Emek Güldoğan1, *, İsmail Okan Yıldırım2, Serkan Sevgi2, Cemil Çolak1
1Department of Biostatistics and Medical Informatics, Faculty of Medicine, İnönü University, Malatya, Turkey - 2Department of Radiology, Faculty of Medicine, İnönü University, Malatya, Turkey
Background: Radiological techniques integrated with artificial intelligence (AI) are a promising diagnostic tool for the rapidly increasing number of COVID-19 cases today. In this study, we intended to construct an artificial intelligence-assisted prediction of COVID-19 status based on thorax computed tomography (CT) scans using a proposed meta-learning strategy.
Methods: A public dataset including 1252 positive and 1230 negative thorax CT scans of SARS-CoV-2 was used in the current study. The CT images for COVID-19 status were analyzed by 26 transfer learning (TL) models. The stacking ensemble learning was used to obtain more consistent and high-performance prediction results by combining the prediction results of 26 TL models with an embedded XGBoost algorithm.
Results: Mobile had the best prediction with an accuracy of 0.946 (95% CI: 0.93-0.962) among the TL models. The Meta-learning model yielded the best classification accuracy of 0.993 (0.98-1), which outperformed MobileNet, the most successful architecture among TL architectures.
Conclusions: The proposed meta-model that can distinguish CT images between COVID-19 positive and abnormal/normal conditions due to other etiology of COVID-19 negative may be beneficial in such pandemics. The AI application in this study can be used in mobile, desktop, and web-based platforms to have facilitating and complementary effects on classical reporting and the current workload in radiology departments.
Computed tomography, COVID-19, transfer learning models, meta-learning, XGBoost.