Kassumova M.K., Tikchonov E.P.
Russia, Saint-Petersburg, “MEDI” Joint-Stock Company, email@example.com
ADAPTIVE METHOD OF FORMING THE SIGNS FOR AUTOMATIC CLASSIFICATION OF BIOMEDICAL IMAGES.
The problem of forming the sign complex for classification of biomedical images is being discussed in the report. It offers an adaptive method of forming the complex of signs, which are invariant against dislocation, rotation, scale changes and the total brightness of segment pictures. The method is based on the principal of a radial-circular count using the neuron networks.
Касумова M.K., Тихонов Э.П.
Россия, Санкт-Петербург, ЗАО “Меди”, firstname.lastname@example.org
АДАПТИВНЫЙ МЕТОД ФОРМИРОВАНИЯ ПРИЗНАКОВ ДЛЯ
АВТОМАТИЧЕСКОЙ КЛАССИФИКАЦИИ БИОМЕДИЦИНСКИХ ИЗОБРАЖЕНИЙ.
В докладе рассматривается проблема формирования комплекса признаков для классификации биомедицинских изображений. Предлагается адаптивный метод формирования комплекса признаков, инвариантных к сдвигу, вращению, изменению масштаба и суммарной яркости изображения сегментов. Основу метода составляет принцип радиально-кругового считывания с использованием нейронных сетей.
Automation methods of presentation and processing the information dominate in the fields of human informational activity and embrace new and new once, including the spheres, which hardly fell under automation because of complicated problems to be resolved. The tasks of automatic processing of visual information in medical and biological studies, particularly in radio-densitometry for dental aims, in clinical human morphology, in histochemistry, etc. are concluded to this grate number of problems.
A swift spread of relatively cheap devices of introduction the optic information in recent years as well as the possibility of their further processing on the base of personal computers have lead to appearance of according program means.
However, the development of program means is connected closely with the search of new methods and algorithms of image description using the so-called signs. Ideally a sign is the simplest feature of an image. The main task of image description by a complex of signs is the decrease of the reference dimensions of poly-segmental images under its hierarchical presentation in the form of some objects, fragments and segments. To simplify the process of analysis and interpreting of images, especially at an automatic mode, the requirement of minimizing the sign dimensions under maintaining the given quality of recognizing results is natural. One of the ways to meet this requirement is to reveal the signs bearing the features of invariance, for example against deformities of pictures (objects, fragments, segments), against their spatial location (displacement, orientation), against illumination degree, etc. Digitized image processing is also associated with fundamental difficulties. One of them is an enormous volume of information included into the double-sized vector signal called the image. That is why the problem of development and introducing the methods and algorithms directed on the reduction of excess and minimizing of a priori information under preserving a reliable and true description, identification and classification of images becomes topical.
In order to classify the poly-segmental pictures a set of signs owing high discriminative features and being simply countable must be selected. Forming the signs of poly-segmental images is carried out hierarchically. The hierarchical principle of forming the signs means that the image signs are formed taking in account the signs of objects, fragments and segments. On the other hand, by developing the object signs the signs of fragments and segments are being taken in consideration and, at least, the signs of fragments include the segment signs.
As the most complete set of signs for the description of a segmented image a multitude is used which consists of geometrical, densitometrical and textural features.
In the report the principle of forming the signs invariant to a dislocation, rotation, scale changes and summary brightness of the fragment image is considered. This principle concludes that by means of an adaptive algorithm the interpolation (approximation) of a segment with the so-called primitives is accomplished, i.e. with the simplest restorative functions owing a restorative error, which doesn't exceed the size given in the established sense. Adaptability of the algorithm allows keeping the invariance of primitive parameters against the segment being analyzed independently of above transformations of the image, which includes the analyzed segment. After the measured primitives describing the total of the analyzed segment one can build new signs or determine the known ones, for example, geometrical signs concluding in particular the form factor.
The essence of the adaptive algorithm of constructing the primitives includes the following:
Mathematically the construction of above primitives varies considerably. The report presents various iteration algorithm modifications of primitive construction, each of them having its own benefits and shortages. Conditions are adduced and the coincidence rate of algorithms is estimated. For example, under a bit-lined contour restoring which describes the analyzed segment the primitives represent triangles, their number and sum of areas under fixed brightness being independent on the center location of the radial-circular readings of information. As a result there becomes not necessary to search for the gravity center of an analyzed segment in order to keep the invariant features under proper description of the segment. To reduce the time of sign determining after the suggested method and to increase the interference stability the discriminate function of adaptive algorithms is offered to be realized using neurons having been regulated previously on the proper function of brightness during the teaching process.
A metrological examination of suggested algorithms after the example of determining the geometrical signs of analyzed images is adduced which demonstrates their sufficiently high interference stability.