A research pipeline skips expensive 3D vertebra segmentation and still catches cervical spine fractures as reliably as methods that do the full volumetric work.
The system chains three stages: a YOLOv8 detector localizes the spine from multi-view projections, hitting a 3D mean Intersection over Union of 94.45%; a DenseNet121-Unet estimates vertebra masks from 2D sagittal and coronal projections with a mean Dice score of 87.86%; and an ensemble of 2.5D CNN-Transformer models classifies fractures from back-projected volumes. The result is a vertebra-level AUC of 91.62 and a patient-level AUC of 90.95 — two distinct measures, because a patient can have fractures at multiple vertebrae. Patient-level F1 reached 82.26, with a patient-level precision-recall AUC of 92.00.
The result matters because CT-based spine fracture diagnosis is time-sensitive and radiologist attention is finite. If projection-derived masks are genuinely good enough — and the paper argues they are, benchmarked against a full 3D segmentation baseline — then the segmentation bottleneck shrinks without sacrificing diagnostic signal. That is a real efficiency gain in trauma workflows, not a marketing claim.
Automatic cervical fracture detection has been a stubborn problem; prior work leaned heavily on full 3D pipelines that are computationally expensive and require annotated volumetric data. This approach trades dimensional completeness for tractability — and the numbers suggest the tradeoff holds.