Chest X-ray Classification Using Deep Learning for Automated COVID-19 Screening

Abstract

In today's world, we find ourselves struggling to fight one of the worst pandemics in the history of humanity known as COVID-2019 caused by a coronavirus. When the virus reaches the lungs, we observe ground-glass opacity in the chest X-ray due to fibrosis in the lungs. Due to the significant differences between X-ray images of an infected and non-infected person, artificial intelligence techniques can be used to identify the presence and severity of the infection. We propose a classification model that can analyze the chest X-rays and help in the accurate diagnosis of COVID-19. Our methodology classifies the chest X-rays into four classes viz. normal, pneumonia, tuberculosis (TB), and COVID-19. Further, the X-rays indicating COVID-19 are classified on a severity-basis into mild, medium, and severe. The deep learning model used for the classification of pneumonia, TB, and normal is VGG-16 with a test accuracy of 95.9 %. For the segregation of normal pneumonia and COVID-19, the DenseNet-161 was used with a test accuracy of 98.9 %, whereas the ResNet-18 worked best for severity classification achieving a test accuracy up to 76 %. Our approach allows mass screening of the people using X-rays as a primary validation for COVID-19.

Supplementary information: The online version contains supplementary material available at 10.1007/s42979-021-00695-5.

Keywords: COVID-19; Chest X-ray; Severity-based classification.

Fig. 1

Proposed system flow diagram. When a person passes through the X-ray screening machine, his X-ray image is passed as an input for further processing. The input chest X-ray image is passed through a VGG-16 model and labeled as either normal, tuberculosis, or pneumonia. Further, the pneumonia affected X-ray images are passed through a DenseNet-161 model and categorized as normal pneumonia or COVID-19. The COVID-19 images are passed through a ResNet-18 model and classified as severe, medium, or mild

Fig. 2

Configuration and performance of VGG-16 a Model of VGG-16. The input size of the image is 64 $\times$ 64 $\times$ 1. After this, the image passes through a 2D convolution layer with dimensions 224x224. Then it passes through the next layer with dimensions 112 $\times$ 112. After this, it passes to consecutive convolution layers along with MaxPooling with dimension changes 56 $\times$ 56, 28 $\times$ 28, and 14 $\times$ 14. Then it goes to a Max Pooling layer with dimension change 7 $\times$ 7. After a series of 2500, 4096, and 4096 fully connected layers, the X-rays are classified into normal X-ray, Tuberculosis affected X-ray, and Pneumonia affected X-Ray. b The graph of the learning rate versus a loss in the range of 0.8 and 1.7 (red dot shows selected learning rate). c A total of 322 images were tested. 127 were correctly labelled as COVID-Pneumonia, 104 were correctly labelled as normal and 78 were correctly labelled as Tuberculosis

Fig. 3

Configuration and performance of DenseNet-161 a Model of DenseNet-161. The input size of the image is 64 $\times$ 64 $\times$ 1 after which the image passes through a 2-D convolutional layer with dimensions 112 $\times$ 112. Then it passes through the next layer with dimensions 56 $\times$ 56 After passing through the consecutive layers the dimensions change from 56 $\times$ 56 to 28 $\times$ 28, 14 $\times$ 14, and finally 7 $\times$ 7. After 1000 fully connected layers, the X-rays are classified into normal pneumonia and COVID-19. b The graph of the learning rate versus a loss in the range of 0.5–0.9 (red dot shows selected learning rate). c Confusion matrix of the output for DenseNet-161. A total of 235 images were tested, out of which 232 were correctly classified as COVID, and 3 were found to be wrongly classified. From 150 images, 149 were correctly categorized as pneumonia whereas 1 was wrongly classified

Fig. 4

Configuration and performance of ResNet-18 a Model of ResNet-18. The input size of the image is 64 $\times$ 64 $\times$ 1. After this, the image passes through a 2D convolution layer with dimensions 112 $\times$ 112 $\times$ 64. Then it passes through the next layer with dimensions 56 $\times$ 56 $\times$ 64. After this, it passes to consecutive 2D convolution layers with dimension changes 28 $\times$ 28 $\times$ 128, 14 $\times$ 14 $\times$ 256, and 7 $\times$ 7 $\times$ 512. After 1000 fully connected layers, the X-rays are classified into severe, medium, and mild. b The graph of the learning rate versus a loss in the range of 0.9 and 1.6 (red dot shows selected learning rate). c Confusion matrix of the output for ResNet-18. A total of 25 images were tested, out of which 7 were correctly labelled as a medium, 8 were correctly labelled as mild and 4 were correctly labelled as severe