Abstract:

Abstract The use of chest X-ray (CXR) in diagnosing respiratory diseases such as covid-19 and viral pneumonia has gained popularity due to its safe and non-invasive nature. However, interpreting CXR images as it is highly dependent on the expertise and experience of the radiologists may lead to inter-observer variability. A thorough analysis which in most cases takes some time then needs to be performed to get a final correct decision on whether a patient is indicated for a respiratory disease or not. Thus, research on devising more effective and efficient schemes to assist radiologists in the identification of respiratory diseases from CXR images is considered extremely important. In the previous studies, the convolutional neural network (CNN) based models demonstrate their capability to detect respiratory diseases not only they are found faster and reliable but also they are able to give a considerable high classification accuracy. In an attempt to obtain the general CNN based model capable of giving the best classification accuracy, sensitivity, and specificity on respiratory diseases from CXR images, in this paper powerful pretrained deep CNN models namely VGG16, DenseNet121, InceptionV3, Xception, and InceptionResnetV2 are computationally experimented on three different datasets. Dataset 1 is generated by combining images from Covid19 Radiography Database and Chest X-ray COVID-19 Pneumonia Dataset. Dataset 2 is created by combining images from Curated Chest X-ray Image Dataset, COVID-19 Pneumonia Normal Chest X-ray Images, and Chest X-ray COVID-19 Pneumonia Dataset. Meanwhile dataset 3 is taken from COVID-QU-Ex Dataset. State of the art accuracy of the pretrained CNN models is achieved by fine tuning the parameters of the convolutional layers of the base models and followed by feeding the high-level feature maps extracted from each corresponding base model into global average pooling (GAP) layer prior to classifying the respiratory diseases by fully connected layers. The highest average testing accuracy score as high as 99.2% is achieved by the InceptionV3 on multi-class CXR images.

Keywords: