Thesis defense of Feline Heinzelmann

Pediatric brain tumors are increasingly treated with proton therapy (PT). However, late treatment-related brain injury—often presenting initially as imaging changes (ICs) on follow-up MRI—raises concern that variable proton relative biological effectiveness (RBE) may lead to biological overdosage of normal brain tissue. This thesis explored whether clinically unaccounted RBE variability is linked to the occurrence and spatial distribution of ICs in childhood brain tumor patients using logistic-regression normal tissue complication probability (NTCP) modeling. Modeling was performed in a retrospective, registry-based mixed-entity pediatric brain tumor cohort and in a pediatric ependymoma subcohort. In the latter, the impact of IC endpoint definition was additionally evaluated. Across cohorts, voxel-wise analyses revealed an evident association of IC risk and dose. Models accounting for variable RBE consistently outperformed constant-RBE dose models for discriminating IC from non-IC voxels, with a particularly clear dose–linear energy transfer (LET)–response relationship in the mixed-entity cohort and one ependymoma subcohort. These results underscore the need for standardized LET reporting and for considering of variable RBE in pediatric PT. Increased radiosensitivity of the periventricular region was not observed, suggesting that findings from adult cohorts may not translate directly to pediatric populations.
Manual contouring of the brainstem and ventricular system requires substantial time and expertise and may introduce variability in the resulting contours. In a separate project, nnU-Net–based deep learning models were developed to segment these two important structures on MRI automatically. Quantitative performance was assessed via well-established segmentation measures, and qualitative performance was assessed prospectively via Likert-scale ratings in routine clinical workflow. The segmentation framework produced mostly clinically acceptable contours with minor edits, with the largest discrepancies near the craniocervical junction after mapping to CT, and could reduce the overall contouring workload by approximately 50 %.
Overall, the established NTCP modeling workflow and the auto-segmentation framework streamline key steps in pediatric PT planning, enable quantitative risk evaluation to inform plan optimization, and facilitate future radiobiological studies.