Session Item

Brainstem necrosis is a rare, yet severe side-effect of paediatric radiotherapy. The cause is likely multifactorial, with one possible contributor being uncertainty in the relative biological effectiveness (RBE) of protons. A constant RBE=1.1 is assumed clinically, but the RBE is known to vary with linear energy transfer (LET), as well as radiosensitivity of tissue and dose fractionation. Variable RBE-based predictive models are therefore needed for treatment plan optimisation. The aim of this study was thus to investigate the effect of LET and variable RBE in normal tissue complication probability (NTCP) modelling of brainstem necrosis following paediatric ependymoma.

CT and dose data from a case-control cohort (n=28, 1:3 case-control ratio) of symptomatic brainstem necrosis was selected from 954 paediatric patients having undergone passive scattered proton therapy. The matching was based on gender, diagnosis, treatment-related factors and RBE1.1 dose data. The FLUKA Monte Carlo code was used to calculate LET and variable RBE (using Rørvik, McNamara and dose-averaged LET (LETd)-weighted dose). LETd was studied both for full structure volumes, as well as restricted to volumes with high dose thresholds. Logistic regression (LR) was used to fit NTCP models. Dose, LETd and volumes were included in the models based on odds-ratios (ORs) from univariate conditional LR and Spearman rank coefficients. Machine learning algorithms were fitted for comparison. Model evaluation included leave-one-out cross validation (LOOCV), area under receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC), and Brier score.  

Variable RBE led to increased RBE-weighted dose (figure 1). Most structures presented with 0.5–1.5keV/µm higher average L50%, L10% (the LETd to the respective volume percentages) and L_max (L_0.1cc) in cases compared to controls. These differences translated to ORs in the range 1.5–7 for full volumes and approaching 30 including dose thresholds. The difference in D50%, D10% and D_max (D_0.1cc) across cases and controls also increased from constant to variable RBE, with 0.5–1.5 Gy(RBE) higher dose to cases with the McNamara model. The selected NTCP models were a univariate and bivariate LR models achieving high performance scores (table 1). The univariate model considered brainstem L10% (D>54 Gy(RBE) from RBE1.1), while the bivariate model also included the anterior pons volume which was negatively correlated with toxicity. The LR models had comparable performance measures to most machine learning algorithms, excluding the bivariate neural network with similar LOOCV accuracy only.

The most pronounced risk was associated with increased LETd in volumes above 54Gy(RBE). The univariate model incorporating brainstem L10% was the preferred model based on simplicity and solid performance. The bivariate model, while more complex, is strengthened by the inclusion of a parameter independent of dosimetry.