COVID-19 Detection Systems Based on Speech and Image Data Using Deep Learning Algorithms

COVID-19 is a worldwide epidemic that seriously affected the lives of people. Since its inception, physicians have tried their best to trace the virus and reduce its spread. Several diagnostic approaches have been reported to detect the coronavirus in research, clinical, and public health laboratories. Although the existing systems aid medical experts in the diagnosis, they still lack precise detection and may fail to detect COVID-19 in a timely manner. Therefore, in this study, we recommend two approaches i.e., the first approach is based on the VGGish network that focuses on vocal signals, such as breathing and coughing, and the second approach is based on ResNet50, which takes chest X-rays as input. With the help of VGGish, the patient's cough, voice, and respiration audios have been classified as patient and non-patient achieving an accuracy of more than 98%. We also assessed the performance of several methods for X-ray classification, such as ResNet50, VGG16, VGG19, Densnet201, Inceptionv3, Darknet, GoogleNet, squeezeNet, and Alex-Net. TheResNet50 outpaced all supplementary CNN models with a precision of 94%. However, when we took both types of inputs simultaneously, the accuracy for detection was increased to 99.7%. After extensive experimentation, we believe that our proposed hybrid method is robust enough to take X-rays and audio as mel-spectrograms and identify COVID-19 at early stages, attaining an accuracy of 99.7%.