Development and Evaluation of an AI System for COVID-19

Cheng Jin; Weixiang Chen; Yukun Cao; Zhanwei Xu; Xin Zhang; Lei Deng; Chuansheng Zheng; Jie Zhou; Heshui Shi; Jianjiang Feng

Features were extracted in the attentional region determined by Guided Grad-CAM. We also extracted the same feature in normal lung in controlled cases for comparison. Due to no valid lesions attentional region for controlled cases is computed by Guided Grad-CAM, we used the shape of attentional region of COVID-19 cases and randomly choose positions within lung area as the attentional regions of controlled cases. We did not use shape features because the shape of attentional regions between COVID-19 and controlled cases are the same. We extracted radiomics features which are widely used in lesion diagnosis these years. These features are composed of different image transforms and feature matrix calculations. We adopted three image transforms: original image, transformed image by Laplacian of Gaussian (LoG) operator, and transformed image by wavelet. For each image after the operation of a transform, six series of features are extracted, including first order features, Gray Level Co-occurrence Matrix (GLCM), Gray Level Size Zone Matrix (GLSZM), Gray Level Run Length Matrix (GLRLM), Neighboring Gray Tone Difference Matrix (NGTDM), Gray Level Dependence Matrix (GLDM). Radiomics analysis was performed using python version 3.6 and the "pyradiomics" package [26] . We designed three other features which are distance feature and fractal features of 2D contour and 3D gray level mesh of attentional region. The distance feature was defined as the distance between the center of gravity of the region of interest (obtained by the classification network after Grad-GAM) and the edge of the lung (obtained by the edge of the lung automatically segmentation results). Besides, 2D contour fractal dimension and 3D grayscale mesh fractal dimension of the attentional region was extracted. The fractal dimension describes the de- gree of curvature of the curve and surface. These three extra features were only extracted from the CT images of COVID-19 patients and were not analyzed and compared on the controlled cases. LASSO logistic regression model, heat map of cluster and correlation coefficient matrix were used to extract, select and verify the radiological features of the attentional region in the original CT images, which can interpret AI system. LASSO analysis was performed using python version 3.6 and the "scikit-learn" package.