NEURAL NETWORK CLASSIFIER OF CHEST X-RAY IMAGES FOR DETECTING SIGNS OF COVID-19 PNEUMONIA

Background. This study presents the development of a neural network-based binary classifier for detecting COVID-related pneumonia in chest X-ray images. Arguments are given in favor of using X-ray as an alternative to computed tomography in detecting abnormalities in the lungs, associated with COVID-19. An analysis of existing publications on automatic classification of X-ray images with signs of COVID-19 pneumonia is conducted. Materials and methods. The author's dataset consisting of 1240 chest X-ray images was used to train and test the model. The training part of the dataset was subjected to the augmentation procedure. An original fourteen-layer classifier model was proposed and trained over 20 epochs. Results. The classification quality was assessed using standard metrics. The following metric values were obtained: Sensitivity (Recall) – 95,4 %, Specificity – 97,8 %, Accuracy – 96,7 %, Precision – 96,6 %, F1-scope – 96 %. Supplementary testing on 228 images from the COVID-19 Radiography Database of the Kaggle platform demonstrated consistent performance: Sensitivity (Recall), Specificity, Accuracy – 96 %, Precision – 93 %, F1-scope – 94 %. Conclusions. The quality of classification of chest X-ray images by the developed model corresponds to the current level and is close enough to the medical one. The developed classifier can be used in clinical radiology practice as an AI-assistant for radiologists.