SOFTWARE OF AUTOMATIC CLASSIFICATION OF A CHEST X-RAY BASED ON HYBRID CLASSIFIERS

Abstract

The purpose of this research is to develop a method of classification of pathological formations, intended for the automated processing of x-ray images of the chest. Method to build classifiers with the aim of establishing the optimal parameters were applied to the Fourier descriptors, in order to form the space of informative features. Hybrid technology of classification of chest x-ray based on three-level hierarchic system is suggested. «Weak» classifiers, which are based on two ways of data analysis, are generated on the first level. The first approach of building the «weak» classifier is based on an analysis of amplitude of Fourier spectrum at a sliding window. X-ray pattern is scanning sequentially by windows of different scale. Each window determine the amplitude of Fourier spectrum with reference to which «weak» classifier is building. It assigns image segment, which is captured by sliding window, to certain class. The second approach of building the «weak» classifier is based on descriptors, resulting from approximation of intensity histogram on analysis window. The number of «weak» classifiers, based on two ways of analysis, is that of a number of the chosen scales of the windows. At the second hierarchic level, the solutions of the «weak» classifiers are combines in each way of analysis of the first hierarchic level. The definitive decision is taken by a final classifier, which aggregates the solutions of two classifiers of the second hierarchic level. The classifier built based on the neural networks of direct distrib. Results. Evaluated ution of trainees by implementing the algorithm of back propagation errors the quality of classification of the morphological formations in the images. Generated validation criterion of classification quality based on the number of incorrectly classified pixels in a given class to the total number of pixels of that class in the reference image. It is established that the proposed method of multi-window spectral transformation allows to perform differential diagnosis of pneumonia and oncological morphological formations by the criterion of not less than 15 %. Conclusion: The results of these studies can be used to build intelligent systems of decision support for the diagnosis and prediction of diseases.