Analysis of COVID-19 Using Imaging and Audio Modalities

The outbreak of coronavirus (COVID-19) resulted in numerous deaths and several significant negative impacts on many levels of human life including disruptions of schools, universities, vocational education segments, global economic recession, and increasing of poverty level [1]. Several COVID-19 diagnosis mechanisms currently appear in the scene such as Polymerase Chain Reaction (PCR) tests. The rate of false negatives for a PCR test varies depending on how long the infection has been present in the patient. Studies shows that the false-negative rate was 20% when testing was performed five days after symptoms began, but much higher (up to 100%) in earlier infection stages. Although the PCR test could be considered relatively accurate, it is also quite costly, ranging from 125 to 250 USD. Moreover, it takes the test results a long time to get out. Given the sensitivity of the situation, this delay in results would be quite risky. The aim of this research is to contribute to the discovery and analysis of COVID-19 invariants in order to assist medical diagnosis of the disease and to utilize deep learning for social good by implementing an aiding screening tool for COVID-19 testing that is accurate, cheap, and fast. Cheaper testing options such as X-ray and Computerized Tomography (CT) lung scans and cough audio records have been targeted for examination of promising results. Two Convolutional Neural Network (CNN) models were developed. One has been pre-trained with 38,000 CT and X-ray lung scans dataset to identify if the CT or X-ray lung scan is of a healthy person, COVID-19, or pneumonia patient. This model achieved an accuracy of 95.9%. Transfer learning has been applied to this model to test its generalizability beyond the given training datasets. For the second CNN model, about 2000 cough audio records have been converted into Mel-spectrograms and used to pre-train the model to identify if the cough audio Mel-spectrogram results are from a COVID-19 patient or not. This model achieved an accuracy of 82.1%.