Deep Learning Blood Dosimetry Predicts Severe Lymphopenia and Survival After Craniospinal Irradiation

Purpose: Craniospinal irradiation (CSI) is a standard treatment for pediatric brain tumors and is increasingly used for leptomeningeal metastases. Although effective, CSI often causes radiation-induced severe lymphopenia (RISL), which is associated with poor outcomes across various cancer types. We developed and validated a novel platform for personalized dosimetry of circulating blood cells (CBCs) and assessed its predictive value for RISL. Methods and Materials: We combined a prospective pediatric CSI registry (initiated April 2018; n = 48) and retrospectively collected patients (January 2010-November 2024; n = 168). RISL was defined as grade 4 lymphopenia (<200/µL) during or within 1 week after CSI. The platform integrated deep learning–based whole-body segmentation, individualized blood volume modeling (HEDOS), patient-specific hemodynamics (cardiac output), and treatment delivery parameters to estimate CBC dose-volume histograms, which quantify the cumulative radiation dose distribution to circulating blood during CSI. CBC dose metrics were analyzed using LASSO, Random Forest, and XGBoost models, with and without cross-validation and Synthetic Minority Over-sampling Technique, to identify RISL predictors. Results: Among 216 patients (median age 12 years; range, 2-68), proton beam therapy was used in 163 (75.5%). RISL occurred in 92 patients (42.6%). Among various CBC dosimetric parameters, CBC D5% consistently emerged as the most important predictor of RISL across all machine learning models. Stratification by CBC D5% with 3.0 GyE effectively differentiated overall survival and event-free survival outcomes (all P < .05). Conclusions: The current deep learning–based personalized blood dosimetry platform enables personalized CBC dosimetry during CSI. CBC D5% is associated with the development of RISL and survival outcomes. Personalized blood dosimetry may serve as a potential tool to identify patients at high risk for RISL, which may inform future efforts to optimize CSI planning.