A machine learning method based on lesion segmentation for quantitative analysis of CT radiomics to detect COVID-19

Abstract

Coronavirus disease (COVID-19) since late December 2019 became an epidemic all over the world so that widely spread throughout. Computed tomography (CT) imaging can be effective in isolating the infected persons and controlling this epidemic. Radiomics is an image quantitative analysis procedure that can quantify imaging by extracting specific features from CT images. We aimed to develop a machine learning (ML) method based lesion segmentation for quantitative analysis of CT radiomics to detect COVID-19. The current study was carried out on two groups of patients including 98 patients with confirmed COVID-19 and 96 with suspected COVID-19. A total of 755 radiomics features were extracted, including 594 gray level co-occurrence matrix features (GLCM), 56 intensity direct features, 49 intensity histogram features, 33 gray level run length matrix features (GLRLM), 18 shape features, and 5 neighbor intensity difference features. Two feature selection procedures including Pearson Correlation (PC) and Recursive Feature Elimination (RFE) were used. As well as, we examined three classifiers including Random Forest (RF), Decision Tree (DT), and K-Nearest Neighbor (KNN). The performance of the feature selection and classification procedures was obtained using 6 criteria. We have obtained the best performance as the accuracy of 98%, recall of 99%, and the area under the curve (AUC) value of 100% for the feature selection procedure RFE and RF classifier. As a result, it can be concluded that the radiomics features of the lung lesions based on ML can be used to differentiate COVID-19 patients.