Patients
This retrospective study was approved by our institutional review board and patient consent was waived. This study retrospectively collected patients with viral pneumonia diagnosed in 15 hospitals in this province from March 1, 2015 to March 15, 2020. RT-PCR assays were performed to identify influenza A virus, influenza B virus, respiratory syncytial virus, parainfluenza virus, adenovirus, SARS coronavirus, SARS-CoV-2, Epstein-Barr virus, measles virus, and other viruses from nasopharyngeal swabs or bronchoalveolar lavage fluid. The study only included pneumonia patients infected with single virus, and patients with multiple respiratory viruses or bacterial or fungal infections were excluded. A total of 375 viral pneumonia patients were diagnosed in this study. The further selection process for viral pneumonia patients is shown in Figure 1. Patients completed the chest CT examination within 48 hours after admission. According to the virus type found in the lungs, the patients were divided into two groups: COVID-19 and Non-COVID-19. The number of cases included in each hospital is summarized in Supplementary Table 1.
CT examination
HRCT examination: CT scanners with 16 or more detector rows (Siemens, Germany; Philips, the Netherlands; and GE, USA) were used. The patient was scanned in the supine position while holding his or her breath after inspiration. The scanning range was from the thoracic inlet to the costophrenic angles. Scanning parameters: detector collimation width 64×0.6 mm or 128×0.6 mm, tube voltage 120 kV, adaptive tube current, high-resolution algorithm reconstruction, reconstruction layer thickness 1 or 1.5 mm, and layer spacing 1.5 mm.
Chest CT signs analysis
Three Chinese radiologists were blinded to the RT-PCR results, all patient information, and type of viral pneumonia. First, two experienced radiologists in the cardiothoracic group independently read the radiographs. When their opinions were inconsistent, they discussed them and reached a consensus, which was reviewed and confirmed by the third senior radiologist in the cardiothoracic group. The signs of the first CT examination after admission were analyzed. The CT imaging evaluation included lesion location (left upper lobe, left lower lobe, right upper lobe, right middle lobe and right lower lobe) and signs [GGO (ground-glass opacities), partial consolidation, consolidation (multifocal consolidation, focal consolidation), fibrous stripes, septal thickening, intralobular interstitial thickening, subpleural lines, crazy-paving pattern, tree-in-bud, bronchial wall thickening, bronchiectasis, air bronchogram, halo sign, reversed halo sign, mediastinal lymphadenectasis, pleural thickening, and pleural effusion] [9-12,20]. The window width and level were set to 1600/-600 HU.
CT image processing and volume of interest (VOI) segmentation
The Lung Kit software (GE Healthcare, Version LK2.2) was used for pneumonia lesion segmentation. All the CT images were firstly resampled into isotropic 1 mm ×1 mm ×1 mm voxel size. The five anatomic lung lobes were firstly automatically segmented. Then pneumonia lesion volume of interest (VOI) was automatically segmented and the margin of the VOI was manually comfirmed by experienced thoracic radiologist. The distributed lesions were considered as a whole VOI in the next analysis steps.
Radiomics feature extraction and selection
A total of 1316 radiomics features were extracted from segmented VOIs by using open source of Python package Pyradiomics[21]. The extracted radiomics features were categorized into five groups: (1) First-order features including 18 intensity statistics and 14 shape features; (2) 75 multi-dimensional texture features including 24 Gray Level Co‐occurrence Matrix (GLCM), 16 Gray Level Size Zone Matrix (GLSZM), 16 Gray Level Run Length Matrix (GLRLM), 14 Gray Level Dependence Matrix (GLDM) and 5 Neighboring Gray Tone Difference Matrix (NGTDM) Features; (3)1209 Transformed first-order and textural features including: 744 wavelet features in frequency channels LHL, LLH, HHH, HLH, HLL,HHL, LHH and LLL; 186 LoG filtered features with sigma of 2.0 and 3.0; 279 local binary pattern (LBP) filtered texture features.
The radiomics feature data was firstly preprocessed by replacing missing values with median values, and z-score normalization was followed. The whole dataset was randomly divided into training and test cohort at the ratio of 7:3. And the radiomics features in the training set was further screened for classification model construction. Firstly, the redundant collinear features were reduced by correlation analysis at a cut-value of 0.7. Then the features without statistical differences between COVID-19 and Non-COVID-19 groups were excluded by Mann-Whitney U test. The significant level was p<0.05. The univariate logistic analysis was used to select the potential classification indicators with P value less than 0.05. Next, the least absolute shrinkage and selection operator (LASSO) logistic regression method with 10-fold cross validation was applied for further feature selection and regularization to improve the model accuracy and avoid overfitting. The minimum mean square error for model fitting among the 10 folds was utilized to determine the optimized lambda values. The remaining features with non-zero coefficients at such lambda values were kept for model construction.
Classification model construction
The logistic regression model was constructed using the selected radiomics features to differentiate COVID-19 from Non-COVID-19 and the Radscore for each patient was calculated based on the regression coefficients. In addition, the independent predictors among CT signs were also selected by using Chi-square test (or Fisher exact test), univariate and multivariate logistic regression methods. These selected CT signs were further combined with radiomics features to construct combined model using logistic regression method. The nomogram of such combined model was also established.
The radiomics model and combined model constructed based on the training set were validated in the test cohort. The classification performances were evaluated by receiver operating characteristic (ROC) curve. The area under the curve (AUC), accuracy, sensitivity and specificity were derived. In addition, the calibration curves and decision curve analysis (DCA) curves were calculated to assess the models’ classification performance and their clinical benefits.
Statistical analysis
The continuous variables or ordinal variables were compared by t-test or Mann-Whitney U test. The distribution of different CT signs was compared by Chi-squared test or Fisher exact test when small sample sizes existed. For ROC analysis, the cut-off value in the training set at the maximum of Youden index of each model was calculated and the confusion matrix and sensitivity, specificity, accuracy in the training and test cohorts were derived at such cut-off value. The Delong test was used for comparison of ROC curves between different models. The reported statistical significance levels were all two-sided with the statistical significance set as p< 0.05. The statistical analyses were performed with SPSS Software (Version 25, IBM, Chicago, IL) and R software (Version: 3.6.1, https: www.r-project.org). The following R packages were mainly involved including: “glmnet” for logistic regression including LASSO regression; “pROC” for ROC analysis; “rmda” for DCA analysis.