Selection of Robust Regional Computed Tomography (CT) Radiomics Features for COVID-19 for AI Based Classification

The lung CT images of COVID-19 patients can be characterized by three different regions - Ground Glass Opacity (GGO), consolidation and pleural effusion. Quantitative characterization of these regions using radiomics can facilitate accurate diagnosis, disease progression and response to treatment. However, according to the knowledge of the author, regional CT radiomics analysis of COVID-19 patients has not been carried out. This study aims to address these by extracting and selecting the radiomics features that can characterize each of the regions separately and can distinguish the regions from each other. Two approaches were implemented to select the features that can differentiate between lung regions - 1) one way ANOVA for finding statistical significance difference (p<0.05) between the regions and 2) Z-score and correlation heatmaps. Radiomics features that show agreement for all cases (statistical significance test, Z-score and correlation) were selected as suitable features. The features were then tested on separate 52 CT images containing GGO and consolidation. 10 radiomics features were found to be the most suitable among 44 features These features in turn can help for more accurate diagnosis, staging the severity of the disease and allow the clinician to plan for more successful personalized treatment for patients.