The Ethics of Committees of People’s Hospital of Deyang City approved this retrospective study, and informed consent for this retrospective study was waived.
Patients with confirmed COVID-19 were reviewed coming from Deyang City from January 26, 2020, to June 15, 2020. The diagnostic criteria were based on the preliminary diagnosis and treatment protocol issued by the National health commission of the People’s Republic of China:
Suspected cases were considered if having any one of the epidemiological histories or meeting any two of the clinical manifestations: (1) travel history or residence history in Wuhan area or other areas with continuous local case transmission within 14 days before the onset of the symptoms; (2) history of having contact with other patients with fever or respiratory symptoms coming from Wuhan or other areas where local cases continued to spread within 14 days before the onset of illness; (3) aggregative onset or epidemiological association with COVID-19 patients; (4) fever; (5) having the imaging characteristics of COVID-19; (6) the total number of white blood cells is normal or decreased, or the lymphocyte count is decreased in the early stage of onset.
Cases confirmed with real-time fluorescent polymerase chain reaction (RT-PCR) of respiratory specimens or blood specimens proved detection of the novel coronavirus; virus gene sequencing of respiratory or blood specimen was highly homologous to the new coronavirus (11). Patients met the following conditions could be discharged: (1) body temperature returned to normal for more than 3 days; (2) significant improvement in respiratory symptoms; (3) lung imaging showed a marked improvement in acute exudative lesions of the lungs; (4) two consecutive sputum, nasopharyngeal swabs, or other respiratory specimens were negative for nucleic acid testing, and the sampling interval was more than 24 h .
All consecutive 18 patients from **** City were admitted to the People’s Hospital of **** City, the regional medical center of **** City, and isolated in the separate infectious disease hospital area for treatment. No patients were excluded.
Chest CT protocol
Non-contrast CT scans were performed using a single-source CT scanner (Emotion 16, VB41A, Siemens Healthcare, Forchheim, Germany), which was dedicated to a separated COVID-19 ward. The patients were scanned with a single inspiratory phase in the supine position. A specially tailored low-dose protocol was employed for COVID-19 cases for reducing the amount of patient exposure to ionizing radiation. The tube voltage was 110 kVp, while reference mAs was 70 mAs with automatic tube current modulation. A matrix size of 512 × 512 (1.5-mm slice thickness and 1.5-mm increment) was performed to reconstruct CT images. B90f kernel and B31f kernel were used for lung kernel and mediastinal kernel, respectively. The iterative reconstruction was not available in this version of CT system. Window level of - 600 600 Hounsfield unit (HU) and window width of 1600 HU were set for lung window setting, while window level of 40 HU and window width of 350 HU for mediastinum window setting. The mean dose-length product was 115.6 ± 22.9 mGy2cm.
CT images review
Anonymous data were transmitted to the Dr. Wise system (version v220.127.116.11, Beijing Deepwise & League of PhD Technology Co. Ltd, China; the system is currently under the China Food and Drug Administration review process.) for automatic detection and segmentation. The Dr. Wise system performed lung segmentation, lesion segmentation and detection, and volume calculation using a dedicated multi-task deep learning algorithm developed for pulmonary pneumonia based on both regular-dose protocol and low-dose protocol. All those algorithms are based on supervised training of deep neural networks, i.e., based on training neural networks on large-scale annotated data (over ten thousand cases), which include a wide range of varieties in the sense of manufactures, acquisition parameters, reconstruction kernel, and slice thickness. Thus, the trained model is robust to all these changes. A variant of the convolutional MVP-Net  is exploited to achieve automatic detection of the pneumonia-related symptom regions, like GGO and CO. The categorization of GGO and CO is not based on thresholding Hounsfield units. It is based on supervised training with deep neural networks. Experienced radiologists firstly label GGO and CO areas, and then the deep neural networks are trained with acquired labels. Since radiologists tend to inspect multiple windows to obtain an accurate diagnosis, the MVP-Net takes advantage of such domain knowledge and employs a multi-view feature pyramid network to extract features from images rendered with varying window widths and window levels. Afterward, 3D U-Net  with pseudo-3D convolution was introduced to segment voxels that represented the abnormality in the detected regions. Thus, we could acquire the delineation of the pneumonia-related symptom regions.
For lung lesions identified on Dr. Wise, the reviewers recorded volumes of ground-glass opacity (GGO), consolidative opacity (CO), and the total abnormalities volume, respectively. The abnormality volumes as a percentage of total lung volume also were recorded for a series of CT examinations from admission to discharge. Comparisons of the percentage of the volume of GGO and CO on the first CT scan, the last CT scan, and the CT scan at the peak stage of the disease course were performed.
The CT images were transmitted to the institutional digital database system (Infinitt, Shanghai, China). Two cardiothoracic radiologists (C.M with 15 years of experience and HBZ) subjectively evaluated images on a 30-in. 8M color LCD. If the two radiologists failed to reach a consensus, a third cardiothoracic radiologist (B.M) with 20 years of experience determined the final decision. Based on the abnormalities of lobes related to COVID-19 (e.g., GGO, CO, reticulation, nodules, interlobular septal thickening, and fibrosis), reviewers assessed lung involvement in lung window images using a semi-quantitative scoring system . Visually scoring each of the five lobes was performed as 0 for none; 1 for 1–25% involvement; 2 for 26–49% involvement; 3 for 50–75% involvement; and 4 for 76–100% involvement. The scores for each lobe were summed to reach a total severity score ranging from 0 to 20. All follow-up CT images were evaluated for all patients to observe the lung change over time until recovery. The spent time was recorded from the start of scoring to completion for every case.
Depending on the normality of distribution, the Student t test, ANOVA, Wilcoxon test, or Mann–Whitney U test was performed for continuous variables assessed by the Shapiro–Wilk test. Continuous data were presented as mean ± standard deviation (minimum-maximum) or median (minimum-maximum). Pearson correlation was used to evaluate correlations between scores and the abnormalities volume percentage. SPSS 22.0 (version 22.0, IBM, Armonk, NY, USA) and GraphPad Prism 7 (GraphPad Software Company, San Diego, CA, USA) were used in statistical analyses. A two-sided P value of less than 0.05 was considered a significant difference.