Deep Learning Audio Spectrograms Processing to the Early COVID-19 Detection

Abstract

The objective of the paper is to provide a model capable of serving as a basis for retraining a convolutional neural network that can be used to detect COVID-19 cases through spectrograms of coughing, sneezing and other respiratory sounds from infected people. To address this challenge, the methodology was focused on Deep Learning technics worked with a dataset of sounds of sick and non-sick people, and using ImageNet's Xception architecture to train the model to be presented through Fine-Tuning. The results obtained were a precision of 0.75 to 0.80, this being drastically affected by the quality of the dataset at our availability, however, when getting relatively high results for the conditions of the data used, we can conclude that the model can present much better results if it is working with a dataset specifically of respiratory sounds of COVID-19 cases with high quality.

Artificial Intelligence for Detecting COVID-19 With the Aid of Human Cough, Breathing and Speech Signals: Scoping Review

“…The type of data collected and stored as datasets were as follows: cough recordings (eg, recording for 3 s) [14] , [15] , [17] , [23] , [25] – [27] cough, breathing and speech recording (reciting a sentence) [16] , [19] , [22] , [27] – [30] , speech collection [11] , [20] , [21] , [24] and sound samples (eg, recording using certain sounds) [18] . Moreover, some studies also collected a questionnaire based on a medical history or COVID-19 symptoms (eg, underlying conditions, age, sex, temperature) [15] , [17] .…”

“…10 of the studies used features like MFCCs [11] , [13] – [18] , [25] – [30] while 19 studies included CNN [11] , [13] – [16] , [18] , [21] – [30] to extract features from the datasets. The data size was between 1000 and 9999 in half of the studies (n=11) [14] , [16] , [19] , [22] , [25] – [27] , [28] – [30] whereas seven of the studies had the sample size greater than 10000 [15] , [18] , [20] , [21] , [28] . The Supplementary Material includes the datasets of all the examined studies.…”

Review on the Evaluation and Development of Artificial Intelligence for COVID-19 Containment

“…The COVID-19 detection method was created by Rodriguez et al [ 115 ] using spectrograms of coughing, sneezing, and other respiratory sounds. The dataset contains two labels, sick and not sick, gathered by the pharmaceutical manufacturer Pfizer in the United States.…”

AUCO ResNet: an end-to-end network for Covid-19 pre-screening from cough and breath

“…Typical Covid-19 symptoms such has wet and dry coughs, croup, pertussis and bronchitis coughs have been considered when acquiring undetermined coughs. A similar approach has been adopted on the Pfizer dataset of “Sick Sound” [30] , in this case audio signals were converted into images of Spectrograms by using a Short Time Fourier Transform and the resulting images were fed into a Xception deep neural network [31] . A final accuracy of 75% was achieved with no inter-patient separation scheme: the result is consistent with those already found in literature (e.g.…”