Distinguishing novel coronavirus influenza A virus pneumonia with CT radiomics and clinical features

Objectives To differentiate novel coronavirus pneumonia (NCP) with influenza A virus (IAV) pulmonary infection based on computed tomography (CT) radiomics features combined with clinical feature. Methods A total of 292 patients were enrolled, as NCP determined with reverse-transcription polymerase chain reaction (RT-PCR) and abnormal chest CT findings and IAV pulmonary infection confirmed by nucleic acid test with pneumonia lesion in the chest CT, retrospectively. The dataset was randomly divided into 233 cases in the training set and 59 cases in the validation set according to the ratio of 8:2, and there were 107 cases collected for verification as external test set. Firstly, voxel-based gray-level discretization (binWidth = 25) and Z-Score normalization were applied to preprocess the patient's ROI and normalize the extracted features. Then, the most predictive radiomic features were selected and their corresponding coefficients were evaluated using the correlation coefficient algorithm and the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. Subsequently, univariate logistic regression was employed to screen for clinically discriminative features from the patient's clinical characteristics. Finally, constructing the radiomics model and clinical model using support vector machines and logistic regression methods, respectively. And then combining these features of the two to construct a combined model. Receiver operating characteristic curve (ROC), calibration curve, and decision curve were performed to evaluate the classification of the radiomics model, clinical model and combined model. Area under ROC curve (AUC) were calculated to evaluate the diagnostic efficiency, and Delong’s test was used to compare the AUC between different models. Results Age, white blood cells, neutrophils, lymphocytes, and basic diseases reached statistical significance in the training set. After LASSO, 16 optimal radiomics features were retained. In the validation set and external test set, the SVM radiomics model achieved AUCs of 0.818 and 0.808 for automatic classification of NCP and IAV pulmonary infection,; and the clinical classification model shad AUCs were 0.676 and 0.669; finally, the 5 clinical features and the 16 selected radiomics features were used to construct the combined model with the AUCs of 0.821 and 0.820. After incorporating clinical features, the clinical model’s discriminatory and predictive efficacy further improved in testing sets (AUC,