Ultrasound in Med. & Biol., Vol. 34, No. 2, pp.

183–195, 2008
Copyright © 2008 World Federation for Ultrasound in Medicine & Biology
Printed in the USA. All rights reserved
0301-5629/08/$–see front matter

doi:10.1016/j.ultrasmedbio.2007.07.023

● Original Contribution

VOLUME SEGMENTATION AND RECONSTRUCTION FROM FREEHAND

THREE-DIMENSIONAL ULTRASOUND DATA WITH APPLICATION TO
OVARIAN FOLLICLE MEASUREMENT

MARK J. GOODING,* STEPHEN KENNEDY,† and J. ALISON NOBLE*

*Wolfson Medical Vision Laboratory, Dept. Engineering Science, University of Oxford, Parks Road, Oxford, UK;
and †Nuffield Dept. of Obstetrics and Gyneacology, John Radcliffe Hospital, Headington, Oxford, UK

Received 5 Mar 2007; revised 29 May 2007; in final form 25 Jul 2007.

Abstract—This article presents a semi-automatic method for segmentation and reconstruction of freehand
three-dimensional (3D) ultrasound data. The method incorporates a number of interesting features within the
level-set framework: First, segmentation is carried out using region competition, requiring multiple distinct and
competing regions to be encoded within the framework. This region competition uses a simple dot-product based
similarity measure to compare intensities within each region. In addition, segmentation and surface reconstruc-
tion is performed within the 3D domain to take advantage of the additional spatial information available. This
means that the method must interpolate the surface where there are gaps in the data, a feature common to
freehand 3D ultrasound reconstruction. Finally, although the level-set method is restricted to a voxel grid, no
assumption is made that the data being segmented will conform to this grid and may be segmented in its
world-reference position.The volume reconstruction method is demonstrated in vivo for the volume measurement
of ovarian follicles. The 3D reconstructions produce a lower error variance than the current clinical measure-
ment based on a mean diameter estimated from two-dimensional (2D) images. However, both the clinical
measurement and the semi-automatic method appear to underestimate the true follicular volume. (E-mail:
gooding@robots.ox.ac.uk) © 2008 World Federation for Ultrasound in Medicine & Biology.

Key Words: Three-dimensional, Freehand, Ovarian follicle, Volume reconstruction.

INTRODUCTION diameters is taken as the measure of follicular size. It is

Measurement of ovarian follicles
Measurement of ovarian follicles is a routine part of
many fertility treatments, necessary for correct treatment
management. A follicle is a fluid filled volume within the
ovary in which an oocyte develops. In fertility treat-
ments, the size of the follicle is used as an indicator of
oocyte maturity and, is, therefore, used to assess the
optimum time for aspiration for in vitro fertilisation
(IVF) or for inducing ovulation. It has also been shown
that follicular size is predictive of final IVF treatment
outcome (Saith et al. 1998).
Currently, the clinical measurement takes the form
of a mean diameter. This is typically assessed from a
single two-dimensional (2D) B-mode ultrasound image
where two diameters are measured approximately or-
thogonally as illustrated in Fig. 1. The mean of these
Address correspondence to: Dr. Mark J. Gooding, Wolfson Medical
Vision Laboratory, Dept. Engineering Science, University of Oxford,
Parks Road, Oxford, UK, OX1 3PJ. E-mail: gooding@robots.ox.ac.uk
this measurement which is then used within the treatment
management process. However, diameter is not represen-
tative of the true follicle size, its volume, for non-spher-
oid follicles (Penzias et al. 1994). Manual measurement
of the follicular volumes from three-dimensional (3D)
ultrasound have been shown to be more reproducible and
a better estimate of the true follicular volume than esti-
mates based on the mean diameter measurement (Kyei-
Mensah et al. 1996b; Forman et al. 1991; Kyei-Mensah
et al. 1996a). Unfortunately, manual measurement from
3D ultrasound is slow and still subject to observer error,
therefore, we wish to perform this measurement auto-
matically.
Most previous methods for the segmentation of
follicles from ultrasound data have only considered sin-
gle 2D B-mode images. Sarty et al. (1998) used graph-
searching techniques within a manually defined region-
of-interest to find the follicle walls within the B-mode
image. The user-defined region was made of two con-
centric circles, one within the follicle antrum and the

183
184 Ultrasound in Medicine and Biology
Fig. 1. An example ultrasound image of ovarian follicles. The
follicles appear as dark homogenous regions. The measurement
currently used in clinical practice is the mean diameter of each
follicle found from a single 2D B-mode ultrasound image. For
the follicle with measurements shown, the measurement re-
corded in the patient’s notes would be 16 mm.
other outside the follicle wall. The algorithm then as-
sumed that the best delineation of the inside wall of the
follicle is the path of highest intensity gradient within
this region, where there is low intensity inside and high
intensity outside. The outside edge of the follicle wall
was found in a similar fashion. Krivanek and Sonka
(1998) extended this by introducing a watershed segmen-
tation algorithm to find the follicles automatically, al-
though user interaction was still required to choose the
appropriate follicle for further delineation. Potoc̆nik and
Zazula (2002a) performed automated detection in single
ultrasound images by smoothing and thresholding to
detect dark homogeneous regions. Segmentation was
achieved by enlarging these regions using a region grow-
ing method, based on both intensity and intensity gradi-
ent, to find the follicle boundaries. Finally, rule-based
selection was used to remove false-positive detections.
The authors extended that work to consider sequences of
images using a Kalman filter to improve the detection
and segmentation over the sequence by predicting the
boundary position in each image from its neighbouring
images (Potoc̆nik and Zazula 2002b). Cigale et al. (2006)
use support vector machines to classify regions of an
ultrasound image as either being follicle or background.
Rather than generating an explicit segmentation results
are presented in terms of the percentage of follicles
detected.
Zimmer et al. (1996) and Muzzolini et al. (1993)
proposed generic ultrasound segmentation methods
which they demonstrate on 2D ultrasound images of
ovarian follicles. However, their lack of explicit interest
in the application mean that no quantitative validation of
segmentation accuracy is performed.
The work of ter Haar Romeny et al. (1999) is unique
in considering segmentation of follicles from 3D ultra-
sound, captured using a 3D probe. Within the 3D image,
Volume 34, Number 2, 2008
the centre of each follicle was found using the 3D topol-
ogy of the intensity information. From each of these
centre points, between 200 and 500 “rays” were sent
radially outwards until a significant edge was detected
using a multiscale one-dimensional (1D) filter. These
points were then used to produce a 3D reconstruction of
the follicle using spherical harmonic models. A limited
validation was performed but good agreement was found
with magnetic resonance imaging (MRI) based measure-
ments and true anatomy as comparators for three bovine
follicles.
Volume segmentation and reconstruction from freehand
3D ultrasound data
Despite the increasing availability of mechanical
and 2D phased-array 3D ultrasound machines, research
continues in the development of freehand 3D ultrasound
systems. This is partially as a result of availability of
freehand 3D systems within the research setting and the
need for integration with existing ultrasound machines in
the clinical setting. However, freehand ultrasound also
offers its own advantages over mechanical and phased-
array systems. These include the ability to image vol-
umes of arbitrary size, accommodating large organs, the
flexibility for the operator to select good views and to
increase data acquisition density in important areas and
the ability to compound data to reduce problems of
speckle, acoustic shadowing and other artifacts. A complete
review of freehand 3D ultrasound systems is beyond the
scope of this article and we confine ourselves to considering
only the state-of-the-art for segmentation and volume mea-
surement from freehand 3D ultrasound. Those interested in
a comprehensive study of freehand 3D ultrasound systems
should refer to (Fenster and Downey 1996; Rohling and
Gee 1996; Rohling et al. 1999; Prager et al. 2002; Gee et al.
2003; Mercier et al. 2005).
Generally, the process for object reconstruction
from freehand 3D ultrasound data is a process of seg-
mentation followed by surface fitting, as illustrated in
Fig. 2. Segmentation of each 2D ultrasound image is
carried out separately, whether manually or automati-
cally and 3D image and surface reconstruction proceeds
on this classified data. Volume measurement can be
made from the surface reconstruction or from knowledge
of the object area in each image and the image position.
Volume reconstruction methods following this approach
generally used manual segmentation [e.g., (Legget et al.
1998; Treece et al. 1999, 2000, 2001; Bogush and Tuz-
ikov 2005)] with some authors reporting use of semi-
automated or automated segmentation methods prior to
volume reconstruction [e.g., (Ye et al. 2002; Ahmad et
al. 2006)].
Manual segmentation of multiple images is a time
consuming task and is unlikely to be accepted as a
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al. 185

This was performed along the length of the vessel at

20 voxel spacing. A deformable model was then cre-
ated from the boundary locations and evolved with
terms attracted to gradient and to make a smooth
surface. Song et al. (2002) create simulated ultrasound
images from a surface model for each real (cardiac)
ultrasound image. The surface model is then evolved
to maximise the similarity between the simulated im-
ages and the real images. This approach is only ap-
propriate for organs for which a surface model can be
generated. Similarly, it is necessary to have a good
method for ultrasound simulation, which may result in
a high computational overhead to the method. Zhang
et al. (2002) use radial basis functions to interpolate
the pixel data in 3D space, without restricting the
reconstructed image to a voxel array. A simple iso-
surface segmentation and object reconstruction is per-
formed for ultrasound phantom data with high contrast
boundaries in the images.
We adopt this approach of performing image seg-
mentation and object reconstruction after image recon-
struction to exploit the additional spatial information
Fig. 2. Figure showing the process that most systems have used available to the follicle segmentation process. Following
to reconstruct objects. Segmentation is performed before image
(Zhang et al. 2002), we endeavour to keep the image data
reconstruction.
in a world reference space, rather than confining it to a
voxel array. However, we opt for a region-based ap-
proach to image segmentation.
clinical solution to the problem of volume measure-
ment for tasks with high patient throughput. Conflict-
ing boundary locations may exist, from both manual or
automated 2D segmentations, on to which the surface
must be interpolated. These conflicting boundaries are
treated equally since the boundary position provides
little information about the underlying image, al-
though potentially a confidence in boundary position
could be included.
In contrast to segmentation being performed in
2D, a number of authors have used the process shown
in Fig. 3, whereby segmentation occurs on the 3D
image after image reconstruction has been performed
from the freehand ultrasound data. In such an ap-
proach, image compounding and 3D spatial informa-
tion can aid object segmentation and surface interpo-
lation. For example, Barry et al. (1997) and Allott et
al. (1999) use a commercially available tool to seg-
ment freehand 3D ultrasound data in phantoms and in
rabbit cardiac data. The 3D ultrasound data is recon-
structed into a voxel based 3D image to enable seg-
mentation using this tool. Similarly, Zahalka and Fen-
ster (2001) first reconstructed the freehand ultrasound
image on a voxel array when segmenting images of the
carotid. From a user selected seed point within the Fig. 3. Figure showing a process that performs object recon-
carotid a 2D plane, rays are extended at 5° within the struction on compounded data. Segmentation is performed after
plane on which potential boundary points are located. image reconstruction.
186 Ultrasound in Medicine and Biology
Overview of this article
In this article, we present a method for the segmen-
tation and surface reconstruction of follicles from free-
hand 3D ultrasound, to allow volume measurement.
First, we present details of the algorithm used, which
incorporates segmentation and surface fitting into a sin-
gle step. Segmentation is performed using intensity in-
formation and a dot-product based similarity measure.
The ultrasound data is processed in the 3D domain to
take advantage of the additional spatial information com-
pared with 2D ultrasound but is not necessarily restricted
to a voxel array. However, freehand 3D ultrasound often
results in gaps in the data meaning that the surface must
be interpolated where there is not data to segment. Then,
details of the implementation are presented, including
how region-competition between neighbouring follicles
is achieved within the level-set framework. The Experi-
mental Methods and Results sections detail the valida-
tion of this method in a clinical setting in comparison
with manual measurement and aspirated follicle vol-
umes. The Discussion largely focuses on explanations
for measurement error. Finally, we conclude by consid-
ering possible directions for future research.
METHODS
Our segmentation and object reconstruction is
posed within the level set framework, a powerful method
which finds application in many fields including medical
image segmentation and object reconstruction (Sethian
1999). The main equation solved by the method is:
⭸⌽
⭸t ⱍ ⱍ
⫹ F ⵜ⌽ ⫽ 0 (1)
where the embedded function, ␾(t), is evolved over time
using a speed function, F, such that the zero level set, ␾
⫽ 0, at time T ⫽ ⬁ is the optimal solution for the
application of interest; in this case, the segmentation and
surface interpolation of freehand 3D ultrasound data.
This method has been applied to many different appli-
cations by varying the choice of the speed function F.
In this research, we opted to use freehand 3D ultra-
sound, primarily on account of equipment availability
but also because freehand 3D offers a number of advan-
tages such as integration with existing systems. How-
ever, freehand ultrasound does have the disadvantage
that the physical space may not be uniformly scanned
leading to gaps in the data acquired. Therefore, the aim
of the 3D object reconstruction method is two-fold: to
segment the image data into different regions and to
interpolate the surface where there is no image data. We
also desire a smooth object surface as this reflects the
physical object, namely the follicle, that we are model-
ing. Therefore, the speed function used to drive the level
Volume 34, Number 2, 2008
set method is a weighted combination of these three
desires. This can be expressed as the speed function:
F ⫽ ␣Fimage ⫹ ␤Fsurface ⫹ ␥Fsmooth (2)
where Fimage drives the image segmentation, Fsurface
drives the surface interpolation and Fsmooth is a regulari-
sation term used to keep the resulting object surface
smooth. Parameters ␣, ␤ and ␥ are used to weight the
data driven and model driven terms and, also, to generate
a similar evolution rate between the surface interpolation
and the image segmentation. We will now consider the
choice for each of these terms separately.
Image segmentation term
The choice of segmentation term used will be mo-
tivated by the particular application intended, in this case
ovarian follicle delineation. Modification of this term
will be necessary for other applications. An extensive
review of ultrasound segmentation methods by clinical
application is given in (Noble and Boukerroui 2006).
For follicle segmentation, a region-based approach
is proposed using image intensity values to guide the
segmentation because there is a clear distinction in both
intensity and texture between the follicle and the sur-
rounding tissue. This can been seen in Fig. 1. Edge
information has not been incorporated at this stage for
simplicity but may be considered in future development.
The region-based segmentation term is similar to those in
(Zhu and Yuille 1996; Paragios and Deriche 2000; Kadir
and Brady 2003), where the difference in probability of
region membership of each voxel is used to guide the
segmentation. In (Zhu and Yuille 1996) and (Paragios
and Deriche 2000), the probability of region membership
is defined a priori by modeling the region probability
density functions (PDFs) using a parametric mixture
model. We modify this approach for use with freehand
3D ultrasound data by using a nonparametric approach as
in (Kadir and Brady 2003) and choose to estimate the
region PDF at each iteration from the current segmenta-
tion. A nonparametric model is used because the true
distribution of the intensities will depend on factors such
as the particular ultrasound machine used and its settings
(Tao et al. 2006). For example, variations in the latter
might occur to take account of the particular patient
being scanned. It is easier to use a nonparametric model
rather than to try to estimate a suitable parametrisation
for all possible machine/patient variations. We model the
region PDFs by taking a histogram of the data for that
region given the current segmentation. These PDFs are
updated at each iteration of the level set evolution. We
chose to use one bin per intensity value since the rela-
tionship between the differing intensity values is unclear
and may vary between cases as the ultrasound machine
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al.
settings are changed. For region competition, it is nec-
essary to compare points on the boundary of each region
with the neighbouring regions. A small spatial window
around each point on the boundary is used to estimate the
local PDF. The use of a small spatial window also
enables the data to be considered in its real-world posi-
tion rather than constraining it to a voxel array (although,
calculation of the level-set update does occur on a voxel
array). It is assumed that coincident pixels from multiple
images will not be in identical positions in the world
co-ordinate system due to calibration and sensor errors.
Instead, they will be close to each other. However, the
use of a small spatial window to calculate the local PDF
means that all values close to a particular location will be
included. This may also include neighbouring pixels
from the same image in a similar fashion to (Zhu and
Yuille 1996; Paragios and Deriche 2000; Kadir and
Brady 2003). The window PDFs are compared with each
of the region PDFs by calculating the probability of
region membership to be the probability of all of the
observed intensities occurring in union. So the probabil-
ity of belonging to region r given a window, W is given
by
p(rⱍW) ⫽ 兿 p (I(x))
r (3)
x僆W
[Zhu and Yuille (1996); Paragios and Deriche (2000);
Kadir and Brady (2003) express this in terms of the
summation of the negative log probability, following
Leclerc (1989) by considering information encoding.]
Given that region probability distribution is based on the
current segmentation, this may lead to a situation where
there exists an intensity within the window which has not
yet been observed in the region, which should belong to
the region. Despite all other evidence within the window
indicating membership of such a region, taking the prod-
uct leads to p(r|W) ⫽ 0. As such, this calculation of
region membership is more suited to situations where the
population distribution is known a priori.
An alternative approach might be to take the mean
probability that a sample from the window belongs to the
region to which it is being compared. The mean proba-
bility of membership of region r for window W is given
by
p(r|W) ⫽
1
nw 兺 pr(I(x))
x僆W
where nW is the number of points in the window.
However, this measure favours less uniform distri-
butions of data within window. To illustrate this, con-
sider the PDF of each region or sample window, as a
vector. We will call this vector the PDF vector and use
the symbol pdfv(r) to represent the PDF vector of region
187
r. The mean probability of membership of a region can
be seen to be the scalar product of the window PDF
vector with the region PDF vector.
1
p(r|W) ⫽
nw 兺 p (I(x)) r (5)
x僆W
⫽ 兺 p (i)p (i)
w r (6)
∀i
⫽pdfv(W) . pdfv(r) (7)
⫽㛳pdfv(W)㛳 㛳 pdfv(r)㛳cos␪ (8)
where pdfv(W) is the PDF vector of the window and ␪ is
the angle between the two vectors. Consider when the
window is being compared with two different regions.
The norm of the window PDF vector is common to both
values while the “similarity” of the two PDFs is mea-
sured by the angle between the vectors. It follows that
where the length of the region PDFs are different, the
similarity will favour the region with the PDF, which has
the largest 2-norm. A region following a uniform distri-
bution has a smaller PDF magnitude and would, there-
fore, have a lower mean probability than a peaked dis-
tribution that is an equal angle away from the window
PDF.
A better similarity measure between the window
and region distributions is found by normalisation of the
similarity with respect to the magnitude of the PDFs.
Normalisation of the window PDF is not required be-
cause this is common to any comparison between alter-
native regions. This becomes a “nearest neighbour” clas-
sifier in terms of the angle between the two PDF vectors,
where the similarity, S(x) at point x between region r and
window W is given by
pdfv(W) . pdfv(r)
S(x) ⫽ (9)
㛳pdfv(r)㛳
The speed of the curve at point x in the level set evolu-
tion is then taken to be the difference in the similarity of
the window with the current region of the point and any
competing region (including the background). The speed
function for the level set evolution used is:
(4)
Fimage ⫽
pdfv(W) . pdfv(i)
㛳pdfv(i)㛳
⫺ max
㛳pdfv(j)㛳
j⫽i
冉
pdfv(W) . pdfv(j)
冊
(10)
It should be noted that this similarity measure is one of
many that could be used for comparing nonparametric
distributions. A closely related measure which has been
used in image processing is the Bhattacharyya distance
(Rathi et al. 2006), B(x) ⫽ 公pdfv(W) · 公pdfv(r).
188 Ultrasound in Medicine and Biology
The Bhattacharyya distance gives greater weight to
differences in low probability intensities within a distri-
bution and thus favours broad distributions over narrow
ones on the boundary. In comparison, the proposed dot-
product measure favours narrow distributions. In prac-
tice, there is very little difference between the proposed
measure and the Bhattacharyya distance for this applica-
tion.
Surface interpolation term
Where there is no ultrasound data to drive the level-
set update, the surface must be interpolated to fill the
gaps. Zhao et al. (2000) present a method for surface
reconstruction from unordered boundary points. That
method handles sparse data, interpolating smoothly be-
tween points. The speed function produces a weighted
minimal surface to this edge data. The speed function of
the level set is given by
冋
F ⫽ ⵜd ·
ⵜ⌽
ⱍ ⱍ
ⵜ⌽
⫹ ␳d ⵜ ·
ⵜ⌽
ⵜ⌽ ⱍ ⱍ 册 (11)
where d is the distance to the nearest data point and ␳ is
a weighting factor controlling the smoothness of the
ⵜ⌽
surface. The first term in the speed function,ⵜd· , ages, we would not expect irregular boundaries and a
|ⵜ⌽|
minimises the distance between the surface and the data
ⵜ⌽
points while the second term ␳dⵜ· , minimises the
|ⵜ⌽|
surface energy. The scaling of the curvature term,
ⵜ⌽
ⵜ· , according to distance from the data points means
|ⵜ⌽|
that the surface is more flexible near to data points,
whilst being more rigid away from the data.
This speed function is suitable for interpolation of
our segmentation surface in the absence of data. How-
ever, since image segmentation is being performed at the
same time as the surface interpolation, there will not be
a fixed set of edge points available. This means that edge
points and the corresponding distance function, d, must
be updated at each iteration. The edge points are consid-
ered to be where the zero level set, i.e., our current
boundary, intersects an image plane at the start of each
iteration. From this, the distance function is computed
using the fast sweeping method (Tsai et al. 2003).
The interpolation is largely governed by the first
ⵜ⌽
term of this function, ⵜd· with the second term
|ⵜ⌽|
acting to keep any solution smooth. The addition of the
regularisation term also helps to maintain a smooth so-
lution. However, it is necessary to maintain the ability to
smooth the interpolated surface independently to the
global regularisation since the global regularisation must
be appropriately weighted against the strength of the
Volume 34, Number 2, 2008
segmentation term. Since this term is only being used to
keep the interpolated part of the surface smooth, its exact
value is unimportant so long as the solution is plausible
and stable. The choice of ␳ and the density of the data
will therefore affect the choice of time step used in the
level set iteration in order to keep the solution stable. It
was found empirically that ␳ ⫽ 101 was suitable for the
freehand data being used. Although it is possible to use
higher values of ␳, e.g., ␳ ⫽ 21 , in areas far from the data,
where d is large, a small time step would be required to
keep the evolution stable. A low time step will lead to
more iterations being required to reach the desired solu-
tion.
We, therefore, used
冋 册
ⵜ⌽ d ⵜ⌽
Fsurf ⫽ ⵜd · ⫹ ⵜ·
ⱍ ⱍ ⱍ ⱍ
(18)
ⵜ⌽ 10 ⵜ⌽
where d is the distance from the point of evolution to the
nearest point where the current boundary intersects an
image plane.
Surface regularisation term
Segmentation of noisy images often leads to image
boundaries, which are not smooth. In many natural im-
regularisation term is used to “smooth” the segmentation.
For segmentation methods based on the level set method,
evolution according to level set curvature has been com-
monly used for this purpose. This, in part, is probably as
a result of the historical development of the level set
method as a method for studying solely curvature-depen-
dant speed (Osher and Sethian 1988). Curvature does
provide some measure of smoothness, as flow under
curvature is equivalent to minimising the length of the
curve. However, using curvature alone causes three
problems with regard to object surface smoothing. First,
once the surface is optimally smooth, eg, a sphere, cur-
vature is non-zero and the shape will continue to shrink
(Delingette 2001). Second, curvature is not scale-invari-
ant in that a small object exhibits higher curvatures and
small deviations from a small object will also have
higher curvatures than similar deviations from a larger
object. Finally, depending on the application, the surface
may have areas of high curvature which require preser-
vation, and as such anisotropic smoothing may be re-
quired (Tasdizen et al. 2002). Given that follicles are
nonrigid fluid filled structures, the use of isotropic
smoothing in this case is appropriate, therefore any regu-
larisation term developed need not consider this to be an
issue. The problem of the surface collapsing has been
examined by a number of authors. The simplest solution
is given in (Sapiro and Tannenbaum 1995), whereby the
mean curvature of the entire surface is subtracted from
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al.
the local curvature. So for a sphere, where the curvature
is constant, the mean curvature is the same as the local
curvature and the speed function is zero. Deviations from
the mean curvature are penalised.
Including this term still leaves a measure, which is
proportional to object size. A small deviation from the
mean curvature of a small object will be penalised more
that a small deviation from the mean curvature of a large
object. For a single object, the weighting parameter be-
tween smoothness and the data driven term can be ad-
justed such that the terms balance at an appropriate
result. However, in the application of follicle measure-
ment there are normally several follicles of different
sizes for which the data driven term will have a similar
strength and it is unlikely that it would be possible to find
a single weighting term that would be satisfactory for all
of them. If smoothing is too strong, smaller follicles will
not be segmented (the level set will collapse) and, if the
smoothness weighting is too weak, large follicles will not
be smooth. Our solution is to apply a degree of scale
invariance by scaling the regularisation term proportion-
ally to 兹 3
Vobject where Vobject is the current volume
estimate for the object. Clearly, this assumes spherical
objects and becomes less appropriate as the object differs
from a spherical shape.
Therefore, the final choice of regularisation term is
given by
冉 冊
冑
Fsmooth ⫽ 3
Vobject
Vmax
␬⫺
兰
surface
兰
␬ds
ds
surface
(20)
where ␬ is the curvature of ␾, Vmax is a constant and
Vobject is the current object volume. The constant Vmax is
used to keep the surface regularisation term, Fsmooth, in a
similar numeric range independent of the stage of the
menstrual cycle. This weighting could also be in the
weighing constant, ␥ of eqn 2 but we believe it has more
intuitive meaning if considered in Fsmooth.
IMPLEMENTATION DETAILS
Use of a multi-class level set method
In our application, we wish for the multiple follicle
regions to compete during the segmentation process.
Therefore, they must be given separate class labels.
However, in its basic form, the level set method can
represent two classes, ␾ ⬎ 0 and ␾ ⬍ 0, separated by the
boundary ␾ ⫽ 0. Four main solutions to this problem
have been proposed in the literature (Paragios and De-
riche 2000; Vese and Chan 2002; Lie et al. 2003; Kadir
and Brady 2003). All of these methods achieve the end
goal of being able to separate multiple regions within an
image but require a different number of level set func-
tions to be used. As the number of level set functions is
189
reduced, the complexity of the method increases to some
extent. However, the reduction in the number of level set
functions is an important consideration in terms of mem-
ory required, as implementation of these methods for 3D
analysis will also increase the memory required to store
the level set function(s).
To address this, we employed an extension of the
method of (Kadir and Brady 2003). That approach only
requires a single level set function regardless of the
number of regions to be encoded. A class label for each
region is found at each iteration with nonconnected fore-
ground classes assigned different labels. A thin region of
background class, ␾ ⬎ 0, is maintained between differ-
ent regions. If regions collide and should not merge (as
will occur in our application), the speed function of the
statistically weaker region is set such that the region
shrinks locally. This method has advantages in preserv-
ing the ease with which regions can be merged, while
preventing class overlap and significantly reducing mem-
ory cost. Their method needs modification to be suitable
for our application. First, the binary speed function used
in (Kadir and Brady 2003) does not allow for easy
implementation of a curvature-based regularisation term.
A binary speed function is used in order to guarantee that
the surface will move a single voxel at most. This is done
to create an easy implementation, where merging and
competing regions can be controlled simply by adjusting
the sign of the speed function. An extension to nonbinary
speeds can be achieved by adjusting the time step in the
evolution process, in order to ensure that the maximum
change in any position on the surface is less than one
voxel each iteration. Competition is managed by detect-
ing locations where regions are competing, then both
regions have their speed functions set to move the
boundaries apart at this point. The region with the stron-
ger data term will then fill into the gap faster than the
competing region. Competition progresses in this way,
allowing the stronger region to push back the weaker
region. Competing regions can be forced back, in fol-
lowing with (Kadir and Brady 2003), by setting the speed
as a constant. It has been found empirically that any
discontinuity in the level set function is quickly rectified
by the curvature regularisation term. Maintenance of ␾
as a signed distance function to the surface is discussed
later.
Second, finding region labels at each iteration
proves a time-consuming part of the implementation in
(Kadir and Brady 2003). To keep track of region
labels, they opt to use a connected components algo-
rithm to build a new region label map at each iteration.
In practice, this is unnecessary, and it is possible to
maintain a label map by detecting class changes, thus
avoiding this time consuming step. Only two instances
of label change can occur. Detecting a change from
190 Ultrasound in Medicine and Biology
foreground to background is simple; at each iteration
any voxel where ␾ ⬎ 0 should be set to the back-
ground label on the region label map. By maintaining
a single background voxel between competing fore-
ground regions, when a background voxel changes to
a foreground region, the region to which it belongs can
be set by detecting the class to which any neighbour-
ing foreground voxels belong.
Efficient storage of ultrasound data
Each of the data sets produced using the freehand
ultrasound system detailed in the next section occupies
approximately 400 MB of memory (180 images; 640 ⫻
480 pixels, of which about 40% are masked off as
nonultrasound content; 1 byte per pixel intensity ⫹ 3 ⫻
4 bytes for floating point position of each pixel in the
world co-ordinate space). Although loading this data into
memory directly would be possible for high-end com-
puters, for easier implementation, some efficient storage
method was required. However, we can take advantage
of the particular PDF based segmentation method we
use. Firstly, only the PDFs of a local window and of each
region are required to calculate the similarity measure.
Secondly, calculation of the speed function only occurs
at each node of the level set array. Therefore only the
intensities associated with each node are stored, rather
than the pixel positions. When calculating the region
PDF, all points associated with a particular level set node
are attributed to the region to which the node belongs.
This moves away from the goal of segmentation occur-
ring in a world reference space since we have essentially
confined the image data to the same voxel array as the
level set function. However, this is a result of implemen-
tation need rather than a requirement of the speed func-
tion defined, and as such the speed function could still be
applied in the world reference space.
Narrowband implementation
The level set method can be quite slow for calcu-
lating curve, or surface, evolution, particularly when
compared with active contours. We adopt the so-called
narrowband method, as described in (Peng et al. 1999),
to speed up the processing of each iteration. This method
uses the fact that we are only interested in the position
and evolution of the zero level set and the assumption
that the other level sets are moving parallel to the zero
level set as a signed distance function. Rather than cal-
culating the update of the level set function over its entire
domain, it is only updated on a narrow band around the
zero level set. This band is then moved such that it
always contains the zero level set.
To avoid re-initialisation of the level set function,
our implementation of the narrowband method follows
(Peng et al. 1999), requiring two bands to be considered:
Volume 34, Number 2, 2008
the narrowband on which the level set function is up-
dated and an extended narrowband on which the signed
distance function is maintained iteratively. The extended
narrowband is slightly wider than the narrowband. A
narrowband width of four grid spaces is sufficient to
encapsulate the zero level set at all times as we are
aiming for the zero level set to be moved by less than one
grid space at each iteration. The extended narrowband is
about two grid points wider than the narrowband since it
is only required to keep the gradient smooth at the edge
of the narrowband, such that the level set function can be
updated accurately all the way to the edge of the nar-
rowband.
EXPERIMENTAL METHOD
Freehand 3D ultrasound acquisition
The freehand ultrasound system used in this work
consisted of a Faro arm (Faro Technologies Inc., Lake
Mary, FL, USA) rigidly attached to a 5 to 7 MHz
transvaginal ultrasound probe connected to a Toshiba
PowerVision SSA-270 (Toshiba Medical Systems,
Otawara-shi, Tochigi-ken, Japan). A PC workstation
(Dell Inc., Round Rock, TX, USA) was used to record
position information while the ultrasound machine re-
corded the images independently. The images were
transferred in a digital format to the workstation after the
scan for processing. Temporal calibration between the
ultrasound machine and the workstation was performed
using the method described in (Gooding et al. 2005b).
Spatial calibration was performed using a cross-wire
phantom as described in (Atkinson et al. 2001). One
hundred eighty images were acquired during each scan at
approximately 12 Hz using the ultrasound machine cine
function. Further details of the freehand ultrasound sys-
tem can be found in (Gooding et al. 2005b).
Clinical testing protocol
Testing the accuracy of the reconstruction for in
vivo data requires that ground truth is identified to com-
pare with the reconstruction results. Comparison with
manual segmentations would reveal the accuracy of the
boundary compared with the clinician’s belief of where
the follicle boundary is in each image (which may not the
correct boundary). Additionally, the manual segmenta-
tions would require reconstruction into a volume in order
to compare volume measurements. To assess the volume
measurement accuracy, it is necessary to find an accurate
and independent measurement of follicle volume. Folli-
cles are aspirated within the routine practice of IVF to
collect the oocytes from each follicle. Measurement of
the fluid at the time of aspiration provides a reasonable
ground truth measurement for the follicle volume (Brun-
ner et al. 1995; Kyei-Mensah et al. 1996b). Although
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al.
error is possible if the follicle is not fully aspirated or, if
extra fluid/blood is taken, the volume of aspirate is still a
good reference for comparison. Reference volumes could
be obtained using MRI (ter Haar Romeny et al. 1999),
however, this method is not used within routine practice
for fertility treatment, making this method unacceptable
for use with patients.
The accuracy of the reconstruction system was
compared with aspirated follicle volumes and the mean
diameter estimate. Local ethics committee approval was
obtained for this study. Ovaries from 11 consenting
patients undergoing IVF treatment were scanned prior to
oocyte recovery using the 3D system. The mean diameter
measurement was also made during this scan. All folli-
cles within the scanned volume were reconstructed after
the scan. To aid identification of each follicle at the time
of aspiration, three methods were employed: (1) the
clinician who would perform follicle aspiration was
present at the time of the 3D scan: (2) labeled diagrams
were drawn for each scan by the scanning clinician and
(3) still images were printed of a number of image planes
showing significant follicles. Although, ideally, all folli-
cles would be measured in each ovary, if an ovary
contained more than about six follicles, identification of
follicles at aspiration is very difficult. Therefore, only the
follicles for which the aspirating clinician could be con-
fident of a correct identification had volumes measured
and recorded during aspiration. This was typically one
follicle per ovary. Aspirated volumes were recorded to
the nearest 0.5 ml. The assumptions made in recording a
mean diameter are that the follicle is spherical and that
the diameter is representative of the follicle size and,
hence, its volume. To test this assumption, the volumes
of the follicles were estimated from the clinical diameter
measurements using a spherical model, for comparison
with the aspirated volume.
Parameters used in reconstructions
For the experiments presented in this article, recon-
structions were performed using a level set grid spacing
of around 0.8 to 1 mm, depending on the size of the
ovary. Tests performed at a range of resolutions indi-
cated that this order of resolution seemed to provided the
best segmentation speed without quantising volume mea-
surements noticeably (1/1000th of a ml per voxel). The
reconstruction method was initialised manually with a
small sphere within each follicle. The reconstruction
parameters were varied on a patient by patient basis as
they are dependent the ultrasound acquisition. The ultra-
sound machine settings were left to the sonographer to
avoid interference with the clinical exam and as a result
were not consistent, leaving the tuning of parameters
unavoidable in this instance. The typical values (and
range) were 1.0 (kept constant), 5.0 (5 to 10) and 3.0 (1
191
to 5) for ␣, ␤ and ␥, respectively. Initially, the typical
values were used. ␤ was generally increased in cases
where the contrast between follicle and background was
low. Low contrast could lead to regions with weak
boundary strength. Increasing the smoothing would help
to constrain the surface in these areas, whilst still en-
abling the data to be segmented in areas where the
boundary contrast was stronger. ␥ was decreased in cases
where the data was dense and interpolation over sparse
areas was not required. Only integer values of each
parameter were used. A detailed scanning protocol may
remove the need for tuning ␤ or it may be possible to
automate their tuning based on analysis of the image/data
sparsity.
RESULTS
The data acquired consisted of 23 follicles from 11
patients for which aspirated volume, reconstructed vol-
ume and diameter measurements were available. A fur-
ther 15 follicles were available for which there were only
diameter and aspirated volumes because the entire ovary
had not been acquired during the 3D scan. This resulted
from sonographer in-experience at using 3D scanning
equipment and is not considered a drawback of the
segmentation method. However, it demonstrates the need
for experience with the new scanning methods. The
relatively low number of follicles in this study is a result
of the difficulty in identifying follicles at aspiration once
the first follicle has been aspirated. This problem was
also identified in (Kyei-Mensah et al. 1996b), where only
the largest follicle was measured in each of 25 patients.
Two of the follicles for which all measurements were
available were subsequently excluded from the analysis
because their reconstructed volume was much larger than
their aspirated volume indicating that the follicle was not
fully aspirated.
The results of volume estimation from the diameter
measurement and from the reconstruction system are
shown against the aspirated volume in Fig. 4. Qualita-
tively, it can be seen that the measurement made by the
3D system has a lower spread than the clinical data.
However, it is also apparent that both measurement
methods underestimate the true volume by about 25% to
30%. The reasons for the underestimate are considered in
the Discussion section. This underestimation was found
to be significant for both the clinical diameter based
volume estimation (p ⫽ 1.3 ⫻ 10⫺4, F ⫽ 19.9) and the
computer based method (p ⫽ 1.3 ⫻ 10⫺6, F ⫽ 46.1).
Taking this underestimate into account, the range of
variation of the clinical data is similar to that found by
(Forman et al. 1991).
The variances of the measurements from the linear
fit to the data were found to be 0.43 ml2 and 1.89 ml2 for
192 Ultrasound in Medicine and Biology Volume 34, Number 2, 2008

Fig. 4. Graph showing the reconstructed volume measurement and the clinical volume estimate (based on a spherical
model) plotted against the aspirated volume for ovarian follicles. The 3D reconstruction can be seen to have lower
variation; however, both underestimate the true volume.

the computer-based estimate and clinical estimate, re- DISCUSSION

spectively. The 3D reconstruction system was found to
The segmentation and object reconstruction ap-
have a significantly different variance to the clinical
peared qualitatively good, however, both the computer-
measurements (p ⬍ 0.001) using an F-test (F ⫽ 2.67).
based measurements and clinical measurements tended
(Neither set would be rejected as a normal distribution to underestimate the true follicular volume. Although
using a ␹2 test.) The variance from the best fit for the 3D lower error variance of the computer-based measure-
reconstruction system is about quarter that of the clinical ments suggest that this method would be useful in the
estimate and, therefore, can be considered to be out- measurement of follicles, it would be desirable to under-
performing the current clinical measurement method in stand and correct for the underestimation.
this respect. It should also be noted that the variance of Since this underestimation occurs with both the
the measurement error will be affected by the quantisa- clinical and computer-based measurements, the cause of
tion of the aspirated volume, and that both are of the the measurement error must affect both measurement
same order of magnitude. Measurement of the aspirate
volume was made to the nearest 0.5 ml in order to
minimise any disruption to the oocyte collection proce-
dure. This quantisation will need to be reduced in future
validation studies to fully assess the variance of the
measurement error.
Figure 5 shows the reconstruction of the ovary
shown in Fig. 1. The foreground follicle in Fig. 5 is the
one for which measurements are shown in Fig. 1 and the
diagonal line illustrates the plane in which Fig. 1 was
acquired. The nonspherical shape of such a follicle and
the difficulty in finding the maximum diameter, cause the
mean diameter measurement to be less representative of
the follicle volume. Figure 6 show a selection of images
planes with the segmentation overlayed together with the
corresponding follicle volume reconstructions for that Fig. 5. This figure shows the 3D object reconstruction for the
ovary to illustrate the quality of the computer-based ovary shown in Fig. 1. The diagonal plane illustrates the
segmentations. position of the image of Fig. 1.
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al.
Fig. 6. A selection of computer-based segmentation results
overlayed on the appropriate image plane. The corresponding
3D reconstruction of all follicle in the ovary is shown adjacent
to each image. The position of the delineated boundary can be
seen to follow the limit of the antral fluid, which we now
believe to represent an under estimation of the follicle volume.
methods. Factors relating to the reconstruction system,
such as calibration error or choice of level set grid size,
cannot account for the observed underestimation since
they do not affect the clinical measurement.
The measurement of the follicle aspirate is common
to the assessment of both measurement methods. Whilst
it is plausible that the follicles could retain some fluid
during aspiration, this effect would lead to the ultrasound
measurements being overestimated. It is very unlikely
that extra fluid is systematically taken during aspiration
and, therefore, this is not considered as a possible cause
for this error.
The other main factor affecting both clinical and
computer-based measurement is the ultrasound acquisi-
tion itself. The B-mode images are formed within the
ultrasound machine using the assumption that the speed
of sound within soft tissue is 1540m/s. If this assumption
is wrong, the size of objects within the ultrasound image
will be affected. However, in (Gooding et al. 2005a), the
speed of sound within follicular fluid was found to be
193
1550 m/s, which would only account for a 2% error in
volume.
In contrast to the physical causes considered
above, it is thought that the most plausible explanation
for the underestimation is a clinical misunderstanding
of the location of the true follicular fluid boundary
within the ultrasound image. Any error in clinical
understanding will be reflected in the computer-based
segmentation since the segmentation method de-
scribed was developed with the guidance of ultra-
sonographers such that the boundary targeted by the
segmentation is that which the ultrasonographers con-
sider to be the follicular fluid boundary. The clinicians
involved in this research measure to the limit of the
antral fluid within the ultrasound image, whereas
Kyei-Mensah et al. measured the follicular boundary
as “the rim of the follicle wall so that the follicle
margin was the margin of annotation” (personal cor-
respondence). As expected, the position of the delin-
eated boundary from the computer-based segmenta-
tion can be seen (in Fig. 6) to follow the limit of the
antral fluid. This difference in boundary delineation
position is sufficient to explain the error in volume
found, and it is noted that (Kyei-Mensah et al. 1996b)
found good agreement between volume measurement,
made by manual delineation of 3D ultrasound, and
aspirate volume.
Correcting for this error requires modification of
the segmentation term within the level set method
because the follicular wall has a different appearance
to the follicular fluid and does not fit the assumption of
visual similarity used by the region growing method.
However, it should be noted that the presence of this
bias is not a barrier to use in a clinical context. The
apparently constant scaling between the true volume
and the reconstructed volume means that the recon-
structed volume is still a good measure of follicular
size and the 3D system has lower error variance than
the clinical measurement. A preliminary study into the
use of this system in a clinical setting is presented in
(Gooding et al. 2004).
CONCLUSION
In this article, we have presented a method for the
reconstruction of follicle volumes from freehand 3D
ultrasound. Surface reconstruction is performed using a
level set based region competition method within the 3D
domain. Performing surface reconstruction and segmen-
tation within the 3D domain allows the benefits of 3D
spatial information and data compounding to be used.
The method also incorporates a term which enables for
surface interpolation in areas where image data was not
acquired during the 3D scan.
194 Ultrasound in Medicine and Biology
Our initial investigation into the accuracy of the
volume measurement using the system found that both
the computer-based measurement and the clinical mea-
surement consistently underestimate the volume of fol-
licular aspirate, as a result of a clinical misunderstanding
of the true location of the follicular boundary. However,
the computer-based system had a lower variance than the
clinical estimate, showing it to provide a better estimate
of follicular size.
Areas for current and future work include develop-
ing a correction for the volume under-estimation and the
clinical study of follicular volume growth.
Acknowledgment—The authors thank Toshiba Medical Systems Eu-
rope, for the donation of a PowerVision ultrasound machine for this
project; the staff of the Fertility Unit, John Radcliffe Hospital, UK, for
their help in the acquisition of clinical results. M.J.G. was funded by
the EPRSC as part of the MIAS-IRC grant (GR/N14248).
REFERENCES
Ahmad A, Cool D, Chew BH, Pautler SE, Peters TM. 3D segmentation
of kidney tumors from freehand 2D ultrasound. Medical Imaging
2006, Proc. of SPIE. 2006;6141:227–237.
Allott CP, Barry CD, Bramley J, et al. Noninvasive 3-D ultrasound of
atherosclerotic plaques in the watanabe rabbit. Ultrasound Med
Biol 1999;25:905–914.
Atkinson D, Burcher M, Declerck J, Noble JA. Respiratory motion
compensation for 3-D freehand echocardiography. Ultrasound Med
Biol 2001;27:1615–1620.
Barry CD, Allott CP, John NW, et al. Three-dimensional freehand
ultrasound: image reconstruction and volume analysis. Ultrasound
Med Biol 1997;23:1209 –1224.
Bogush AL, Tuzikov AV. 3D object volume measurement using free-
hand ultrasound. Lecture Notes in Computer Science. 2005;3540:
1066 –1075.
Brunner M, Obruca A, Bauer P, Feichtinger W. Clinical application of
volume estimation based on three-dimensional ultrasonography.
Ultrasound Obstet Gynecol. 1995;6:358 –361.
Cigale B, Lenic̆ M, Zazula D. Segmentation of ovarian ultrasound
images using cellular neural networks trained by support vector
machines. Lecture Notes in Artifical Intelligence 2006;4253:515–
522.
Delingette H. On smoothness measures of active contours and surfaces.
In: IEEE Work-shop of Variation and Level Set Methods 2001.
2001. pp. 43–50.
Fenster A, Downey DB. 3-D ultrasound imaging: A review. IEEE
Engineering in Medicine and Biology. 1996;15:41–51.
Forman RG, Robinson J, Yudkin P, et al. What is the true follicular
diameter: An assessment of the reproducibility of transvaginal
ultrasound monitoring in stimulated cycles. Fertil Steril. 1991;56:
989 –992.
Gee AH, Prager RW, Treece GM, Berman L. Engineering a freehand
3D ultrasound system. Pattern Recognition Letters. 2003;24:757–
777.
Gooding MJ, Barber D, Kennedy SH, Noble JA. The effect of follicle
volume measurement on clinical decisions. In: Rueckert D, Hajnal
J, Yang GZ (Eds.), Proceedings of Medical Image Understanding
and Analysis (MIUA) 2004. 2004. pp. 121–124.
Gooding MJ, Barber D, Kennedy SH, Noble JA. Measurement of the
speed of sound in follicular fluid. Human Reproduction. 2005a;20:
497–500.
Gooding MJ, Kennedy SH, Noble JA. Temporal calibration of freehand
three-dimensional ultrasound using image alignment. Ultrasound
Med Biol. 2005b;31:919 –927.
Volume 34, Number 2, 2008
Kadir T, Brady JM. Unsupervised non-parametric region segmentation
using level sets. In: IEEE International Conference on Computer
Vision. 2003. pp. 1267–1274.
Krivanek A, Sonka M. Ovarian ultrasound image analysis: follicle
segmentation. IEEE Trans Med Imag. 1998;17:935–944.
Kyei-Mensah A, Maconochie N, Zaidi J, et al. Transvaginal three-
dimensional ultrasound: reproducibility of ovarian and endometrial
volume measurements. Fertil Steril. 1996a;66:718 –722.
Kyei-Mensah A, Zaidi J, Pittrof R, et al. Transvaginal three-dimen-
sional ultrasound: accuracy of follicular volume measurements.
Fertil Steril. 1996b;65:371–376.
Leclerc YG. Constructing simple stable descriptions for image parti-
tioning. Internation Journal of Computer Vision. 1989;3:73–102.
Legget ME, Leotta DF, Bolson EL, et al. System for quantitative
three-dimensional echocardiography of the left ventricle based on a
magnetic-field position and orientation sensing system. IEEE Trans
Biomed Eng. 1998;45:494 –504.
Lie J, Lysaker M, Tai XC. A variant of the level set method and
applications to image segmentation. Tech. Rep. CAM 03–50,
UCLA, Computational and Applied Mathematics. 2003.
Mercier L, Lang␸ T, Lindseth F, Collins LD. A review of calibration
techniques for freehand 3-D ultrasound systems. Ultrasound Med
Biol. 2005;31:143–165.
Muzzolini R, Yang YH, Pierson RA. Multiresolution texture segmen-
tation with application to diagnostic ultrasound images. IEEE Trans
Med Imag. 1993;12:108 –123.
Noble JA, Boukerroui D. Ultrasound image segmentation: A survey.
IEEE Trans Med Imag. 2006;25:987–1010.
Osher S, Sethian JA. Fronts propagating with curvature dependent
speed: Algorithms based on hamilton-jacobi formulations. Journal
of Computational Physics. 1988;79:12– 49.
Paragios N, Deriche R. Coupled geodesic active regions for image
segmentation: A level set approach. In: European Conference on
Computer Vision (2). 2000. pp. 224 –240.
Peng D, Merriman B, Osher S, Zhao HK, Kang M. A PDE-based fast
local level set method. Journal of Computational Physics. 1999;
155:410 – 438.
Penzias AS, Emmi AM, Dubey AK, et al. Ultrasound prediction of
follicle volume: is the mean diameter reflective? Fertil Steril. 1994;
62:1274 –1276.
Potoc̆nik B, Zazula D. Automated analysis of a sequence of ovarian
ultrasound images, part i: Segmentation of single 2D images. Image
and Vision Computing. 2002a;20:217–225.
Potoc̆nik B, Zazula D. Automated analysis of a sequence of ovarian
ultrasound images, part ii: Predicion-based object recognition for a
sequence of images. Image and Vision Computing. 2002b;20:227–
235.
Prager RW, Gee AH, Treece GM, Berman L. Freehand 3D ultrasound
without voxels: volume measurement and visualisation using the
stradx system. Ultrasonics. 2002;40:109 –115.
Rathi Y, Michailovich O, Malcolm J, Tannenbaum A. Seeing the
unseen: Segmenting with distributions. In: IASTED Conf on Signal
and Image Processing. Acta Press. 2006.
Rohling RN, Gee AH. Issues in 3-D free-hand medical ultrasound
imaging. Tech. Rep. CUED/F-INFENG/TR 246, Cambridge Uni-
versity Engineering Department. 1996.
Rohling RN, Gee AH, Berman L. A comparison of freehand three-
dimensional ultrasound reconstruction techniques. Medical Image
Analysis. 1999;3:339 –359.
Saith RR, Srinivasan A, Michie D, Sargent IL. Relationships between
the developmental potential of human in vitro fertilization embryos
and features describing the embryo, oocyte and follicle. Human
Reproduction Update. 1998;4:121–134.
Sapiro G, Tannenbaum A. Area and length preserving geometric in-
variant scale-spaces. IEEE Trans Patt Anal Mach Int. 1995;17:67–
72.
Sarty GE, Liang W, Sonka M, Pierson RA. Semiautomated segmenta-
tion of ovarian follicular ultrasound images using a knowledge-
based algorithm. Ultrasound Med Biol. 1998;24:27– 42.
Sethian JA, Level set methods and fast marching methods, 2nd Edition.
Cambridge University Press. 1999.
 Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al.
Song M, Haralick RM, Sheehan FH, Johnson RK. Integrated surface
model optimization for freehand three-dimensional echocardiogra-
phy. IEEE Trans Med Imag. 2002;21:1077–1090.
Tao Z, Tagare HD, Beaty JD. Evaluation of four probability distribu-
tion models for speckle in clinical cardiac ultrasound images. IEEE
Trans Med Imag. 2006;25:1483–1491.
Tasdizen T, Whitaker R, Burcahrd P, Osher S. Geometric surface
smoothing via anisotropic diffusion of normals. In: IEEE Visual-
ization 2002. 2002. pp. 125–132.
ter Haar Romeny BM, Titulaer B, Kalitzin S, et al. Computer assisted
human follicle analysis for fertility prospects with 3d ultrasound.
1999;1613:56 – 69.
Treece GM, Prager RW, Gee AH, Berman L. Volume measurement
from sequential free-hand 3D ultrasound. Lecture Notes in Com-
puter Science. 1999;1613:70 – 83.
Treece GM, Prager RW, Gee AH, Berman L. Surface interpolation
from sparse cross sections using region correspondence. IEEE
Trans Med Imag. 2000;19:1106 –1114.
Treece GM, Prager RW, Gee AH, Berman L. 3D ultrasound measure-
ment of large organ volume. Medical Image Analysis. 2001;5:41–
54.
Tsai YR, Cheng LT, Osher S, Zhao HK. Fast sweeping algorithms for
a class of hamilton-jacobi equations. SIAM Journal on Numerical
Analysis. 2003;41:673– 694.
195
Vese LA, Chan T. A multiphase level set framework for image seg-
mentation using the Mumford and Shah model. International Jour-
nal of Computer Vision. 2002;50:271–293.
Ye X, Noble JA, Atkinson D. 3-D freehand echocardiography for
automatic left ventricle reconstruction and analysis based on mul-
tiple acoustic windows. IEEE Trans Med Imag. 2002;21:1051–
1058.
Zahalka A, Fenster A. An automated segmentation method for three-
dimensional carotid ultrasound images. Phys Med Biol 2001;46:
1321–1342.
Zhang Y, Rohling R, Pai DK. Direct surface extraction from 3D
freehand ultrasound images. Proceedings of IEEE Visualization
2002. 2002.
Zhao HK, Osher S, Merriman B, Kang M. Implicit and non-parametric
shape reconstruction from unorganized data using a variational
level set method. Computer Vision and Image Understanding.
2000;80:295–319.
Zhu SC, Yuille AL. Region competition: Unifying snakes, region
growing, and bayes/MDL for multiband image segmentation. IEEE
Trans Patt Anal Mach Int. 1996;18:884 –900.
Zimmer Y, Tepper R, Akselrod S. A two-dimensional extension of
minimum cross entropy thresholding for the segmentation of ultra-
sound images. Ultrasound Med Biol. 1996;22:1183–1190.