The study was approved by the institutional review board of the First Affiliated Hospital of China Medical University. Informed consent was waived due to the nature of the retrospective study.
Between February 10, 2020, and March 10, 2020, the clinical data and CT images of COVID-19 patients at Fangcang Shelter Hospital in Hongshan Gymnasium, Wuhan, Hubei, were reviewed retrospectively. All cases were mild from the onset and during the course of hospitalization, as defined by no hypoxemia or respiratory distress (respiratory rate≥30 breaths/min, requirement for oxygen treatment or mechanical ventilation, or SpO2≤93% on room air) . Patients were included if they met the following criteria: (1) No abnormal clinical symptoms (fever and severe respiratory symptoms) for more than 3 days. (2) Underwent RT-PCR tests at least 3 times after abnormal clinical symptoms disappeared. (3) The first RT-PCR tests were performed between 3-5 days after abnormal clinical symptoms disappeared. (4) Underwent chest CT exams within 2 days after the first RT-PCR test. Patients with inconsistent results in the first two consecutive RT-PCR tests were excluded (Fig. 1A, B). Novel coronavirus 2019-nCoV nucleic acid detection kit (fluorescence PCR method) (Sansure Biological Technology Co., Ltd., Changsha, China, Serial Number: 20150036) was used for RT-PCR tests.
The enrolled patients were divided into two groups: RT-PCR-negative and RT-PCR-positive groups (Fig. 1A, B). Inclusion criteria for the RT-PCR-negative group were: (1) All RT-PCR tests were negative; (2) No worsening clinical symptoms during hospitalization and the 2-week isolation after discharge. Inclusion criteria for the RT-PCR-positive group: the first two RT-PCR tests were positive.
We collected 20 available clinical characteristics, including general characteristics (age, gender, time interval from symptoms onset to CT exams), comorbidities, vital signs on the CT scan day and laboratory tests on admission. Comorbidities included diabetes, hypertension, cardiovascular disease, chronic obstructive pulmonary disease, chronic liver disease and cancer. Vital signs on the CT scan day included heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, and blood oxygen saturation. Laboratory tests include white blood cell count, neutrophil count, lymphocyte count, platelet count, hemoglobin and neutrophil/lymphocyte ratio (NLR) (NLR= neutrophil counts/lymphocyte counts).
The first RT-PCR tests for all enrolled patients were performed between 3-5 days after abnormal clinical symptoms disappeared. Then, all patients underwent CT exams within 2 days after the first RT-PCR test. Chest CT scanning used a mobile cabin CT (CT-NeuVz Prime, Neusoft) with a single breath-hold in the supine position. The scan parameters are as follows: tube voltage of 120 kVp, tube current of 100-200 mA, detector collimation of 64 or 128× 0.625 mm, field of view of 350 mm×350 mm, and matrix size of 512 × 512. Imaging data were reconstructed using a medium sharp reconstruction algorithm with a slice thickness of 5 mm and an interval of 1 mm.
Image segmentation and feature extraction
CT image analysis was performed on a dedicated workstation - Lung intelligence Kit (LK) Version V2.1.1. R (GE Healthcare, China). The main processes included data import and preprocessing, lung lobe segmentation, lesion segmentation and feature extraction (Fig. 2). Lung lobes were segmented with the purpose of improving the accuracy of lesion segmentation and calculating the proportion of lesions in each lung lobe.
Lobe and lesion segmentation
Before lung lobe and lesion segmentation, the images were resampled to voxel size 1 × 1 × 1 mm3, and a Gaussian filter was applied for denoising. Then, a fully automatic segmentation of three-dimensional lung lobes and lesions based on deep learning algorithms was performed. In cases of unsatisfactory lung lobe and lesion segmentation, two thoracic radiologists (with 5 and 15 years of experience, respectively) blinded to the clinical information and RT-PCR results manually adjusted the contour and resolved discrepancies by consensus.
Quantitative feature extraction
After segmentation, 86 CT quantitative parameters were automatically calculated: the statistical results of lung lobe and lesion (volume, volume percentage, pneumonia score, average density, standard deviation of density) and the component analysis of the lesion (partial solidity, solidity and total lesions) (Supplementary Data 1).
Radiomic feature extraction
After segmentation, 120 radiomic features of 7 categories were automatically calculated: (1) first-order features (n=19); (2) 2D and 3D shape features (n=26); (3) gray level cooccurrence matrix features (n=24); (4) gray level run length matrix features (n=16); (5) gray level size zone matrix features (n=16); (6) neighboring gray tone difference matrix features (n=5); and (7) gray level dependence matrix features (n=14). Detailed names and definitions of all 120 features can be found in Supplementary Data 2.
Missing values were replaced by the median, and the data were standardized by the following formula: standardized value = (original value-average value)/standard deviation.
The patients were randomly assigned at a 7:3 ratio to either the training cohort or the testing cohort. All patients in the training cohort were used to build the predictive model, while patients in the testing cohort were used to independently evaluate the model’s performance. To obtain the strongest features that were significantly associated with negative RT-PCR results in the training cohort, we performed univariate logistic regression analysis, and features with a p-value <0.10 were used for subsequent analysis. Then, Spearman correlation analysis was used to remove the features highly correlated with others; here, the |r| value was 0.9.
Model Establishment and Evaluation
We constructed a multivariate logistic regression model to identify a strategy to best classify RT-PCR-negative patients in the training dataset. Radiomics scores (Rad-scores) were calculated in each patient through a linear combination of the extracted features with their respective coefficients. The predictive performance was evaluated in terms of discrimination-receiver operating characteristic (ROC) curve, calibration-calibration curve and clinical application-decision curve.
Categorical variables are presented as the number and percentage of the total. The normality of continuous variables was evaluated by using the Shapiro-Wilk test. Normally distributed variables are shown as the mean ± standard deviation or the median (25% percentile, 75% percentile). The differences in variables between different subgroups were assessed by the t test or Mann-Whitney U test as appropriate. The chi-squared test was used to compare the significance of the differences between categorical variables. All statistical analyses for the present study were performed with R 3.5.1 and Python 3.5.6. A two-tailed p-value <0.05 indicated statistical significance.