The research was approved by Medical Ethical Committee (Approved Number. 2020037). Our institutional review board waived written informed consents for this study, and got consent from patients.
Patients and data source
All the patients in our study had been diagnosed of COVID-19 according to the guideline of 2019-nCoV (Fifth Trial Edition) issued by the National Health Commission of China . A total of 127 patients (68 men and 59 women; mean age, 57.7 years; age range, 20-83 years) with confirmed SARS-CoV-2 were identified who had undergone at least two chest CT studies at Wuhan Leishenshan Hospital between Feb 12, 2020 and Apr 10, 2020 (see more details in Table 1). These patients underwent the first chest CT using the conventional manual positioning and centering method, and an AI-based automatic positioning and centering method in the follow-up CT examination. The patients in our study were limited to the ones without the need for life-supporting tubes and other equipment and could follow verbal command. The interval time between the two scans was 5-8 day. Based on the different positioning methods, patients were categorized into the conventional manual positioning (MP) group and AI-based automatic positioning (AP) group, and all CT images and clinical data between the two groups were compared.
CT image acquisition and reconstruction
The imaging workflows for the MP and AP groups are shown in Figure 1. A and B. The chest CT scanning was performed on a Revolution Maxima CT equipped with an AI-based automatic patient centering and anatomic positioning software (GE Healthcare, Waukesha USA) from the apex pulmonis to diaphragm. Both groups used the same scan protocol with the following parameters: tube voltage, 120 kVp; gantry rotation time, 0.4 second; pitch, 1.375:1; scan field-of-view (SFOV), 50cm; slice thickness, 5 mm; tube current (mA), automated tube current modulation (ATCM) to obtain a noise index of 11.57; All axial images were reconstructed using a standard reconstruction algorithm with the standard kernel; reconstruction display field-of-view (DFOV), 35-50 cm; reconstruction thickness, 1.25 mm.
AI-based automatic patient positioning and centering
The AI-based automatic positioning uses a fixed, ceiling mounted, off the shelf, 2D/3D video camera that can determine distances to points in its field of view. It displays standard RGB video images on the CT system’s existing gantry‐mounted touchscreens (Figure 2 A, B). Information from the standard output of the camera is used, along with precise spatial information of the individual CT system’s gantry and table installation geometry, to determine the anatomical landmark location and the start and end locations for the scout scan(s). The scan protocol structure on the scanner contains a field for the anatomical reference. The 8 supported anatomical references for the automatic positioning method are: Orbital Meatal baseline (OM), Sternoclavicular Notch (SN), Xyphoid (XY), Iliac Crest (IC), Left and Right Knee (KN), Left and Right Ankle Joint (AJ), as shown in Figure 2 C. The automatic positioning software uses two deep learning algorithms (RGBLandmarkNet network and DepthLandmarkNet network) with different inputs that produce comparable outputs to identify all 8 of the anatomical landmarks on the patient’s body. All 8 of these identified landmarks are used to determine the patient orientation (head or feet first). In our study, the SN and IC landmarks were used for the chest scan. The RGBLandmarkNet network uses 2D video images as inputs and outputs all eight of the predefined landmark locations in X and Z. In parallel, the DepthLandmarkNet network uses the 3D depth data from the camera to also produce all eight of the predefined landmark locations. The 3D depth images are used to generate a “point cloud” on a mesh of points on the patient surface contour as determined from the depth information. The point cloud is then segmented to produce the body contour. The body contour is used to deterministically calculate the vertical geometric center of the patient. The center point location is then used to calculate the required table elevation for patient centering. With patient on the CT scanning table, the patient position and centering can be performed automatically with the one-touch button on the console in the control room.
Assessment of image quality
The image quality was analyzed by three radiologists (H.B.X., J.X.H., Y.D.G) at a standard pulmonary display window setting (window level, −700 and window width, 1500). the pulmonary lesions and the locations of ROI for these lesions were established by consensus. The mean CT value and standard deviation (SDev) in Hounsfield Units (HU) of the aorta, trachea and erector spinae in the upper and middle thorax areas were measured by placing a 50 mm2 region-of-interest (ROI) on a homogeneous-appearing area of these structures, as shown in Figure 3 C. Three consecutive images were measured in each ROI area for each study, and the average value was determined. The mean and SDev of CT values within pulmonary lesions were also measured, as shown in Figure 3 A, B. The pulmonary lesions mainly included ground glass opaciﬁcation, consolidation opaciﬁcation and interstitial thickening. Other radiographic abnormality (hydrothorax, nodule or lump, cavitation or calcification, bronchiole or bronchiectasis and emphysema) were also noted. The pulmonary segments were defined by referring to the branching patterns of bronchi [10-12]. If a lesion was located in the outer one third of the lung, it was defined as peripheral, otherwise, it was defined as central. The signal-to-noise ratio (SNR) of the lesions was calculated based on the formula: SNR = Mean CT values/SDev. The image noise was represented using the SDev value.
The positioning time was recorded by the CT technologist for each study. The positioning time was deﬁned as the time from the patient lying on the CT examination bed to technologist finishing positioning and starting scanning.
Off-center distance and positioning accuracy
The patient off-center distance was measured using an axial CT image in the following steps: (ⅰ) select a transverse image containing manubrium and draw a horizontal line that passes through both armpits. (ⅱ) locate the center of the display field of view (DFOV) for the image by displaying the grid and selecting the center cross over point of the grid. (ⅲ) record the vertical distance from the center of DFOV to the horizontal line (Fig 3 D). For the positioning accuracy, a complete coverage should contain the apex pulmonis and diaphragm. thus, if the images of apex pulmonis and diaphragm were fully covered, the patient positioning was considered successful, otherwise, it was defined incomplete or inaccurate (Fig 4 A, B).
The volume CT dose index (CTDIvol in mGy) and dose length product (DLP in mGy-cm) were recorded from the dose report image by the CT technologist for each study. The effective dose (ED in mSv) of the patient was calculated based on the formula: ED = DLP × Cf, where the Cf represents the conversion factor for chest CT (=0.014mSv/mGy-cm).
Continuous variables were expressed as mean ± SD and compared using paired-sample t tests when the data were normally distributed; otherwise, the Wilcoxon signed-rank tests was used; Data distribution was tested with Shapiro-Wilk test. The categorical variables were expressed as number (percentage %) and compared with McNemar's test. A two-tailed P value of less than 0.05 was considered statistically significant. All statistical analyses were conducted with IBM SPSS software (version 22.0).