3703 - Predicting Gross Tumor Volume Change in Personalized Ultra-Fractionated Stereotactic Adaptive Radiotherapy (PULSAR)

Purpose/Objective(s): Personalized ultra-fractionated stereotactic adaptive radiotherapy (PULSAR) is an innovative treatment approach. However, early decision-making in PULSAR faces challenges in feature selection and outcome prediction, which goes beyond just classification. This study aims to develop a regression model to predict the change in Gross Tumor Volume (GTV), facilitating data-driven decision-making and more personalized adaptation.

Materials/Methods: A retrospective study encompassing 69 brain metastasis lesions treated with PULSAR was conducted. Radiomics and dosiomics features were extracted from MRI scans and dose maps, respectively, from both pre-treatment and intra-treatment magnetic resonance images and dose maps. Delta-omics was computed based on the relative changes between pre- and intra-treatment features. For feature selection, we applied variance threshold, linear correlation analysis, and least absolute shrinkage and selection operator (Lasso) algorithm. Support vector regression (SVR) models were then constructed to predict GTV change at 3-month follow-up. To manage the small sample size and mitigate overfitting, we employed a five-fold cross-validation procedure with 10 repeats.

Results: The combined radiomics and dosiomics features yields superior prediction performances compared to individual models. Among different SVR kernel functions, the radial basis function (RBF) kernel outperforms linear, polynomial, and sigmoid alternatives, achieving a coefficient of determination (R²) of 0.743 and a relative root mean square error (RRMSE) of 0.022.

Conclusion: Integrating radiomics and dosiomics enhances the predictive power of individual-omic features, resulting in a more accurate GTV regression model for PULSAR-treated lesions. Compared to classification, regression provides a continuous and nuanced prediction of tumor volume changes, capturing subtle variations in response rather than categorizing outcomes into two labels (“responder” vs. “non-responder”). By leveraging multi-omics and SVR, our approach can help transform early decision-making in PULSAR from empirical assessment to a data-driven strategy, enabling personalized radiation therapy, optimizing patient management, and minimizing the risks of under- or over-treatment.