MRI-Based Deep Learning for Predicting Vertebral Fractures Risk in Patients with Low Bone Mass: a Multicenter Validation Study (n=1182)

Motivation: Patients with low-energy vertebral fractures are typically at high risk for future fracture. Lumbar MRI-based deep learning (DL) can help improve prediction of fracture risk. Goal(s): To develop and validate a DL model for fracture prediction based on vertebral and paraspinal muscular MR images. Approach: Establish a DL model to predict fracture risk in low bone mass patients, including subgroup analyses for osteopenia and osteoporosis. Data was collected from three hospitals with an external validation set. Results: In osteopenia, the combined vertebral-muscle model achieved an AUC of 0.648, outperforming the vertebral-only model's 0.582 on the external validation set. Impact: The paraspinal muscles, as one of the key structures in maintaining spinal stability, work synergistically with the vertebrae in predicting vertebral fracture risk, especially in osteopenia patients.