Image Anal Stereol 2003;22:81-89
Original Research Paper
MORPHOLOGICAL QUANTIFICATION OF AORTIC CALCIFICATION
FROM LOW MAGNIFICATION IMAGES
J ES ÚS A NGULO1, T HAO N GUYEN -K HOA2, Z IAD A. M ASSY2, T ILMAN D R ÜEKE 2 AND
J EAN S ERRA1
1
Centre de Morphologie Mathématique, Ecole des Mines de Paris, 35, rue Saint-Honoré, 77305 Fontainebleau,
France, 2 Laboratoire INSERM U 507, Hôpital Necker, 161, rue de Sèvres, 75015 Paris, France
e-mail: angulo,serra @cmm.ensmp.fr,massy@necker.fr
(Accepted May 14, 2003)
✁
ABSTRACT
Atherosclerotic and medial vascular calcifications are frequent in chronic renal failure patiens and predict their
increased cardiovascular mortality. Experimental models for mice have been recently developed in order to
study these disorders. The aim of this paper is to present the morphological image processing algorithms
developed for the semi-automated measurement of calcification from sections of aorta stained using von
Kossa’s silver nitrate procedure and acquired at low magnification power ( 2 5) on colour images. The
approach is separated into two sequential phases. First, the segmentation is aimed to extract the calcification
structures and on the other hand to demarcate the region of the atherosclerotic lesion within the tissue. The
segmentation yields the image data which is the input to the second phase, the quantification. Calcified
structures are measured inside and outside the lesion using a granulometric curve which allows the calculation
of statistical parameters of size. The same operator computes the shape of the lesion. The relative proportion
of the area of calcification is also calculated respectively for the atherosclerotic lesion area and the area outside
such lesions. In conclusion, the here developed method allows quantification of vascular calcified deposits in
mouse aorta. This method will be useful for the quantitative assessment of pathological vascular changes in
animals and man.
✂
✄
Keywords: aortic calcification, automation in bioimaging, low magnification histology, mathematical
morphology, nephrology, quantitative image analysis.
INTRODUCTION
widely used for detection of calcifi cation (Lardenoye
et al., 2002).
Atherosclerotic and medial vascular calcifi cations
are frequent in chronic renal failure patients and
predict their increased cardiovascular mortality.
Experimental models for mice have been proposed
in order to study these disorders (Muntzel et al.,
2002). The detection of coronary artery calcium,
using mainly fluoroscopy, has been shown to be of
value in predicting artery disease (Margolis et al.,
1980; Reinmuller and Lipton, 1987; Tanenbaum et
al., 1989). Recently, electron beam tomography has
been introduced for the detection of coronary calcium
(Agatston et al., 1990; Morgan-Hughes et al., 2002).
In previous studies, mathematical morphology
has shown its ability for solving bioimaging
problems from low magnifi cation power images
in haematological cytology (Angulo and Flandrin,
2003). Such morphological approaches have also been
applied successfully to other issues in nephrology
(Moreso et al., 1994; Seron et al., 1996).
MATERIALS AND METHODS
STAINED AORTIC LOW MAGNIFICATION
IMAGE
The purpose of the present work is to propose an
image-based method of quantifi cation for the uremiaenhanced vascular calcifi cation in the aortic root,
which allows the evaluation of previously established
biological models.
The model of apolipoprotein E gene
knockout mice has been initially created by
homologous recombination in embryonic stem
cells. This genetically engineered mouse generates
The images were derived from thin sections
atherosclerotic lesions within weeks after birth that
of aortic tissue. The aim is the extraction and
are similar to those found in humans (Zhang et al.,
quantifi cation of calcifi cation deposits. Confi rmation 1992; Ishibashi et al., 1994). We created chronic renal
of the presence of calcifi cation is provided by staining failure in these mice by cortical electrocauterization
in one kidney and ablation of the contralateral kidney
using von Kossa’s silver nitrate method which is
81
A NGULO J ET AL : Morphological quantification of aortic calcium
two weeks later. After sacrifi ce, the aortic arch was cut
into serial sections in a cryostat. Sections were stained
with von Kossa’s silver nitrate method. The animals
were treated according to the recommendations of
animal care committees, under anesthesia (Protocol:
short term anesthesia of a duration of 30-40 min; one
IP injection composed of 75 µl Rompun 2% in 25 ml
vial, 300 µl Ketamine CLORKETAM 1000 in 10 ml
vial, 1.53 ml 0.09% NaCl; the amount of the anesthetic
fluid administered 100 120 µl / 20 g body weight).
(a)
(b)
Fig. 1. Images under two magnification powers
10) of a section of the aortic
( 2 5 and
sinus showing the elements of interest: tissue,
atherosclerotic lesion and calcification structures.
✁
✁
✂
In Fig. 1 are depicted two colour images under
two magnifi cation powers ( 2 5 and 10) of a
section of the aortic sinus showing the elements of
interest: tissue, atherosclerotic lesion and calcifi cation
structures. In this study, the series of colour images
were acquired under very low magnifi cation power at
2 5. The main advantage is that in one image fi eld all
the tissue to be examined is included; however there
is a considerable drawback since the quality of the
structures is mediocre. Therefore the automation of the
segmentation and quantifi cation procedures involves
the development of specifi c algorithms using advanced
techniques of image analysis. This is the rationale
behind the use of mathematical morphology operators.
As one can see from Fig. 1, other alternatives are
possible. For instance, working at the magnifi cation
power of 10 and using a motorized microscope is
an alternative to acquire several subimages to cover
the whole area of interest. This approach has however
serious drawbacks including defi nition of lesion at
10, extraction of the whole calcifi cation region and
merger of quantifi ed data (overlapping of images).
✁
(c)
✁
✂
✁
✂
(d)
✁
(e)
✁
Fig. 2. Example of microscopic field image from a
section of aorta stained using the von Kossa method
at low magnification power ( 2 5): (a) RGB colour
image fRGB , (b) red component fR , (c) green component
fG , (d) blue component f B , (e) saturation component
fS .
The size of the image is 672 538 pixels on a
rectangular raster with 256 grey tone levels for each
red, green and blue colour channels. We consider for
quantifi cation that at 2 5 magnifi cation and for the
resolution used, the size of a pixel is 5 5 µm2 pixel.
✁
✁
✂
✁
✂
✄
✂
82
Image Anal Stereol 2003;22:81-89
–
Fig. 2 is an example of colour image to be
processed, fRGB . The red fR , green fG and blue
fB colour component images are shown and beside
this, the saturation component f S of a 3D-polar
coordinate colour representation✁ (Hanbury and Serra,
✁
2002).
✁ ✁ For✁ a colour
✁ pixel c r✂ g ✂ b ✄ ; i.e. fRGB ✁ c ✄☎
fR c ✄✆✂ fG c ✄✆✂ fB c ✄✝✄ the saturation coordinate f S c ✄ is
obtained by the simple expression
✁
✁
fS c ✄✞ max r✂ g ✂ b ✄
✁
min r✂ g ✂ b ✄
✂
–
–
–
In order to detect the tissue in the fi eld image the
simplest means is to take f S where the biological
structures are clearly defi ned, despite the shading
effect (non-uniform illumination).
–
✁
f x
y✄
☞
✁
✁
opening: γ B f ✄✑ δB ✒ εB f ✄✔✓
✁
✁
closing: ϕ B f ✄✞ εB ✒ δB f ✄✔✓
The morphological openings (closings) fi lter out light
(dark) structures from the images according to the
predefi ned size and shape criterion of the structuring
element.
Using the spectral properties of the staining procedure
on the tissues (calcifi cations appear black and tissues
red-pink), we use the colour component images in the
following way:
The calcifi cation is relatively more contrasted
against the rest of tissues in f R , therefore this
component is used for extracting the calcifi cations.
✎ ✠
y B
The two elementary operations of erosion and dilation
can be composed together to yield a new set of
operators having desirable feature extractor properties
which are given by:
(1)
–
✁ ✁
erosion: ε B f x ✄☛✄✞ inf ✏
The top-hat transformation is a powerful operator
which permits the detection of contrasted objects on
non-uniform backgrounds (Meyer, 1977). There are
two versions:
–
–
For segmenting the region of the atherosclerotic
lesion a combined method of thresholding and
manual interaction is achieved from f G .
white top-hat: The
✁ residue✁ of the initial
✁ image f
and an opening γ f ✄ ; i.e. ρ f ✄✕ f γ f ✄ , extracts
bright structures,
✁
black top-hat: The residue
of ✁ a closing ϕ f ✄ and
✁
the initial image f ; ρ ✖ f ✄✗ ϕ f ✄ f , extracts dark
structures.
Usually, the top-hat is accompanied by a thesholding
operation, in order to binarise the extracted structures.
Through this choice of the colour components,
the color shading effect is negligible. Anyway,
a method for shading correction of color images
(Tomaževič et al., 2002; Derganc et al., 2003) can
be used in applications with important multispectral
inhomogeneities.
A granulometry is the study of the size
distributions of the objects of an image. Formally, a
granulometry
can be defi ned as a family of openings
✁
γn ✄ n ✘ 0 such that ✙ n ✚ 0 ✂✛✙ m ✚ 0 ✂ γn γm γm γn
Γ
γmax ✜ n ✢ m✣ . Moreover, granulometries by closings (or
anti-granulometry) can also ✁ be defi ned as families
ϕ ✄ n ✘ 0 . Performing the
of increasing closings Φ
MORPHOLOGICAL IMAGE PROCESSING granulometric analysis of an nimage
f with Γ is
AND QUANTIFICATION
equivalent to mapping
each opening of size n✁ with
✁ ✁
✤
☛
✄
✄
γ
f
of
the opened image. ✤
f ✄ is
a
measure
n
First introduced as a shape-based tool for binary
the
area
of
f
in
the
binary
case
(number
of
pixels)
images, mathematical morphology has become a very
and the volume in the grey scale case (sum of pixel
powerful nonlinear image analysis technique with
values).
The size distribution or ✁pattern spectrum of
operators capable of handling sophisticated image
f
with
respect
to Γ, denoted PSΓ f ✄ is defi ned as the
processing tasks in binary, grey-scale, colour and
following
(normalised)
mapping
multiresolution imaging modalities. A tutorial in the
technique can be found in (Serra, 1982, 1988; Coster
✁
✁
PSΓ f ✂ n ✄✑ PS f ✂ n ✄✞
and Chermant, 1989). In this section we briefly review
✁ ✁
✁
✁
the basic morphological operators used in this work.
✤ γn f ✄✝✄ ✥✤✁ γn✦ 1 f ✄✝✄
✂ n ✚ 0 (2)
m f✄
In the framework of digital grids, a grey-tone
✁
image can be represented by a function f : D f ✟ T ,
f ✄ maps each size n to
The
pattern
spectrum
PS
2
Γ
where D f is a subset of Z and T ✡✠ tmin ✂ ☛✂ tmax ☞ is an
some measure of the bright image structures with
ordered set of grey-levels. Let B be a subset of Z2 and
this size (loss of bright image structures between✁ two
s ✌ N a scaling factor. sB is called structuring element
successive openings). The pattern spectrum PSΓ f ✂ n ✄
(shape probe) B of size s. The basic morphological
is a probability density function (a histogram): a large
operators are:
impulse in the pattern spectrum at a given scale
✁ ✁
✁
indicates the presence of many image structures at that
– dilation: δ B f x ✄☛✄✍ supy ✎ B ✠ f x y ✄ ☞
✂
✂
✂
✂
83
A NGULO J ET AL : Morphological quantification of aortic calcium
the convergence to the u’s. An alternative to make
the choice of ✒ u1 ✂ u2 ✓ easier is a method which relies
on a double thresholding combined with a geodesic
reconstruction; the technique is known as thresholding
by hysteresis (Soille, 1999). For instance, in order
to extract light structures we have to take u 2 tmax
and let uTlow be a low threshold value and uThigh be a
high threshold value. Using this double threshold,
two
✁
binary images are obtained: Iwide T✡ uTlow ✢ tmax ☛ f ✄ and
✁
Inarrow T✡ uThigh ✢ tmax ☛ f ✄ . The fi nal binary image is given
by the reconstruction of Iwide using Inarrow as a marker;
i.e.
scale. It is also possible to use standard probabilistic
defi nitions to compute the moments of PS.✁ The fi rst
µ 1 ∑n nPS f ✂ n ✄ , the
moment µ is given by µ
k-th pattern ✁ spectrum moment,
k ✚ 2, is computed
✁
k
n
µ
f
n
.
In
particular, the fi rst
PS
as µ k
✄
✂
✄
∑n
four moments: mean µ , variance µ 2 , skewness µ 3 and
kurtosis µ 4 are often used.
Using a pair of an opening γ and a closing ϕ as
primitives, an operator of contrast enhancement can
be obtained (Serra, 1989). This toggle mapping or two
states contrast κ is generated by the following criterion
applied to each point x,
✁✂✂✂
✁
✁
✂✂✂ γ f ✁ x ✄✝✁✄
✂✂✂
✁
✁
κ f x ✄✝✄
ϕ f x ✄✝✄
if
✁
f x ✄✝✆
✁
✁
✁
f x✄
γ f x ✄✝✄
✄ ϕ ✁ f ✁ x ✄✝✄ if
✂✂✂
✁ ✁
✁
✁
ϕ f x ✄✝✄
f x ✄✝✞ f x ✄
✂✂✂
✂✂✂☎ ✁
γ f x ✄✝✄
f x ✄ ✁ ✁ if
ϕ f x ✄✝✄
✁
f x ✄✕
✁
f x✄
✁
✁
✁
✁
I
low
γ f x ✄✝✄
such that
f ✄✑
δgi ✦ 1
✁
✁
✁
(6)
We start by binarising the saturation component f S
for detecting the tissue presented in the fi eld image
by using a simple thresholding operation at uT1 (a low
value; e.g. uT1 10, has shown to be suitable for this
kind of images), see Fig. 3(a),
f ✄ (idempotence).
I x ✄✑ T✡ u1 ✢ u2 ☛ f x ✄☛✄✍✌☞
✂
(4)
The thresholding transformation of the image f
between the grey levels
u1 and u2 (typically u1 tmin
✁
or u2 tmax ), T✡ u1 ✢ u2 ☛ f ✄ , yields the binary image I such
that for each pixel x the binary value is given by
✁
✁
✁
rec
☛ f ✄✞ γ Inarrow ✂ Iwide ✄
Detection of tissue in the field image
✁
γ rec g ✂ f ✄✑ δgi f ✄
✁
uThigh ✢ tmax
The approach is separated into two sequential
phases. First, the segmentation is aimed on the one
hand to extract the calcium structures and on the other
hand to demarcate the region of the lesion on the tissue.
The segmentation yields the image data which are
the input to the second phase, the quantifi cation. The
calcifi cation structures are measured inside and outside
the lesion using a granulometric curve which allows to
calculate some statistical parameters of size. The same
operator is used to compute the shape of the lesion.
The relative proportion of area of calcifi cation is also
calculated.
A morphological tool that complements the
opening and closing operators for feature extraction
(extract the marked particles) is the morphological
reconstruction, implemented using the geodesic
dilation operator based on restricting the iterative
dilation of ✁ a function marker
f by B✁ to a function
✁
✏ ✁
mask g, δgn f ✄✍ δg1 δgn 1 f ✄ , where δg1 f ✄✍ δB f ✄✠✟ g.
The reconstruction by dilation is defi ned by
δgi
✎✑✏
ALGORITHMS
(3)
The closing and the opening may be replaced by a
dilation and an erosion of f .
✁
T✡ hyst
uT
Is
✁
1 u1 ✍ f x ✄ ✍ u2
0 otherwise
.
(5)
✁
T✡ uT1 ✢ tmax ☛ fS ✄
✂
(7)
In order to remove the noise and mistakes due to the
small pieces of tissue (histology artefact), the image
is then fi ltered by applying a reconstruction using an
opening as the marker,
The choice of ✒ u1 ✂ u2 ✓ determines the set of grey
levels associated with the object of interest. The
histogram summarises the grey-level contents of an
image and typically, the optimal threshold values can
be obtained from an analysis of the histogram, see in
(Angulo and Flandrin, 2003) the method of automated
thresholding which combines the classical selection of
the threshold value by minimising the sum of within
class variances with a morphological technique for
selecting the central mode values which speeds up
Itissue
✁
✁
γ rec γs1 B Is ✄✆✂ Is ✄ ✂
(8)
with s1 such that s1 B is larger than the size of the noise
and the small artefact pieces of tissue (good results
were obtained with s1 5); the structuring element B
is a circle. See the result in Fig. 3(b).
84
Image Anal Stereol 2003;22:81-89
(a)
(a)
(b)
(b)
Fig. 3. Detection of tissue in the field image:
(a) Binary image after thresholding saturation
component, Is ; (b) cleaned binary mask of tissue, Itissue .
(c)
Interactive segmentation of the region of
atherosclerotic lesion
In the tissue, the atherosclerotic lesion zone
squares with a texture of tissue less thick than the
rest, recognising visually, but its precise morphological
defi nition is almost impossible. Therefore due to the
fact that the automated segmentation of the lesion on
the tissue can not be achieved without some mistakes
and in order to avoid the propagation of these errors to
the quantifi cation step, we preferred to developed an
interactive approach.
(d)
In a fi rst step the human expert must perform a
manual dot-marking of the region of interest (ROI)
using the computer mouse on the colour image. This
zone associated to the lesion has to be closed, see
Fig. 4(d), but it is possible for the human user to
demarcate several closed partial ROI’s: the defi nitive
ROI is the union of the partial ones and is represented
by the binary image Imrk .
(e)
Fig. 4. Interactiv segmentatio of the region of
lesion: (a) Simplified green component using an
opening by reconstruction, f G ; (b) followed by closing
by reconstruction, fG ; (c) negative of the binary
mask without empty image zones, I g ; (d) manual dotmarking region of interest, Imrk ; (e) binary mask of
lesion inside the region of interest, Ilesion .
The automated step works on the green component
fG . This image is simplifi ed by means of an opening by
reconstruction (which simplifi es the light structures),
fG
✁
✁
γ rec γs2 B fG ✄✆✂ fG ✄✆✂
(9)
85
A NGULO J ET AL : Morphological quantification of aortic calcium
followed by a closing by reconstruction (which
simplifi es the dark structures),
✁
✁
ϕ rec ϕs3 B fG ✄✆✂ fG ✄✆✂
fG
(10)
where the sizes of the opening and the closing
15 and s3
have been empirically fi xed to s2
10 respectively (B is a circle). The corresponding
example of tissue image after simplifi cation is shown
in Fig. 4a - b. The image f G must be binarised for
extracting the empty zones (zones of the fi eld without
tissue); the optimal threshold value uT2 for each image
is obtained by the algorithm presented in Angulo (02).
The result of this thresholding process at uT2 ,
✁
T✡ tmin ✢ uT2 ☛ fG ✄✆✂
Ig
(a)
(11)
is a fi rst binary mask which is then restricted to the
manual defi ned Imrk to obtain the binary mask of the
region of lesion,
Imrk ✟ Ig
Ilesion
(b)
(12)
✂
See the result in Fig. 4e.
Extraction of calcification
In the red component of the aortic section colour
image fR the calcium appears as dark structures upon
a bright background. In order to enhance the contrast
of the calcifi cations against the background a toggle
mapping is taken,
✁
κ fR ✄✆✂
fR
(c)
(13)
Fig. 5. Extraction of calcification: (a) Contrast
enhancement of calcification on red component, fR ; (b)
extracted calcification by means of a dual top-hat, f c ;
(c) binary mask of calcification after thresholding by
hysteresis, Icalci f .
where the two primitives of κ are an erosion and a
dilation of size 3, see Fig. 5(a). From this image,
a black top-hat of size s4 extracts the calcifi cation
(s4 corresponds to the size of the biggest calcium
structures which can be found); experimentally we
have fi xed s4 25 with a circular structuring element,
fc
✁
ρ ✖ fR ✄
Quantification of atherosclerotic lesion
✁
(14)
✂
✁
I ✄ be the surface area; i.e., number of pixel
Let
to ones, of the binary image I. We start by calculating
the size of the tissue and the lesion regions,
✁
Atissue
On the image fc , Fig. 5(b), a thresholding by
hysteresis is performed to provided the binary mask
of calcifi cations (see Fig. 5(c)),
Icalci f
T✡ uhyst
T
low
✎✑✏
uThigh ✢ tmax ☛
✁
f✄ ;
✁
Itissue ✄ ✂
✁
Alesion
✁
Ilesion ✄
✂
(16)
We propose to characterise the shape of the lesion
region
by means of a pattern spectrum curve,
✁
PSlesion . The structuring elements are
PS Ilesion ✂ n ✄
circles of increasing size n 2 (in fact, the practical
shape is an octagon: isotropic approach to a circle
4 to n
70. The
in the square grid), from n
granulometric curves describe in a compact way the
different thickness of the wall of the aortic valves: each
(15)
✂
the choice of the threshold values is not so critical (the
top-hat facilitates just the thresholding); e.g. u Tlow 50
and uThigh 70.
86
Image Anal Stereol 2003;22:81-89
peak corresponds to the thickness of a sector of the
wall. These histograms of shape can be parametrised
by using their statistical moments. The surface area in
pixels of an octagon of size n is given by the formula
7n2 4n 1. In order to
(Serra, 1982), Noctagon
2
obtain the sizes in µm we have to multiply the area
in pixels by 5 5 µm2 pixel for all the measurements.
✂
✂
✄
✂
(a)
Quantification of calcification
In order to quantify the calcium, we compute fi rst
the absolute and the relative surface of calcifi cations
inside and outside the lesion,
✁
✁
Acalci f i
Acalci f o
✁
✁
Icalci f
Icalci f
Acalci f
Acalci f
o
Ilesion ✄ ✂
✁
Itissue Ilesion ✄✝✄✆✂
✟
✟
i
✁
✁
✁
Acalci f i
✁
✂
Ilesion ✄
Acalci f
Itissue
Fig. 6. Two examples of aortic section images:
(a) with low calcification content (mouse 35); (b)
riches in calcification (mouse 40). On the left, the
initial images and on the right, summary of segmented
structures: the background in black, the tissue in gray,
the lesion in white and the calcification in red.
(18)
o
Ilesion ✄
(b)
(17)
✂
(19)
Besides these parameters, the calcifi cation structures
inside the lesion✁ are quantifi ed by using a size
Ilesion ✂ n ✄
PScalci f i (the
distribution, PS Icalci f
structuring elements of openings are again circles of
increasing size n 2, from n 3 to n 23) and its
moments.
Pattern spectrum of lesion
0.16
Mouse 35
Mouse 40
0.14
✟
Normalised size (area)
0.12
✂
RESULTS AND DISCUSSION
0.1
0.08
0.06
0.04
0.02
In Fig. 6 two examples of segmented aortic section
images are depicted: one corresponding to a mouse
with low calcifi cation content (control mouse) and
another to high calcifi cation (uremic mouse). In Fig. 7
their associated pattern spectra are shown.
0
0
10
20
30
40
n (size of opening)
50
60
70
(a)
22
(b)
Size distribution of calcification
0.8
Mouse 35
Mouse 40
0.7
In Table 1 the obtained parameters of the lesion
are included. For mouse 35 the area of lesion is
larger than for mouse 40, but above all, the thickness
of the walls is considerably greater. Besides being
used for determining the relative amount of calcium
located inside, the size of the atherosclerotic lesion
may be signifi cant a priori. However, due to the
fact that during the tissue preparation procedure the
shape of the atherosclerotic lesion may be modifi ed
involuntarily, the practical usefulness of the parameters
from PSlesion remains doubtful. Other techniques of
tissue preparation could be envisaged in order to
achieve more reproducible sections in such a way
that the parameters associated to PSlesion will be very
important.
Normalised size (area)
0.6
0.5
0.4
0.3
0.2
0.1
0
2
4
6
8
10
12
14
n (size of opening)
16
18
20
Fig. 7. Quantification of size and shape using
morphological granulometries (curves associated to
the examples of Fig. 6): (a) Pattern spectrum of the
region of lesion; (b) Size distribution of calcification
structures inside the lesion.
87
A NGULO J ET AL : Morphological quantification of aortic calcium
Table 1. Size (Atissue and Alesion ) and shape (first four moments of PSlesion ) parameters of lesion zone for the
examples of Fig. 6.
Mouse 35
Mouse 40
Atissue
120859
160483
Alesion
59079
78696
µ
38.55
51.17
Acalci f o
170
37137
Acalci f i
0.0139
0.5476
The approach has however several limitations. On
the one hand, in order to minimise the errors of
segmentation, an interactive algorithm was developed
which involves a necessary human action for each
image. Obviously, the task is simple (some “clicks”
of mouse) but time consuming. The balance between
automation and precision leads sometimes to this
kind of approach. On the other hand, working on
low magnifi cation microscopic images entails that
the image structures can be very small, limiting the
resolution of the methods. For instance in Fig. 7 (right),
notice that the size distribution of calcium for the
control mouse (small calcifi cation structures) is limited
to the openings of size n 3 and 5. The classical
problem of reproducibility of histologic preparations
(sectioning, staining, etc.) must be also taken into
account in order to prevent artefacts, like dust or
others.
Acalci f o
0.0014
0.2314
(a)
Mouse 35
Mouse 40
µ
2.19
5.59
µ2
µ3
0.39
74.21
0.48
12.92
µ4
208019.51
720299.94
✂
Table 2. Surface area parameters and first four
moments of PScalci f i of lesion zone for the examples
of Fig. 6.
Acalci f i
821
43094
µ3
-1176.89
-206.51
has been used. The results of the interactive human
segmentation of the atherosclerotic lesion and the
extraction of calcifi cation have been subjected to
evaluation by another human grader in order to fi nd
out serious mistakes: only 3 cases among the set of 173
have been rejected and manually corrected (1 7%). In
view of these results we can state that the behaviour of
the image analysis algorithms is quite satisfactory.
Regarding the parameters of calcifi cation,
summarised in Table 2, the use of this set of parameters
yields an easy way to identify the different levels
of calcifi cation. Obviously, the most interesting
parameters are the relative surfaces of calcifi cation
and the two fi rst moments of size distribution: the
mean size of calcium structures and the variance of
size (which gives an idea of size dispersion and is
helpful for distinguishing large compact calcifi cations
from large disintegrated calcifi cations).
Mouse 35
Mouse 40
µ2
293.03
647.21
µ4
0.32
729.99
(b)
The robustness of an algorithm can be defi ned
with respect to changes in the parameters or to
image quality. The present algorithms have seven
confi gurable parameters: uT1 (threshold value for
the tissue determination on f S ), s1 (size of the
tissue fi ltering on fS ), s2 and s3 (size of the
tissue simplifi cation on fG ), s4 (size of the calcium
enhancement on fR ), uTlow and uThigh (threshold values
for the calcium extraction on f R ). The threshold value
uT2 is obtained automatically for each image. The
values proposed for the other parameters have been
set after empirical tests on a random selection of 10
images which covers the different levels of pathology
and image quality. Someone who would like to use
these methods with a different magnifi cation or a
different camera has to start by using a training set
of images (representative of its problem) in order to
adjust the values of parameters. In any case, some tests
of images at 10 have shown that the approach is quite
robust.
CONCLUSION
The robustness and accuracy of segmentation
results allowed us to consider the subsequent
quantifi cation as a correct procedure. The
morphological parameters obtained have been
analysed and correlated to several biomedical
parameters (Massy et al., 2003). We also envisage
to apply this quantifi cation approach to the study
of calcifi cation using non-invasive techniques, for
instance by obtaining the morphological information
by means of electron beam tomography or other
imaging technique. In the previous studies based
on computed tomography (Agatston et al., 1990,
Morgan-Hughes et al., 2002), the quantifi cation
of calcifi cation has not been attempted using
morphological techniques.
✁
Once these parameters were fi xed, a deep study has
been performed on the basis of the present methods.
A database of 173 images corresponding to 45 mice
To summarise the algorithmic discussion above,
we conclude that quantitative measurements of
88
Image Anal Stereol 2003;22:81-89
histologic sections by morphological image analysis
software are very useful. They provide a powerful tool
for improving and automating experimental studies
on structural-functional correlations in tissues with
histopathologic changes.
Meyer F (1977). Constrast features extraction. In: Chermant
JL, ed. Quantitative Analysis of microstructures in
Materials Science, Biology and Medecine. Stuttgart:
Riederer Verlag, 374-80.
Moreso F, Ser´on D, Vitri´a J, Griny´o JM, Colom´e-Serra FM,
Par´es N, Serra J (1994). Quantification of interstitial
chronic renal damage by means of texture analysis.
Kidney 46:1721-7.
ACKNOWLEDGMENTS
The authors thanks the comments and suggestions
of the anonymous reviewers.
Morgan-Hughes GJ, Roobottom CA, Marshall AJ (2002).
Aortic valve imaging with computed tomography: A
review. J Heart Valve Disease 11(5):604-11.
REFERENCES
Agatston AS, Janowitz WR, Hildner FJ, Zusmer NR,
Viamonte M, Detrano R (1990). Quantification of
Coronary Artery Calcium Using Ultrafast Computed
Tomography. J Am Coll Cardiol 15:827-32.
Muntzel M, Massy ZA, Ruellan N, Descamps-Latscha
B, Lacour B, Drueke TB (2002). Chronic renal
failure increases oxidative stress and accelerates
atherosclerosis in apolipoprotein-E knock-out (EKO)
mice (Abstract). Nephrol Dial Transplant 17:46.
Angulo J, Flandrin G (2003). Automated detection of
working area of peripheral blood smears using
mathematical morphology. Anal Cell Pathol 25:37-49.
Reinmuller R, Lipton MJ (1987). Detection of coronary
artery calcification by computed tomography. Dynam
Cardiovasc Imag 1:139-45.
Coster M, Chermant JL (1989). Pr´ecis d’analyse d’images,
2nd. ed. Paris: Les Presses du CNRS.
Derganc J, Likar B, Bernard R, Tomaževič D, Pernuš F
(2003). Real-time automated visual inspection of color
tablets in pharmaceutical blisters. Real-Time Imaging
9:113-24.
Ser´on D, Moreso F, Gratin C, Vitri´a J, Colom E, Griny´o
JM, Alsina J (1996). Automated Classification of Renal
Interstitium and Tubules by Local Texture Analysis and
a Neural Network. Anal Quant Cytol Histol 18:410-9.
Hanbury A, Serra J (2002). A 3D-polar Coordinate Colour
Representation Suitable for Image Analysis. Submitted
to Comput Vis Image Und.
Serra J (1982). Image Analysis and Mathematical
Morphology. Vol I. London: Academic Press.
Serra J (1988). Image Analysis and Mathematical
Morphology. Vol II: Theoretical Advances. London:
Academic Press.
Ishibashi S, Herz J, Maeda N, Goldstein JL, Brown
MS (1994). The two-receptor model of lipoprotein
clearence: Tests of the hypothesis in ”knockout”
mice lacking the low density lipoprotein receptor,
apolipoprotein E, or both proteins. Proc Natl Acad Sci
USA 91:4431-5.
Serra J (1989). Toggle mappings. In: Simon JC, ed. From
Pixels to Features. Amsterdam: North Holland, 61–72.
Soille P (1999). Morphological image analysis. Berlin,
Heidelberg: Springer-Verlag.
Lardenoye J, de Vreis M, lowik C, Xu C,Dhore C, Cleutjens,
J van Hinsberg V, van Bockel J, Quax P (2002).
Accelerated atherosclerosis and calcification in vein
grafts. Circ Res 91:577-84.
Tanenbaum SR, Dondos GT, Veselik KE, Prendergast MR,
Brundage BH, Chomka EV (1989). Detection of calcific
deposits in coronary arteries by ultrafast computed
tomography and correlation with angiography. Am J
Cardiol 63:870-2.
Margolis JR, Chen JTT, Kong Y, Peter H, Behar VS, Kisslo
JA (1980). The diagnostic and prognostic significance
of coronary artery calcification. Radiology 137:609-16.
Tomaževič D, Likar B, Pernuš F (2002). Comparative
evaluation of retrospective shading correction methods.
J Microsc-Oxford 208:212-23.
Massy Z, Nguyen-Khoa T, Angulo J, Munzel M, Ivanovski
O, Szumilak D, Ruellan N, Descamps-Latscha B,
Lacour B, Drüeke T (2003) Uremia accelerates
calcification in apolipoprotein E knock-out (EKO)
mice. To be presented in 17th Congress of the
International Society of Nephrology – 40th Congress of
the European Renal Association, Berlin, June 8-12.
Zhang SH, Reddick RL, Piedrahita JA, Maeda N
(1992). Spontaneous hypercholesterolemia and arterial
lesions in mice lacking apolipoprotein E. Science Oct
16;258(5081):468-71.
89