A machine learning-based framework for diagnosis of COVID-19 from chest X-ray images

Abstract

Corona virus disease (COVID-19) acknowledged as a pandemic by the WHO and mankind all over the world is vulnerable to this virus. Alternative tools are needed that can help in diagnosis of the coronavirus. Researchers of this article investigated the potential of machine learning methods for automatic diagnosis of corona virus with high accuracy from X-ray images. Two most commonly used classifiers were selected: logistic regression (LR) and convolutional neural networks (CNN). The main reason was to make the system fast and efficient. Moreover, a dimensionality reduction approach was also investigated based on principal component analysis (PCA) to further speed up the learning process and improve the classification accuracy by selecting the highly discriminate features. The deep learning-based methods demand large amount of training samples compared to conventional approaches, yet adequate amount of labelled training samples was not available for COVID-19 X-ray images. Therefore, data augmentation technique using generative adversarial network (GAN) was employed to further increase the training samples and reduce the overfitting problem. We used the online available dataset and incorporated GAN to have 500 X-ray images in total for this study. Both CNN and LR showed encouraging results for COVID-19 patient identification. The LR and CNN models showed 95.2-97.6% overall accuracy without PCA and 97.6-100% with PCA for positive cases identification, respectively.

Keywords: Artificial neural network; COVID-19; Computer-aided diagnosis; Image classification; Principal component analysis.

Fig. 1

Workflow of proposed system

Fig. 2

Samples of X-ray image dataset used for proposed system; a images of COVID-19 affected cases, b images of healthy individuals

Fig. 3

Architecture of the proposed CNN

Fig. 4

Visual representation of neurons and weights in CNN

Fig. 5

Visual representation of kernel convolving with input vector and then applying 2 × 2 max-pooling

Fig. 6

Core function of LR

Fig. 7

Training and testing loss/accuracy graphs of CNN network when a variance = 1, b variance = 0.99 and components = 147, c variance = 0.98 and components = 126, d variance = 0.97 and components = 108, e variance = 0.96 and components = 96, f variance = 0.95 and components = 84, g variance = 0.90 and components = 48, h when variance = 0.85 and components = 30

Fig. 8

ROC curves for logistic regression model when a variance = 1, 0.99, 0.98, 0.96, 0.95 and 0.90, b variance = 0.97, c variance = 0.85

Fig. 9

Confusion matrix for a CNN model at variance of 1, 0.98, 0.97, 0.96, 0.95, and 0.90, b CNN model at variance of 0.99, c CNN model at variance of 0.85, d LR model at variance of 1, 0.99, 0.98, 0.96, 0.95 and 0.90, e LR model at variance of 0.97, f LR model at variance of 0.85