Study Design: Cross-sectional. Objective: To evaluate the performance of a deep learning model designed to detect column-specific cervical spine (C-spine) fractures using CT imaging. Summary of Background Data: Accurate and timely diagnosis of C-spine fractures is essential in trauma cases to prevent complications and improve patient outcomes. Although CT imaging is the gold standard for fracture detection, it is resource-intensive and prone to human error. Advancements in artificial intelligence (AI) offer the potential to diagnose and classify C-spine fractures with little human input. Methods: A total of 398 C-spine CT studies were retrospectively analyzed. Ground-truth labels were established through expert annotations of MRI images. Fractures were categorized into vertebral body (or anterior arch and dens), posterior element, left, and right transverse process fractures. Of the 398 cases, 80% (318) were used for training, 10% (40) for validation, and 10% (40) for testing. The deep learning model was developed using a ResNet-based architecture. Model performance was evaluated using sensitivity, specificity, the area under the receiver operating characteristic curve (AUROC), and accuracy. Results: Males comprised the majority of cases (318 patients), with a mean age of 54.3±16.0 years. Among the 398 cases, C1 through C7 fractures were identified in 49, 79, 63, 78, 107, 161, and 201 cases, respectively. The AI model demonstrated strong performance with an AUROC of 0.8841 and an accuracy of 0.8179 for overall C-spine fracture detection. Subgroup AUROCs were 0.8691 for vertebral body fractures, 0.8766 for posterior fractures, 0.8694 for left transverse process fractures, and 0.9009 for right transverse process fractures. Conclusion: The deep learning model demonstrated high diagnostic accuracy for detecting column-specific C-spine fractures on CT imaging. These findings highlight the potential of AI to enable efficient, precise C-spine fracture classification in trauma settings, ultimately improving patient outcomes.
Lee et al. (Fri,) studied this question.