Abstract
According to the World Health Organization, more than 17 million people die annually in the world from diseases of the circulatory system, half of them die from coronary heart disease and cerebral stroke. Stroke is a structurally complex disease based on various pathogenetic mechanisms. Taking into account the multicomponent nature of this pathology, as well as its complex structure, the medical community has developed various evaluation algorithms based on the recognition of various symptoms. Determining the effectiveness of these algorithms is recognized as the most important. Incorrect symptoms appear as a result of inaccuracies made by the radiologist in the process of manual annotation of computed tomography images. A convolutional neural network is used to perform image classification in a collection of brain stroke data. Since the data set is small, training the entire neural network does not give good results, therefore, to obtain more accurate results, model training uses the concept of transfer learning. Transfer learning is a method in which a model for a specific task is used as a starting point for another task. In particular, the Inception V3 model with Imagenet scales is used for the current task. The developed neural network was developed using the Tensorflow library of the Python programming language. Using machine learning, a data set with computed tomographic images of 2,515 normal and stroke-damaged areas of the brain was obtained. The task of a neural network is to classify a given image, that is, to determine whether it is normal or damaged. Using this algorithm, accuracy increased from 65 percent to 99.2 percent, and losses decreased from 7,532 to 0.756 percent. Key indicators: accuracy 99.6%, recall 99.2%, F1-the price was 99.1%.