Participants
A total of 214 cases of ARFs diagnosed by chest CT examination due to acute blunt chest trauma were collected retrospectively from July 2018 to February 2019 in our hospital, including 123 males and 91 females, age range from 21 to 89 years, with an average of (54±14) years old. The inclusion criteria were as follows: 1) A clear history of acute chest trauma; 2) certain or suspected rib fractures were reported by the radiological diagnoses. Exclusion criteria: Image quality undiagnostic due to motion artifacts,15 unconsciousness or lack of self-control, etc. The Affiliated Hospital of Shaanxi University of Chinese Medicine institutional review board reviewed and approved the protocol and provided continuing oversight. All participants provided written informed consent.
CT Examination
All cases were scanned with a 64-row spiral CT (Discovery HD 750, GE Healthcare). Scan position: supine position with upper arm lifted (some patients were naturally placed on both sides of the body due to injury in the shoulder or upper limb). Scanning range: from the thoracic opening to the 12th rib lower edge. Scanning parameters: tube voltage: 120 kVp, tube current: 20mA-500mA, pitch: 0.85, reconstruction convolution kernel: Standard or Bone, slice thickness: 1.25mm.
DL-CAD System
A commercial and easy-to-use DL-CAD system (InferRead CT Bone Research, Infervision, Beijing, China), which extracted image features via an artificial convolutional neural network to automatically detect rib fractures were used in our study. An information list containing all detected rib fractures for each patient were provided by the DL-CAD, and the location of each fracture on the CT images was labeled with a box.
Reading Experiment
2 intern radiologists (who have less than 1 year of experience) and 2 attending radiologists ( who have 7 years of experience) participated in this reading experiment. The task was performed in two reading sessions at an interval of 4 weeks apart. At each session, all 214 cases were randomized into 7 groups, including 31 patients in groups 1, 3, 5, and 7, and 30 patients in groups 2, 4, and 6. All data sets were presented to readers in randomized order, and orders were different for every reader. In order to reduce the adverse impact from fatigue, which was usually caused by long-time consecutive reading work, the readers performed one group of reading experiment per day and thus they finished each session in a week. All readers interpreted all the cases on a picture archiving and communication system (PACS) independently in the first session. And after a memory washout period of 4 weeks, the second session were implemented, in which the reader re-interpreted all cases with the assistance of DL-CAD in concurrent reading mode. During the reading procedure, the readers could adjust the window width/level and zoom in/out, or use maximum intensity projection (MIP) or volume rendering (VR) if needed.
In both reading sessions, all readers were instructed to focus on detecting rib fracture. Considering that each rib has a possibility of fracture, 24 ribs for each patient were evaluated. And a 5-point Likert-scale was used to evaluate the diagnostic confidence of each rib: 1, definitely absent; 2, probably absent; 3, indeterminate; 4, probably present; 5, definitely present. For each patient, the fracture locations and diagnostic confidence score of each rib and the time-consumption for reading procedure were recorded. All readers had received DL-CAD system knowledge training before reading.
Gold Standard
Two senior radiologists (Mr. JIA and Mr. DUAN, both with more than 15 years of experience in chest CT diagnosis) reviewed all the cases and their consensus diagnoses were referred as surrogate gold standard. Additionally, in order to evaluate the diagnostic performance of complete rib fracture (CRF) and occult rib fracture (ORF), each fracture was classified as either complete or occult by the two senior radiologists. A CRF was confirmed if the fracture line run through the entire cortical bone with the cortex continuity completely interrupted. An ORF was referred to the bone density increasing, folding, warping, partial unconnected at external and/or internal bone cortex.16 In order to improve diagnostic accuracy, the electronic medical record and follow up CT examination images would be reviewed in this gold standard session if needed.
Statistical Analysis
The diagnostic sensitivity, specificity, positive predictive value (PPV), diagnostic confidence and average time-consumption per case were calculated and compared between two reading sessions. Receiver operating characteristics (ROC) analysis was performed and area under the receiver operating characteristic curve (AUC) was calculated and compared. The paired sample t-test was used to compare quantitative data, Chi-square test was used to compare the Constituent ratio data, and nonparametric Wilcoxon test was used to compare the ranked data. The P value of less than 0.05 was considered statistically significant. Statistical analysis was performed using SPSS® Statistics 19.0 (IBM Corporation, Armonk, NY; formerly SPSS Inc., Chicago, IL).