An AI model can flag broken neck vertebrae almost as well using flattened 2D images as it can by processing entire three-dimensional CT volumes.
Researchers built a multi-stage pipeline that first locates the cervical spine in a CT scan, then estimates masks for each of the seven cervical vertebrae (C1 through C7) using optimized two-dimensional projections rather than full volumetric segmentation. A YOLOv8 object detector handles the initial spine localization, hitting a 3D mean Intersection over Union of 94.45%. A DenseNet121-Unet then segments the vertebrae from energy-based sagittal and coronal projections, scoring a mean Dice of 87.86%. Those 2D masks get back-projected into approximate 3D volumes, which feed an ensemble of hybrid CNN-Transformer models for the actual fracture call. The system reached an area under the ROC curve of 91.62 at the vertebra level and 90.95 at the patient level.
Cervical spine fractures are time-sensitive emergencies, and radiologists reading CT scans face the brutal math of large 3D volumes full of subtle findings. If a 2D-projection approach can stand in for full volumetric segmentation without meaningful accuracy loss, it meaningfully lowers the barrier to deploying automated triage tools in settings without high-end GPU infrastructure. The researchers also ran saliency-based explainability analysis, a nod toward the interpretability demands regulators and clinicians increasingly place on diagnostic AI.
The F1 scores here - 68.15 at the vertebra level, 82.26 at the patient level - are competitive but not yet good enough to replace a radiologist. This looks more like a credible triage assist than a diagnostic endpoint.