TECHNICAL FIELD:

The invention relates to the field of computational intelligent technique and in particular to medical image understanding.

BACKGROUND OF THE INVENTION :

In order to understand the medical images more accurately, there are three main themes to achieve this: the solid experience and cognitive intuition of a physician and the medical imaging devices used in addition to the aid techniques used to understand the medical images. As for the experience of the physicians are required, whether in building medical devices or aid techniques used to understand the medical images, in other word, precisely, a physician decision is still play final crucial role in the diagnosis.

Recently, there has been massive increase in the field of medical equipment and information technology in diagnostic imaging, which is observed in the last 30 years, this development is mostly obvious in diagnosis of diseases and disorders thus helping physicians to determine appropriate medical procedures. This leap in diagnostic tools techniques is promoted by the continuous evolution in diagnostic medical instruments side by side with information analysis sciences and image processing.

As for medical imaging devices, and based on the variety of modern of them, a physician can obtain different kind of images that present interesting organs and analyzed tissues in many different forms. One can use typical x-ray images (like mammography, dentist pantomography, computed tomography (CT) visualizations, ultrasound (US) presentations, nuclear images or thermograph plots - and many others)

As for aid techniques used, the field of biomedical informatics is a crossway of numerous academic fields simultaneously. It provides an interface between medical sciences and information technology. These aid techniques are included the most recent computational intelligence, statistical, data structures, and techniques means in analysis of medical images and thus diagnosing diseases and disorders on basis of establishing new algorithmic models for understanding medical images. These computational intelligence techniques have been employed in various applications of medical imaging, since it showed their ability to improve in more image processing phases like image acquisition, enhancement, segmentation, visualization, registration, reconstruction, comparison, analysis and classification of medical images compared with traditional methods.

SUMMARY OF THE INVENTION :

In this invention, the computational intelligence (CI) units for medical imaging are developed, at some units, a CI with a traditional technique is used, and in others, a full hybrid CI technique is used. It is according to the complexity at this phase. For example, to separate a pectoral muscle and label from mammogram image, pulse coupled neural networks (PCNNs) is used for image segmentation with k-nearest neighbor (KNN) for shape recognition system. For mass and micro-calcification clusters segmentation, a full hybrid CI technique is used (Fuzzy-PCNNs): maximum fuzzy entropy and fuzzy rule inference are used to improve the performance of the PCNNs, which make it more suitable for the user and more accurate in terms of performance of mammogram images segmentation. And for classification, a full hybrid CI technique is used (FAMNN with ICHCGA) to get a higher accuracy mean high TP and low FP in detection.

The system (31) comprising:

1. Image preprocessing unit (32)

To separate a pectoral muscle and label from mammogram image, image enhancement by histogram equalization (unit 71 ); pulse coupled neural networks (unit 72) is used for image segmentation with k-nearest neighbor (KNN unit 76) for shape recognition system and tracking boundary contours (unit 73). Unit 74 is continues until the ROIs complete, use the unit (77), which based on a result of KNN, consequence, the ROIs is removed or not, the unit (79) is used for remove the undesirable ROIs. In unit (75) the binary resulting image from shape recognition unit is reconstructed using filtering it from original image, and uses this image as input unit to Fuzzy-PCNNs unit. The main aim of unit (32) is to identify the objects in mammogram depending on the shape and not by the location only. In other words, if there some of ROIs are near the pectoral muscle or label. In this case and depending on location only, remove of the objects is considered a mistake

2. ROIs segmentation unit (33)

It consists of 3 units as follows: pre-defined unit (100), which is used to define the tissue types, pre Fuzzy-PCNNs segmentation unit (101) and Fuzzy-PCNNs segmentation unit (102).

In unit (100) multilayer perceptron neural networks (MLPNNs) classifier (unit 43) is used for recognize a type of tissue, in unit (44) image enhancement based on AHE in specific range. The unit (101 ) is comprised from image normalization unit (51 ), 2D image histogram unit (52) and fuzzy partition unit (53).

A Fuzzy-PCNNs unit (102) is used for mammogram images segmentation, and it is comprised from the following sub-units: fuzzy entropic threshold unit (54) that computes by use the output of unit ( 100). Unit (55) is an inverse of 2D Laplacian of Gaussian filter of the resulted images from unit (100), which is used as sharpening or high-pass filter, allow high frequencies pass and reduce the lower frequencies, and consequence is extremely sensitive to shut noise. To construct a high-pass filter, the kernel coefficients should be set positive near a center of kernel and in the outer periphery negative. Unit (56) is represents the Fuzzy-PCNNs feeding and linking. Unit (57) represents a fuzzification of linking coefficient (β). Unit (58) represents fuzzy rule of linking coefficient. Unit (59) is a defuzzification .unit (60) is the linking modulation of Fuzzy- PCNNs. With unit (91 ) the initialize thresholding from table (61 ) is selected according to tissue types. Unit (62) is Fuzzy-PCNNs thresholding, which is computed according to fuzzy max entropic. Unit (63) is Fuzzy-PCNNs pulse generator. Unit (64) is sequence of resulted images. In unit (65) polygon mask is used to fill the ROIs that are resulted from Fuzzy- PCNNs (102). Unit (66) is as input for feature extraction unit (34).

According to the tissue recognition system (unit 100), if tissue type is not dense tissue, in the other hand, one from first four types see table 2, and mammogram image enhancement by AHE (unit 44), which is applied on a specific range of gray levels. The main aim of unit (33) shown as follows: J

Where a tumor (cancer) has high density in most tissues, therefore, the AHE (unit 44) may show insignificant regions as significant regions, or it may combine the ROI with other is as insignificant region exactly in high density tissue. In other hand, most the content of a mammogram image has small intensity, especially in the surround area of breast boundary (background) or soft surround area of ROI boundary. This problem make from AHE is not efficiently, if a ROI location is near to a boundary of the breast, and therefore an enhancement of ROI is coming backfire than expected. And to avoid this effect, the enhancement region is within specified range of density. Mammogram images segmentation by Fuzzy- PCN s (unit 102):

Fuzzy-PCNNs new model is development of PCNNs model [ 1] and used to solve several main obstacles to segment mammogram images lays are explained as follows:. 1. Low contrast between normal and abnormal tissues and the noise in such images. In addition, the different in gray scale from tissue to another. 2. Unfortunately, the ROI boundary is uncertain and it shape is ill-defined. Moreover, the imprecision of gray values and vagueness in various mammogram image definitions make the segmentation problem more difficult to manage with deterministic or stochastic image processing schemes. 3. The mammogram images, which are different in contrast to each other and also of the same type. This reason makes a PCNNs task with these type of images is very difficult. And the appropriate values of the PCNNs parameters are changed constantly by the user.

One from the large problems is the density tissue (as ROIs background), which makes from PCNNs using without benefit. This type of tissues is affecting on the firing of ROI from the first iterations. The ROIs and density tissue often have higher gray scale values than other surround area in a mammogram image, this means a parts from this tissue are fired before the ROIs are fired or within it .consequently the shape of the tumor is affected. The tumor shape may appear incomplete or vice versa may be surrounded by a layer from this tissue (the ROI shape is hidden). To solve such as these problems Fuzzy-PCNNs unit is proposed, it is a hybrid model from PCN s with fuzzy set and fuzzy rule inference. Fuzzy set and fuzzy rule inference are used to adaptive the PCNNs parameters to make it is more efficiency for mammogram images segmentation.

3. Feature extraction (34)

In this invention, various special methods are used to extract the ROI features. Textural features such first order statistics, second order statistics features of well-known gray level co-occurrence matrixes (GLCMs), gray level run length matrixes (GLRLMs) features, fractional dimension features and multi-level wavelet decomposition features. Shape features and density features also are extracted.

4. Classification unit (35)

A full hybrid CI technique (FAMNN unit 85 with ICHCGA units (81 , 82, 83, 84, 86, 87, 88, 89)) that is used for assigning a class to the mammogram image to one from three classes (normal, benign and malignant cases). The ICHCGA is used for feature selection while FAMNN with ROC is used for classification (used for measure fitness of ICHCGA). Through this coupled, the most discriminating features among the classes are selected that will be used later in order to the ROIs recognition.

Integer-CHC Genetic Algorithm (ICHCGA) is proposed to attain a best balance between the exploration and exploitation. This is accomplished by maintaining diversity in the population and allowing the algorithm to focus in several areas of search space simultaneously. A CHC algorithm is developed to solve the problems of premature convergence that genetic algorithm frequently suffers, and it uses a conservative strategy of selection. In ICHCGA, integer-coded is used in lieu of binary coded, because the last one require a decodification step to apply the fitness function and also does not fit well when the number of features is fixed. FAMNN (unit 85) is one of the incremental learning algorithms are presented by Carpenter et al [2], in response to stability-plasticity dilemma (the catastrophic forgetting phenomenon through neural network learning). This unit is characterized by the following:

1. The FAMNN (nonlinear separability): able to build decision boundaries that separate classes of any shape and size.

2. The FAMNN (overlapping classes): creates decision boundaries to minimize the misclassification for all overlapping classes. In other words, there is no overlap between hyperboxes of different classes.

3. The FAMNN (training time): needs only one pass to learn and refine its decision boundaries.

5. Screen display unit (36)

In this unit, the report of final diagnosis is display within mammographic image. BRIEF SESCRIPTION OF THE DRAWINGS :

Fig.1 schematically shows a block diagram of an exemplary embodiment of the system 30 for medical images understanding.

Fig.2 illustrates pre-processing of medio-lateral oblique-view mammographic images.

Fig.3 shows the ROIs segmentation system by MLPNNs, AHE and Fuzzy- PCNNs unit.

Fig.4 shows the classification unit included (FAMNN unit 60 with ICHCGA unit 61 ).

Fig.5 shows user interface screen (pre-processing view).

Fig.6 shows user interface screen (enhancement and segmentation).

Fig.7 some images are fatty tissues have low contrast and other have high contrast.

Fig.8 the first binary resulting image (micro-calcification) from Fuzzy-PCNNs unit.

Fig.9 shape of malignant tumors with a surrounding region, each image has size (32x32).

Fig.10 shape of benign tumors without a surrounding region, each image has size

(32x32).

DETAILED DESCRIPTION OF EMBODIMENTS :

Figure 1 schematically shows a block diagram of an exemplary embodiment system 31 for medical images understanding, the system comprising:

• Data set unit 30 In this invention, mammographic images are considered, which are chosen from the Mini-Mammographic Image Analysis Society (MIAS). MIAS is an organization of research groups interested in the understanding of mammograms situated in UK, has produced a digital mammography database.

For example, the database contains left and right breast images from 161 patients with ages ranging from 50 to 65. In total, it counts 322 images mediolateral oblique (MLO)- view mammograms, belonging to three types, namely normal, benign, and malignant. There are 208 normal, 63 benign and 51 malignant (abnormal) images. Each image is categorized by an expert radiologist as one of three breast types namely fatty, fatty- glandular, or dense-glandular; In addition, the abnormal cases are further divided in six categories: micro-calcification, circumscribed masses, spiculated masses, ill-defined masses, architectural distortion, and asymmetry. An important characteristic of this database is that each abnormal image comes with a consultant radiologist's truth information, i.e., the locality of the abnormality is given as the co-ordinate of its center and an approximate radius (in pixels) of a circle enclosing the abnormality and breast \position (left or right). The mammogram image are represented as a two-dimensional matrix of intensity values (pixels) 1024 x 1024. According to mammography data set available for us, 200 image from most common types of breast cancer tissue are used. • Pre-processing unit 32

In Mediolateral Oblique (MLO) view of digital mammographic, the pectoral muscle, label and background (non-informative regions) are problem. The pectoral muscle, label and background have the same of density or larger than the tumor. Consequently, these parts affects a negatively of the impact on the work of the Fuzzy-PCNNs for segmentation. Therefore cleanup these parts from a mammogram image are recommended. The mammogram image pre-processing unit shown as follows (see figure 2):

1. In unit (71) original mammogram image enhancement by histogram equalization unit.

2. In unit (72) original mammogram image segmentation: the non-informative regions in mammogram image are separated from image by standard PCN s unit.

3. In unit (76) shape recognition system: run-length based algorithm is used for tracing boundary contours of all regions (label, pectoral muscle and ROIs) in a binary resulting image from first iteration of PCNNs. Then these regions are labeled using Haralick, Robert label components algorithm. After that the label components are passed to shape recognition system later. In mammogram image, the label has a rectangle or square shape, whereas the pectoral muscle has a triangle. And as for the ROI has a round, oval, or lobular, so that, KN classifier is used to remove a pectoral muscle and label.

4. In unit (75) the binary resulting image from shape recognition unit is reconstructed using filtering it from original image, and uses this image as input unit to Fuzzy- PCNNs unit. In unit (77) if a shape is rectangle, square, or triangle, a system cleanup them from original mammogram image by using image filtering unit (78) automatically In unit (73): tracing boundary contours of ROIs and feature extraction.

• Segmentation unit 33

According to figure 3, this unit is divided to three sub-units shown as follows:

1. Unit (100)

In unit (100), the tissue recognition unit based on multilayer perceptron neural networks (MLPNNs).

The input unit (90) is taken from first binary resulting image in unit (72) and resulting image in unit (75), the features vector is extracted. These features shown as follows:

• A size of ROIs in binary resulting image at first iteration (in unit 72).

• 9 density features from resulting image (unit 75), 19 features are derived from each GLCMs.

The features vector is used as dataset to MLPNNs classifier (in unit 43). In this invention, the image dataset is divided to 6 classes according to tissue type and its contrast see table

2. According to the tissue recognition system, if tissue type is not dense tissue, in the other hand, one from first four types, and mammogram image enhancement by AHE unit (44) is applied on a specific range of gray levels.

2. Unit (101 )

The unit (101) is represented as pre-processing unit to Fuzzy-PCNNs unit (102), this unit comprising from the three sub-units as follows:

2.1. Image normalization unit (51 ) Normalization of the image resulted image from unit (91) if tissue type is not dense tissue, in the other hand, one from first four types, or from AHE unit (44), this image having gray levels ranging from l _min to l _max can be modeling as an array of fuzzy number, each element in the array is the value representing the degree of brightness of gray level between 0 and 1 , using the following min-max normalization formula.

2.2. 2D histogram unit (52)

2D histogram of resulted image from unit (91 ) if tissue type is not dense tissue, in the other hand, one from first four types, or from AHE unit (44). Each the bin of the 2D histogram represent a frequency of occurrence of each (level, local average gray level) pair. The bins form a surface with ideally two peaks corresponding to background and object regions. Thus, the pixels interior to the object or background are found mainly to the near-diagonal bins of a 2D histogram and off-diagonal bins being contributed by edges and noise in the region. For an n grey-level region there are obviously ² bins. By means of two thresholds S and T a 2D histogram is divided into 4 quadrants. Since the shaded quadrants of 2D histogram will in general contain information only about edges and noise that are ignored in the calculation. The quadrants 0 and 1 contain the distributions corresponding to the background and object classes. For more detail see [3].

2.3. fuzzy partition unit(53)

In this unit, fuzzy entropy is employed for image thresholding based on both intensity distribution and local information among pixels. The purpose of this unit is to automatically determine the fuzzy region and optimal decay threshold parameter, therefore matrix of threshold in Fuzzy-PCN s unit (102), which is based on fuzzy entropy principle, given 2D histogram array of region N X M and K gray levels. The

2D histogram is divided to three regions: background region, fuzzy region and bright region. The background region is defined as the region with left top point (0, 0) and right button point (c, c). The overlapping region denoted by fuzzy-region, which starts at point (a, a) and ends at point(c, c). For more details see [3, 4, 5, and 6].

3. Unit ( 102)

Firstly, fuzzy pulse-coupled neural networks (Fuzzy-PCNNs) as developed model of PCNNs shown as follows: F _tj [n] = e- ^{E 5 n} .F^n-\]+s _n + VF∑ _kl M _!jUl Y _kl [n - 1 ( 1) kl

U _H [n] = F [n].* (l +¾ (n) .* L _:IJ [n]) (3)

Where ^defuzzification (centroid method (unit 59) of as shown in fuzzy rules unit (58).

Where

θδη = eq. 6 in unit(62).

The main purpose of this unit is separation the mammogram image regions well and full robotic. This unit consists of 13 sub-units shown as follows:

3.1. Fuzzy en tropic thresholding unit (54 )

In unit (102) the fuzzy entropic thresholding according to [3, 4, 5, 6], we can obtain the optimal threshold. Multi-level segmentation is performance by searching for local maxima of fuzzy entropy with predetermined window (fuzzy region) shifted along the histogram. For memberships of both fuzzy subsets: background and object.

Unit (55) is an inverse of 2D Laplacian of Gaussian filter of the resulted images from unit (100), which is used as sharpening or high-pass filter, let high frequencies pass and reduce the lower frequencies, and consequence is extremely sensitive to shut noise. To construct a high-pass filter, the kernel coefficients should be set positive near a center of kernel and in the outer periphery negative. Unit (56) is represents the Fuzzy-PCNNs feeding and linking. Unit (57) represents a fuzzification of linking coefficient (β). Unit (58) represents fuzzy rule of linking coefficient. Unit (59) is a defuzzification .unit (60) is the linking modulation of Fuzzy-PCNNs. With unit (91) the initialize thresholding from table (61) is selected according to tissue types. Unit (62) is Fuzzy-PCNNs thresholding, which is computed according to fuzzy max entropic. Unit (63) is Fuzzy-PCNNs pulse generator. Unit (64) is sequence of resulted images. In unit (65) polygon mask is used to fill the ROIs that are resulted from Fuzzy- PCNNs (102). Unit (66) is as input for feature extraction unit (34).

3.2. Fuzzy-PCNNs filters unit (55 )

Fuzzy-PCNNs filters unit (55) is an inverse of 2D Laplacian of Gaussian filter of the resulted images from unit (100). Which is used as sharpening or high-pass filter, let high frequencies pass and reduce the lower frequencies, and consequence is extremely sensitive to shut noise. To construct a high-pass filter, the kernel coefficients should be set positive near a center of kernel and in the outer periphery negative.

3.3. Fuzzy-PCNNs feeding and linking unit(56)

Unit (56) is represents the Fuzzy-PCNNs feeding and linking see eq. (1 , 2).

3.4. Fuzzification of linking coefficient (β) unit (57)

In this Unit a fuzzification of linking coefficient (β) is presented. The pixels in n X n nieghbarhood region X surrounding each pixel (i, j) from feeding inputs, x _; is gray level of (i, j) pixel in X. Let μ _χ (x _t , ) denote the membership value represents the degree of coefficient between (i, j) pixel (F^ in PCNNs) with (L _; - in PCNNs) in X. A fuzzy membership of region set X is mapping μ from X into interval [0, 1 ]. For membership function, the homogeneity and edgeness [7] measures are computed as fuzzy rules.

3.5. Fuzzy rule of linking coefficient unit(58)

In this unit, fuzzy rule of linking coefficient is represented see eq. [3, 4, 5, and 6].

3.6. Defuzzification of linking coefficient unit(59)

In this unit, a defuzzification of fuzzy rule of linking coefficient, see [3, 4, 5, and 6].

3.7. Fuzzy-PCNNs linking modulation unit(60)

In this unit, the linking modulation of Fuzzy-PCNNs is represented see eq.3. 3.8. Initialize thresholding table unit(61)

This unit is represent the initialize thresholding from unit (61) is selected according to tissue types see table 2.

3.9. Fuzzy-PCNNs thresholding unit(62)

In this invention, in Fuzzy-PCNNs unit (102) an optimal decay thresholding parameter αθδ ^η is calculated as follows:

θδη = max ((t _ttes + i _maxfn ),max. (μ, _αχ ) ) (6)

Where i _maxfn is the global maximum fuzzy entropy of a normalize image (maximum entropy of fuzzy number), t _¾iss is the coefficient based on the type of mammogram tissue see table 2, which is determined based on the experiments. And the max(u _max )) is the maximum fuzzy entropy of ii _max matrix. Where u _max is a matrix of membership for optimal thresholding (maximum fuzzy entropy) for feeding with its surrounding neighborhood L .

3.10. Fuzzy-PCNNs pulse generator unit (63)

See eq.4.

3.1 1. Resulting images unit (64)

In this unit (64), a sequence of binary resulted images. See eq.4.

3.12. Filling of ROIs unit (65)

In unit (65) polygon mask is used to fill the ROIs that are resulted from Fuzzy- PCNNs (102). In binary image ,the regions of interest are detected by polygon mask(tracing boundary contours) and fill it using filtering the ROI from original image, which returns an image that consists intensity values for pixels in locations where ROI image contains 1 's, and unfiltered values for pixels in locations where ROI image contains 0's. Then save the each region of interest as image.

3.13. Input for feature extraction unit(34)

Unit (66) is as input for feature extraction unit (34). Finally, Fuzzy-PCNNs algorithm for mammogram mass segmentation passes through various sub-units as shown in follows (also see figure 3):

1. Before use Fuzzy-PCNNs unit for Mammogram mass segmentation, a system of tissues recognition is worked, a MLPNNs classifier (see unit (43)) is applied to know the tissue type.

2. If tissue type is not a dense tissue, in the other hand, one from first four types as those explained in table 2, a mammogram image is enhanced by AHE method on a specific range of gray levels.

3. 2D histogram is calculated one only.

4. Set initial values of all Fuzzy-PCNNs parameters and matrixes.

5. _;,j (n) is calculated at each iteration. The coefficient parameter P _:,j (n) is different from F; _j with its surrounding pixels in the same region to other. Thus the coefficient degree is determined based on the strong relationship between the pixels with its surrounding pixels.

6. Given an eq. 6 an optimal decay thresholding parameter αθδ ^η is obtained base on a maximum of local maximum fuzzy entropy matrix M _ma:x , the global maximum fuzzy entropy of the normalize image (maximum entropy of fuzzy number H _ma:i f _n ), and a value of t _iss parameter, which is based on MLPNNs unit result and the experimental results see table 1.

7. The binary images are resulted using the Fuzzy-PCNNs unit. These images are included the ROIs (exactly a first binary image). all the ROIs are detected by polygon mask to draw each ROl separately. Therefore a new binary image for each ROl is created separately based on values of its boundary (by tracking boundary is aforementioned). Each boundary of ROl has same the location in original image. The ROl is filled using a filtering it from original image unit.

• Feature extraction unit 34 In this invention, various special methods are used to extract the ROI features. Textural features such first order statistics, second order statistics features of well-known gray level co-occurrence matrixes (GLCMs), gray level run length matrixes (GLRJLMs) features, fractional dimension features and multi-level wavelet decomposition features. Shape features and density features also are extracted.

• Classification unit 35

In figure 4 the classification unit shown as follows:

1.1. Feature selection by ICHCGA

Integer-CHC Genetic Algorithm (ICHCGA) is proposed to attain a best balance between the exploration and exploitation. This is accomplished by maintaining diversity in the population and allowing the algorithm to focus in several areas of search space simultaneously. ICHCGA based on CHC (cross-generational elitist selection, heterogeneous recombination, and cataclysmic mutation) is used to force diversity onto a population, when it may have become trapped around a sub-optimal solution.

1.1.1. Integer coded

For feature subset selection integer coded is not require a decodification step to apply the fitness function and does fit well when the number of features is fixed.

1.1.2. CHC units (81, 82, 83, 84, 86, 87, 88, 89))

ICHCGA based on CHC (cross-generational elitist selection, heterogeneous recombination, and cataclysmic mutation) is used to force diversity onto a population, when it may have become trapped around a sub-optimal solution [8].

1.2. Classification unit(85) as fitness function of ICHCGA

In this unit, a fuzzy artmap neural network (FAMNN) with receiver operating characteristics (ROC), FAMNN for training and testing, and ROC for evaluate the performance of FAMNN.

• Display( report) unit 36

In this unit, in mammogram image a location of ROIs of abnormal cases are shown (recognition of ROIs in mammogram image as normal, malignant or benign). Or with attached report.

Figures 5, 6 shown graphics user interface of this system (unit 31). Figure 7 shown the output mammogram images of unit (32).

Figures 8, 9, 10 shown the micro-calcification detection, the output of unit (33).

Tables 1 , 2 shown the types of mammogram image tissues and classification of them according to size and contrast.

References:

1 . Lindblad, Th.; and Kinser, J.M. (1998). Image Processing using Pulse- Coupled Neural Networks, Perspectives In Neural Computing. Springer- Verlag Limited. ISBN 3-540-76264-.

2. Carpenter G.A. and Grossberg S," A massively parallel architecture for a self-organizing neural pattern recognition machine, Computer Vision", Graphics, and Image Processing 37, 54- 1 1 5, 1987a.

3. H. D. Cheng, Yen-Hung Chen, Fuzzy partition of two-dimensional histogram and its application to thresholding, Pattern Recognition, Volume 32, Issue 5, May 1 999.

4. R.L. Kirby and A. Rosenfeld, A note on the use of (gray-level, local average gray-level) space as an aid in threshold selection. IEEE Trans. Syst. Man Cybernet. SMC-9 1 2 ( 1979), pp. 860-866.

5. B. Riecan, D. Markechova, The entropy of fuzzy dynamical systems, general scheme and generators, Fuzzy Sets and Systems, Volume 96, Issue 2, 1 June 1 998, Pages 1 91 - 199.

6. H. D. Cheng, Jim-Rong Chen, "Automatically determine the membership function based on the maximum entropy principle" Information Sciences, Volume 96, Issues 3-4, February 1997, Pages 1 63-182.

7. Moser B, Haslinger P, Kazmar Τ.'Όη the Potential of Hermann Weyl's Discrepancy Norm for Texture Analysis". International Conference on Computational Intelligence for ModeIling,Control and Automation Vienna, AUSTRIA, DEC 10-12, 2008.

8. L. Eshelman, The CHC Adaptive Search Algorithm, How to Have Safe Search When Engaging in Non-traditional Genetic Recombination, Morgan Kaufman, S. 265 - 283- 1991 .