Author: Yukio Fujita, Syoma Ide, Kei Ito, Tomohiro Kajikawa, Satoshi Kito, Keiko Murofushi, Yujiro Nakajima, Yuhi Suda, Kentaro Taguchi, Naoki Tohyama, Fumiya Tsurumaki 👨🔬
Affiliation: Komazawa University Graduate School, Tokyo Metropolitan Cancer and Infectious Diseases Center Komagome Hospital, Department of Radiology, Kyoto Prefectural University of Medicine 🌍
Purpose: Stereotactic body radiotherapy (SBRT) for spine metastases is more effective for pain relief and local control than conventional radiotherapy. However, it is associated with vertebral compression fractures (VCF). We herein aimed to create a model for accurately predicting VCF occurrence using clinical, radiomic, and dosiomic features.
Methods: We retrospectively analyzed 162 spinal segments, including 44 VCF cases, in 114 patients with spine SBRT. Radiomic and dosiomic features were extracted from the clinical target volume using computed tomography (CT), T1-weighted, T2-weighted MR images (T1WI, T2WI), and dose distributions. The features of interest were selected using the leave-one-out procedure, the least absolute shrinkage selection operator, and pearson’s correlation coefficient. Conventional model was created on the basis of clinical charactersitics (age, dose prescription, etc.) to serve as clinical model. Clinical, radiomic, and dosiomic models were developed using the random forest method. Six prediction models (clinical, CT, T1WI, T2WI, dose distribution, and clinical/CT/T1WI/T2WI/dose distribution) were created to predict VCF occurrence, and their performance was assessed using the area under the curve (AUC) of the receiver operating characteristic (ROC). The ROC curves were compared using DeLong’s method (95% confidence intervals).
Results: For the optimal classification model, 11 features were selected for the clinical/CT/T1WI/T2WI/dose distribution model, including three clinical, one T1WI, and five T2WI features and two dose distributions. The AUC (95% confidence interval) for the clinical, CT, T1WI, and T2WI features and the dose distribution and clinical/CT/T1WI/T2WI/dose distribution was 0.84 (0.76–0.92), 0.72 (0.63–0.81), 0.79 (0.71–0.87), 0.81 (0.74–0.89), 0.76 (0.68–0.85), and 0.88 (0.81–0.94), respectively. The clinical/CT/T1WI/T2WI/dose distribution model was significantly more accurate than the other models except clinical model (p < 0.05 for each comparison).
Conclusion: These findings indicated that combining VCF prediction models based on clinical, radiomic, and dosiomic features improved the accuracy of the prediction.