2624 - Pre-Treatment Vascular Profiling on Multiparametric MRI for Improved Radiation Necrosis Prediction after Brain SRS

Purpose/Objective(s): Radiation necrosis (RN) is a dose-limiting toxicity of stereotactic radiosurgery (SRS) for brain metastases, with dose-volume constraints failing to account for brain heterogeneity. While its exact etiology remains unclear, evidence links RN to vascular injury and blood-brain barrier disruption, with susceptibility varying spatially. We hypothesize that pre-existing vascular heterogeneity influences RN risk and aim to develop a predictive model using routine MRI to isolate vascular contributions and guide personalized treatment strategies.

Materials/Methods: 74 brain metastases from 16 patients treated with SRS were analyzed. RN was confirmed in 24 lesions via biopsy (50%) or follow-up imaging over at least six months. Pre-treatment multiparametric MRI (T1, T1-Gd, FLAIR, T2, and ADC) were preprocessed and deformably transformed into a common template space. A vascularity map was generated by subtracting pre-contrast T1 from post-contrast T1 with a 10% intensity threshold. Three regions of interest (ROIs) were defined: the planning GTV (ROI1), GTV with 1 cm peritumoral expansion (ROI2), and the peritumoral region alone (ROI3). For RN prediction, mean values from each MRI sequence, planning dose, and vascularity map within ROIs were input into a logistic regression model. Additionally, 187 radiomic intensity and texture features from the most predictive sequence were extracted and incorporated into Elastic Net (EN) regression for potential enhanced prediction (5-fold cross validation applied).

Results: As shown in Table 1, metrics derived from GTV plus peritumoral regions (ROI2) improved predictive performance compared to GTV alone, highlighting the role of the peritumoral environment in RN development. Based on ROI2, mean vascularity outperformed other MRI metrics and dose in RN differentiation (Area under the curve (AUC) = 0.64, p = 0.02). EN regression incorporating vascular radiomic features yielded further improved prediction (AUC = 0.78).

Conclusion: We proposed a vascular profiling pipeline for pre-treatment RN susceptibility prediction using routine clinical MR scans. Our findings demonstrated the vascular role in RN development, providing the potential to identify RN risks before treatment. These metrics may assist in patient stratification for personalized treatment. Larger studies are warranted for improved performance and validation.