Automated Deep Learning Method for Rib Fracture Detection and CWIS Classification
Overview
This study developed an automated deep learning approach using nnDetection and nnU-Net frameworks to detect rib fractures and classify them according to the Chest Wall Injury Society (CWIS) system. The method addresses challenges in manual rib fracture classification, improving reliability and potentially enhancing clinical decision-making.
Background
Traumatic rib fractures are common in thoracic trauma, accounting for 10% of trauma admissions and occurring in 10–40% of trauma patients. These fractures can lead to significant morbidity and mortality, especially when combined with complications such as hemothorax or pneumonia. The CWIS classification system categorizes rib fractures by type, displacement, and position, but manual classification suffers from interobserver variability. Computed tomography (CT) is the gold standard for diagnosis, yet manual detection misses up to 26.8% of fractures and is time-consuming. An automated, reliable classification method is needed to improve treatment strategies.
Data Highlights
The automated approach consists of three components: fracture detection, classification according to CWIS categories (type, displacement, position), and rib numbering. Detection and classification use the nnDetection framework with Retina U-Net architecture, while rib numbering is performed by a dedicated nnU-Net. The method was trained on CT scans with annotated fractures represented as spherical objects with a 10 mm radius centered on the fracture site.
Key Findings
The automated method uses three separate nnDetection models to classify fracture type (simple, wedge, complex), displacement (undisplaced, offset, displaced), and position (anterior, lateral, posterior).
Retina U-Net architecture enables multi-scale feature extraction, improving detection of fractures of varying sizes.
The approach includes an automated rib numbering system using nnU-Net to assign rib segments to detected fractures.
Manual rib fracture classification shows significant interobserver variability, which the automated method aims to reduce.
CT remains the gold standard imaging modality, but manual detection misses up to 26.8% of fractures, highlighting the need for automation.
Clinical Implications
The automated detection and classification system can standardize rib fracture evaluation, reducing variability and diagnostic errors in clinical practice. This may facilitate more consistent application of SSRF guidelines and improve communication among clinicians. Ultimately, the method has the potential to optimize treatment decisions and patient outcomes by providing reliable, rapid fracture assessment.
Conclusion
This study presents a novel, automated deep learning approach for rib fracture detection and CWIS classification that addresses limitations of manual assessment. The method promises to enhance diagnostic accuracy and support clinical decision-making in thoracic trauma care.
