(Retracted) Review of artificial intelligence-assisted COVID-19 detection solutions using radiological images

Abstract

A new virus called coronavirus has been affecting the world since 2019 and has killed millions of people. Even though vaccines for the virus have been developed and the mortality rate is decreasing across the world, many countries are still struggling with the pandemic. Artificial intelligence (AI) methods have been regarded as fast techniques and powerful tools for screening this disease. We reviewed papers that used AI-based systems for the diagnosis of COVID-19 using radiological images, such as X-rays and computed tomography (CT) images. This survey focuses specifically on deep learning (DL)-based systems for screening COVID-19 patients. Privacy and accuracy of diagnosis are of paramount importance in a clinical environment. In most surveys, the privacy issue is not taken into consideration. In this regard, we categorize recent work into three taxonomies: federated learning (FL) models (privacy-guarding methods), ensemble machine learning (ML) models, and other ML and DL models. A summary of the selected articles is presented; parameters such as the modality, experimental tools, data sources, number of classes, and positive and nonpositive aspects of each model and work, as well as evaluation measures, are depicted. In fact, we compare papers and their experimental results to find more accurate and privacy-guarding methods. Also, the type of data and tools that are helpful for more accurate prediction were investigated. Finally, we refer to some limitations of ML methods and provide useful insight for future researchers. In this survey, 45% and 41% of papers used X-ray and CT images for experiments, respectively. Using multiple datasets was the preference of 61% of researchers, and 45% of papers considered binary classification. The average accuracy of 95.71%, 97.09%, and 93.38% was obtained for federated ML, ensemble ML, and other ML models, respectively. To sum up, X-ray images were the favorite of most articles. Also, most researchers employed multiple databases for their experiments, and binary classification was the method of choice for most of them. Among the three categories, ensemble learning-based systems demonstrated the best performance in terms of all evaluation metrics. Therefore, these systems can be used to screen COVID-19 patients.