Development of methods and algorithms of an estimation of a cell's functional
state. Kayira V.V.
Scientific adviser: Ph.D., c.t.s. assistant prof. V.G.Adamov
ABSTRACT
Table of contents
The analysis of the image represents scientific area which deals with geometrical
and densimetrical measurements
on images, received from various sources. The main scope of similar measurements
and calculations — the quantitative microscopy allowing quickly and precisely to
receive statistically significant results and consequently superseding
traditional and subjective qualitative methods. Examples of similar measurements
can be definition of a volume fraction of various phases in alloys or
in geological test, numerical and dimensional distribution of the polluting
particles filtered from the air or liquid environment, or distribution of
integrated optical density inside nucleus of the painted biological cells.
Now in laboratories and on manufacture there is a necessity of the
digital pictures analysis
containing the same objects,
probably distinguished from each other in some parameters.
The decision of the given problem encounters the certain problems of
realization existing developed methods of the analysis intended for
single objects in a picture. New techniques of solving such problems and also possible procedures of the analysis of the received data are in
the view.
Within the framework of a task, existing software products
are observed.
For the first time the method of an image analysis has appeared as a
mean ready to use in 1963 together with development QTM (QTM
— the Quantitative Television Microscope), created by firm Metals Research Ltd.,
becoming subsequently
a part of company "Leica". The specified device
intended for use in metallurgical laboratories — especially for the quantitative
control over cleanliness of steel and for other microstructural measurements,
obvious utility of this device in other areas has soon become evident. One of its first applications in biology began measurement of the size of
air spaces in lungs (that was required for the quantitative description of a
lungs lesion level) and for calculation of grains amount of silver in an
autoradiography.
Since then the advanced technics of an image analysis has found the application almost in all scientific and technical areas of
natural sciences, from anatomy and zoology, and has
expanded the abilities to including such functions of mathematical
processing, as a filtration and amplification of images.
Now the most appreciable and significant program in the given field is the
program of above mentioned firm "Leica" (Leica Imaging Systems Ltd.) –
LEICA Qwin versions 1.56.
QWin is the set of applied programs of company "Leica" based on
technical opportunities Windows for the analysis of the image. The given set is
under the control of standard operational environment Microsoft Windows for the
industrial purposes.
Program QWin allows to carry out measurements on several classes of accuracy,
starting from measurement of objects in a dialogue manual mode up to
the fully-automated analysis where as an example it is possible to result
definition of inclusions in steel. As examples of use also it is possible to
name such as:
· Measurement of length, distances up to object and the areas
· Percentage of a phase, the contents of fraction in the
area and volume.
·
Calibrating densitometry.
· Definition of the form of particles and the analysis of their sizes.
· Profiling of brightness levels.
However this program in practice has shown, that its work results
have very big error. Qwin has very big scope, but with the decision of a
concrete medical task of definition cells necrosis, and its smooth borders
delineation
of a living cell and sharp differences of the wrinkled membrane of cell that already
died, this program consults badly since its
methods are too "common" and do not take into account concrete conditions
of shooting and cell.
As example the medical computerized complex of the image analysis can serve
"Diamorph", used in medical establishments and scientific
institutes. Specialized complex
"Diamorph"
provide automatic input of microscopic images, allocation of objects of a
picture (cells, nucleus, sites of different painting or brightness). The
advanced toolkit for carrying out of measurements in a picture is provided:
the linear sizes, perimeter, the area, optical parameters, position of objects.
The statistical subsystem carries out mathematical processing results of
measurements with automatic construction of a wide set of histograms, schedules,
tables.
The complex software in an automatic mode provides the following
functions of the quantitative and qualitative analysis of the image:
On group of objects: amount of objects, total perimeter, the total
area, total integrated optical density.
On each object: perimeter, the area, the factor of the form, diameter of the
circle equal on the area, the minimal diameter, the maximal diameter, sizes of
projections to axes, coordinates of "the center of gravity", a corner between
a direction of the maximal diameter and an absciss axis, color, average value of intensity and
its dispersion, average optical density, average value optical transmission, integrated optical density.
Input and the specialized processing of roentgenograms, morphological images,
smears with the purpose of increase of the diagnostic importance of
researches,
and also for archiving
and conducting a database.
It is necessary to note, that this complex does not assess objects in
functional sense, and gives only their parameters. Besides the entrance high
quality image is necessary, that in real conditions is frequently very
labour-consuming and it is not always possible. The texture analysis is absent.
Let's consider program Hesperus intended for processing and
visualization of two-dimensional sets of the numerical data of any nature. It is a
product of Applied Mathematics Laboratory of the Moscow State University.
The package provides the big set of functions of processing, such as a
filtration, a stretching and turn, the texture analysis, calculation of a
spectrum, classification, etc.
From the mathematical point of view image processing and visualization
of results is one of the best, however its applicability for so
narrow problem is limited.
Lacks:
1) only the expert well understanding mathematics can give conclusions on
calculated parameters;
2) there is no processing of groups of objects, such segmentation, that also
complicates the analysis;
3) the complex is not automated, so all sequence of actions and the
responsibility lays on the operator;
Let's consider examples of other automated complexes.
Organization Global automated image analysis gives similar systems:
«Analysis of
scleral collagen fibrils»,
«Analysis
of the tear breakup in the human eye»,
«Automated
Analysis of Ultrasonic Images of the Human Carotid Artery»,
and others.
On used methods are similar to complex "Diamorph",
are only more narrowly specialized, that gives them advantage in the decision of
typical problems.
At the decision of problems(tasks) of such class generally use the following
structure of diagnostic system:
Fig. 1 – Structure of the specialized computer system.
As the most authentic and exact technique of definition of parameters the
method of construction of the generalized òåêñòóðíî-planimetric model of objects
on the image can serve. We shall consider the circuit of a subsystem of
preliminary processing and allocation of researched parameters.
Fig. 2 – the Block diagram of a subsystem of preliminary
processing and allocation of researched parameters.
In this system subsystems of preliminary processing and allocation of
researched parameters and formations of the conclusions are difficultly sold.
At processing images a lot of information is concluded in contours of
objects, but also the analysis of a structure is very important. On a texture
of those or other sites of objects it is possible to make assumptions or
even sometimes concrete conclusions. For the texture analysis are intended the
block of construction of texture model. For construction of any texture model,
preliminary it is required to break the analyzed image into sites with a
homogeneous texture (make texture segmentation), and also, it
is desirable to define the type of a texture.
2.2 Methods of the planimetric analysis.
Algorithm Susan.
Basic idea SUSAN it that neighbours of each point in homogeneous area have
brightness close to it, and near to border number of neighbours with
identical brightness decreases. Except for borders this method finds out also
other features on the image (corners, thin lines, etc.). This principle
illustrates figure 3.
Around of each pixel of the image the mask is constructed
which central pixel refers to as a
nucleus (in work round mask with radius 3.4 pixels which includes 37 pixels or a
traditional square mask 3x3). Pixels within the
limits of a mask, having brightness comparable to a nucleus, form area USAN
(Univalue Segment Assimilating Nucleus – a homogeneous segment,
àññèìèëèðóåìûé
a nucleus). For detection of bidimentional features and borders the size, the
centre of gravity and second moments USAN are used. Such approach of detection
of features differs from known methods that does not use derivatives of the
image and, hence, there is no necessity for preliminary suppression of noise.
Fig. 3 – the Different masks imposed on the image;
Area USAN is maximal, when the nucleus is in homogeneous (or nearly so
homogeneous) areas of the image, it decreases up to half of this maximum
near to direct border and decreases even more near to a corner and
reaches local minima precisely on border and in corners. This property
of area USAN is used as the main criterion of presence of borders and
two-dimentional features. In figure 4 are shown USAN, apparently, axis Z (area
USAN) is directed aside reduction!
Fig. 4 – the Three-dimensional schedule, showing change USAN for a sample of the grey image.
Comparing SUSAN, for example, with one of most widely used Canny borders detector (association of gradient operators and Gauss
smoothings) is possible to note the following features and differences:
- Algorithm Canny finds out unique border, i.e. on
the image such as Ò – crossed borders – it will select one way of a contour while SUSAN will find out a
corner and even will allocate with its falling of area USAN.
- Algorithm Canny – because of use of a derivative
smooths borders and corners (it certainly plus at a faltering contour since
it will close it); At enough high quality of the processable
image Susan can show ideally correctly all features of the form of object
(objects!) images.
There are more differences, however it is enough of these to prefer the
detector of borders SUSAN processing the image of a fabric (a layer of cells).
One of the most promising technique for the
description of the form is based on Fourier descriptors for borders of the image.
Let's assume,
that N points are available on an area border. We can consider area as
placed on a complex plane with ordinate as imaginary ordinate and
absciss which is real absciss. Then, coordinates x-y each point of an
analyzed contour can be submitted as complex numbers (x + jy). The sequence for
borders then can be written down as complex sequence Zi:
Zi = Xi +j*Xi, i
= 0,1,2, …, N-1,
where j – imaginary unit.
We receive Z = z(i).
Let's put, that z(i) – a continuous curve. The continuous curve is simple
curve (Jordan arch) if it will consist
of a uniform branch and does not contain multiple points; it means, that is not
present such various j1 and j2 for which it is fair z(j1) = z(j2).
Also we shall put, that an arch is
rectifiable(has
a limit of length).
Complex planimetric integral:
Then points z0, z1, …, zn are located one by one along a contour With.
Each of points xi lays on a site of a curve [zi-1, zi] and can coincide with
one of its ends.
Then generally (for z = u (x, y) + j×v (x, y)): for
our valid argument (u (x, y) = x (i) +0, v (x, y) = y (i) +0)), it is
received
,
Applying Fourier transformation, we close a
contour on themselves, having made its period 2p.
,
c(v) – comlex coefficients of Fourier transfirmation.
Then in a discrete variant:
(5)
i=0,1,2, … N-1
To coefficients for an analysis may be applied theorems (2.4.2).
Fractals meet everywhere where correct forms of Euclidean
geometry come to an end. Everything, that is created by the person, limited to
planes. If there is a natural object, at the first view it is visible that to describe its form with all roughnesses it is possible only
approximately. Here fractals come to the aid.
According to Benua Mandelbrot, a word the fractal occurs from latin words
fractus – fractional and frangere – to break, that reflects essence of a fractal
as the "broken", irregular set and designates the set having fractional dimension.
Mandelbrot has given
strict mathematical definition of a fractal as sets, which Hausdorff dimension is strict more than topological
dimension. However he has not been satisfied with this definition because it
does not include some sets considered by many mathematicians as
fractals. In the given work the variety of fractals and spheres of their
appendix is not considered, therefore we shall consider only the basic
concepts for definition of the form of object.
Hausdorff dimension (dimA) of set A is
defined as:
(6)
Box dimension is a little more simple concept. If A some
compact set, and N (r) is the minimal number of spheres of radius r covering A
and if there is a limit
lim (log N (r) / log (1/r)),
(7)
at r aspiring to zero this limit refers to as cubic dimension of set A. It is
known, that Hausdorff dimension does not surpass box dimension, and for self-similar fractals they are coincide.
Topological dimension, accepting exclusively the whole values, will be
coordinated to intuitive representation about dimension of set. So dimension of
one-dot set is equal to zero, a piece and a straight line – unit,... Dimension of
a n-dimensional cube is equal n. More strictly:
Topological dimension of set A is equal to zero if for any point of set A
there will be as much as small vicinity which border is not crossed with A;
Topological dimension A is equal n if for any point of this set there will be
as much as small vicinity which border is crossed with A on set of dimension
n-1, and besides n there is the least positive number for which this condition
is executed.
In Hesperus use both the histogram, and a matrix of spatial dependence
(GLCM).
Matrix of spatial dependence (GLCM – gray level co-occurrence matrix) –
the histogram of the second order showing probability of joint occurrence of two
certain values pixels on the set distance and in
a certain direction. The sizes of matrixes depend on quantity of
gradation of color taken in consideration. In Hesperus matrixes in the size
256x256 elements that corresponds to 256 shades of grey color are
applied. On the basis of a matrix of spatial dependence the big number of
textural characteristics is calculated.
The structure can be allocated on the basis of various criteria. Application
of any of these criteria to the image gives other image where intensity of
every pixel reflects size of conformity to
this particular criterion in a concrete point of the entrance image on
an output. Results of the textural analysis are usually treated as one
multichannel image, and can be sent on an input to the standard
qualifier which groups structures on classes.
To apply a certain particular criterion, it is required to be set
by a number of parameters. Usually it is the size of a window which is examined
around of every pixel,
and also a direction and displacement. Last two parameters are used for the
analysis of structures which differ in various directions (for example, a brick
wall where the distance between bricks is more in one direction, than in another).
Difficult methods for recognition of all areas with a similar texture can be
necessary because of the criteria are sensitive to orientation of a texture in
the image. For example, on the ploughed field the texture along furrows is distinct from a structure across
furrows,
hence two adjacent areas with furrows, focused
under 90 degrees relatively each other, will not be classified as identical if
the criterion dependent on orientation used for them.
Let's consider a method focused on frequency characteristics of a texture,
which are invariant to rotation, displacement and partially scale.
A spectrum image is calculated by application to the image a
discrete Fourier
transformation. The result of transformation
is the complex image. It is possible to
define the main frequencies of the image on the amplitude spectrum, characteristics and
dependences should be studied.
Coefficients of multi-dimensional Fourier transform (for our two-dimensional case):
(8)
aj<tj<aj+Tj,
j=1,2.
Such description of images possesses the essential advantages useful at
recognition:
1. The module of spectral function: |F (u, v) | does not depend on
displacement of function f (x, y), i.e. the description is
invariant to displacements of the image in planes of supervision.
2. The description of images possesses the certain noise stability. When
spectrum of recognizabl
images and an additive handicap are various, it is possible to increase the
attitude(relation) signal / noise with the help of a spatial filtration.
3. A rotation of the image around any point results in rotaion of spatial
spectrum F (u, v) around the beginning of coordinates with
respective alteration of phases making (change of phases does not influence on
|F(u, v)|).
4. if the image f (x, y) has spectrum F (u,
v), the image f (ax, ay), connected by transformation of
similarity with f (x, y) where a – the constant
factor, has a spectrum
The radial line in a frequency plane corresponds to an unique direction
in the image which includes all frequency components. A set of sizes
,
(9)
characterizes all directions, if Qj (j=1,2..., n) cover sector from 0
up to 3600. In practice it is used a sample window
of a wedge-like
form. Such window defines the contribution from a small number of
adjacent directions and possesses that advantage that reduces number of necessary
excerpt and reduces influence of little changes. The set of sample windows
of a wedge-like
form
allows to receive the description of a power spectrum along a radial direction.
This method of digitization is not sensitive to scale of the image.
Fig. 6 – the Set of the optical windows used at processing of
power spectrum
The one-dimensional frequency (power) characteristic of a contour.
Normalized Fourier descriptor of the contour.
Fractal
dimension of the contour.
two-dimensional power spectrum and its characteristic.
In the given work few possible(probable) characteristics of objects are
resulted. The primary goal is their generalization and carrying out of
statistical researches on real objects. Preliminary results (in work are not
resulted) have shown good accuracy and reliability. The big field for
development is the presence(finding) of the generalized characteristic of a
bidimentional spectrum. Wide application in this sphere can find
âåé âëåòû with which help it is possible to consider
more in detail frequency characteristics as objects and a structure.
Table of contents
Literature
1. G.Korn, T.Korn. Mathematical handbook // M.: "Science", 1974, 831
with.
2. A user manual for program LEICA QWin.
3. A user manual for program Heperus.
4.
V.V.Zhikov. Fractals. Mathematics 1996.
5. Smith and Brady. «SUSAN – A New Approach to Low
Image Processing», 1995.
6. Sven Loncaric. A survey of shape
analysis techniques.
7. Jenkins, WK. " Fourier Series, Founer Transforms, and theDFT " Digital
Signal Processing Handbook Ed. Vijay TO Madisetti and Douglas IN Williams
BocaRaton CRC Press LLC, 1999.
|