solution strategy enables real time estimation of shape parameters hence facil-
13
1.4
1.2
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
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)
triangle mesh list containing 4098 vertices. The triangle mesh was the direct
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
The CT data sets in our database were segmented, and surface models of the
between points on the surface of the bones in the training database was initial-
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
shape, making our method size invariant. The remaining points are added uni-
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.
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-
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
estimate with mean error of 1.5mm obtained with as few as 6 digitized points
(Aspert et al., 2002). The second correspondence scheme does not use the im-
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
of our method and showed that we can extrapolate a three dimensional shape
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
to get a much better estimate. Our formulation makes ρ a factor of the num-
points are digitized. To evaluate the influence of the ρ factor we studied the
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
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
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
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
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
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
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
Nine different dry cadaver femur bones were chosen for this validation study.
models were generated. The experiment trials were carried out in the CT
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
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
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
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
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-
istration error of 0.2 mm for this experiment. The registered surface models
To initialize as well as provide surface points for our bone deformation method
20
Cadaver Mean error [mm] w.r.t. # of points
3 27 54 3 27 54
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
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
21
Statistics Cumulated From 9 Cadaver Bones
2.4
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
0.8 seconds of computation for each ultrasound image frame and has a mean
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-
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
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
5 Discussion
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
a Mahanalobis weighted least squares error minimization, and the novel so-
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
this case. Hence the ρ factor seems not to contribute much as was observed
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
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
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
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
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
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
safety, often reduced radiation exposure, and improved surgical reality through
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
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
USA for insightful discussions about modeling and shape prediction. The CT
26
datasets were provided by Frank Langlotz from MEM Research Center,Bern
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
(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-
Brechbühler, C., Gerig, G., Kübler, O., 1995. Parametrization of closed sur-
faces for 3-D shape description. Computer Vision, Graphics, Image Process-
Cootes, T., Taylor, C., Cooper, D., Graham, J., 1995. Active shape models -
61, 38–59.
Cootes, T. F., Edwards, G. J., Taylor, C. J., 1998. Active appearance models.
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
Davies, R. H., Twining, C. J., Cootes, T. F., Waterton, J. C., Taylor, C. J.,
Davies, R. H., Twining, C. J., Cootes, T. F., Waterton, J. C., Taylor, C. J.,
Dryden, I., Mardia, K., 1998. Statistical Shape Analysis. John Wiley and Sons.
Fleute, M., Lavallée, S., 1998. Building a complete surface model from sparse
Gong, J., Bächler, R., Sati, M., Nolte, L.-P., 1997. Restricted surface matching:
Heitz, G., Rohlfing, T., Maurer, C., 2005. Statistical shape model generation
1411–1421.
Kelemen, A., Székely, G., Gerig, G., October 1999. Elastic model-based seg-
Kendall, D., 1989. A survey of the statistical theory of shape. Statistical Sci-
Kowal, J., Amstutz, C., Nolte, L., 2001. On b-mode ultrasound based regis-
28
CAOS.
Kowal, J., Amstutz, C., Nolte, L., 2003. On the development and comparative
Surgery, 107–119.
Livyatan, H., Yaniv, Z., Joskowicz, L., 2003. Gradient-based 2-d/3-d rigid reg-
22 (11), 1395–1406.
Rajamani, K., Gonzalez Ballester, M., Nolte, L., Styner, M., 2005. A novel
Rajamani, K. T., Joshi, S., Styner, M., 2004a. Bone model morphing for en-
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
Stindel, E., Briard, J., Merloz, P., Plaweski, S., Dubrana, F., Lefevre, C.,
Troccaz, J., 2002. Bone morphing: 3d morphological data for total knee
Styner, M., Brechbühler, C., Székely, G., Gerig, G., 2000. Parametric estimate
Styner, M. A., Rajamani, K. T., Nolte, L. P., Zsemlye, G., Szekely, G., Taylor,
29
model building. IPMI. pp. 63–75.
Turk, M., Pentland, A., 1991. Eigenfaces for recognition. Journal of Cognitive
Wang, Y., S., P. B., Staib, L. H., 2000. Shape-based 3d surface correspondence
30
A Comparison Study Assessing the Feasibility of
Ultrasound-Initialized Deformable Bone Models
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
2 Methods
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.
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
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.
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
2.5 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.
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.
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
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.
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
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.
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
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
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
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
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
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
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
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.
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
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
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.
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
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
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
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
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
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
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
one set of points, and the procedure is iterated until convergence. Normally, when trying to register a set
needs to be computed. However, as the mean shape in our case is described by a dense point surface, a
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
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
(P − H) PH × GH z×x
x= , z= , y= ; R1 = [x y z ] (4)
|| ( P − H ) || || PH × GH || || 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:
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
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
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
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
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:
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
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
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 → ∞ ,
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
Another advantage of using TPS is that the affine transformation A in eq. (12) is automatically
⎧ 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 )
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
L[t ] = γ T Kγ + θT Kθ + ω T Kω (15)
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
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
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,
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
1
drmse (T ' , T ) =
| T' | ∫∫
v'∈T '
d( v ' , T ) 2 dT ' (18)
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
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.
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.
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
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
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
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
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. 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
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
M.E. Müller Institute for Surgical Technology and Biomechanics, University of Bern,
Switzerland,
kumar.rajamani@MEMcenter.unibe.ch,
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
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
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
( )
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
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
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
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.
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.
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.
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
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.
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
1 Introduction
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.
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.
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
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
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
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.
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.
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).
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
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
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.
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
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