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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spinal Diagnostics

Background:

• Accurate cervical spine fracture detection is critical for trauma care.
• Current deep learning methods lack anatomical consistency and require extensive manual annotation.
• This limits the reliability and generalization of existing AI approaches.

Purpose of the Study:

• To develop a robust, data-efficient, and anatomically consistent AI framework for cervical spine fracture detection.
• To improve automated segmentation and classification of cervical vertebrae.
• To enhance AI-assisted diagnosis in spinal imaging.

Main Methods:

• A three-stage framework integrating self-supervised learning, structure-aware segmentation, and vertebra-level classification.
• Utilized a bone-window 3D masked autoencoder for feature learning from unlabeled CT scans.
• Employed a structure-aware 3D U-Net with priors and regularization for segmentation, followed by classification with positional embeddings.

Main Results:

• Achieved 89.23% Dice score for cervical vertebra segmentation on the RSNA dataset.
• Reached an AUC of 0.969 for vertebra-level fracture classification (C1-C7).
• Outperformed existing CNN- and transformer-based methods in segmentation and classification.

Conclusions:

• The proposed framework offers a robust, interpretable, and data-efficient solution for automated cervical spine fracture detection.
• Combines self-supervised 3D pretraining with anatomy-aware modeling for reliable AI-assisted diagnosis.
• Presents a practical pathway toward clinical application in spinal imaging.