Clinical Context Stereotactic body radiotherapy (SBRT) is a precision radiation technique delivering high doses in few sessions to treat pulmonary oligometastases -- a limited number of metastatic tumors in the lung. While effective for many patients, local failure (tumor regrowth) occurs in a significant minority.
Study Design This study analyzed 223 tumor lesions from 146 patients treated with SBRT to determine which pretreatment CT radiomics features extracted from the gross tumor volume (GTV) and the surrounding peritumoral region (pGTV) could predict which lesions would achieve durable local control.
Best Model Performance Model-GPC, combining GTV radiomics, peritumoral radiomics, and clinical features, achieved AUC 0.902 on external validation -- a strong performance indicating reliable discrimination between cases likely to achieve local control versus those prone to local failure.
Key Clinical Predictor Among clinical factors, receiving a curative radiation dose (biologically effective dose at 10 Gy fractionation sensitivity, BED10 of 100 Gy or higher) was the strongest independent predictor of local control, confirming dose-response principles.
SBRT Mechanism SBRT delivers ablative radiation doses (often 3-5 fractions of 10-15 Gy each) that kill tumor cells directly and also damage tumor blood vessels, triggering an immune response that adds to local control. The high dose per fraction distinguishes SBRT from conventional fractionated radiotherapy.
BED10 Concept Biologically effective dose (BED) accounts for tumor repair capacity. A BED10 of 100 Gy or higher is considered the threshold for curative intent SBRT, and this study confirmed that lesions receiving BED10 below 100 Gy had significantly worse local control.
Peritumoral Region Significance The tissue immediately surrounding a tumor reflects the local invasiveness and microenvironmental context of that tumor. Radiomics features from this region capture the biological aggressiveness of the tumor-host interface that CT appearance of the tumor alone misses.
Heterogeneity of Oligometastases Not all metastatic lung lesions respond equally to SBRT. Tumor characteristics such as histology, metabolic activity, and internal heterogeneity influence radioresistance. Imaging-based models can capture these properties non-invasively.
VOI Segmentation For each lesion, the GTV was manually contoured and the pGTV was automatically generated by expanding the GTV contour outward by a fixed margin (typically 5-10 mm) while excluding chest wall, airways, and vessels. Both regions were analyzed separately.
Feature Extraction From each volume of interest, hundreds of radiomics features were extracted including shape descriptors, first-order statistics, and texture features from gray-level co-occurrence matrices, run-length matrices, and other quantitative approaches.
Feature Selection Pipeline LASSO regression reduced the high-dimensional feature space to a smaller set of non-redundant predictors. Separately selected features from GTV and pGTV were then combined in models.
SHAP Analysis SHapley Additive exPlanations (SHAP) values were used to interpret which features drove individual predictions. This analysis revealed that peritumoral heterogeneity features consistently ranked among the top predictors for local failure risk.
Model Architecture The best-performing Model-GPC used a Multilayer Perceptron (MLP) -- a type of neural network -- trained on the combination of GTV radiomics, pGTV radiomics, and clinical features including BED10 dose, primary tumor histology, and time from initial diagnosis.
Training and Validation Models were developed on an internal cohort and validated on an independent external cohort of patients treated at a different institution. This two-institution validation is a stronger test of generalizability than internal cross-validation alone.
Comparative Performance Model-GPC (AUC 0.902 on external validation) outperformed models using only GTV radiomics (AUC 0.823), only pGTV radiomics (AUC 0.841), and only clinical features (AUC 0.798), demonstrating that the combination of all three information sources was essential.
Calibration and Decision Curve Analysis Calibration plots confirmed good agreement between predicted and actual local control rates. Decision curve analysis showed the combined model provided positive net benefit across clinically meaningful risk thresholds.
Dose Escalation Guidance For lesions predicted to have high local failure risk based on the model, clinicians could consider dose escalation (increasing to BED10 greater than 100 Gy) when anatomically feasible and normal tissue tolerances allow.
Monitoring Intensity Lesions with higher predicted failure risk could be imaged more frequently after SBRT -- for example, with CT every 2 months rather than 3 months -- to detect early local failure when salvage reirradiation or surgery is still possible.
Combined Modality Planning The model could identify lesions unlikely to achieve durable local control with SBRT alone, potentially prompting consideration of combined SBRT plus immunotherapy approaches that could enhance the radiation response.
Patient Counseling Quantitative risk scores from the model allow oncologists to provide patients with individualized probability estimates for local control, enabling more informed discussions about expected treatment outcomes.
Retrospective Design The study was retrospective, meaning treatment decisions and follow-up intervals were not standardized. Prospective studies with predefined SBRT protocols would provide cleaner data for model training.
Histology Heterogeneity Pulmonary oligometastases arise from many primary tumor types (colorectal, breast, kidney, etc.), each with different radiation sensitivity. The model may need to be separately validated or adapted for different primary tumor histologies.
Segmentation Standardization Manual GTV contouring introduces inter-observer variability. Different radiation oncologists may contour the same tumor slightly differently, potentially changing the radiomics features extracted.
Dynamic Radiomics Future studies could use changes in radiomics features between the planning CT and mid-treatment or post-treatment CT as early response biomarkers, rather than relying solely on pre-treatment prediction.