Abstract
Stereotactic body radiation therapy (SBRT) is the standard of care for medically inoperable non-small cell lung cancer (NSCLC) patients. Lung SBRT offers excellent local control rates for NSCLC. However, the regional and distal failure remains high, at approximately 30%. This study aims to design and compare two novel dosiomics models incorporating the interplay between tissue density and the dose to predict potential regional or distant failures following lung SBRT.
We retrospectively collected 140 primary NSCLC patients treated with SBRT and calculated five matrices representing the interaction of the planning CT with the physical dose or biologically effective dose (BED). We used two distinct image sets: each set had seven images of interest—the CT image, dose distribution (physical or BED), and their five corresponding interaction matrices. We extracted 2205 features from each image set using radiomics mathematics. We selected the most important and non-redundant features and used them to build predictive models for each set.
Among the 2205 extracted features in each set, we used the 14 top-important features to build the respective predictive models using XGBoost, of which 11 and 10 features are from the physical dose-based and BED-based interaction matrices, respectively. For the BED-based predictive model, the AUC for the test dataset was 0.84, with a 95% confidence interval of (0.60, 1.00), and that for the physical dose-based model was 0.80, with a 95% confidence interval of (0.53, 1.00). The physical dose-based model demonstrated prediction accuracy, sensitivity, and specificity of 0.857, 0.60, and 1, respectively, whereas that for the BED-based model was 0.857, 0.84, and 0.89, respectively.
The results of our exploratory study indicate that the dose and tissue interaction has the potential to predict treatment outcomes. Further study with a larger patient population and more testing cohorts is warranted to establish a clear advantage of using physical dose or BED to calculate the interaction matrices for building similar predictive models. In conclusion, the novel dosiomics model, incorporating the interaction between the CT and dose, may effectively predict regional or distant failure following lung SBRT treatment.