Martinencko Anna

Recently processing of digital images becomes one of the actual directions of development of computer technologies in medicine: image quality improving, recovery of the damaged images, recognition of separate elements.

In connection with complexity of the analysis of informativeness of the image and owing to subjectivity of human perception, there is no objective standard criterion of image quality. The brightness component plays not the last role in perception of the image, based on it image quality can be improved due to change of brightness of pixels.

Actual method of the solution of an objective are the artificial neural networks (NN) as possess good adaptive properties, and also learning capability for execution of the required functions. NN – represent a network of elements – artificial neurons – connected among themselves by synoptic connections. The network processes input information and in the course of change of the status in time creates set of output signals. Network functioning consists in conversion of input signals in time therefore the internal state of a network changes and output influences are created. However, the success of NN in many respects depends on architecture of algorithm of training and a choice of the properties used in training activity. Unfortunately, determination of architecture of a neural network is a process of tests and errors, algorithms of training shall be set carefully up under data. For optimum selection of architecture of a network and algorithm of its training evolutionary approach, which allows receiving suboptimal result in the minimum periods will be used.

Figure1 – Diagram of neuron (animation; the number of frames – 10 pieces, size – 109 KB)

In this operation, use of the genetic algorithm (GA) for setup and training of the neural network (NN) is offered in case of the solution of the task of improving of quality of medical images. The neural network is set up by means of the genetic algorithm according to contrast level.

1. Review of researches and development

1.1 Review of international sources

The analysis of publications of scientific development of foreign colleagues on similar subject showed that the main development are also connected using the genetic algorithm for creation of architecture of a neural network and its training. Examples of such articles are:

1) Eric Kant Notch’s and Chandrika Kamaza’s article – the Evolving neural networks for classification of galaxies. The purpose of article is to show that the GA can successfully solve some common problems in application to NN such as training of a network, a choice of the appropriate network topology and a choice of the appropriate properties [4].

2) Nikolay Kazabov's article – Model of the evolving neural network for verification of the personality on the basis of combinations of the speech and images. In article application of GA for creation of NN is also described. This article enters the method based on Evolutions of the Downlink System (ECOS) for tasks of verification of the personality. The method allows development of models of persons and their continuous adjustment on the basis of new speech and front images. Some pilot models of verification of the person based on speech and front features are developed on the basis of this method where the voice information and information on the image of the person is integrated at the level of art simulation of each person. It is shown that integration of speech and graphic functions considerably improves the accuracy of human verification model when there is a comparing with use only of picture data of the person or the speech [5].

3) Derek James's and Filip Tuker's article – the Evolving neural network the Active System of Vision for distinguishing of the form. In article the evolving neural network based on the active system of vision which is able to move a 2D surface in any direction along with ability to change the scale of the image and to rotate it is provided. It is shown that the system with such features can correctly classify the forms provided to it despite distinction in location, scale and rotation [6].

1.2 Review of national sources

Examples of operations in Ukraine are:

1) E.V. Mantula's article – the Predicting neural network with variable structure for monitoring of indices of environmental pollution. In article the analysis of possibility of use in tasks of environmental monitoring for prediction of nonstationary time series of the polynomial neural networks which are characterized by the high speed of training, and MGAA-networks (the Method of the Group Accounting of Arguments) which have variable structure with possibility of change during training is carried out. The neural network which integrates advantages of a multi-layer perseptron and the MGAA-network to training on the basis of small selection and numerical simplification of training activity of a network is offered [7].

2) O.K. Tishchenko's, I.P. Plis's, D. S. Koapalin's article – the Hybrid cascade optimized neural network. The new architecture and algorithms of training for a hybrid cascade neural network with optimization of a pool of neurons in each stage is offered. The offered hybrid cascade neural network provides computing simplicity and is characterized by the properties both tracing, and filtering [8].

3) B. B. Nesterenko's article, M. A. Novotarsky's – Simulation modeling of cellular neural networks. In operation the short review of the classical principles of the organization and functioning of the discrete cellular neural networks is this. The problem definition of the solution of the operator equation is considered by the local and asynchronous method oriented on implementation in homogeneous computing structures. Approach to creation of simulation models for cellular neural networks with different types of interneural interaction is described. Imitative algorithms of the synchronous, asynchronous and aggregate interaction in simulation models of cellular neural networks are offered [9].

1.3 Review of local sources

Examples of operations of students and employees of DonNTU:

1) V. N. Balabanov's article – An evolutionary method of rational planning of cuttings of coiled steel in production of electro welded pipes. In article the task of rational cutting of rolls arising in production of electro welded pipes is considered. The multi criteria model which except losses of rolled material in a withdrawal allows to consider production losses and a row of important technological limits is offered. The approximate metaheuristic method of the solution of the cutting task using the device of evolutionary computation is developed [10].

2) Dr.Sci.Tech., the prof. Yu.A. Skobtsov's article – Evolutionary computation in technical tasks. The new direction is provided to theories of artificial intelligence - the evolutionary computation including genetic algorithms, genetic programming, evolutionary strategy and evolutionary programming. Use of this new device on the example of tasks of design automation and processing of the image is considered [11].

3) Yu.A. Skobtsov's, V. Yu. Skobtsov's, K.M. Nasser Iyad's article – the Checking crosstalk tests of failure on the basis of evolutionary methods. The problem of creation of the checking tests for "cross failure" of the Crosstalk type, characteristic for deep submicronic design of element basis of the modern computer systems is considered. In case of the solution of an objective the many-valued logical simulation and the genetic algorithm of generation of the checking tests for this failure is used [12].

2. Development of a method of improving of quality of medical images

2.1 Improving of quality of medical images

The main idea consists in change of brightness of pixels of the source image for increase of its contrast range. Thus the image is processed pixel-by-pixel. We will assume that there is some function which executes conversion of each pixel of the image in such a way that the processed image is optimum from the point of view of some criterion. We will consider that change of brightness of pixel is made on the basis of information on statistical characteristics distributions of brightness in some neighborhood of radius of R these pixels, and also with use of information on value of average brightness of the source image. Thus change of brightness of pixel can be presented in a general view the following conversion of the checking tests for this failure:

where T – some function which executes conversion of each pixel; Lavg – the average brightness of the source image; L (x, y), L * (x, y) – current and new value of brightness of pixel; RL (x, y)) it is described by the following expression:

where RL (x, y) – convolution operation; F (x, y, c) – function of the filter; c – standard deviation of the filter.

In this operation the gaussian filter is used.

If the color image is processed, values of brightness of pixels beforehand are calculated, and then processing of the received monochrome image is carried out. Recovery of the color information is carried out as follows:

where the three (R * (x, y), G * (x, y), B * (x, y)) and (R (x, y), G (x, y), B (x, y)) represent according to a vector of the processed and initial color components of pixel (x, y).

For an assessment of image quality we will use the following factors offered in [1]:

1. Number of pixels on boundaries between areas with different values of brightness. Than more pixels will be on boundaries of areas of different brightness, especially the processed image will be contrasting.

2. Number of levels of gradation of brightness. It is used to avoid possible "degeneration" of the image in the binary.

This quality of the processed image is characterized by the value calculated on the following formula [1]:

where a and b – respectively width and height of the image in pixels, µ– number of pixels on boundaries between areas with different values of brightness, li – a share of pixels of the processed image with i-m the brightness level. The first item in a formula (6) is necessary for maximizing quantity of "boundary" pixels, and the second – for increase of number of levels of gradation of brightness.

2.2 Application of a neural network for improving of quality of medical images

We will formulate the task of neural network improving of quality of images as the task of approximation of some unknown function T from the algorithm described above. The NN approximating conversion (1) shall has 3 inputs (in case of the color image) and one output. The assessment of the processed images is made taking into account number of pixels on boundaries between areas with different values of brightness and number of the levels of gradation of brightness, which are present on the image (6).

2.3 Evolutionary approach to creation and training of a neural network

The generalized algorithm of GA [2]:

1. Installation of the GA parameters (probability of a crossover, mutation);

2. Generation of initial population. Population represents a set of chromosomes. Each chromosome corresponds to the certain trained NN;

3. Values of fitness function of individuals are estimated at populations;

4. Application of genetic operators;

5. Check of criterion of break. In case of its execution transition to a step 6, differently a step 3;

6. A choice of the best decision in the last population.

In the developed algorithm for creation and training of NN by means of GA, as an individual the neural network is considered. The fragment of coding of an individual is provided in a figure 2.

Figure 2 – Representation of an individual of GA

That is as an individual the three-dimensional array containing information on a structure of a neural network and its weight factors is used.

For an assessment of an individual fixed function of an assessment of an error of training of NN is used [3].

Genetic operators: selection, crossover, mutation.

Selection is executed by a standard roulette. Process of transposition is provided in a figure 3.

Figure 3 – Transposition process example

As a result of transposition two descendants who inherit the general neurons and communications of parents with redistribution of weight factor turn out, i.e. weights of parent individuals recombine with use of a 2-dot krossingover. Different neurons and communications are terminated between descendants in a random way. The dotted line in a figure 3 provided the breaking-off communications, by the whole line – remaining.

Operator of a mutation. The type is in a random way selected: 1 – accidental selected neuron is deleted, 2 – the accidental neuron is added, 3 – the weight factor of the existing neuron changes, 4 – communication of accidental neuron, 5 – any combination from previous four changes.

As criterion of break the acceptable error of training of NN is used.

Conclusion

The method of improving of quality of medical monochrome and color images is developed. We will mark, operation in this direction is only begun, its program implementation is now executed. Further it is planned to carry out:

1. Research of neural network image processing on each color channel separately;

2. Image processing with noise and distortions;

3. Research of dependence of results of operation of the received neural network decisions on a choice of the GA parameters.

The list of the used literature

1. Чернявский А.В., Цой Ю.Р., Спицын В.Г. Нейроэволюционное улучшение качества изображений // Современная техника и технологии: Труды XII Междунар. научнопракт. конф. молодых ученых. – Томск, 2006. – Т. 2. – С. 213–215.

2. Скобцов Ю.А. Основы эволюционных вычислений // Учебное пособие. – Донецк:ДонНТУ – 2008 – 326с.

3. Саймон Хайкин. Нейронные сети: полный курс, 2-е издание.: Пер. с англ. – М. : Издательский дом «Вильямс», 2006. – 1104 с.

4. Erick Cantu´-Paz, Chandrika Kamath. Evolving Neural Networks for the Classiﬁcation of Galaxies // Интернет ресурс – режим доступа: https://computation.llnl.gov/casc/sapphire/pubs/147020.pdf.

5. Akbar Ghobakhlou, David Zhang, and Nikola Kasabov. An Evolving Neural Network Model for Person Verification Combining Speech and Image // Интернет ресурс – режим доступа: http://www.geo-informatics.org/publications/ghozhakasiconip04.pdf.

6. Derek James, Philip Tucker. Evolving a Neural Network Active Vision System for Shape Discrimination // Интернет ресурс – режим доступа: http://anji.sourceforge.net/docs/james_gecco05.pdf.

7. Е.В. Мантула. Прогнозирующая нейронная сеть с переменной структурой для контроля показателей загрязнения окружающей среды// БИОНИКА ИНТЕЛЛЕКТА. – 2013. – № 1(80). – С. 112-116.

8. Тищенко О. К., Плісс І. П., Копаліані Д. С. Гібридна каскадна оптимізована нейронна мережа// Радіоелектроніка, інформатика, управління. – 2014. – № 1. – С. 129-134.

9. Б.Б. Нестеренко, М.А. Новотарский. Имитационное моделирование клеточных нейронных сетей // Искусственный интеллект. – 2005. – № 3. – С. 296-307.

10. В.Н. Балабанов. Эволюционный метод рационального планирования раскроев рулонной стали в производстве электросварных труб // Сборник научных трудов. Тематический выпуск «Информатика и моделирование». – 2010 – № 15. – С. 4 -9.

11. Ю.А. Скобцов. Эволюционные вычисления в технических задачах// Пленарні доклади. Інформаційні управляючі системи та технології та комп'ютерний моніторінг (ІУС та КСМ 2010). – 2010. – С.30-34.

12. Ю.А. Скобцов, В.Ю. Скобцов, К.М. Нассер Ияд. Проверяющие тесты crosstalk неисправностей на основе эволюционных методов. // Вестник «Информатика и моделирование» – 2010. – С. 170-176.