Resident Baylor College of Medicine Neurosurgery Houston, TX, US
Introduction: Neurosurgical evaluation is required in the setting of metastatic vertebral compression fractures (VCF). Prompt diagnosis and triage is critical in determining management. Commonly, VCF are first seen on CT of the chest/abdomen/pelvis (CAP), but may be missed as these studies are primarily ordered for diagnosis and staging of solid organ tumors. We developed a CT-radiomic-based classifier to identify VCF on CT CAP.
Methods: A single institution, adult all-comer cohort of patients with tumor infiltrated vertebral bodies, spanning T1 through L5, with at least 1 year of follow up were included. Individual vertebral bodies were manually segmented using respective CT CAPs. All PyRadiomics feature classes were extracted from the original images/masks. Wilcoxon-rank sum tests were used to find significantly variable features, and recursive feature elimination was conducted using a random forest algorithm yielding the features used in the final random forest classifier for fracture identification.
Results: The cohort included 185 fractured vertebrae (23%) and 622 non-fractured vertebrae (77%) across 122 patients. In the follow-up period, 43 new fractures developed, with 7 (16%) missed on CT CAP. Feature selection yielded a set of 28 radiomic features that maximized AUC-ROC. Upon 10-repeat 10-fold cross-validation, tuning of hyperparameters resulted in an AUC-ROC of 0.85 (95% CI: 0.81-0.89) with a 74% sensitivity (95% CI: 66%-81%) and 81% (95% CI: 78%-84%) specificity in detecting VCF.
Conclusion : We have created an AI solution for diagnosing pathologic compression fractures on CT CAP. With further training and validation on external datasets, this tool could aid radiologists with prompt diagnosis of VCF resulting in earlier triage of these patients.
.jpg)
