[Application of deep learning-based chest CT auxiliary diagnosis system in emergency trauma patients]

Objective: To investigate the diagnostic efficacy and potential application value of deep learning-based chest CT auxiliary diagnosis system in emergency trauma patients. Methods: A total of 403 patients, including 254 males and 149 females aged from 16 to 100 (50±19) years, who received emergency treatment for trauma and chest CT examination in the Eastern Theater General Hospital from September 2019 to November 2019 were retrospectively analyzed. Dr. Wise Lung Analyzer's chest CT auxiliary diagnosis system was applied to detect 5 types of injuries, including pneumothorax, pleural effusion/hemothorax, pulmonary contusion (shown as consolidation and ground glass opacity), rib fractures, and other fractures (including thoracic vertebrae, sternum, scapula and clavicle, etc.) and 6 other abnormalities (bullae, emphysema, pulmonary nodules, stripe, reticulation, pleural thickening). The diagnostic reference standards were labeled by two radiologists independently. The sensitivity and specificity of the auxiliary diagnosis system were evaluated. The imaging diagnostic reports were compared with the results of the auxiliary diagnosis system, and the diagnostic consistency between the two was calculated by using the Kappa test. Results: According to the reference standards, among the 403 patients, 29 were pneumothorax, 75 were pleural effusion/hemothorax, 131 were pulmonary contusion, 124 were rib fractures, and 63 were other fractures. The sensitivity and specificity of the auxiliary diagnosis system for detection of pneumothorax, pleural effusion/hemothorax, rib fractures, and other fractures were 96.6%, 97.6%, 80.0%, 99.7%, 99.2%, 83.9%, 84.1%, and 99.7%, respectively. The sensitivity of detecting lung contusion was 97.7%. There was a high consistency between the auxiliary diagnosis system and imaging diagnosis in the diagnosis of injuries, in which the kappa values of pneumothorax, pleural effusion, rib fracture and other fractures were 0.783, 0.821, 0.706 and 0.813, respectively (all P<0.001). Two cases of pneumothorax, three cases of pleural effusion/hemothorax, nine cases of rib fractures, and six cases of other fractures missed by imaging diagnosis were all detected by the auxiliary diagnosis system. The detection sensitivity of the auxiliary diagnosis system was higher for emphysema, pulmonary nodules and stripe (all>85%), but lower for bullae, reticulation and pleural thickening. Conclusions: The deep learning-based chest CT auxiliary diagnosis system could effectively assist chest CT to detect injuries in emergency trauma patients, which was expected to optimize the clinical workflow.