Fully Automatic Deep Learning Model for Spine Refracture in Patients with OVCF: A Multi-Center Study

Area of Science:

• Radiology
• Artificial Intelligence
• Orthopedics

Background:

• Current artificial intelligence (AI) models for predicting spinal refracture lack specific indicators for osteoporosis and bone quality.
• Existing methods rely on bone mineral density, X-rays, and conventional lab tests, which have limitations.

Purpose of the Study:

• To develop a novel predictive model for spinal refracture using bone turnover markers and computed tomography (CT) imaging.
• To identify patients at higher risk of spine refracture.

Main Methods:

• Retrospective collection of CT images and clinical data from 383 patients with osteoporotic vertebral compression fractures (OVCF).
• Utilized U-net for automatic region of interest (ROI) segmentation and Densenet 121-3D for predictive modeling.
• Developed a T-NIPT prediction model and assessed diagnostic performance using ROC curves, calibration curves, and decision curve analysis (DCA).

Main Results:

• The 3D-Full model demonstrated superior calibration performance and predictive accuracy compared to models using only clinical features or 3D_RoiOnly.
• T_P1NT was identified as an independent risk factor for refracture.
• The 3D-Full model integrated with clinical data showed significant improvement (p < 0.05) in predicting refractures.

Conclusions:

• T_P1NT is a significant independent risk factor for spinal refracture.
• The developed 3D-Full AI model outperforms existing methods and clinicians in predicting high-risk populations for spine refracture.
• The model's automatic segmentation and detection capabilities support real-world clinical application.