Practical Machine Learning Techniques for COVID-19 Detection Using Chest X-Ray Images

Abstract

This paper presents effective techniques for automatic detection/classi-fication of COVID-19 and other lung diseases using machine learning, including deep learning with convolutional neural networks (CNN) and classical machine learning techniques. We had access to a large number of chest X-ray images to use as input data. The data contains various categories including COVID-19, Pneumonia, Pneumothorax, Atelectasis, and Normal (without disease). In addi-tion, chest X-ray images with many findings (abnormalities and diseases) from the National Institutes of Health (NIH) was also considered. Our deep learning approach used a CNN architecture with VGG16 and VGG19 models which were pre-trained with ImageNet. We compared this approach with the classical machine learning approaches, namely Support Vector Machine (SVM) and Random Forest. In addition to independently extracting image features, pre-trained features obtained from a VGG19 model were utilized with these classical machine learning techniques. Both binary and categorical (multi-class) classification tasks were considered on classical machine learning and deep learning. Several X-ray images ranging from 7000 images up to 11500 images were used in each of our experi-ments. Five experimental cases were considered for each classification model. Results obtained from all techniques were evaluated with confusion matrices, accuracy, precision, recall and F1-score. In summary, most of the results are very impressive. Our deep learning approach produced up to 97.5% accuracy and 98% F1-score on COVID-19 vs. non-COVID-19 (normal or diseases excluding COV-ID-19) class, while in classical machine learning approaches, the SVM with pretrained features produced 98.9% accuracy, and at least 98.2% precision, recall and F1-score on COVID-19 vs. non-COVID-19 class. These disease detection models can be deployed for practical usage in the near future.