ES2428215A1 - Computer-implemented method for recognising and classifying abnormal blood cells, and computer programs for performing the method - Google Patents

Abstract

The invention relates to a method consisting in classifying cells on the basis of automatic processing techniques and analysis of blood samples, which includes acquiring digital images of abnormal blood cells from blood cells, and then: segmenting said digital images of abnormal cells, providing identified regions of the nucleus, cytoplasm and outer area of the cell of said abnormal blood cells of said digital images calculating intrinsic characteristics of each of said identified regions of the nucleus, cytoplasm and outer area of the cell of said abnormal blood cells, comprising calculating the geometric characteristics of said identified regions automatically recognising and classifying abnormal blood cells on the basis of said calculated intrinsic characteristics of said identified regions and using said recognised and classified abnormal blood cells for carrying out a diagnostic orientation of hematological diseases.

Description

Computer implemented method for cell recognition and classification
abnormal blood and software to carry out the method

Technical sector

The present invention concerns in a first aspect, in general, a method implemented by computer for recognition and classification of abnormal blood cells, and in particular a method comprising using said abnormal blood cells recognized and classified to perform a diagnostic orientation of diseases. Hematological

A second aspect of the invention concerns computer programs adapted to perform some of the steps of the method of the first aspect.

Prior art

Peripheral blood (SP) is an easily accessible organic fluid, so its study represents the initial analytical link in the diagnosis of most hematological or non-hematological diseases.

The diagnosis of more than 80% of hematological diseases is achieved by
Morphological studies that have SP as a starting point. Based on automated image processing techniques using an artificial neural network, equipment has been developed that performs a preclassification of SP nucleated cells taking into account hundreds of calculations based on morphological aspects, such as coloring, size, shape and texture. of the cells, among others. However, although these analyzers represent a technological advance of great interest, they are not able to identify and preclassify pathological blood cells, especially lymphoid cells, which can circulate in peripheral blood in certain lymphoid neoplasms, whose morphological identification is usually a complex task.

Due to the difficulty of correct automated preclassification of abnormal lymphoid cells, few papers have been published using different methods of digital image processing with satisfactory results, and some of them still up to date with the presentation of the present invention are still under study. [one]. The problem has been addressed by extracting a significant number of measures and parameters that describe the morphological characteristics of interest in the cells, together with systems of

pattern recognition for the classification of different cells into categories [2], [3],

[4].

For example, Chronic Lymphoid Leukemia cells or LLC are typically described as small lymphocytes with pooled chromatin and poor cytoplasm, on the contrary, Tricoleukemia or HCL cells have a low basophilic cytoplasm and abundant irregular or hairy edges. Therefore, the morphological differentiation between various types of lymphoid cells is not trivial, due in large part to the lack of objective values to define cytological variables, which requires high experience and skill.

[one].

F. Scotti, "Robust Segmentation and Measurements Techniques of White Cells in Blood

Microscope

Images, " 2006 IEEE Instrumentation and Measurement Technology

Conference Proceedings, Dec. 2006, pp. 43-48.

[2].

Bergmann, M., Heyn, H., Müller-Hermelink, H.-K., Harms, H., Aus, H.M. (1990)

Automated recognition of cell images in high grade malignant Iymphoma and reactive

Follicular hyperplasia. Analytical Cellular Pathology, 2: 83-95.

[3].

Foran, 0.1, Comaniciu, D., Meer, P. et Goodell, L.A. Computer-Assisted Discrimination

Among Malignant Lymphomas and Leukemia Using Immunophenotyping, Intelligent

Image Repositories, and Telemicroscopy. IEEE Trans.

on Information Technology in

Biomedicine, 4 (4): 265-273,2000.

[4].

Juan, 1, Sigaux, F., Flandrin, G. (1985) Automated Classification of Lymphoid Cells.

Analytical and Quantitative Cytology and Histology, 7: 38-46.

Summary of the Invention

Therefore, there is a general interest in being able to identify descriptors or intrinsic characteristics of the lymphoid cell that make it possible to classify different pathologies by improving the recognition and automatic classification of said cells, and thus allowing greater accuracy in the detection and recognition of hematological diseases such as different types of leukemias, including leukemized lymphoproliferative neoplasms.

Accordingly, the invention provides, according to a first aspect, a computer-implemented method for recognition and classification of abnormal blood cells, which, like known techniques, comprises classifying cells based on automatic processing techniques. imaging and blood sample analysis techniques, which includes acquiring digital images of abnormal or atypical blood cells, for example lymphoid or blast cells, from a plurality of blood cells.

In a characteristic way, and contrary to previously known solutions

In the state of the art, the proposed method comprises performing the following stages:

a) segmenting said digital images of abnormal cells by providing identified regions of the nucleus, cytoplasm and external cell area of said abnormal blood cells of said digital images;

b) calculate intrinsic characteristics of each of said identified regions of the nucleus, cytoplasm and external cell area of said abnormal blood cells comprising and calculating at least the geometric characteristics of said identified regions;

c) automatically recognize and classify abnormal blood cells based on said intrinsic characteristics calculated from said identified regions; and d) using said recognized and classified abnormal blood cells to perform a diagnostic orientation of hematological diseases.

To improve the quality of the images of the blood cells, preferably, according to one embodiment, and prior to performing said segmentation, a step of preprocessing said digital images of acquired abnormal blood cells is performed.

On the other hand, according to another embodiment, for the segmentation of the cells the variable form of them is taken into account, with different shades and textures both in the nucleus and at the cytoplasm level, and that the periphery of the cytoplasm can show extensions. Cell segmentation can be done by two different techniques. For example, either by using an active contour technique in said digital images of abnormal blood cells using Gradient Vector Flow (GVF) or conversely, by using a technique of grouping the components of different spaces of color and Watershed transformation in such digital images of abnormal blood cells.

Oe according to an exemplary embodiment, the calculation of the intrinsic characteristics of said identified regions of the nucleus, the cytoplasm and the external region of said abnormal blood cells further comprises extracting statistical characteristics of the first order,

for example, the mean, standard deviation, statistical asymmetry, etc. of said identified regions based on a histogram of the identified region. Oe according to another embodiment, the calculation of the intrinsic characteristics of said identified regions of the nucleus, cytoplasm and the external area of the cell of said

Abnormal blood cells further comprise calculating second-order statistical characteristics, for example, contrast, homogeneity, entropy, etc. of those regions

identified based on a co-occurrence matrix of each region identified.

According to another embodiment, said calculation of the intrinsic characteristics of the identified regions of the nucleus, cytoplasm and the external area of the cell of said abnormal blood cells comprises extracting characteristics of granularity of the nucleus and of the cytoplasm identified by the application of a Mathematical morphology technique.

According to another embodiment, said first and second order statistical characteristics, and the granularity characteristics of the nucleus and cytoplasm are calculated in different components of various color spaces.

Finally, according to another embodiment, said calculation of the intrinsic characteristics of said regions of the nucleus, cytoplasm and external area of the cell of said abnormal blood cells comprises calculating a parameter of hairiness of the identified cytoplasm, wherein said segmentation is performed by using a technique of grouping the components of different color spaces and the Watershed Transformation into said digital images of abnormal blood cells.

The villus parameter is calculated, according to a preferred embodiment, using a threshold segmentation of a green component of the abnormal blood cells by calculating the total number of pixels of said segmented green region of the digital image.

The classification of abnormal blood cells comprises at least one classification into five different groups. In a preferred way these five groups will be: N

or healthy patients; HCL or patients with Tricoleukemia; LLC or patients with Chronic Lymphoid Leukemia; MCL or patients with lympoma of the leukemic mantle; and B-PLL or patients with Prolymphocytic B-cell leukemia.

Finally, said steps up to cl will be repeated as many times as digital images of abnormal blood cells are acquired.

The invention, according to a second aspect provides several programs

software to carry out the method of the first aspect of the present invention.

Brief description of the drawings

The above and other advantages and features will be more fully understood by

From the following detailed description of some embodiments with reference to the attached drawings, which should be taken by way of illustration and not limitation, in which: Fig. 1 shows a flow chart that includes the different stages of the method proposed by the present invention for an exemplary embodiment; Fig. 2 illustrates a flow chart of the general scheme of the method proposed by the invention, in the case of applying a Watershed transformation; Fig. 3 illustrates a flow chart of the general methodology for the segmentation of lymphoid cells; Fig. 4 shows an example of the segmentation methodology stage, where the region that limits a normal lymphoid cell is obtained;

Fig. 5 shows an example of the different steps to obtain the Watershed transformation of the whole cell through the imposition of minimum markers on the gradient of the Y component, in a lymphoid cell in a patient with MCL;

Fig. 6 shows an example of several cells of a patient with MCL and its corresponding labeled image obtained after applying the Watershed transformation to obtain the complete regions of each cell;

Fig. 7 shows the complete segmentations of the cells used as examples

in the previous figures;

Fig. 8 shows the general scheme that is followed to perform the extraction of intrinsic characteristics or descriptors according to an embodiment of the present invention;

Fig. 9 shows the image of a lymphocyte with its respective histogram; Fig. 10 shows the creation of the co-occurrence matrix, with a distance of 1 pixel, according to an embodiment example; Fig. 11 shows the example of a complete granulometric curve for the case of the normal lymphoid cell of Fig. 9; Y

Fig. 12 shows the steps to calculate the villus descriptor of a Tricoleucocyte according to an embodiment of the present invention: where in a) the calculated Watershed lines are shown; in b) the region of interest to describe the villi, in c) the resulting histogram for the region of interest; and in d) the pixels between intensities 70 and 215 of the green component.

Detailed description of an embodiment example

A flowchart depicting the basic stages of the proposed method is shown in Fig. 1. These basic steps to perform digital cell image processing include: acquisition, preprocessing, segmentation, description (extraction of the main morphological characteristics), and recognition and classification.

As discussed above, the present invention provides a

new mechanism implemented and / or executed on a computer to identify descriptors

or intrinsic characteristics of abnormal blood cells, such as the lymphoid cell, that allow the classification of different diseases or pathologies, for example neoplasms. For this, the proposed methodology develops a segmentation that separates the regions of the nucleus, cytoplasm and periphery from the outer edge of the cytoplasm or external area of the cell, from which the calculations of the different variables can be performed. The analysis of the variables is preferably carried out by means of a linear discriminant analysis, where the classes referring to each of the different pathologies are separated.

The first step to perform, like the techniques known in the state of the art, is the preparation of blood samples and the acquisition of digital images. There are different factors that affect the quality of the acquisition and processing of the images, derived from the SP smear: a) staining process; b) variations in lighting; c) optical geometric distortions due to the type of lens and microscope; d) format in which images are saved; e) random noise; f) lack of contrast between tone levels, etc. To improve the quality of blood cell images, different techniques can be applied, such as: spatial and frequency filtering, color transformations, histogram manipulations and others.

Subsequently, a staining of the blood cells is performed for which various techniques can be used, among which the staining of May-Grünwald-Giemsa (MGG), which highlights the basic components of the cell. It is of great importance to develop a precise and automated methodological process to achieve high quality SP smears in terms of staining, since it is the first step in the acquisition process, and is essential for the development of subsequent processing algorithms. Thus, low quality extensions will result in degraded images, that is, with the inclusion of noise to the entire analysis system. For the observation of blood cells the most widely used method is by means of an optical microscope and with immersion objectives of 50 or 100 magnifications (500 or

1000 total increases). However, as discussed above,

recently automated techniques have been developed for the acquisition and

Preclassification of blood cell images. These teams have a built-in

Motorized microscope, as well as a digital camera. An example of this equipment is the

CellaVision DM96.

5

Once the images of the cells have been acquired, according to an example of

embodiment and in a characteristic manner of the present invention, a step of

preprocessing that attenuates or enhances the characteristics of the image, such as edges,

limits or you can also modify the contrast. This procedure does not increase the information

present in the data, but it does increase the dynamic range of the properties of an image

1 d

to facilitate handling. Some of the preprocessing techniques that can be used

are the filtering of images by applying Gaussian filters, medium filters,

Wiener filter and / or convolution with Wavelets, in order to reduce the noise level. By

on the other hand, the images stained with MGG are bluish and purple, so also

it is possible to apply a color treatment of the cell images on them

fifteen

blood, which can separate, highlight or extract the best profiles.

Next, segmentation of abnormal cell images is performed

to separate the regions of the nucleus, cytoplasm and periphery from the outer edge of the cytoplasm, to

from which the calculations of the variables can be made. For segmentation it

Keep in mind that: the shape of the cells is variable, with different shades and textures

2D

both in the nucleus and at the cytoplasm level and that the periphery of the cytoplasm can show

extensions.

Segmentation in the proposed method can be done according to two

different alternatives. For example, either using an active contour segmentation or

using color segmentation by grouping and Watershed transformation. In the first

25

case, the method recognizes the cellular components using color information from the

original image of the cell and by applying contour detection techniques in

an H component of the HSV color space is obtained in the entire cell. Likewise

mode, the entire cell nucleus is also obtained using a color space

RGB

3D

In the second case, the method uses the composition of the different colors in the

image of the cell due to staining applied to the blood sample, especially

in the colors blue, magenta and pink, seen as a first perceptual approach. The

developed technique involves segmentation in a preferred way

color grouping to obtain the nucleus directly and indirectly for the segmentation of the entire lymphoid cell. Fig. 3 shows the proposed general methodology, which in turn contains sub-algorithms that are used throughout the segmentation. The most important procedures of the mentioned methodology are mentioned below:

•

Cells only: this algorithm takes the original image as input and produces as an output a labeled image where the regions of the lymphoid cells are separated from the rest. The main part of this algorithm uses the Kernel Spatial Fuzzy C-means (KSFCM) grouping technique to the Cyan and Magenta components of the CMYK color space to isolate the red blood cells from the lymphoid cell. Subsequently, the Watershed transformation is applied to obtain each region that contains cells.

•

RG approach of the nucleus: its input is the original image and produces a simple approximation of the lymphocyte nucleus, using a combination of the red and green components of the RGB space. Its usefulness is that it can be located approximately where the cells of interest are located.

•

XVZ Core: This procedure mainly uses the KSFCM grouping technique on the XYZ color space to isolate the nucleus, cytoplasm and background groups. To precisely segment the nucleus, it requires as input the original image and the region of interest of the lymphoid cells (obtained in Cell Only). As a result, it produces a binary mask of the nucleus (s) of the lymphoid cell (s), the gradient of the Y component of the XYZ color space, as well as the background group of the cluster. First, the original image is preprocessed through a smoothing filter

but that retains the edges of the objects. Next, a binary mask is generated from the "real" region where lymphocytes (extO) are found, from the approximate mask of the nucleus (obtained by combining the R and G components) and the labeled image that separates the cells (product of Only cells). Fig. 4 shows the previous process for a normal lymphoid cell: Fig. 4a represents a normal lymphoid cell, Fig. 4b is the binary mask that approaches the nucleus, Fig. 4c shows the labeled matrix that separates the cells in different regions and Fig. 4d is the resulting binary mask that encloses the lymphocyte of interest. The perimeter of the latter region that limits the lymphoid cell (s) is used as the external marker to be used in the Watershed transformation.

The next stage of the process consists in determining the real nucleus and the consequent generation of the internal marker of the WT for the cytoplasm. Fig. 5 shows the different steps applied to a cell (Fig. Sa) of a patient with Mantle Cell Lymphoma. The binary nucleus mask is achieved by the XYZ Core subroutine (Fig. Sb), in which the original image and the region surrounding the cells of interest (extO) have been used. Subsequently, the core mask is thinned if it touches the external marker and is

{{improvement l1 is taken as the final segmentation of the nucleus, with some operations of

mathematical morphology This same binary mask is used as an internal marker of the WT that will determine the cytoplasm. Fig. Sc shows the overlap of both markers. Thus, the external and internal marker are imposed as minimum in the gradient of the component y (Fig. Sd and Fig. Se), to then perform a Watershed transformation and thus generate a labeled image where the whole cell (s) is located ( s) lymphoid (s), as shown in Fig.sf.

At this point, the three regions of interest have already been generated indirectly: the nucleus, the entire cell and the region surrounding the cell. However, the final objective is to separate the binary masks of each region for each cell, since it can be presented that there are several lymphocytes per image. In this case, and in an improvement of the present invention, the above procedure is similar but the different regions must be separated by cell. Fig. 6 shows an example with four lymphocytes of a patient with MCL and its corresponding image labeled product of all the aforementioned stages. Therefore, another stage is created in which the masks are separated from the possible different cells. This procedure is highlighted by the shaded block Masks of Fig. 3, whose final product is the binary masks of the cell, the nucleus and the region surrounding the cell. Here are some important items that the algorithm preferably performs at the stage of the Masks block:

•

It joins the entire cell regions when there are several nuclei from the image labeled product of the WT.

•

Or determine the number of nuclei for each cell region and the number of levels labeled for each zone according to the image labeled by the WT.

•

Create the label areas corresponding to the regions that enclose each region produced by the WT (regardless of the background).

•

It determines the remains of cytoplasms that cross the borders of the regions that delimit each lymphoid cell.

•

Gets the lines that delimit the cell, the nucleus and the region outside the cell.

•

Obtain the masks for each cell by separating the cases: (1) there is only one nucleus per cell, (2) several nuclei per cell are presented in which case (a) there are several cells, each with a nucleus or (b) They have a cell with several nuclei.

•

For the resulting masks: those cells at the edges of the image are removed and

5 red blood cells that may occur in the region surrounding the cell are eliminated through the red blood cell region of the background group (from space XYZ). Fig. 7 compares the segmentation results for the three lymphoid cells

shown above.

The segmentation process described and summarized in Fig. 3 further comprises

10 the creation of a database of binary masks (1 within the object, Ofuera thereof) corresponding to the three regions (nucleus, whole cell and region surrounding the cell) for the whole set of cells. Once the segmentation has been carried out, the method comprises the extraction of characteristics or descriptors according to a scheme like the one shown in Fig. 8.

15 Description is the stage at which information about the objects in the image to be analyzed is extracted. The characteristics that can be calculated depend on both qualitative and quantitative reasoning and the evaluation of abstract mathematical descriptors that can provide information. Table 1 shows all the descriptors used in a characteristic manner by the present invention according to

20 several embodiments.

Type of descriptor

Quantitative descriptor

Cytoplasmic profile descriptor (1)

Villus estimation

Geometric Descriptors (10)

Cell area Cell diameter Conical eccentricity of the cell Perimeter of the cell Area of the nucleus Diameter of the nucleus Conical eccentricity of the nucleus Perimeter of the nucleus Core-cytoplasm ratio Eccentricity of the nucleus with respect to the cytoplasm

Cytoplasm

Nucleus

~ ooM u> <ID "<!," OC ') f /) ~ ID ~' "or ~ _ :: J 0 ..> <~ IDu_ f /) IID or

First-order statistical descriptors (12 x 3) Mean Standard Deviation Asymmetry (Skewness) Kurtosis Energy Entropy Mean Standard Deviation Asymmetry (Skewness) Kurtosis Energy Entropy

Second-order statistical descriptors (14 x 3)

Contrast Homogeneity Contrast Homogeneity

Correlation Energy Entropy Variance Variance variance

Correlation Energy Entropy Variance Variance variance

Granulometry descriptors: calculated on the granulometric curve (8 X 3)

Mean Standard Deviation Asymmetry Kurtosis Mean Standard Deviation Asymmetry Kurtosis

..

Table 1. Descriptors of the blood cell. A set of 113 descriptors were used: 1 of the cytoplasmic profile, 10 geometric, and 34 of texture applied to the three color planes of the Lab space

The main descriptors calculated or used in a preferred manner in the present invention are detailed below:

Geometric Descriptors:

5 Area: defined as the internal area of the object in question, it is calculated by counting the

number of pixels that the region contains. Diameter: defined as the diameter of a circle with the same area as the region. Conical eccentricity: calculated as the eccentricity of an ellipse whose second

moment is equal to that of the region. An eccentricity equal to 1 represents a segment

10 rectilinear, while if it is O it represents a circle. Perimeter: is the length of the entire edge of the region. Core / cytoplasm ratio: For the specific case studied, the core-cytoplasm ratio is calculated by dividing the area of the nucleus by the area of the cytoplasm. Color descriptors -texture:

15 These descriptors allow you to extract information about the texture of the image. These were applied to each color plane of the space L * a * b *, on the region of the nucleus and cytoplasm, obtaining a total of 96 descriptors. The CIE L * a * b * color space is the most complete color space created by the International Commission of Illumination (Commission internationale de I'éclairage). This describes all the visible colors I host

20 human and was created as a reference model independent of the device. Since Lab was created to approximate human vision, it is perceptually uniform and its L component is quite close to the human perception of Luminosity. The three coordinates of the CIELAB model represent the luminosity of the color (L = Oproduce black and L = 100 represents a diffuse white; white specula can be of higher value), the

25 position between red / magenta and green (negative a values indicate green, while

positive values represent magenta) and the position between yellow and blue (negative values of b produce blue colors and positive values indicate yellow. Each descriptor is explained below according to its statistical interpretation or obtaining:

-

First-order statistical descriptors: The histogram is a function that shows for each pixel value i, the number of pixels H (i) (proportional to the frequency) that the image has that value. In the graphical form of the histogram, i is represented by the horizontal axis, and H (i) by the vertical. Fig. 9 is an example of the histogram of the green component of an image of a peripheral blood lymphocyte. Since the image is a discretized rectangular matrix, the pixels take the value i in a range [O, Ll] (L is not necessarily the number of values that the pixels can take is, but the number of containers under which they are calculated histogram frequencies). The histogram can also be interpreted as the probability density function. For example, p (i = 3) = H (3) / 16, is the probability that a pixel, which belongs to an image of 4x4 elements, has a value of 3. That is, that p (i) is the probability density , calculated as the frequency of the respective level, divided by the number of containers (L). Table 2 shows the first order statistical parameters: mean, standard deviation, asymmetry, kurtosis, energy and entropy, obtained from the histogram.

-

Second-order statistical descriptors: Second-order descriptors can provide more information about the texture of the objects to be analyzed. These parameters are defined from the occurrence matrix. This matrix arrangement represents the joint probability that two pixels have an intensity value "j" and 'T, respectively, at a distance d in a given direction. The co-occurrence matrix considers not only information on intensity levels, but also the position of pixels with similar intensity values.

Name

Formula

Half

¡, -, p = ¿ipU), = 0

fa

Standard deviation

(J = ¿(1 -p) 'p (i)

i = tl

Statistical asymmetry

"-1¿ (í-, u)" p (í) i ~ (J '

Kurtosis

"_1 ¿(í -¡ti 'p (') i ~". '

L-;

Energy

¿P '(í)

J = lI

L-j

Entropy

Lag, (p (íJ) I'U]

¡= TI

Table 2. First-order statistical descriptors

In Fig. 10, according to a preferred embodiment example, a possible

5

matrix to be used by the proposed 4x4 size method with three intensity levels. To the

side be

sample the co-occurrence matrix, with distance d = 1 And diagonal direction that

It points to the northwest. Its dimension is 3 x 3, since the intensity levels are: D, 1 and 2. The

element at position (1,1) of the co-occurrence matrix indicates that level "D" is next to the

level "D" 2 times, in the diagonal northwest direction. The element in position (2,3), relates

1O

level ull1 next to level u211 in the same direction, presenting itself once. Table 3 shows

mathematically define the second order descriptors used, from the matrix

of co-occurrence

-Descriptors

from Granulometry: The granulometry determines the distribution from

particle sizes

in a image, for that generally uses morphology

fifteen

mathematics to estimate the distribution of particles indirectly, without having to identify

and measure each of these on the image.

Formula name

; Ir. = 2,3, ... 2N

=

; k =, \ '- 1

=

Table 3. Second order descriptors, obtained with the co-occurrence matrix.

Since clear particles with some regular shape are brighter than the

5 background of the image, the Granulometry consists of applying openings increasing the size of the structural element of this morphological operation. The basic idea is that opening operations with a particular size should have a greater effect on regions of the image that contain clear particles with a similar size. In each step the sum of the pixel intensity values in the image resulting from operating with the opening is calculated.

10 This sum (when normalized by the sum of the intensities of the pixels of the original region is called the size distribution function for the particle size) should decrease as much as the size of the structural element is increased, since openings reduce the intensity of the shining particles. This procedure produces a vector where each element is equal to the sum of the intensities of the image pixels

15 operated (with opening) and each location of the vector represents a size of the structural element. To emphasize the changes between successive openings, the difference between adjacent elements of this vector 10 is calculated. To visualize the results you are differences (if it has been normalized is the density distribution function for the particle size) they are plotted, where Each peak is an indication of the dominant size distribution over the particles in the image.

If the entire procedure described above is performed with the morphological operation of c / osing, the method is called antigranulometry. The main difference is that the size distribution of the dark particles is measured. That is to say, as a new c / osing is executed (increasing the size of the structural element) there is a greater effect of the operation on the dark granules.

Granulometry produces information about bright particles, while anti-granulometry produces information about dark particles. One way to take advantage of both methodologies is to join the distribution densities into one. The result is the Granulometric Curve, which reflects the distribution of closing operations, placing them in the negative coordinates (size of the structural element for closing). The distribution for the iterations with opening is left the same and they are placed in the positive coordinates (size of the structural element for the opening). Finally, the density value for size zero is zero, so as not to affect the total contribution of the densities in subsequent quantitative measurements. The result is a graph as shown in Fig. 11, with some symmetry and covering all sizes from closing to opening.

The proposed method uses, according to an exemplary embodiment, preferably four measures for the extraction of information on the granulometric curve of the nucleus and the cytoplasm of the lymphocytes: the mean, the standard deviation, the statistical asymmetry and the kurtosis.

Cytoplasmic profile descriptor: A descriptor has been implemented that can estimate the amount of villi present on the outer edge of the cytoplasm by analyzing the region outside the cell, which has been previously segmented next to the nucleus and the cell. The number of pixels in this region that "do not belong to the background" is estimated by the number of pixels of a new thresholding segmentation between, for example, intensity levels 70 and 215 of the green component. In Fig. 12, the different steps for calculating the cytoplasmic profile descriptor of a Tricoleucocyte are shown.

Finally, all calculated descriptors are stored in a row vector for each cell. That is to say, that for a given set of cells there is a matrix with as many rows as cells to be analyzed. Once the characteristics of each cell have been

calculated, these characteristics are used to classify the abnormal cells into 5 different groups. Generally, these five groups will be: N; HCL; LLC; MCL or B-PLL. According to a preferred embodiment, the proposed method has been made on 1500 digital images of SP lymphoid cells stained with May-Grünwald-Giemsa in

5 the CellaVision DM96 (Cellavision AB, Lund, Sweden). The images correspond to: 180 images characteristic cells of healthy patients (H), 301 images of lymphoid cells characteristic of patients with Tricholeukemia (HCL), 542 images of lymphoid cells characteristic of patients with Chronic Lymphocytic Leukemia (LLC), 401 images of cells characteristics of mantle cell lymphoma (MCL) and 75 images of lymphoid cells

10 characteristics of prolymphocytic leukemia B (B-PLL). Thus, a database of descriptors in the form of a matrix of 1500 (cells) x 44 (descriptors) has been created. In order to analyze the characteristics of each of the aforementioned pathologies, the supervised classification algorithm Linear Discriminant Analysis (LOA) has been implemented. With the LOA classification algorithm, a cross-validation of 10 folders is performed

15 (10-fold cross-validation). This technique divides the data set into 10 equally spaced subsets. Then, a single subset is used to validate while the remaining nine are used for training. In this way, this process is preferably repeated 10 times, using each subset as the validation data. The proportion of poorly classified cells was only 1.93%. Table 4 shows the matrix of

20 confusion corresponding to the whole process.

Known

Predicted

H

HCL CLL MCL B-PLL ·/.Right

H

179 OR OR OR 2 98.9

HCL

5 294 OR OR 2 97.7

CLL

5 OR 536 OR one 98.9

MCL

4 OR OR 391 6 97.5

B-PLL

3 one OR OR 71 94.7

""

Table 4. Matm of confuslon of cross-walled technology (LO-fold) using LDA.

The scope of the invention is defined in the following set of appended claims.

Claims (19)

 1. A computer-based method for recognition and classification of abnormal blood cells, which comprises classifying cells based on automatic processing techniques and blood sample analysis techniques, including acquiring digital images of abnormal blood cells from a plurality of blood cells, characterized in that it comprises performing the following steps: a) segment said digital images of abnormal cells by providing regions
identified from the nucleus, cytoplasm and external area of the cell of said abnormal blood cells of said digital images;  b) calculate intrinsic characteristics of each of said identified regions of the nucleus, cytoplasm and external cell area of said abnormal blood cells comprising calculating at least the geometric characteristics of said identified regions;  c) automatically recognize and classify abnormal blood cells based on said intrinsic characteristics calculated from said identified regions; and d) using said recognized and classified abnormal blood cells to make a diagnosis of hematological diseases.

 2.

Method according to claim 1, characterized in that it comprises performing, prior to said segmentation, a preprocessing step of said digital images of acquired abnormal blood cells.

 3.

Method according to claim 1, characterized in that said segmentation is performed using an active contour technique in said digital images of abnormal blood cells using the Gradient Vector Flow (GVF).

 Four.

Method according to claim 1, characterized in that said segmentation is performed using a technique of grouping the components of different color spaces and transforming Watershed into said digital images of abnormal blood cells.

 5.

Method according to claim 1, characterized in that said calculation of the intrinsic characteristics of said identified regions of the nucleus, cytoplasm and the external region of said abnormal blood cells further comprises extracting first order characteristics of said identified regions based on a histogram of the region identified.

 6.

Method according to claims 1 or 5, characterized in that said calculation of the intrinsic characteristics of said identified regions of the nucleus, cytoplasm and of the external area of the cell of said abnormal blood cells further comprises calculating second order characteristics of said regions identified on the basis of a co-occurrence matrix of each region identified.

 7.

 Method according to claims 1, 5 or 6, characterized in that said calculation of the intrinsic characteristics of the identified regions of the nucleus, cytoplasm and of the external area of the cell of said abnormal blood cells further comprises extracting characteristics of granularity of the nucleus and of the identified cytoplasm. by applying a mathematical morphology technique.

 8.

Method according to any of claims 5, 6 or 7, characterized in that said first order, second order and granularity characteristics are calculated in different components of various color spaces.

 9.

Method according to claim 1, characterized in that said calculation of the intrinsic characteristics of said regions of the nucleus, cytoplasm and external area of the cell of said abnormal blood cells further comprises calculating a villus parameter of the identified cytoplasm, wherein said segmentation is performed by the use of a technique of grouping the components of different color spaces and Watershed Transformation in said digital images of abnormal blood cells.

 10.

Method according to claim 9, characterized in that said villus parameter is calculated using a threshold segmentation of a green component of the abnormal blood cells by calculating the total number of pixels of said segmented green region of the digital image.

 eleven.

Method according to claim 1, characterized in that said classification of the abnormal blood cells comprises at least one classification in five different groups.

 12.

Method according to claim 1, characterized in that it comprises repeating said steps a) to c) as many times as digital images of abnormal blood cells are acquired.

 13.

Method according to any of the preceding claims, characterized in that said abnormal blood cells are lymphoid cells.

 14.

Method according to any of the preceding claims 1 to 12, characterized in that said abnormal blood cells are blast cells.

 fifteen.

Computer computer program comprising means of computer program code that executed in a computer implement the method according to steps a), b) and c) of claim 1.

 16.

Computer computer program comprising means of computer program code that executed in a computer implement the method according to step d) of claim 1.

 Fig. L   i  Fig. 2   '1 ", ¡- ;; $ :; h;" éY "e" O " I eleven I eleven I eleven 1 I cell  Fig. 3 ftll) 'J1ib ,, \ tm birurf, l; j: t, t; cgi1ín qu ;: (t)! UfCn ;; f'ouJe l'ry'tJj Fig. 4 iaWT  Fig. 5 tl "the IVT  Fig. 6 (b j C¡jiuiu Cüfl MCL Fig. 7 ] I I I I I . I I I I . EC . C) . LL Green cell component "IJ> Green component cell histogram V> .. !! '"  -'"'C. "t: l '" or  '"~ AND

-

: J Z

Gray levels Fig. 9 1 o Or 2 eleven (.) Address of the elect (c) Cirocurre Matrix: nda  Fig. 10   0.18 ~ 0.16 or "OR N '" ; :: 0.14 '" ' " or -5 0.12 " ~ '" x ", 0.1 '<I CL "O" 0.08 " ' " or  ~ 0 + 06 o: 0.04 '" "OR '" ?; "0.02 O //  r '· - "//

-

./

\ i \ ~ / r I \ one\ \\. } \ .. ,, / \ -25 -20 -15 -10 -5 O 5 10 15 20 particle radius (pixel is)  Fig. 11   green Fig. 12