Preprocessing Filters Improve Performance of Ct-biomarker-based Prediction of Metastatic Vertebral Body Compression Fractures

Introduction: Neurosurgical evaluation is required in the setting of spinal metastases that are at high risk of developing a vertebral body fracture (VBF). Understanding fracture risk is critical in determining management. Although spinal metastases are typically diagnosed initially by CT, few tools exist to quantify fracture risk from this imaging. CT-based biomarkers have shown promise in predicting VBF, but achieving state-of-the-art performance has been hindered by low training data. We sought to maximize radiomic performance on a small dataset utilizing a suite of preprocessing filters to achieve highest in-class performance.

Methods: A multi-institutional, 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. Subsequently, eight preprocessing filters were applied (wavelet, square, square root, logarithm, exponential, Laplacian of Gaussian, gradient, local binary pattern 3D) and all PyRadiomics feature classes were extracted from these images, as well original images. Predictive modeling was performed using an ensemble pre-optimized ensemble classifier using neural network, random forest, and support vector classifier algorithms, and predictive performances were compared across models using area-under-the-curve at receiver operating characteristics analysis (AUC-ROC) using a 15% holdout sample.

Results: Forty-three patients with 212 vertebrae (35 fractures) were included. For predicting future fracture, statistically significant improvements up to 0.92 AUC-ROC (95% CI: 0.83-0.96) were achieved when all image preprocessing filters were used compared to AUC-ROC 0.77 (95% CI: 0.70 – 0.82) with only non-filtered images (p = 0.032). Individually, the square root filter AUC-ROC 0.93(95% CI: 0.75-0.97) trended toward the best performance but did not reach significance compared to no filters (p = 0.091).

Conclusion : Preprocessing filters can critically increase performance for this radiomic task and is an effective method for low training data problems. Custom filters may further enhance performance and should be subject of future studies.