SpineMAE: A bone-window self-supervised and structure-aware framework for 3D cervical vertebra segmentation and fracture classification

Accurate segmentation and classification of cervical spine fractures are essential for timely diagnosis and clinical decision-making in trauma care. Existing deep learning approaches often require extensive manual annotations and struggle to maintain anatomical consistency across vertebral levels, limiting their reliability and generalization.

We propose a unified three-stage framework that integrates self-supervised representation learning, structure-aware segmentation, and vertebra-level classification. A bone-window 3D masked autoencoder first pretrains the encoder to learn bone-sensitive volumetric features from unlabeled CT scans. The pretrained encoder is then transferred to a structure-aware 3D U-Net that incorporates bone-likelihood and edge priors with geometry-consistency regularization for anatomically coherent vertebral segmentation. Finally, a vertebra-level classification head performs fracture prediction using ROI features with positional embeddings to encode anatomical context.

Evaluated on the RSNA Cervical Spine Fracture Detection dataset, the proposed method achieved a Dice score of 89.23% for cervical vertebra segmentation. For vertebra-level fracture classification, it achieved an AUC of 0.969 with sensitivity of 0.788 and specificity of 0.962 (macro-averaged across C1-C7), outperforming CNN- and transformer-based baselines.

The proposed framework provides a robust, interpretable, and data-efficient solution for automated cervical spine fracture detection and localization. By combining self-supervised 3D pretraining with anatomy-aware modeling, it offers a practical pathway toward clinically reliable AI-assisted diagnosis in spinal imaging.