Professional Documents
Culture Documents
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
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.
183
184 Ultrasound in Medicine and Biology Volume 34, Number 2, 2008
settings are changed. For region competition, it is nec- r. The mean probability of membership of a region can
essary to compare points on the boundary of each region be seen to be the scalar product of the window PDF
with the neighbouring regions. A small spatial window vector with the region PDF vector.
around each point on the boundary is used to estimate the
1
local PDF. The use of a small spatial window also
enables the data to be considered in its real-world posi-
p(r|W) ⫽
nw 兺 p (I(x)) r (5)
x僆W
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
⫽ 兺 p (i)p (i)
w r (6)
∀i
images will not be in identical positions in the world
co-ordinate system due to calibration and sensor errors. ⫽pdfv(W) . pdfv(r) (7)
Instead, they will be close to each other. However, the
⫽㛳pdfv(W)㛳 㛳 pdfv(r)㛳cos (8)
use of a small spatial window to calculate the local PDF
means that all values close to a particular location will be where pdfv(W) is the PDF vector of the window and is
included. This may also include neighbouring pixels the angle between the two vectors. Consider when the
from the same image in a similar fashion to (Zhu and window is being compared with two different regions.
Yuille 1996; Paragios and Deriche 2000; Kadir and The norm of the window PDF vector is common to both
Brady 2003). The window PDFs are compared with each values while the “similarity” of the two PDFs is mea-
of the region PDFs by calculating the probability of sured by the angle between the vectors. It follows that
region membership to be the probability of all of the where the length of the region PDFs are different, the
observed intensities occurring in union. So the probabil- similarity will favour the region with the PDF, which has
ity of belonging to region r given a window, W is given the largest 2-norm. A region following a uniform distri-
by bution has a smaller PDF magnitude and would, there-
fore, have a lower mean probability than a peaked dis-
p(rⱍW) ⫽ 兿 p (I(x))
r (3)
tribution that is an equal angle away from the window
x僆W
PDF.
[Zhu and Yuille (1996); Paragios and Deriche (2000); A better similarity measure between the window
Kadir and Brady (2003) express this in terms of the and region distributions is found by normalisation of the
summation of the negative log probability, following similarity with respect to the magnitude of the PDFs.
Leclerc (1989) by considering information encoding.] Normalisation of the window PDF is not required be-
Given that region probability distribution is based on the cause this is common to any comparison between alter-
current segmentation, this may lead to a situation where native regions. This becomes a “nearest neighbour” clas-
there exists an intensity within the window which has not sifier in terms of the angle between the two PDF vectors,
yet been observed in the region, which should belong to where the similarity, S(x) at point x between region r and
the region. Despite all other evidence within the window window W is given by
indicating membership of such a region, taking the prod-
uct leads to p(r|W) ⫽ 0. As such, this calculation of pdfv(W) . pdfv(r)
S(x) ⫽ (9)
region membership is more suited to situations where the 㛳pdfv(r)㛳
population distribution is known a priori.
The speed of the curve at point x in the level set evolu-
An alternative approach might be to take the mean
tion is then taken to be the difference in the similarity of
probability that a sample from the window belongs to the
the window with the current region of the point and any
region to which it is being compared. The mean proba-
competing region (including the background). The speed
bility of membership of region r for window W is given
function for the level set evolution used is:
by
p(r|W) ⫽
1
nw 兺 pr(I(x)) (4)
Fimage ⫽
pdfv(W) . pdfv(i)
㛳pdfv(i)㛳
⫺ max
㛳pdfv(j)㛳
j⫽i
冉
pdfv(W) . pdfv(j)
冊
x僆W
(10)
where nW is the number of points in the window.
However, this measure favours less uniform distri- It should be noted that this similarity measure is one of
butions of data within window. To illustrate this, con- many that could be used for comparing nonparametric
sider the PDF of each region or sample window, as a distributions. A closely related measure which has been
vector. We will call this vector the PDF vector and use used in image processing is the Bhattacharyya distance
the symbol pdfv(r) to represent the PDF vector of region (Rathi et al. 2006), B(x) ⫽ 公pdfv(W) · 公pdfv(r).
188 Ultrasound in Medicine and Biology Volume 34, Number 2, 2008
The Bhattacharyya distance gives greater weight to segmentation term. Since this term is only being used to
differences in low probability intensities within a distri- keep the interpolated part of the surface smooth, its exact
bution and thus favours broad distributions over narrow value is unimportant so long as the solution is plausible
ones on the boundary. In comparison, the proposed dot- and stable. The choice of and the density of the data
product measure favours narrow distributions. In prac- will therefore affect the choice of time step used in the
tice, there is very little difference between the proposed level set iteration in order to keep the solution stable. It
measure and the Bhattacharyya distance for this applica- was found empirically that ⫽ 101 was suitable for the
tion. freehand data being used. Although it is possible to use
higher values of , e.g., ⫽ 21 , in areas far from the data,
Surface interpolation term where d is large, a small time step would be required to
Where there is no ultrasound data to drive the level- keep the evolution stable. A low time step will lead to
set update, the surface must be interpolated to fill the more iterations being required to reach the desired solu-
gaps. Zhao et al. (2000) present a method for surface tion.
reconstruction from unordered boundary points. That We, therefore, used
冋 册
method handles sparse data, interpolating smoothly be-
tween points. The speed function produces a weighted ⵜ⌽ d ⵜ⌽
Fsurf ⫽ ⵜd · ⫹ ⵜ·
ⱍ ⱍ ⱍ ⱍ
(18)
minimal surface to this edge data. The speed function of ⵜ⌽ 10 ⵜ⌽
the level set is given by
where d is the distance from the point of evolution to the
冋
F ⫽ ⵜd ·
ⵜ⌽
ⱍ ⱍ
ⵜ⌽
⫹ d ⵜ ·
ⵜ⌽
ⵜ⌽ ⱍ ⱍ 册 (11)
nearest point where the current boundary intersects an
image plane.
where d is the distance to the nearest data point and is Surface regularisation term
a weighting factor controlling the smoothness of the Segmentation of noisy images often leads to image
ⵜ⌽ boundaries, which are not smooth. In many natural im-
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 regularisation term is used to “smooth” the segmentation.
ⵜ⌽ For segmentation methods based on the level set method,
points while the second term dⵜ· , minimises the evolution according to level set curvature has been com-
|ⵜ⌽|
surface energy. The scaling of the curvature term, monly used for this purpose. This, in part, is probably as
ⵜ⌽ a result of the historical development of the level set
ⵜ· , according to distance from the data points means method as a method for studying solely curvature-depen-
|ⵜ⌽|
that the surface is more flexible near to data points, dant speed (Osher and Sethian 1988). Curvature does
whilst being more rigid away from the data. provide some measure of smoothness, as flow under
This speed function is suitable for interpolation of curvature is equivalent to minimising the length of the
our segmentation surface in the absence of data. How- curve. However, using curvature alone causes three
ever, since image segmentation is being performed at the problems with regard to object surface smoothing. First,
same time as the surface interpolation, there will not be once the surface is optimally smooth, eg, a sphere, cur-
a fixed set of edge points available. This means that edge vature is non-zero and the shape will continue to shrink
points and the corresponding distance function, d, must (Delingette 2001). Second, curvature is not scale-invari-
be updated at each iteration. The edge points are consid- ant in that a small object exhibits higher curvatures and
ered to be where the zero level set, i.e., our current small deviations from a small object will also have
boundary, intersects an image plane at the start of each higher curvatures than similar deviations from a larger
iteration. From this, the distance function is computed object. Finally, depending on the application, the surface
using the fast sweeping method (Tsai et al. 2003). may have areas of high curvature which require preser-
The interpolation is largely governed by the first vation, and as such anisotropic smoothing may be re-
ⵜ⌽ quired (Tasdizen et al. 2002). Given that follicles are
term of this function, ⵜd· with the second term nonrigid fluid filled structures, the use of isotropic
|ⵜ⌽|
acting to keep any solution smooth. The addition of the smoothing in this case is appropriate, therefore any regu-
regularisation term also helps to maintain a smooth so- larisation term developed need not consider this to be an
lution. However, it is necessary to maintain the ability to issue. The problem of the surface collapsing has been
smooth the interpolated surface independently to the examined by a number of authors. The simplest solution
global regularisation since the global regularisation must is given in (Sapiro and Tannenbaum 1995), whereby the
be appropriately weighted against the strength of the mean curvature of the entire surface is subtracted from
Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al. 189
the local curvature. So for a sphere, where the curvature reduced, the complexity of the method increases to some
is constant, the mean curvature is the same as the local extent. However, the reduction in the number of level set
curvature and the speed function is zero. Deviations from functions is an important consideration in terms of mem-
the mean curvature are penalised. ory required, as implementation of these methods for 3D
Including this term still leaves a measure, which is analysis will also increase the memory required to store
proportional to object size. A small deviation from the the level set function(s).
mean curvature of a small object will be penalised more To address this, we employed an extension of the
that a small deviation from the mean curvature of a large method of (Kadir and Brady 2003). That approach only
object. For a single object, the weighting parameter be- requires a single level set function regardless of the
tween smoothness and the data driven term can be ad- number of regions to be encoded. A class label for each
justed such that the terms balance at an appropriate region is found at each iteration with nonconnected fore-
result. However, in the application of follicle measure- ground classes assigned different labels. A thin region of
ment there are normally several follicles of different background class, ⬎ 0, is maintained between differ-
sizes for which the data driven term will have a similar ent regions. If regions collide and should not merge (as
strength and it is unlikely that it would be possible to find will occur in our application), the speed function of the
a single weighting term that would be satisfactory for all statistically weaker region is set such that the region
of them. If smoothing is too strong, smaller follicles will shrinks locally. This method has advantages in preserv-
not be segmented (the level set will collapse) and, if the ing the ease with which regions can be merged, while
smoothness weighting is too weak, large follicles will not preventing class overlap and significantly reducing mem-
be smooth. Our solution is to apply a degree of scale ory cost. Their method needs modification to be suitable
invariance by scaling the regularisation term proportion- for our application. First, the binary speed function used
ally to 兹 3
Vobject where Vobject is the current volume in (Kadir and Brady 2003) does not allow for easy
estimate for the object. Clearly, this assumes spherical implementation of a curvature-based regularisation term.
objects and becomes less appropriate as the object differs A binary speed function is used in order to guarantee that
from a spherical shape. the surface will move a single voxel at most. This is done
Therefore, the final choice of regularisation term is to create an easy implementation, where merging and
given by competing regions can be controlled simply by adjusting
冉 冊
the sign of the speed function. An extension to nonbinary
冑
Fsmooth ⫽ 3
Vobject
Vmax
⫺
兰
surface
兰
ds
ds
surface
(20)
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
where is the curvature of , Vmax is a constant and voxel each iteration. Competition is managed by detect-
Vobject is the current object volume. The constant Vmax is ing locations where regions are competing, then both
used to keep the surface regularisation term, Fsmooth, in a regions have their speed functions set to move the
similar numeric range independent of the stage of the boundaries apart at this point. The region with the stron-
menstrual cycle. This weighting could also be in the ger data term will then fill into the gap faster than the
weighing constant, ␥ of eqn 2 but we believe it has more competing region. Competition progresses in this way,
intuitive meaning if considered in Fsmooth. 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
IMPLEMENTATION DETAILS
as a constant. It has been found empirically that any
Use of a multi-class level set method discontinuity in the level set function is quickly rectified
In our application, we wish for the multiple follicle by the curvature regularisation term. Maintenance of
regions to compete during the segmentation process. as a signed distance function to the surface is discussed
Therefore, they must be given separate class labels. later.
However, in its basic form, the level set method can Second, finding region labels at each iteration
represent two classes, ⬎ 0 and ⬍ 0, separated by the proves a time-consuming part of the implementation in
boundary ⫽ 0. Four main solutions to this problem (Kadir and Brady 2003). To keep track of region
have been proposed in the literature (Paragios and De- labels, they opt to use a connected components algo-
riche 2000; Vese and Chan 2002; Lie et al. 2003; Kadir rithm to build a new region label map at each iteration.
and Brady 2003). All of these methods achieve the end In practice, this is unnecessary, and it is possible to
goal of being able to separate multiple regions within an maintain a label map by detecting class changes, thus
image but require a different number of level set func- avoiding this time consuming step. Only two instances
tions to be used. As the number of level set functions is of label change can occur. Detecting a change from
190 Ultrasound in Medicine and Biology Volume 34, Number 2, 2008
foreground to background is simple; at each iteration the narrowband on which the level set function is up-
any voxel where ⬎ 0 should be set to the back- dated and an extended narrowband on which the signed
ground label on the region label map. By maintaining distance function is maintained iteratively. The extended
a single background voxel between competing fore- narrowband is slightly wider than the narrowband. A
ground regions, when a background voxel changes to narrowband width of four grid spaces is sufficient to
a foreground region, the region to which it belongs can encapsulate the zero level set at all times as we are
be set by detecting the class to which any neighbour- aiming for the zero level set to be moved by less than one
ing foreground voxels belong. grid space at each iteration. The extended narrowband is
about two grid points wider than the narrowband since it
Efficient storage of ultrasound data is only required to keep the gradient smooth at the edge
Each of the data sets produced using the freehand of the narrowband, such that the level set function can be
ultrasound system detailed in the next section occupies updated accurately all the way to the edge of the nar-
approximately 400 MB of memory (180 images; 640 ⫻ rowband.
480 pixels, of which about 40% are masked off as
nonultrasound content; 1 byte per pixel intensity ⫹ 3 ⫻ EXPERIMENTAL METHOD
4 bytes for floating point position of each pixel in the
world co-ordinate space). Although loading this data into Freehand 3D ultrasound acquisition
memory directly would be possible for high-end com- The freehand ultrasound system used in this work
puters, for easier implementation, some efficient storage consisted of a Faro arm (Faro Technologies Inc., Lake
method was required. However, we can take advantage Mary, FL, USA) rigidly attached to a 5 to 7 MHz
of the particular PDF based segmentation method we transvaginal ultrasound probe connected to a Toshiba
use. Firstly, only the PDFs of a local window and of each PowerVision SSA-270 (Toshiba Medical Systems,
region are required to calculate the similarity measure. Otawara-shi, Tochigi-ken, Japan). A PC workstation
Secondly, calculation of the speed function only occurs (Dell Inc., Round Rock, TX, USA) was used to record
at each node of the level set array. Therefore only the position information while the ultrasound machine re-
intensities associated with each node are stored, rather corded the images independently. The images were
than the pixel positions. When calculating the region transferred in a digital format to the workstation after the
PDF, all points associated with a particular level set node scan for processing. Temporal calibration between the
are attributed to the region to which the node belongs. ultrasound machine and the workstation was performed
This moves away from the goal of segmentation occur- using the method described in (Gooding et al. 2005b).
ring in a world reference space since we have essentially Spatial calibration was performed using a cross-wire
confined the image data to the same voxel array as the phantom as described in (Atkinson et al. 2001). One
level set function. However, this is a result of implemen- hundred eighty images were acquired during each scan at
tation need rather than a requirement of the speed func- approximately 12 Hz using the ultrasound machine cine
tion defined, and as such the speed function could still be function. Further details of the freehand ultrasound sys-
applied in the world reference space. tem can be found in (Gooding et al. 2005b).
error is possible if the follicle is not fully aspirated or, if to 5) for ␣,  and ␥, respectively. Initially, the typical
extra fluid/blood is taken, the volume of aspirate is still a values were used.  was generally increased in cases
good reference for comparison. Reference volumes could where the contrast between follicle and background was
be obtained using MRI (ter Haar Romeny et al. 1999), low. Low contrast could lead to regions with weak
however, this method is not used within routine practice boundary strength. Increasing the smoothing would help
for fertility treatment, making this method unacceptable to constrain the surface in these areas, whilst still en-
for use with patients. abling the data to be segmented in areas where the
The accuracy of the reconstruction system was boundary contrast was stronger. ␥ was decreased in cases
compared with aspirated follicle volumes and the mean where the data was dense and interpolation over sparse
diameter estimate. Local ethics committee approval was areas was not required. Only integer values of each
obtained for this study. Ovaries from 11 consenting parameter were used. A detailed scanning protocol may
patients undergoing IVF treatment were scanned prior to remove the need for tuning  or it may be possible to
oocyte recovery using the 3D system. The mean diameter automate their tuning based on analysis of the image/data
measurement was also made during this scan. All folli- sparsity.
cles within the scanned volume were reconstructed after
the scan. To aid identification of each follicle at the time
RESULTS
of aspiration, three methods were employed: (1) the
clinician who would perform follicle aspiration was The data acquired consisted of 23 follicles from 11
present at the time of the 3D scan: (2) labeled diagrams patients for which aspirated volume, reconstructed vol-
were drawn for each scan by the scanning clinician and ume and diameter measurements were available. A fur-
(3) still images were printed of a number of image planes ther 15 follicles were available for which there were only
showing significant follicles. Although, ideally, all folli- diameter and aspirated volumes because the entire ovary
cles would be measured in each ovary, if an ovary had not been acquired during the 3D scan. This resulted
contained more than about six follicles, identification of from sonographer in-experience at using 3D scanning
follicles at aspiration is very difficult. Therefore, only the equipment and is not considered a drawback of the
follicles for which the aspirating clinician could be con- segmentation method. However, it demonstrates the need
fident of a correct identification had volumes measured for experience with the new scanning methods. The
and recorded during aspiration. This was typically one relatively low number of follicles in this study is a result
follicle per ovary. Aspirated volumes were recorded to of the difficulty in identifying follicles at aspiration once
the nearest 0.5 ml. The assumptions made in recording a the first follicle has been aspirated. This problem was
mean diameter are that the follicle is spherical and that also identified in (Kyei-Mensah et al. 1996b), where only
the diameter is representative of the follicle size and, the largest follicle was measured in each of 25 patients.
hence, its volume. To test this assumption, the volumes Two of the follicles for which all measurements were
of the follicles were estimated from the clinical diameter available were subsequently excluded from the analysis
measurements using a spherical model, for comparison because their reconstructed volume was much larger than
with the aspirated volume. their aspirated volume indicating that the follicle was not
fully aspirated.
Parameters used in reconstructions The results of volume estimation from the diameter
For the experiments presented in this article, recon- measurement and from the reconstruction system are
structions were performed using a level set grid spacing shown against the aspirated volume in Fig. 4. Qualita-
of around 0.8 to 1 mm, depending on the size of the tively, it can be seen that the measurement made by the
ovary. Tests performed at a range of resolutions indi- 3D system has a lower spread than the clinical data.
cated that this order of resolution seemed to provided the However, it is also apparent that both measurement
best segmentation speed without quantising volume mea- methods underestimate the true volume by about 25% to
surements noticeably (1/1000th of a ml per voxel). The 30%. The reasons for the underestimate are considered in
reconstruction method was initialised manually with a the Discussion section. This underestimation was found
small sphere within each follicle. The reconstruction to be significant for both the clinical diameter based
parameters were varied on a patient by patient basis as volume estimation (p ⫽ 1.3 ⫻ 10⫺4, F ⫽ 19.9) and the
they are dependent the ultrasound acquisition. The ultra- computer based method (p ⫽ 1.3 ⫻ 10⫺6, F ⫽ 46.1).
sound machine settings were left to the sonographer to Taking this underestimate into account, the range of
avoid interference with the clinical exam and as a result variation of the clinical data is similar to that found by
were not consistent, leaving the tuning of parameters (Forman et al. 1991).
unavoidable in this instance. The typical values (and The variances of the measurements from the linear
range) were 1.0 (kept constant), 5.0 (5 to 10) and 3.0 (1 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.
Our initial investigation into the accuracy of the Kadir T, Brady JM. Unsupervised non-parametric region segmentation
using level sets. In: IEEE International Conference on Computer
volume measurement using the system found that both Vision. 2003. pp. 1267–1274.
the computer-based measurement and the clinical mea- Krivanek A, Sonka M. Ovarian ultrasound image analysis: follicle
surement consistently underestimate the volume of fol- segmentation. IEEE Trans Med Imag. 1998;17:935–944.
Kyei-Mensah A, Maconochie N, Zaidi J, et al. Transvaginal three-
licular aspirate, as a result of a clinical misunderstanding dimensional ultrasound: reproducibility of ovarian and endometrial
of the true location of the follicular boundary. However, volume measurements. Fertil Steril. 1996a;66:718 –722.
the computer-based system had a lower variance than the Kyei-Mensah A, Zaidi J, Pittrof R, et al. Transvaginal three-dimen-
sional ultrasound: accuracy of follicular volume measurements.
clinical estimate, showing it to provide a better estimate Fertil Steril. 1996b;65:371–376.
of follicular size. Leclerc YG. Constructing simple stable descriptions for image parti-
Areas for current and future work include develop- tioning. Internation Journal of Computer Vision. 1989;3:73–102.
Legget ME, Leotta DF, Bolson EL, et al. System for quantitative
ing a correction for the volume under-estimation and the three-dimensional echocardiography of the left ventricle based on a
clinical study of follicular volume growth. magnetic-field position and orientation sensing system. IEEE Trans
Biomed Eng. 1998;45:494 –504.
Acknowledgment—The authors thank Toshiba Medical Systems Eu- Lie J, Lysaker M, Tai XC. A variant of the level set method and
rope, for the donation of a PowerVision ultrasound machine for this applications to image segmentation. Tech. Rep. CAM 03–50,
project; the staff of the Fertility Unit, John Radcliffe Hospital, UK, for UCLA, Computational and Applied Mathematics. 2003.
their help in the acquisition of clinical results. M.J.G. was funded by Mercier L, Lang T, Lindseth F, Collins LD. A review of calibration
the EPRSC as part of the MIAS-IRC grant (GR/N14248). techniques for freehand 3-D ultrasound systems. Ultrasound Med
Biol. 2005;31:143–165.
Muzzolini R, Yang YH, Pierson RA. Multiresolution texture segmen-
REFERENCES tation with application to diagnostic ultrasound images. IEEE Trans
Med Imag. 1993;12:108 –123.
Ahmad A, Cool D, Chew BH, Pautler SE, Peters TM. 3D segmentation Noble JA, Boukerroui D. Ultrasound image segmentation: A survey.
of kidney tumors from freehand 2D ultrasound. Medical Imaging IEEE Trans Med Imag. 2006;25:987–1010.
Osher S, Sethian JA. Fronts propagating with curvature dependent
2006, Proc. of SPIE. 2006;6141:227–237.
speed: Algorithms based on hamilton-jacobi formulations. Journal
Allott CP, Barry CD, Bramley J, et al. Noninvasive 3-D ultrasound of
of Computational Physics. 1988;79:12– 49.
atherosclerotic plaques in the watanabe rabbit. Ultrasound Med
Paragios N, Deriche R. Coupled geodesic active regions for image
Biol 1999;25:905–914.
segmentation: A level set approach. In: European Conference on
Atkinson D, Burcher M, Declerck J, Noble JA. Respiratory motion
Computer Vision (2). 2000. pp. 224 –240.
compensation for 3-D freehand echocardiography. Ultrasound Med
Peng D, Merriman B, Osher S, Zhao HK, Kang M. A PDE-based fast
Biol 2001;27:1615–1620.
local level set method. Journal of Computational Physics. 1999;
Barry CD, Allott CP, John NW, et al. Three-dimensional freehand
155:410 – 438.
ultrasound: image reconstruction and volume analysis. Ultrasound
Penzias AS, Emmi AM, Dubey AK, et al. Ultrasound prediction of
Med Biol 1997;23:1209 –1224.
follicle volume: is the mean diameter reflective? Fertil Steril. 1994;
Bogush AL, Tuzikov AV. 3D object volume measurement using free- 62:1274 –1276.
hand ultrasound. Lecture Notes in Computer Science. 2005;3540: Potoc̆nik B, Zazula D. Automated analysis of a sequence of ovarian
1066 –1075. ultrasound images, part i: Segmentation of single 2D images. Image
Brunner M, Obruca A, Bauer P, Feichtinger W. Clinical application of and Vision Computing. 2002a;20:217–225.
volume estimation based on three-dimensional ultrasonography. Potoc̆nik B, Zazula D. Automated analysis of a sequence of ovarian
Ultrasound Obstet Gynecol. 1995;6:358 –361. ultrasound images, part ii: Predicion-based object recognition for a
Cigale B, Lenic̆ M, Zazula D. Segmentation of ovarian ultrasound sequence of images. Image and Vision Computing. 2002b;20:227–
images using cellular neural networks trained by support vector 235.
machines. Lecture Notes in Artifical Intelligence 2006;4253:515– Prager RW, Gee AH, Treece GM, Berman L. Freehand 3D ultrasound
522. without voxels: volume measurement and visualisation using the
Delingette H. On smoothness measures of active contours and surfaces. stradx system. Ultrasonics. 2002;40:109 –115.
In: IEEE Work-shop of Variation and Level Set Methods 2001. Rathi Y, Michailovich O, Malcolm J, Tannenbaum A. Seeing the
2001. pp. 43–50. unseen: Segmenting with distributions. In: IASTED Conf on Signal
Fenster A, Downey DB. 3-D ultrasound imaging: A review. IEEE and Image Processing. Acta Press. 2006.
Engineering in Medicine and Biology. 1996;15:41–51. Rohling RN, Gee AH. Issues in 3-D free-hand medical ultrasound
Forman RG, Robinson J, Yudkin P, et al. What is the true follicular imaging. Tech. Rep. CUED/F-INFENG/TR 246, Cambridge Uni-
diameter: An assessment of the reproducibility of transvaginal versity Engineering Department. 1996.
ultrasound monitoring in stimulated cycles. Fertil Steril. 1991;56: Rohling RN, Gee AH, Berman L. A comparison of freehand three-
989 –992. dimensional ultrasound reconstruction techniques. Medical Image
Gee AH, Prager RW, Treece GM, Berman L. Engineering a freehand Analysis. 1999;3:339 –359.
3D ultrasound system. Pattern Recognition Letters. 2003;24:757– Saith RR, Srinivasan A, Michie D, Sargent IL. Relationships between
777. the developmental potential of human in vitro fertilization embryos
Gooding MJ, Barber D, Kennedy SH, Noble JA. The effect of follicle and features describing the embryo, oocyte and follicle. Human
volume measurement on clinical decisions. In: Rueckert D, Hajnal Reproduction Update. 1998;4:121–134.
J, Yang GZ (Eds.), Proceedings of Medical Image Understanding Sapiro G, Tannenbaum A. Area and length preserving geometric in-
and Analysis (MIUA) 2004. 2004. pp. 121–124. variant scale-spaces. IEEE Trans Patt Anal Mach Int. 1995;17:67–
Gooding MJ, Barber D, Kennedy SH, Noble JA. Measurement of the 72.
speed of sound in follicular fluid. Human Reproduction. 2005a;20: Sarty GE, Liang W, Sonka M, Pierson RA. Semiautomated segmenta-
497–500. tion of ovarian follicular ultrasound images using a knowledge-
Gooding MJ, Kennedy SH, Noble JA. Temporal calibration of freehand based algorithm. Ultrasound Med Biol. 1998;24:27– 42.
three-dimensional ultrasound using image alignment. Ultrasound Sethian JA, Level set methods and fast marching methods, 2nd Edition.
Med Biol. 2005b;31:919 –927. Cambridge University Press. 1999.
Volume segmentation and reconstruction from 3D ultrasound data ● M.J. GOODING et al. 195
Song M, Haralick RM, Sheehan FH, Johnson RK. Integrated surface Vese LA, Chan T. A multiphase level set framework for image seg-
model optimization for freehand three-dimensional echocardiogra- mentation using the Mumford and Shah model. International Jour-
phy. IEEE Trans Med Imag. 2002;21:1077–1090. nal of Computer Vision. 2002;50:271–293.
Tao Z, Tagare HD, Beaty JD. Evaluation of four probability distribu- Ye X, Noble JA, Atkinson D. 3-D freehand echocardiography for
tion models for speckle in clinical cardiac ultrasound images. IEEE automatic left ventricle reconstruction and analysis based on mul-
Trans Med Imag. 2006;25:1483–1491. tiple acoustic windows. IEEE Trans Med Imag. 2002;21:1051–
Tasdizen T, Whitaker R, Burcahrd P, Osher S. Geometric surface 1058.
smoothing via anisotropic diffusion of normals. In: IEEE Visual- Zahalka A, Fenster A. An automated segmentation method for three-
ization 2002. 2002. pp. 125–132. dimensional carotid ultrasound images. Phys Med Biol 2001;46:
ter Haar Romeny BM, Titulaer B, Kalitzin S, et al. Computer assisted 1321–1342.
human follicle analysis for fertility prospects with 3d ultrasound.
Zhang Y, Rohling R, Pai DK. Direct surface extraction from 3D
1999;1613:56 – 69.
freehand ultrasound images. Proceedings of IEEE Visualization
Treece GM, Prager RW, Gee AH, Berman L. Volume measurement
2002. 2002.
from sequential free-hand 3D ultrasound. Lecture Notes in Com-
puter Science. 1999;1613:70 – 83. Zhao HK, Osher S, Merriman B, Kang M. Implicit and non-parametric
Treece GM, Prager RW, Gee AH, Berman L. Surface interpolation shape reconstruction from unorganized data using a variational
from sparse cross sections using region correspondence. IEEE level set method. Computer Vision and Image Understanding.
Trans Med Imag. 2000;19:1106 –1114. 2000;80:295–319.
Treece GM, Prager RW, Gee AH, Berman L. 3D ultrasound measure- Zhu SC, Yuille AL. Region competition: Unifying snakes, region
ment of large organ volume. Medical Image Analysis. 2001;5:41– growing, and bayes/MDL for multiband image segmentation. IEEE
54. Trans Patt Anal Mach Int. 1996;18:884 –900.
Tsai YR, Cheng LT, Osher S, Zhao HK. Fast sweeping algorithms for Zimmer Y, Tepper R, Akselrod S. A two-dimensional extension of
a class of hamilton-jacobi equations. SIAM Journal on Numerical minimum cross entropy thresholding for the segmentation of ultra-
Analysis. 2003;41:673– 694. sound images. Ultrasound Med Biol. 1996;22:1183–1190.