Human-in-the-loop enhanced COVID-19 detection in Transfer Learning based CNN Models

Abstract

Recently, many countries have concentrated on protecting human health and struggle with COVID-19 outbreak. Now that the virus is showing up in many places worldwide and the numbers of infected people are increasing rapidly, the governments of some countries are going to have to act quickly. The diagnostic tools are expensive, and the diagnostic kits can be challenging to obtain in many countries which makes it more difficult to stop the epidemic. Therefore, we need a fast and straightforward diagnostic tool based on AI technology that will save time, money and resources in the prevention and control of the virus. To achieve this problem, researchers are developing X-Ray scan images based COVID-19 detection models which are based on convolutional neural networks and transfer learning algorithms. These models are trained on visual pictures in order to predict the presence of the virus in the sample. On the other hand, predictive modelling can make some false predictions like false-negative which means that the sample may not be infected with COVID-19, but it will show positive COVID-19. It is important to remark that the wrong diagnostic will lead to an incorrect approach to the treatment and prevention of the virus. For example, if a patient is infected with COVID-19, but the test is negative for the virus, the patient can lose their life. Such an incorrect diagnosis will be very costly for the government. For this reason, model performance and the reliability of the classification result should be considered together. Thus we recommend applying to a medical doctor for low quality estimates by using uncertainty.In this study, we propose uncertainty quantification enhanced transfer learning-based convolutional neural network models for prediction.Closeness and Uncertainty Aware Adversarial Examples Detection in Adversarial Machine Learning

Publication
Computational Intelligence for COVID-19 and Future Pandemics - Emerging Applications and Strategies
Date