SIMULTANEOUS FOCUSING AND CONTOURING OF HUMAN
ZYGOTES FOR IN VITRO FERTILIZATION
Alessandro Giusti, Giorgio Corani, Luca Gambardella
Dalle Molle Institute for Artificial Intelligence, SUPSI and University of Lugano, Switzerland
Cristina Magli
International Institute for Reproductive Medicine, Lugano, Switzerland
Luca Gianaroli
INFERGEN, Lugano, Switzerland
Keywords:
In vitro fertilization, Microscopy, Phase contrast, Hoffmann modulation contrast, Segmentation, Contouring,
Autofocus.
Abstract:
Observation of ovocytes and zygotes plays an important role in In Vitro Fertilization procedures, and is usually
perfomed by means of a microscope equipped with Hoffman Modulation Contrast optics, which produces
images with a complex, side-lit, 3D-like appearance. Our algorithm operates on a Z-stack of such images taken
at different focal planes, and simultaneously identifies: a) a repeatable, meaningful focal plane corresponding
to the cell’s equator line, and b) the external contour of the cell. As the cell is a thick stucture with respect to
the microscope depth of field, the two problems are nontrivial and deeply related. Our algorithm is also robust
to other structures, clutter and artefacts affecting the images and lying at varying focal planes. We describe
implementation details, applications and experimental results of our technique.
1 INTRODUCTION
During In Vitro Fertilization (IVF) procedures, biolo-
gists observe zygotes (fertilized ova) at different times
in order to assess their quality and select the ones
maximizing the implantation success rate (Gianaroli
et al., 2007). Such observations are usually performed
by means of an inverted microscope with Hoffmann
Modulation Contrast and 20x or 40x magnification,
where focusing plays an important role as the depth
of field is quite shallow, so that only a thin slice of
the cell’s volume is clearly visible at a single focal
plane. We solve the problem of automatically find-
ing a repeatable and sensible focal plane from a set
of images taken at different focal planes (Z-stack), by
means of an original approach where high-level scene
information (the cell contour) is recovered simulta-
neously with the correct focal plane (see Figures 1
and 2).
Traditional autofocusing algorithms based on
global or local contrast fail at consistently finding a
meaningful and repeatable focus for the zygote ob-
servation: in fact, many images in a Z-stack exhibit
sharp, strong-contrast features. Structures of interest
such as pronuclei and polar bodies can be found at
different depths, and therefore come into focus at dif-
ferent levels (see Figure 1); a number of other struc-
tures, such as fragments and debris, are also visible at
varying depths, both inside and outside the cell, rep-
resenting strong and confusing clutter for automated
processing.
We maintain that a proper solution to the focus-
ing problem requires an higher-level understanding of
the scene. Our approach is therefore aimed at solving
both of the following problems:
segment the zygote cell, and
find a focal plane where the sharpness of its con-
tour is maximized, which roughly corresponds to
the plane of the cell’s contour generator curve.
The two tasks are deeply related: in fact, the cell’s ac-
tual contour can only be recovered in an image where
it is sharply visible, and such image can only be easily
determined when the contour itself is known.
We solve both problems simultaneously by repre-
senting data from all focal planes in a single graph,
151
Giusti A., Corani G., Gambardella L., Magli C. and Gianaroli L. (2010).
SIMULTANEOUS FOCUSING AND CONTOURING OF HUMAN ZYGOTES FOR IN VITRO FERTILIZATION.
In Proceedings of the Third International Conference on Bio-inspired Systems and Signal Processing, pages 151-158
DOI: 10.5220/0002592501510158
Copyright
c
SciTePress
(a) I
1
(b) I
6
(c) I
14
(d) I
19
(e) I
23
Figure 1: The source stack is composed by a number of images (slices) taken at different focus levels. Our experiments use
23-slice stacks; 5 slices covering the whole range of a typical stack are shown in the figure. Note in-focus debris generating
strong contrast in (d,e), whereas the best focus of the cell is in (c).
then solving a minimum-cost path problem on such
graph. Sharp, in-focus edges are favored by assign-
ing a lower cost to localized, large gradient magni-
tudes. We also take advantage of the complex, 3D-
like sidelit appearance of the zygote in HMC images
for improving the system’s robustness, by enforcing
priors about the cell lighting, which is a predictable
and prominent feature. Finally, the peculiar topology
of the graph on which the minimum-cost path prob-
lem is solved implicitly enforces:
shape priors on the cell contour, which is expected
to be smooth and circular-shaped;
limited and smooth variations in focus along the
contour.
A similar approach has been used for robust zy-
gote segmentation from a single image in (Giusti
et al., 2009), where the cell is assumed to be already
correctly focused, which implies manual intervention
in the image acquisition process.
On the contrary, our approach removes any in-
fluence of the operator in the focusing process; this
is especially important as the focal plane found by
our algorithm roughly bisects the cell in a top and
bottom half. This acts as an important reference for
determining the relative position of other structures
of interest. In addition, the knowledge of the zy-
gote contour is useful to readily compute a number
(a) I
13
(b) I
14
Figure 2: Results for the stack in Figure 1. Our algorithm
detects the contour γ of the zygote (yellow line). It is best
focused (thick line) partly in I
13
(lower-left part) and partly
I
14
(upper-right segment).
of quantitative measures (apparent size, simple shape
descriptors, relative positions) for the cell, which are
not easily judged otherwise. The obtained segmen-
tation and focal plane may also be applied for other
tasks, such as driving an automated microscope for
unattended imaging of zygotes, or providing a ro-
bust, precise initialization for subsequent (automatic
or user-assisted) analysis algorithms, such as those
introduced in (Morales et al., 2008; Karlsson et al.,
2005) for the segmentation of the zona pellucida.
We briefly review related works in the following
Section, then introduce notation and terminology in
Section 3. Our approach is then described in detail
in section 4, and experimentally validated in in Sec-
tion 5. Section 6 concludes the paper and presents
ongoing work.
2 RELATED WORKS
Autofocus systems are widely used in microscopy,
and can be implemented either by means of additional
hardware, or by using a software approach for ana-
lyzing a sequence of camera images taken at varying
focus positions; our system belongs to the latter cat-
egory. Many approaches for software autofocus are
documented in literature (Shih, 2007), mostly based
on the maximization of local or global contrast: the
core difference of our approach is that we seek an
higher level understanding of the scene, thus provid-
ing a repeatable focusing position which is linked to a
specific focal plane which bisects the cell in a top and
bottom half. As the sample is thick with respect to
the depth of field, and many other sources of contrast
exist in different focal planes, the same goal can not
be attained by a lower level analysis.
Our approach is based on the segmentation of the
cell contour, which by itself is not a straightforward
task even if the correct focal plane was known, due to
the complexity of the image.
Classical region-based segmentation algorithms,
including watersheds (Soille and Vincent, 1991), are
BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing
152
not applicable in this context because of the complex
appearance of the cell, including the surrounding zona
pellucida, clutter, and artifacts; this also hinders the
application of straightforward edge-based segmenta-
tion algorithms, as many spurious contours are de-
tected.
Iterative energy minimization methods such as ac-
tive contours (Xu, 1998) and level sets are frequently
employed in biomedical imaging: in this context,
their application is not straightforward because de-
bris are likely to generate several local minima in
the energy function, which makes quick and robust
convergence problematic; for example, in (Morales
et al., 2008) active contours are used for measuring
the thickness of the zona pellucida in embryo images,
but only after a preprocessing step aimed at removing
debris and other artifacts.
In (Beuchat et al., 2008) a semisupervised tech-
nique for measuring various zygote features is used,
where the cell shape is approximated by an ellipse:
in our case, instead, we recover the actual shape of
the cell, which is often not well approximated by an
ellipse.
The technique we are presenting includes a global
energy minimization step, and may be classified as a
specialized graph-cut (Zabih and Kolmogorov, 2004)
approach, where: a) priors on the cell shape are ac-
counted for by operating on a spatially-transformed
image and searching for a minimum-cost path on a
directed acyclic graph; b) priors on the contour ap-
pearance due to HMC lighting are directly integrated
in the energy terms; c) information at different fo-
cal planes is simultaneously represented in a single
large graph. We are therefore extending the approach
in (Giusti et al., 2009) by operating on information
from many focal planes at the same time.
Interestingly, several previous works handled the
peculiar lighting in HMC and DIC images as an ob-
stacle to segmentation (Kuijper and Heise, 2008),
and adopted preprocessing techniques for removing
it, whereas we actually exploit such appearance for
improving robustness.
3 PRELIMINARIES, MODEL AND
NOTATION
Our algorithm is designed to operate on a Z-stack
of N images taken with Hoffman Modulation Con-
trast (HMC) microscopy
1
. We denote the input im-
1
a technique delivering visually similar results is Dif-
ferential Interference Contrast (DIC), which is also a likely
application scenario for our technique.
ages as I
1
,I
2
,...,I
N
, and their respective focal planes
z = z
1
,z
2
..z
N
. Such focal planes can be considered
horizontal slices at different depths of a 3D space
whose cartesian axes are (x,y, z).
HMC is an imaging technique converting optical
slopes to variations of the light intensity: it is rou-
tinely used in IVF labs for observing zygotes, as it
provides a large amount of contrast for transparent
specimens and eases human observation as the objects
appear three-dimensional and side-lit, as if a light
source was illuminating them from a side (apparent
lighting direction).
The underlying imaging model is considerably
complex, especially if the effect of out-of-focus fea-
tures is taken into account. Still, several intuitive prin-
ciples hold, on which we base our approach:
structures which lie on or near the current focal
plane z
i
appear sharp and exhibit strong localized
gradients in the image intensity I
i
;
as the focal plane depth moves farther from the
structure’s depth, the structure image becomes
blurred. Consequently, its gradients of the struc-
ture’s image lose locality and strength, although
the global contrast and visibility of the feature
may not be affected, or may even be emphatised
in some situations
2
.
In this work, the main feature of interest is the cell
contour; as the cell is a 3D object, in order to explain
the appearance of its contour at different focus levels,
we provide the following formalization: Let S be the
surface of the cell, which we assume to be smooth,
in the 3D space (x,y,z). The contour generator curve
Γ is a curve in 3D space, identified by the locus of
points P on S such that the tangent plane to S in P
contains the z direction
3
. Although this definition al-
lows Γ to be composed by several disjoint curve parts,
the regularity of the cell shape, which is convex and
ellipsoid-like, allows us to assume that Γ is a single,
closed curve in the following.
We are interested in detecting the image of the
contour generator curve Γ in our input images I
i
. In
particular, let γ be the 2D apparent contour, i.e. the or-
thogonal projection of Γ on the (x, y) plane. Follow-
ing the principles introduced previously in this sec-
tion, a part of γ is visible and well-focused in an image
I
i
if the corresponding part of Γ is on or near the z = z
i
2
in fact, a slightly defocused feature imaged through a
phase contrast technique may appear more evident to an hu-
man operator than the same feature in perfect focus; this
makes manual focusing inherently operator-dependent and
hardly repeatable.
3
Note that this is similar to the concept of contour gen-
erator curve in projective geometry where an ortographic
camera is considered.
SIMULTANEOUS FOCUSING AND CONTOURING OF HUMAN ZYGOTES FOR IN VITRO FERTILIZATION
153
Figure 3: Representation of model (see text).
plane; in this case, such part of γ will exhibit large, lo-
calized gradients in image I
i
. The gradient intensity is
weaker as Γ gets farther away from the plane z = z
i
;
eventually, if a part of Γ lies far from the plane z = z
i
,
then the corresponding part of γ may be invisible (i.e.
not generating any significant gradient) in I
i
.
Our goal is to identify γ in the (x, y) image coordi-
nates, as well as the I
m
image where γ is most visible,
whose focal plane z
m
corresponds to the depth of Γ.
As described in the following Section, we account for
the fact that different parts of Γ may lie at different
depths, by detecting different parts of γ on different I
i
images.
4 SIMULTANEOUS FOCUSING
AND CONTOURING OF A
ZYGOTE CELL
We divide the segmentation process in two sequential
steps: first, we find the approximate location (x, y) =
(c
x
,c
y
) of the cell center; in doing this, we assume
that a single zygote is visible in the image, which is
always the case as zygotes are kept in separate wells
in clinical practice.
Then, we build a transformed representation of
the whole Z-stack in polar coordinates, constructing
a single graph, then using a minimum-cost path for-
mulation in order to recover the actual zygote contour
and its focus plane.
We briefly introduce the former part, which we
consider of lesser importance and interest, in Section
4.1. The main focus is instead on the latter part, de-
scribed in Section 4.2.
4.1 Approximate Localization of Zygote
Center
First, a representative image for the whole stack is se-
lected, by applying a naive autofocusing algorithm: in
particular, for each slice i we compute a number f (i)
as the average value for the modulo of the gradient of
I
i
. We consider the slice b maximizing such value:
b = argmax
i
( f (i)). (1)
Due to the gradient in the internal part of the
cell, and strong gradients due to debris and additional
structures, I
b
does not represent in general the slice
where γ is most visible (see Section 5). However, it
proves to be a suitable image for applying the same al-
gorithm described in (Giusti et al., 2009) for roughly
detecting the cell centroid.
In particular, in order to find an approximate loca-
tion for the cell centroid, the modulo of the image gra-
dient of I
b
is subsampled to a smaller image, which is
automatically thresholded then regularized by means
of median filtering (see Figure 4 a,b). The largest con-
nected component is isolated and its holes filled (Fig-
ure 4 c); for each point inside the resulting region,
the minimum distance to the region boundary is com-
puted by means of the distance transform; the point
with the maximum distance is finally chosen as the
approximate centroid of the cell.
This preliminary analysis phase is not critical for
the quality of results, as the subsequent processing
tolerates quite large displacements of the detected
centroid; nonetheless, this simple algo- rithm counter-
intuitively proves to be quite robust also in presence
of large artifacts attached to the cell; this is mainly
due to the distance transform, which implicitly can-
cels or reduces the effect of any non-convex artifact
protruding of the border of the cell.
4.2 Contouring and Focusing of the
Zygote
4.2.1 Transformation to Polar Coordinates
Once the zygote centroid (c
x
,c
y
) is detected in one
slice, it is considered valid in all the images of the
stack (as no significant displacements happens be-
tween focus levels). For each image I
i
, a circular
corona centered at such point is transformed using
BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing
154
(a) (b) (c) (d) (e)
Figure 4: Approximate localization of the cell center. (a): original image. (b): binary mask obtained after thresholding the
modulo of the gradient. (c): largest connected component with holes filled. (d): distance transform. (e) the maximum of
the distance transform is considered as the approximate center of the cell. Note that the large artifact on the left does not
significantly displace the maximum of the distance transform.
bilinear interpolation to an image J
i
in polar coordi-
nates:
J
i
(θ,ρ) = I
i
(c
x
+ ρcos(θ),c
y
+ ρsin(θ))
0 θ < 2π ρ
0
ρ ρ
00
i = 1 · · · N.
(2)
In order to account for variations in the cell shape
and errors in the centroid location, the range [ρ
0
÷ ρ
00
]
of ρ values is very conservatively set to [0.3r,1.5r],
where r represents the expected cell radius; this is a
quite large range (see Figure 4e), which allows for
large variations in the actual radius of the zygote, and
for displacements of the estimated centroid (c
x
,c
y
).
ρ and θ values are uniformly sampled in ρ
n
and θ
n
intervals, respectively, which correspond to rows and
columns of each image J
i
. We use ρ
n
= 80, θ
n
= 180
in the following.
4.2.2 Computation of Energies
Images J
i
are then processed in order to associate an
energy to each pixel for each of the N planes. Such
energy will drive the following graph-based formula-
tion. Let α be the direction of apparent lighting due
to HMC, which only depends on the optical setup and
can be assumed known in most scenarios (if it’s not,
it can be easily estimated); we define an energy E
i
for
each pixel of J
i
, regardless on its plane in the stack,
as:
E
i
(θ,ρ) = P
z }| {
cos(θ α) · G
ρ
(J
i
)+
z }| {
sin
2
(θ α) ·
G
ρ
(J
i
)
!
(3)
P(x) =
1 + e
x
k
1
(4)
where G
ρ
denotes the gradient operator along the ρ
axis, and P(·· · ) is a simple decreasing sigmoid func-
tion which conditions the energy values to lie in the
[0 ÷ 1] interval; the scaling parameter k is not critical,
and can be safely set to 1/5 of the image’s dynamic
range.
The first term in (3) dominates where the con-
tour is orthogonal to the apparent light direction,
i.e. where the cell is expected to appear signifi-
cantly lighter (θ α ' 0) or darker (θ α ' ±π) than
the surroundings; large gradient values with a sign
consistent with this assumption lead to lower ener-
gies. The second term takes account for the unpre-
dictability of the contour appearance where the con-
tour is parallel to the apparent light direction, and
just associates lower energies to large absolute values
for G
ρ
(J
i
).
4.2.3 Building of a Directed Acyclic Graph
A single directed acyclic graph is built over the stack
of all J
i
images, by instantiating a node for each pixel
and for each plane (for a total of ρ
n
· θ
n
· N nodes).
Moreover:
arcs are added connecting each node to its three
8-neighbors at the right on the same plane;
every β
n
columns, a set of interfocal arcs are
added. In particular,
The cost of each arc is set to the energy E
i
(θ,ρ) of its
source node.
A single global source node s is added, with zero-
cost arcs that lead to every pixel in the first column
of every plane; also, we add a sink node reached by
zero-cost arcs from every pixel at the last column of
every plane.
4.2.4 Minimum-Cost Path
As the resulting graph is a directed acyclic graph,
efficient algorithms are available for computing the
minimum-cost path from the source node to the sink
node. Such path passes through low-energy arcs, and
is constrained by the graph topology to have a quite
regular shape, because:
SIMULTANEOUS FOCUSING AND CONTOURING OF HUMAN ZYGOTES FOR IN VITRO FERTILIZATION
155
Figure 5: The stack of I
i
slices (a) is transformed to J
i
slices in polar coordinates (b), then an energy value E
i
(c) is computed
for each pixel and a graph is built on each (d). A single global graph is then computed.
the path is forced to steadily move from left to
right (i.e. increasing θ values);
variations along the radial direction ρ are bounded
in slope;
shifts to an adjacent focal plane can only occur
rarely (e.g. every β
n
columns).
After excluding the source and sink nodes (which
have no geometric meaning), the path can be brought
back to cartesian coordinates by using the inverse
transform to (2): the result is a curve in the (x,y, z)
space. If the first and last nodes of the path (almost)
match, the curve can be smoothly closed. If such
points do not match, we have a strong hint that the
image did not contain a zygote, or that the approx-
imate center was very displaced with respect to the
true center of the zygote; we are currently disregard-
ing this possibility as none of our test images exhibits
this issue, although we plan to investigate the related
problem of the zygote detection in future works.
In practice, the resulting path estimates the 3D
contour generating curve Γ. It simultaneously rep-
resents the contour of the cell and identifies its main
focal plane: the interior of the resulting polygon pro-
jected to the 2D plane (x,y) defines the computed bi-
nary mask M. The average value for the focus plane
associated to the nodes in the minimum-cost path de-
fines the focus plane z
m
where the cell contour is
sharpest.
Larger values for ratio θ
n
/ρ
n
, as well as a smaller
β
n
parameter, allow more freedom to the path built
over the graph, which translates to better accommoda-
tion of an irregular cell shape or a displaced centroid
(c
x
,c
y
); at the same time, this reduces the robustness
of the approach, as shape priors are less strongly en-
forced. We found any ratio between 1.5 and 3.0 to be
acceptable, although we keep with θ
n
/ρ
n
= 180/80 =
2.25 in the following; changing β
n
has little effect in
most cases, and should be set such that the ratio θ
n
/β
n
is larger than 8 i.e. there are at least 8 places in the
graph where the interfocal arcs are created which
allows the resulting contour generator curve to span
at most 4 adjacent focal planes. Due to the almost
spherical shape of the zygotes, none of our test im-
ages required a larger variation
4
.
5 EXPERIMENTAL RESULTS
We evaluated the technique on 101 image stacks rep-
resenting 84 unique zygotes from 22 different patients
(some zygotes are acquired twice). Stacks are ac-
quired with a 0.35 megapixel JVC camera attached
to an Olympus IX-51 inverted microscope equipped
with a 20x objective, HMC optics, and a 0.63x cam-
era adapter. Each stack is composed by 23 images,
acquired in a rapid sequence (10 frames per second)
during a regular motion of the focus knob; due to the
manual nature of the stack acquisition, the spacing be-
tween adjacent slices of a stack is not exactly fixed
5
,
but consistently averages between 8 and 12 microns;
the whole stack covers approximately 200 microns,
which amounts at twice the expected diameter of the
cell. The contour generator curve Γ of the cell is al-
ways included in the range of focus, although not in
a predictable position. The cell is always completely
included in the image, but is not well centered in the
image as the biologist acquiring the images uses the
microscope oculars and thus does not have a reference
for the center of the image acquired by the camera.
We can now quantitatively evaluate the effective-
ness of our focusing technique. We compare z
m
, the
average focus plane of Γ determined by our algorithm,
to a reference focus depth z
ref
, obtained by an operator
carefully navigating the stack and choosing the slice
maximizing the overall contrast and precision of the
cell contour. We consider z
ref
as the ground truth.
We also compare two additional estimators for
z
ref
:
z
quick
, obtained by an operator rapidly focusing a
stack (in about 2 seconds on average) by means
4
note that if the cells were perfectly spherical then Γ
would lie on a single focal plane
5
note that fixed spacing is never assumed in our ap-
proach; still, equally-spaced slices, which we plan to ac-
quire by means of automated microscopes, will prove very
useful for performing 3D measurements on the resulting
stacks
BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing
156
(a) (b) (c) (d)
Figure 6: Some images focused and simultaneously segmented by our approach. γ shown in yellow. Thin white lines delimit
the circular corona around the detected approximate center (c
x
,c
y
).
Table 1: Evaluation of focusing performance (see main text). First row reports the error in terms of number of z-stack slices;
second row expresses the approximate value in µm, assuming a fixed distance between adjacent slices of 10µm.
z
m
z
quick
z
autofocus
RMSE [slices] 1.26 3.22 3.90
RMSE [µm] 12.6 32.2 39.0
Table 2: Evaluation of segmentation performance (see main text).
q d
a
d
m
e
a
e
e
pixels fraction of R pixels fraction of R
0.949 2.625 0.020 5.823 0.044 0.017 0.077
of the mouse wheel, with the goal of getting an
acceptable contrast.
z
autofocus
, the measure computed by the naive auto-
focusing algorithm which maximizes the average
modulo of the gradient in the image; we use such
measure in order to find the approximate cell cen-
ter (see Section 4.1). We evaluated other simple
histogram-based autofocusing algorithms, which
performed consistently worse.
The root mean squared error (RMSE) with respect
to z
ref
, computed on all 101 stacks, is reported in Ta-
ble 1 for each of z
m
, z
quick
, and z
autofocus
.
On average, the contour generator curve spanned
a range of 1.71 slices (u 17µm).
In order to evaluate the quality of the segmenta-
tion, we randomly selected 40 of the 101 stacks and
created a ground truth binary mask T for each, by con-
sidering the slice with best focus as returned by our al-
gorithm and manually segmenting the contour of the
zygote on such slice. We then compared this ground
truth value to the segmentation returned by algorithm,
represented by a binary mask M, by considering:
the Jaccard quality metric q =
|T M|
|T M|
0 q 1,
which approaches 1 for better segmentations;
the average distance d
a
and maximum distance d
m
between the true boundary and the computed one;
the relative error in the measured area e
a
;
the absolute error in the measured eccentricity e
e
.
The results, shown in Table 2, are consistent with
those reported in (Giusti et al., 2009), where a sim-
plified version of the same algorithm is applied to one
single image with the contour in good focus.
6 DISCUSSION, CONCLUSIONS
AND RELATED WORKS
We presented an effective technique for simultane-
ously focusing an human zygote cell and recovering
its contour. We maintain that, in this scenario, the
problem of precise and repeatable focusing is strictly
tied to the segmentation problem, as low-level fo-
cus measures which consider the image as a whole
are easily misled by additional structures and debris
which exhibit strong contrast. Our approach segments
the cell and determines its focal plane in a single step,
by operating on a single graph summarizing informa-
tion from all the slices in the stack.
We have shown that our focusing algorithm is able
to determine a repeatable focal plane, which has a
SIMULTANEOUS FOCUSING AND CONTOURING OF HUMAN ZYGOTES FOR IN VITRO FERTILIZATION
157
meaningful interpretation as the average depth of the
contour generator curve of the cell. The precise and
automated measurement of such depth is a fundamen-
tal step for providing accurate and repeatable mea-
sures of the zygote morphology which also account
for 3D information. Obviously, this requires operat-
ing on stacks whose slices have a known depth, which
we are going to obtain in the future. We are currently
improving the system by automatically segmenting,
measuring and locating in 3D other zygote structures
such as the pronuclei. This has important applications
in the clinical practice for In Vitro Fertilization.
REFERENCES
Beuchat, A., Th
´
evenaz, P., Unser, M., Ebner, T., Senn,
A., Urner, F., Germond, M., and Sorzano, C. (2008).
Quantitative morphometrical characterization of hu-
man pronuclear zygotes. Human Reproduction,
23:1983–1992.
Gianaroli, L., Magli, M. C., Ferraretti, A. P., Lappi, M.,
Borghi, E., and Ermini, B. (2007). Oocyte euploidy,
pronuclear zygote morphology and embryo chromo-
somal complement. Human Reproduction, 22(1).
Giusti, A., Corani, G., Gambardella, L., Magli, C., and Gi-
anaroli, L. (2009). Segmentation of human zygotes
in hoffman modulation contrast images. In Proc. of
MIUA.
Karlsson, A., Overgaard, N. C., and Heyden, A. (2005).
Scale Space and PDE Methods in Computer Vision,
chapter A Two-Step Area Based Method for Auto-
matic Tight Segmentation of Zona Pellucida in HMC
Images of Human Embryos, pages 503–514. Springer.
Kuijper, A. and Heise, B. (2008). An automatic cell seg-
mentation method for differential interference contrast
microscopy. In Proceedings of ICPR.
Morales, D., Bengoetxea, E., and Larra
˜
naga, P. (2008). Au-
tomatic segmentation of zona pellucida in human em-
bryo images applying an active contour model. In
Proc. of MIUA.
Shih, L. (2007). Autofocus survey: a comparison of algo-
rithms. Digital Photography.
Soille, L. and Vincent, P. (1991). Watersheds in digi-
tal spaces: an efficient algorithm based onimmersion
simulations. IEEE Transactions on Pattern Analysis
and Machine Intelligence.
Xu, C. (1998). Snakes, shapes, and gradient vector flow.
IEEE Transactions on Image Processing.
Zabih, R. and Kolmogorov, V. (2004). Spatially coherent
clustering using graph cuts. In Proc. of CVPR.
BIOSIGNALS 2010 - International Conference on Bio-inspired Systems and Signal Processing
158