R.A. Tomakova - Dr.Sc. (Eng.), Professor, Southwest State University
M.V. Dyudin - Post-graduate Student, Southwest State University
M.V. Tomakov - Ph.D. (Eng.), Associate Professor, Southwest State University

For statement of the diagnosis of diseases of lungs clinical and radiological symptoms and signs, and also laboratory diagnostics and additional methods of research are used. On an entrance of network model of the analysis and diagnostics of diseases of lungs we have some set of groups of the informative signs received as a result of use of various methods and technologies of diagnostics which need to be integrated into uniform network structure of classification. The technology of the analysis of pathological educations on the fluorogramm thorax includes two stages. At the first stage segments with a uniform brightness are allocated. The method of amplitude scanning with the subsequent morphological processing is for this purpose used. At the second stage it is necessary to check a hypothesis of accessory of the allocated segment to a certain class. Use multilayered perceptrons for classification of such data is difficult therefore at design of neural networks for diagnostics of diseases of lungs we use neural networks with macrolayers. In these neural networks for each pathology , determined by k to group of informative signs, the neural network of direct distribution answers. If are allocated the L diseases of lungs, for each group of informative signs of tk we receive the macrolayer containing so many neural networks of direct distribution, how many diseases of lungs are differentiated. The number of macrolayers is defined by number of groups of informative signs of K used for diagnostics. As decisions on belonging to the set disease are made in each of K layers, the network has to have K+1 a macrolayer. The output layer is intended for aggregation of the decisions made in each layer, and too is carried out in the form of a macrolayer of neural networks of direct distribution. Each neural network in macrolayers is adjusted on algorithm of the return distribution of a mistake. In the beginning neural networks in K the first layers are adjusted. After their control control of neural networks of an output layer is carried out. The database in which three blocks are allocated is necessary for network functioning with macrolayers: base of images, base of models of neural networks and base of training selections. As an example the intellectual system intended for diagnostics of pneumonia is considered. Results of control quality check of classification on control selection showed practical coincidence to results of expert estimation (an admissible error of 3%) that allows to draw conclusions on expediency of use of the received diagnostic rules for the solution of problems of rational management of treatment-and-prophylactic actions at diagnostics and treatment of pneumonia.