Voxel-based model predicts proton-induced brain necrosis using LET and ventricular distance

This study aimed to develop a predictive model for radiation-induced brain necrosis (BN) in adults with skull base tumors following intensity modulated proton therapy (IMPT) incorporating variable relative biological effectiveness (vRBE) based on dose, linear energy transfer (LET), and proximity to ventricles. Four patients with radiation-induced BN at distal edge of beams after proton therapy were analyzed. Necrotic regions were delineated on follow-up MRI, and dosimetric parameters were evaluated. A voxel-level support vector machine (SVM) model was trained using dose, LET, and distance to ventricular system (d_ventricle) to predict necrosis probability. The dose calculated by derived vRBE equation was compared to that using constant RBE (cRBE = 1.1). Necrotic voxels clustered near ventricles (d_ventricle < 10 mm). The SVM model achieved high accuracy (AUC = 0.965) and identified vRBE values significantly higher in the necrotic regions (mean 1.2 vs 1.15 in non-necrotic voxels, p < 0.01). All BN regions contained voxels with necrosis probabilities higher than 80%. Using vRBE, voxels exceeding EQD2 of 90 GyRBE formed an aggregate volume (≥0.1 cc) in the corresponding necrotic regions, while using cRBE, no high-dose volume above 0.01 cc was found in each patient. The validation data from non-necrotic patients also supported this dose constraint. The study highlights the limitations of cRBE in predicting BN and supports vRBE as a supplementary tool for risk mitigation, particularly in periventricular zones. Volume-based constraints (e.g., Doses to 0.1cc (D0.1cc) < EQD2 of 90 GyRBE) are recommended to account for voxel clustering effects. Further validation in larger cohorts is warranted.