INFORMATION TECHNOLOGY FOR THE EARLY DIAGNOSYS OF PNEUMONIA USING CONVOLUTIONAL NEURAL NETWORKS

Over the past few years, pneumonia has become one of the most common and severe lung diseases globally, and its treatment is vital nowadays. Clinical practice has proved that early diagnosis of pneumonia is a crucial factor in its successful treatment. An efficient approach to diagnosing pulmonary diseases, including pneumonia, is automated chest X-ray analysis implemented in clinical recommendation systems. However, it is still unclear what features of pneumonia in an X-ray image correspond to the early stage of the disease according to the automated method of diagnosis. The question of interpreting the results of digital diagnostics also remains open and needs further investigation. Therefore, to address an urgent issue of interpretation in digital diagnosis, we propose an information technology for the visual analysis of X-ray images to explain the results of diagnosing pneumonia. The technology comprises a classification model based on a convolutional neural network to extract mild features of early viral pneumonia and a modified method of distinctive localization to interpret the classification results. The neural network used in the study contains an effective dilated convolutional operation to combine features of various receptive fields. Our method of interpretation is based on applying weighted gradients to class activation maps. It distinguishes lung masks in the X-ray image and imposes thermal maps with a color gradient from blue to bright red. The red color corresponds to the most probable location of the pneumonia features in the radiograph. Such a modification provides excellent localization of abnormal areas on radiographs, removing the mild target features of early pneumonia. According to the computational results, our model surpassed other neural architectures in precision (98,5 %) but slightly conceded in classification accuracy (96,1 %) and recall (93,6 %). Moreover, it shows relatively low false positive and false negative rates, with 1,4 and 6,4 %, respectively. Overall, according to computational experiments, the proposed information technology can be an effective tool for instant diagnosis in the first suspicion of pneumonia.