Analysis of Contrast and Correlation between Deep Learning Algorithms for diagnosis of COVID 19 from Lung Ultrasonography

Abstract

Since the end of 2019 and early 2020s, the world has come across a life-altering virus that little did anyone knows would change the lives of everyone on this planet. The virus is known to affect people of all age groups and create complications in the lungs of human beings. Coronavirus disease 2019 (COVID-19) has affected a rapidly growing patient population worldwide. To effectively manage the disease, physicians need tests or methods [1]. The symptoms and effects really escalate if the person carrying the virus has comorbidity. Even though the symptoms are similar to that of other lung diseases such as pneumonia, the effects are far more deadly. Though there has been a significant amount of research done on the various methodologies to test and identify the COVID-19 virus and symptoms since the inception of the virus, the aim is to compare statistically and via various other scientific technologies in order to provide a solid conclusion. A study analyzing US mortality in March-July 2020 reported a 20% increase in excess deaths, only partly explained by COVID-19. Surges in excess deaths varied in timing and duration across states and were accompanied by increased mortality from non–COVID-19 causes [2].

Identification of COVID-19 in patients from chest CT scan [3] has been the most prevalent approach but it exposes the patient to X-ray radiations and is not a suitable approach for frequent monitoring. Computer analysis of ultrasound pulmonary images is a relatively recent approach that showed a large potential to diagnose pulmonary states that are a profitable and safer alternative to CT scans. Deep learning techniques for computerized analysis of lung ultrasound images offer promising opportunities for screening and diagnosing COVID-19 (Figure 1). It cannot replace the role of radiologists but can provide them with an automated computerized opinion on the condition, highlighting the specific region of interest in the images. In this paper, the developments made in the classification of lung ultrasound images for COVID-19 identification are critically analyzed and reviewed. Lung ultrasound provides data either in the form of images or videos from which relevant frames are extracted. The general process of classifying an image involves extracting features from the image pixel matrix and using them to train a model which can later be used to classify images. Such an approach to image classification is an example of supervised machine learning. Major improvements in the classification accuracy have been done by extending this approach to neural networks forming the basis for deep learning and for image classification particularly, Convolutional Neural Network (CNN) has shown great success. Various researchers, universities, and organizations such as Google and Oxford University have developed state-of-the-art CNN frameworks, many of which have been used for lung ultrasound image classification such as VGG19, InceptionV3, Xception, and ResNet50. Specific purpose frameworks like POCOVID-net [4] were developed for point of care ultrasound (POCUS) devices. Performance of such sophisticated frameworks on COVID-19 infected lung ultrasound image dataset will be analyzed and compared using performance metrics such as confusion matrix, precision, recall/sensitivity, specificity, F1- score, and AUC.