The impact of AI tools on the detection of chest injuries from clinical imaging: a systematic review and meta-analysis
File(s)
Author(s)
Lai, James
Cheng, Ka-Jun
Frith, Dan
Masouros, Spyros
Hettiaratchy, Shehan
Type
Poster
Abstract
Chest injuries are a common trauma presentation to emergency departments. Artificial intelligence (AI) has the potential to detect pathologies using medical imaging through pattern recognition. This study aims to review current literature investigating the diagnostic accuracy of AI tools to detect traumatic chest injuries in trauma patients.
We conducted a systematic review of studies to evaluate the diagnostic performance of chest-injury detection by clinicians with and without the aid of AI. Studies written in English were retrieved from EMBASE, MEDLINE, and IEEE Xplore in March 2024 and included studies evaluating the diagnostic accuracy of detecting pneumothoraces, haemothoraces, lung contusions and rib fractures. The risk of bias was assessed using the Quality Assessment of Diagnostic Accuracy Studies tool. A meta-analysis of the diagnostic performance was summarised using recall (sensitivity) and diagnostic time.
A total of 21 studies were selected for our final review. Two studies evaluated the AI-assisted detection of pneumothoraces and rib fractures on plain film radiographs. The remaining studies evaluated the detection of rib fractures on CT imaging. Twelve studies were included in the meta-analysis, with 3524 and 2885 CT scans read by clinicians with and without AI assistance, respectively. Unassisted recall ranged from 0.61 – 0.83 compared to AI-assisted, ranging from 0.73 – 0.95, representing a mean difference of 0.12 (95% CI 0.08 – 0.16). For diagnostic time, the unassisted performance was 82s–342s, with an assisted performance of 41s–302s, representing a mean difference of -102.12s (-141.00s – -63.25s).
Our study is the first to evaluate the performance of AI tools in patients presenting with chest injuries. Our findings demonstrate that AI-assisted reading of CT imaging is associated with an improved identification of rib fractures and reduced diagnostic time. However, the detection of rib fracture injuries highlights the current limitation of AI model use in the acute trauma-care pathway.
We conducted a systematic review of studies to evaluate the diagnostic performance of chest-injury detection by clinicians with and without the aid of AI. Studies written in English were retrieved from EMBASE, MEDLINE, and IEEE Xplore in March 2024 and included studies evaluating the diagnostic accuracy of detecting pneumothoraces, haemothoraces, lung contusions and rib fractures. The risk of bias was assessed using the Quality Assessment of Diagnostic Accuracy Studies tool. A meta-analysis of the diagnostic performance was summarised using recall (sensitivity) and diagnostic time.
A total of 21 studies were selected for our final review. Two studies evaluated the AI-assisted detection of pneumothoraces and rib fractures on plain film radiographs. The remaining studies evaluated the detection of rib fractures on CT imaging. Twelve studies were included in the meta-analysis, with 3524 and 2885 CT scans read by clinicians with and without AI assistance, respectively. Unassisted recall ranged from 0.61 – 0.83 compared to AI-assisted, ranging from 0.73 – 0.95, representing a mean difference of 0.12 (95% CI 0.08 – 0.16). For diagnostic time, the unassisted performance was 82s–342s, with an assisted performance of 41s–302s, representing a mean difference of -102.12s (-141.00s – -63.25s).
Our study is the first to evaluate the performance of AI tools in patients presenting with chest injuries. Our findings demonstrate that AI-assisted reading of CT imaging is associated with an improved identification of rib fractures and reduced diagnostic time. However, the detection of rib fracture injuries highlights the current limitation of AI model use in the acute trauma-care pathway.
Date Issued
2025-09-29
Citation
2025
Copyright Statement
© 2025 The Author(s).
Source
EUSEM 2025 European Emergency Medicine Congress
Subjects
Chest injuries
Rib fractures
Artificial intelligence
Computed tomography
X-rays