COVID-19 Diagnosis from cough samples using Deep Learning methods

Abstract

Coronavirus disease of 2019 (COVID-19) has affected the lives of millions of people around the globe. Up until July 2022, there were 569.771.691 active cases of COVID-19 globally, and there had been 6.383.776 deaths. The virus is mainly transmitted through droplets generated when an infected person coughs, sneezes, or exhales. The most common occurring symptoms are fever, cough, and fatigue. The current diagnosis method is performed through Reverse-Transcription Polymer Chain Reaction (RT-PCR) testing. However, scarcity, cost, long turnaround time of clinical testing and the fact that they can lead to another infection if done improperly are some downsides of the RT-PCR testing. Furthermore, the in-person testing methods put the medical staff, particularly those with limited protection, at serious risk of infection. Vaccination remains a key component of the approach needed to reduce the impact of SARS-CoV-2. Unfortunately, variants of Covid-19 reduce at some point the effectiveness of vaccines, subsequently leading to reinfections. Therefore, the need for constant testing remains as immunity is often threatened by mutations. The current thesis aims at demonstrating the feasibility of the automatic detection of COVID-19 from cough sounds. This type of screening is non-contact, easy to apply, and can reduce the workload in testing centres as well as limit transmission. Two datasets have been used in this thesis, containing coughs from people from all continents, namely the “Coswara” and the “Cambridge” dataset. Dataset skew was addressed by applying an ensemble learning approach so that the Covid class is not underrepresented. The preprocessing step involves cough detection to identify if and when the cough is present in the raw audio recordings. The datasets are crowdsourced which means that the collected sounds are from differing environments and the quality of the microphone is disputed. As a result, the models could be highly prone to overfitting to unwanted signals. To address this issue, the sound files classified by the cough detector as cough are denoised. Data augmentation is applied to address data scarcity, since Deep Learning Architectures are data hungry. Then, audio samples are converted to mel spectrograms. For the Covid-19 classification task, nine different deep learning architectures are tested and presented in this thesis. Specifically, CNNs combined with bidirectional Long Short-Term Memory (BiLSTM) and bidirectional Gated Recurrent Units (BiGRU) networks in conjunction with an attention mechanism, are implemented. Three pretrained networks on ImageNet and an ensemble model consisting of them are presented as well. VGG-13 and DenseNet Speech, an architecture used to a prior study for voice recognition and keyword spotting, are also implemented. Temporal CRNNs seem to produce promising and consistent results in Covid-19 detection. Multistage Transfer Learning process consists of three stages of transfer learning and uses all of the available datasets. This pretraining on cough related tasks leads to higher classification results for the Cambridge dataset. Eventually, an interpretability attempt of InceptionResnetV2 has been made on mel spectrograms using Local Interpretable Model-agnostic Explanations. The best classification results, obtained through 5-fold cross validation and TCRNNs, have reached an accuracy of 76,67% and an AUC of 76,16%. These results demonstrate that cough can potentially serve as a helpful triage or diagnostic tool for Covid-19 infection. Since this type of cough audio classification is cost-effective and easy to deploy, it is potentially a useful and viable means of non-contact COVID-19 screening.