Method for analysis of digitized x-ray images of the chest for differential diagnosis of infectious diseases of the respiratory system

Abstract

Due to the rapid development of deep learning theory, there has been significant progress in the field of computer vision, namely image classification and object detection. An analysis of current research in this area has shown that to solve the problem of diagnosing pathologies of the respiratory system, it is advisable to use neural network models with subsequent assessment of their effectiveness in real conditions. The purpose of the study is to improve the quality of differential diagnosis of infectious diseases of the respiratory system through the use of neural network models. The article proposes a method for analyzing digitized chest X-rays for the differential diagnosis of infectious diseases of the respiratory system, which consists of using a modified one-stage neural network model using a soft suppression algorithm for non-maxima. ROC analysis was used to assess the quality of the basic and modified neural network models. To solve the problem of using ROC analysis for multi-class classification, the one-vs-all method was used, which consists in creating binary classifiers for each class. During the study, a number of experiments were carried out for two neural network models (basic and modified). Based on the analysis of the obtained values of the quality indicator, as well as visual analysis on the test data set produced as part of testing neural network models, the effectiveness of the studied models was assessed. The use of the proposed method made it possible to take into account the imposition of bounding boxes when detecting objects in images, to minimize false removal of frames and to achieve more accurate results when solving the problem of differential diagnosis of infectious diseases of the respiratory system using digitized X-rays of the chest in comparison with the traditional algorithm for suppressing non-maxima. The analysis showed the high sensitivity of the basic and modified neural network models, however, the number of correctly classified examples for the modified model, which uses a soft non-maxima suppression algorithm, is higher for each of the three classes in comparison with the basic model.