Late radiation-induced contrast-enhancing brain lesions (CEBLs) on magnetic resonance imaging (MRI) after proton therapy of brain tumors have been observed to occur frequently in regions of high linear energy transfer (LET) and in proximity to the ventricular system. We analyzed 110 patients with low-grade glioma treated with proton therapy to determine whether the risk for CEBLs is increased in proximity to the ventricular system and if there is a relationship between relative biological effectiveness (RBE) and LET. We contoured CEBLs identified on follow-up T1-MRI scans and computed dose and dose-averaged LET (LETd) distributions for all patients using the Monte Carlo method. We then performed cross-validated voxel-level logistic regression to predict local risks for image change and to extract model parameters, such as the RBE. From the voxel-level model, we derived a model for patient-level risk prediction based on the treatment plan. Of 110 patients, 23 exhibited 1 or several CEBLs on follow-up MRI scans. The voxel-level logistic model has an accuracy as follows: area under the curve of 0.94 and Brier score of 2.6 × 10-5. Model predictions are a 3-fold increased risk in the 4 mm region around the ventricular system and an LETd-dependent RBE of, for example, 1.20 for LETd = 2 keV/μm and 1.50 for LETd = 5 keV/μm. The patient-level risk model has an accuracy as follows: area under the curve of 0.78 and Brier score of 0.13. Our findings present clinical evidence for an increased risk in ventricular proximity and for a proton RBE that increases significantly with increasing LET. We present a voxel-level model that accurately predicts the localization of late MRI contrast change and extrapolate a patient-level model that allows treatment plan-based risk prediction.