Purpose: Despite improvements in treatment technique, local recurrence (LR) of non-small cell lung cancer (NSCLC) treated with stereotactic body radiation therapy (SBRT) remains a clinical problem. Quantitative parameters to predict LR risk are limited. This study investigates the potential utility of CT radiomics features and dose distribution radiomics features, “dosiomics”, from the GTV and PTV to predict for risk of LR.
Methods: 105 patients with early-stage NSCLC treated with SBRT were retrospectively evaluated. Planning CT images and SBRT dose distributions (calculated with the Varian AAA 15604 algorithm) were analyzed. A total of 284 radiomics and dosiomics features were extracted from these images. Key features selected by a least absolute shrinkage and selection operator method were used to perform a linear regression correlation model to predict LR within two years of treatment. The predictive performance was evaluated by receiver operator characteristic (ROC) curve analysis.
Results: Models based on CT radiomics and dosiomics were both associated with LR status, and the best performance came from models derived from both. The models based on PTV radiomics and dosiomics had areas under the ROC curves (AUCs) of 0.76 and 0.70, respectively. The model based on both radiomics and dosiomics features from the PTV had an AUC of 0.82. Including radiomics and dosiomics features from the GTV improved the AUC to 0.84. Two clinically relevant dosiomics features were correlated with LR: kurtosis in the GTV and minimum dose to the PTV.
Conclusion: Radiomics and dosiomics features correlate with risk of LR after SBRT in patients with NSCLC. Two clinically relevant dosiomics features were identified: minimum dose to the PTV and kurtosis, which relates to the distribution of dose, in the GTV. These features warrant further investigation as possible predictors of LR and might eventually be useful for incorporation into clinical SBRT plan optimization and evaluation.
Image Analysis, Linear Regression Analysis, Lung
Not Applicable / None Entered.