2258 - A Dosiomics Prediction Model of Brain Metastasis Post-SRS Radionecrosis Diagnosis

Purpose/Objective(s):

To develop a dosiomics based machine learning prediction model for differentiating post-SRS radionecrosis (RN) or local failure (LF) in non-small cell lung cancer (NSCLC) patients with brain metastases (BM). The model utilizes the spatial dose distribution information from the SRS plan, making it feasible to predict BM outcomes at the time of SRS planning.

Materials/Methods:

LINAC-based SRS plans from 99 BMs in NSCLC patients were analyzed to develop a dosiomics-based prediction model. Patients underwent either single fraction (various 15, 16, 18, 20, 22Gy prescription) or five fraction (25Gy or 27.5Gy prescription) SRS, with treatment prescribed to the planning tumor volume (PTV), defined as BM plus 1mm margin. The average PTV volume was 11.5cc. SRS plans were designed using dynamic conformal arcs (DCA) or volumetric modulated arc therapy (VMAT) with 1mm isotropic dose grid. A subset of 13 plans employed single-isocenter multitarget (SIMT) approaches, accommodating up to seven targets per plan.

We extracted 90 dosiomic features from the normalized spatial dose distribution. The patient-specific feature vector included the 90 dosiomic features, total radiation dose, and fractionation scheme (92 features total). A balanced random forest (BRF) model was trained to predict RN/LF outcome, using a 7:3 training-to-test ratio. Model tuning includes leaf depth, sample split, and number of estimators. To assess robustness, 100 model iterations were performed with randomized validation sample assignments. Sensitivity, specificity, and accuracy were evaluated, and a receiver operating characteristic (ROC) curve was generated.

Results:

The BRF model demonstrated acceptable predictability for LF, with sensitivity of 0.71±0.18. RN predictability reported a sensitivity of 0.58±0.09, which is a reasonable result with room for improvement. Other performance metrics, including ROC area under the curve (AUC) of 0.67±0.05 and accuracy of 0.61±0.07, were limited by sample size but remained reasonable. Model performance was likely affected by the limited sample size. The model demonstrated consistent trends in performance metrics throughout the 100 iterations. Further refinement with a larger dataset and additional feature selection strategies may improve both RN predictability and overall classification accuracy.

Conclusion: The dosiomics-based BRF model showed promise in predicting local failure following SRS for brain metastases. Future work with a larger dataset, with a possible multi-institution setup, will be essential for enhanced model performance.