POSSIBILITIES OF ARTIFICIAL NEURAL NETWORKS FOR DIAGNOSING LUNG DISEASES BY X-RAY

Abstract

The article reveals the possibilities of artificial neural networks for diagnosing lung diseases on X-rays. It is emphasized that the development of machine learning algorithms provides ample opportunities in the field of automation of medical problems, and computer processing of X-rays increases the accuracy of image analysis, reduces the role of the human factor in decision making, evaluates the effectiveness of therapy and improves quality of life. Studies have confirmed that chest X-rays can be important for diagnosing patients and can also be useful for diagnosing various types of pneumonia, and a neural network architecture has been developed to diagnose lung disease. The structure of the network is described, with the separation of each individual layer, it is emphasized that all information processing in the hidden layer is to use a system of weights, namely, the use of certain coefficients during transmission from one layer to another. The strength of gravity or otherwise the importance of information is determined by the value of this factor, the greater it is, the more important information is transmitted between neurons. It is emphasized that the final information provided by the neural network will depend on many factors: through which neurons will pass information, what the internal architecture of the network looks like, as well as the number of hidden layers in the network architecture. It is noted that the main advantage of the presented method is that the machine of reference vectors can be easily integrated into the architecture of deep learning. The trained model forms feature vectors from the last fully connected model of deep learning and provides grades to each class, which is an important factor in the overall system of diagnosis. It is also emphasized that the use of the machine of reference vectors instead of the classifier allows to increase the accuracy of classification for diagnosing diseases. The article also highlights the shortcomings, which include the fact that deep learning models in combination with models of reference vector machines require at least two training data sets.