Patients
This was a retrospective study that collected data from six centers (TongDe Hospital of ZheJiang Province, The Second People's Hospital of Neijiang, The First Affliated Hospital of Bengbu Medical College, Wenzhou People's Hospital, Anqing Municipal Hospital and The First Affiliated Hospital of Xi'an Jiaotong University). This study was approved by the Ethics Committee of the above hospitals, and written informed consents were wavied because the anonymized study did not alter any diagnosis and treatment of the patients. 294 patients were enrolled from 24, February, 2020 to 1, June, 2020. We included patients who satisfied the following criteria: (a) positive for nextgeneration sequencing or realtime RTPCR of SARS-CoV-2 in throat or nose swabs; (b) complete clinical data; (c) patients underwent CT scans. The exclusion criteria were (a) poor images with heavy breathing artifacts or metal artifacts; (b) patients had history of pulmonary surgery. In this case, the patient has series of CT examinations, the most severe sets of images were included in our study.
Clinical information
The basic data including gender, age, incubation period, comorbidity (hypertension, diabetes mellitus, cardiovascular disease, cerebrovascular disease,COPD, hepatitis B infection, malignancy, chronic kidney disease and immunodeficiency, etc.), severity status (severe or non-severe), laboratory examinations including C reactive protein (CRP), white blood cell count (WBC) and lymphocyte percentage were extracted from medical computerized database for all patients. The severity of COVID-19 includes four types: mild, common, severe, and critical according to the guideline of 2019-nCoV (trial version 7) issued by the China National Health Commission [15]. In this study, we divided all the patients into severe group (including severe and critical) with 38 cases and non-severe group (mild and common) with 256 cases. Of all the patients, there were 52 patients with comorbidity, the remaining 242 patients without comorbidity. Among these patients with comorbidity, 34 (65.4%) patients were with hypertension, 15 (28.9%) with diabetes mellitus, 8 (15.4%) with hepatitis B infection, 4 (7.7%) with cardiovascular disease, 3 (5.8%) with COPD and 2 (3.8%) with cerebrovascular disease. Patient with malignancy, chronic kidney disease and immunodeficiency was 1(1.9%).
CT image acquisition
The non-contrast chest CT were performed using three multi-detector CT scanners with 64 or 128 channels (Somatom Definition AS+, Siemens Healthineers, Forchheim, Germany; GE Medical Systems, China Branch, Beijing, China or Philips Ingenuity Core128, Philips Medical Systems, Best, the Netherlands). The scanning range was from apex to the base of lungs. The detailed parameters for CT acquisition were as follows: tube voltage, 120 kVp; tube current, standard (reference mAs, 60–120) to low-dose (reference mAs, 30) with automatic exposure control; slice thickness, 1.0 or 1.25 mm; reconstruction interval, 1.0-3.0 mm; noise index (NI), 25 and matrix 512×512. A lung window was with a width of 2000 HU and a level of -600HU, and a mediastinal window with a width of 350 HU and a level of 40HU.
CT image segmentation and quantitative analysis
The deep learning method used in our study is Unet neural network [16] which has been reported to have a good performance on the segmentation of the biomedical images. Here, we used a multi-task Unet with a single encoder and two parallel decoders to learn to predict and segment the region of lung and lesions. The decoder containing attention block was used to learn to segment the lesions while the decoder containing stacked dilated convolutions was used to learn the lung segmentation, which provided a more efficient feature encoding and a regularizing effect. This neural network were implemented in Dr. pecker cloud platform (http://www.jianpeicn.com/category/yuepianjiqiren) and our segmentation results were acquired from the platform. Our platform is open and free to all public research institutions in the world.
In order to make the neural network to learn to predict lesion and lung regions, labelled lesion samples lung samples were required. The lesion and lung regions were manually segmented using ITK-SNAP software (version 2.2.0; http://www.itksnap.org) in lung window with a width of 2000 HU and a level of -600 HU. Our segmentation system was pre-trained by 650 annotated CT images (550 in primary dataset and 100 in test dataset) with COVID-19 or community acquired pneumonia. This neural network extracted CT image features, segmented lung and lesions, and classified whether the lesion was consolidation or GGO. The volumes of the lesions as the results in underlying disease group and non-underlying disease group were outputted finally. The general flow of this study was shown in Figure 1.
The specific image segmentation steps were as follows. First, a threshold value of -450 HU was used to distinguish GGO and consolidation (Figure 2&3. B). The margin of the lesion in each axial slice was delineated (Figure 2&3. C-D). Then, a 3D region of interest (ROI) was obtained based on delineated results including lung erosion diagram (Figure 2&3. E-F) and lesion diagram (Figure 2&3. G-H). The SimpleITK software tool (http://www.simpleitk.org) was used to quantify the mean HU of lung and lesions, volumes and numbers of lesions automatically.
Dice coefficient (between 0 and 1) in test dataset was used as an index to evaluate the quantitative segmentation effect of this model. Higher Dice coefficient presents better model.
To assess the accuracy of segmentation on our cohort, two cardiothoracic radiologists (C.Z. and J.W. with 5 and 12 years of experience, respectively) manually marked the infected area of 50 CT scans of 50 patients (15 with comorbidity and 35 without comorbidity) and compared the results with those of segmentation model.
In total, the following 8 quantitative parameters were acquired for further analysis: (a) the volume of the whole lung (LUV); (b) the volume of lesion (LEV); (c) the ratio of volume of lesion to whole lung (LEV/LUV); (d) the volume of consolidation (COV); (e) the ratio of volume of consolidation to whole lung (COV/LUV); (f) the volume of GGO (GGOV); (g) the ratio of volume of GGO to whole lung (GGOV/LUV) and (h) the number of lesions (LEN).
Statistical Analysis
Continuous variables were presented as median (IQR) and categorical variables were shown as n (%). Mann-Whitney U test, χ2 test or Fisher’s exact test were used to compare differences between patients with and without comorbidity. Statistical analyses were performed with SPSS (ver. 22.0; SPSS Inc., Chicago, IL, USA). Two-sided P < 0.05 was considered statistically significant.