CT Radiomics Predict New or Progression of Compression Fracture in Tumor Infiltrated Thoracolumbar Vertebrae

Introduction: Spinal metastases are common and neurosurgical evaluation is required in the setting of spinal lesions that are at high risk of developing a vertebral body fracture (VBF). Virtually all patients are initially diagnosed with spinal metastases via CT, yet few tools exist to quantify fracture risk from this imaging. Understanding fracture risk is critical in determining management, including follow-up timing and prophylactic interventions. Often, pathology and specific treatment status are unknown. We report our results on a computer vision model that predicts development/progression of pathologic compression fractures in tumor infiltrated thoracolumbar vertebrae on CT imaging alone.

Methods: Vertebral levels T1 through L5 containing tumor from adult all-comers with at least 1 year follow up were included. Clinical features of the patient, disease, and treatment were collected. Radiomic features were extracted from the trabecular bone and associated tumor of the vertebral bodies from their earliest CT with tumor or pre-fracture/progression CT. Radiomic features selected by an extra-trees classifier were included in a logistic regression model trained to predict VBF within 1 year. Model performance was assessed with stratified cross-validation.

Results: Within 1 year from initial CT, 25 of 156 vertebrae developed new or progressive VBF. These included 13 pathologies, both pre- and post-chemo/radiation. Voxel volume, shape elongation, shape flatness, maximum 2D diameter row, shape sphericity, first order 10th percentile, mean absolute deviation, first order mean, and first order median were the highest contributing attributes used in the final logistic regression model yielding AUC 0.760 (SD 0.035).

Conclusion : We have developed a generalizable computer vision model, based only on CT radiomic profiles of tumor infiltrated thoracolumbar vertebrae, to predict risk of developing or progressing compression fractures in an all-comer population. If validated, this tool will be useful for fracture risk assessment from initial or follow up CTs.