Uncertainty-Aware Semi-supervised Method using Large Unlabelled and Limited Labeled COVID-19 Data

Alizadehsani, Roohallah; Sharifrazi, Danial; Izadi, Navid Hoseini; Joloudari, Javad Hassannataj; Shoeibi, Afshin; Gorriz, Juan M.; Hussain, Sadiq; Arco, Juan E.; Sani, Zahra Alizadeh; Khozeimeh, Fahime; Khosravi, Abbas; Nahavandi, Saeid; Islam, Sheikh Mohammed Shariful; Acharya, U Rajendra

The new coronavirus has caused more than 1 million deaths and continues to spread rapidly. This virus targets the lungs, causing respiratory distress which can be mild or severe. The X-ray or computed tomography (CT) images of lungs can reveal whether the patient is infected with COVID-19 or not. Many researchers are trying to improve COVID-19 detection using artificial intelligence. In this paper, relying on Generative Adversarial Networks (GAN), we propose a Semi-supervised Classification using Limited Labelled Data (SCLLD) for automated COVID-19 detection. Our motivation is to develop learning method which can cope with scenarios that preparing labelled data is time consuming or expensive. We further improved the detection accuracy of the proposed method by applying Sobel edge detection. The GAN discriminator output is a probability value which is used for classification in this work. The proposed system is trained using 10,000 CT scans collected from Omid hospital. Also, we validate our system using the public dataset. The proposed method is compared with other state of the art supervised methods such as Gaussian processes. To the best of our knowledge, this is the first time a COVID-19 semi-supervised detection method is presented. Our method is capable of learning from a mixture of limited labelled and unlabelled data where supervised learners fail due to lack of sufficient amount of labelled data. Our semi-supervised training method significantly outperforms the supervised training of Convolutional Neural Network (CNN) in case labelled training data is scarce. Our method has achieved an accuracy of 99.60%, sensitivity of 99.39%, and specificity of 99.80% where CNN (trained supervised) has achieved an accuracy of 69.87%, sensitivity of 94%, and specificity of 46.40%.