Radiomics- and Dosiomics-Based Machine Learning Models for Predicting 1- and 2-Year Local Failure after SBRT for Non-Spinal Bone Metastases
Abstract
Purpose
Stereotactic body radiotherapy (SBRT) for non-spinal bone metastases generally achieves high local control; however, approximately 10% of patients experience local failure (LF). Conventional clinical and dose metrics often fail to capture patterns associated with LF. This study aimed to develop and validate machine learning models using radiomic and dosiomic features to predict 1- and 2-year LF.
Methods
We retrospectively analyzed SBRT-treated non-spinal bone metastases, predominantly with 35 Gy in five fractions. Seventy-six lesions (11 LF) were analyzed for the 1-year endpoint and 48 lesions (15 LF) for the 2-year endpoint. Radiomic and dosiomic features were extracted from the planning target volume using CT images and dose distributions. Feature selection used LASSO regression with leave-one-out cross-validation (LOOCV), followed by Pearson correlation filtering. Four models were developed for each endpoint: (1) conventional clinical factors (e.g., age, primary tumor site), (2) CT radiomics, (3) dosiomics, and (4) combined model integrating all features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) with LOOCV, and differences were assessed using DeLong’s test (significance level: 0.05).
Results
For the 1-year endpoint, combined model selected six features (2 clinical, 2 CT radiomics, and 2 dosiomics). The AUCs were 0.50 (clinical), 0.82 (CT radiomics), 0.57 (dosiomics) and 0.82 (combined). For the 2-year endpoint, combined model selected seven features (4 radiomic, 3 dosiomic), yielding AUCs of 0.72 (clinical), 0.84 (radiomics), 0.72 (dosiomics), and 0.84 (combined). The CT radiomics and combined models demonstrated significantly higher AUCs than the clinical model for the 1-year endpoint, with no significant difference for the 2-year endpoint.
Conclusion
Radiomic features derived from CT images demonstrated predictive performance for both 1- and 2-year LF, significantly outperforming conventional clinical factors for 1-year LF prediction. These findings suggest that radiomics-based modeling can facilitate early risk stratification after SBRT for non-spinal bone metastases.