Introduction
In patients with severe trauma, CT scans are increasingly becoming the standard method for screening spinal fractures. While CT scans offer high diagnostic accuracy, their review demands considerable time and effort. Despite significant advancements in the use of AI for image analysis, its application in spinal diagnostics remains limited. The aim of this study was to investigate the diagnostic accuracy of an AI system trained on axial and coronal CT scans specifically for identifying traumatic lumbar fractures.
Methods
A total of 327 consecutive patients who visited a level one trauma center with thoracolumbar fractures were included in the study. The ground truth dataset for these fractures was verified by two experienced musculoskeletal radiologists and one spine surgeon, using MRI for confirmation. The fractures were classified into three types: vertebral body fracture, transverse process fracture, and posterior element fracture. For the development of the AI system, ResNet deep learning networks were utilized. To assess diagnostic accuracy, the area under the curve (AUC) was calculated. We combined the axial and coronal models through an ensemble method, assigning weights to the axial model for optimization.
Results
The AI system trained on axial CT scans achieved an AUC (Area Under the Curve) of 0.9065 for vertebral body fractures, 0.9801 and 0.9800 for left and right transverse process fractures respectively, and 0.9267 for posterior element fractures. Conversely, the AI trained on coronal CT scans demonstrated AUCs of 0.9154, 0.9345, 0.8556, and 0.8540 for the same respective fracture types. Notably, when combining weights from both axial and coronal images in an ensemble approach, the diagnostic accuracy for posterior element fractures improved to 0.9306.
Discussion
The AI system, when trained on both axial and coronal images, exhibited high diagnostic accuracy. In particular, the AI utilizing coronal images displayed a high level of diagnostic precision for vertebral body fractures. Conversely, the AI employing axial images was notably accurate in diagnosing transverse process and posterior element fractures. Additionally, our findings suggest that incorporating weighted inputs from both types of images into the AI could further enhance its diagnostic accuracy.