Cervical spine fracture detection in computed tomography using convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Trauma Care

Background:

• Timely interpretation of computed tomography (CT) scans is crucial for in-hospital trauma management.
• Cervical spine fracture assessment is time-consuming, with risks of under-diagnosis or delayed diagnosis.
• Artificial intelligence (AI) shows promise in assisting radiologists with spinal fracture detection.

Purpose of the Study:

• To develop and evaluate an AI-based segmentation model for detecting cervical spine fractures.
• To compare different image reformation approaches for optimizing fracture detection.
• To improve the accuracy and efficiency of cervical spine fracture identification in trauma CT scans.

Main Methods:

• Trained a U-Net model using a dataset of 195 CT examinations with 454 annotated cervical spine fractures.
• Employed a multi-task loss for segmenting both spine fractures and the spine.
• Compared three data processing approaches: Cartesian, straightened curved planar reformatted (CPR), and spinal canal aligned volumes of interest (VOI).

Main Results:

• The VOI approach yielded the best fracture detection rate and reduced computation time.
• The algorithm achieved an 87.2% detection rate for cervical spine fractures with an average of 3.5 false positives per case.
• On a public dataset, the method achieved 0.9 false positives per cervical spine case.

Conclusions:

• Voxel classification-based fracture detection, particularly using the VOI approach, offers high sensitivity for identifying individual fracture locations.
• This AI tool has the potential to significantly support trauma CT reading workflows by minimizing missed findings.
• The developed segmentation approach enhances diagnostic accuracy and efficiency in critical trauma scenarios.