Information Technologies and Computer Engineering

The object of the study is the process of classifying lung radiographs for the diagnosis of COVID-19. The research is based on the use of deep convolutional neural networks, which make it possible to preserve spatial information and analyse complex images by preventing gradient attenuation. The principle of operation of convolutional neural networks and the advantages of their use in complex images are considered in comparison with artificial neural networks based on a multilayer perceptron. The main assumption of the study is the hypothesis that the use of a deep convolutional neural network for classifying lung radiographs will allow obtaining a sufficiently high accuracy result in diagnosing COVID-19 and will make it possible to automate the diagnostic process. The relevance of the problem of automated diagnosis of COVID-19 based on lung radiographs is considered. High-performance deep convolutional neural network architectures were trained using additional image processing methods to prevent overtraining. The results of the neural networks were compared and statistical information was provided to assess the quality of their work. An analysis of the results of the artificial neural network using the Lime image partitioning method is carried out. The expediency and prospects of using deep convolutional neural networks to automate the diagnosis of COVID-19 based on lung radiographs are substantiated. Common errors of artificial neural networks and possible approaches to their prevention are analysed. The disadvantages of using the considered approaches and the difficulties that may arise during automation are analysed, and possible options for improving the quality of the deep convolutional neural network are proposed

References