Speaker
Description
Brain necrosis after brain and head & neck radiotherapy presents a fundamental inference problem since by the time a lesion is visible on MRI, it has already expanded, remodeled, and erased the evidence of where and why it began. Behind this expansion lies a spatially dynamic process governed by brain architecture and patient-specific biology, which are not captured in clinical dose thresholds. Dosimetric indices showed no consistent voxel-level correlation with necrosis in our cohort, leading us to hypothesize that the dose-outcome relationship is masked by patient heterogeneity. Decomposing this heterogeneity through an expert-augmented Bayesian network revealed that brain anatomy, particularly proximity to ventricles and white matter, governs necrosis risk more strongly than dose alone. Building on this, a 3D cellular automaton incorporating MRI-derived vascular density maps reproduced the anisotropic spatial progression of lesions with AUC 0.87-0.95. Inverting the model's rules allowed backward-in-time simulation to localize lesion initiation sites invisible to standard imaging. A multi-channel Vision Transformer model was developed to integrate 3D dose distributions with crucial nondosimetric factors to predict necrosis risk. Disentangling the biological, anatomical, and dosimetric drivers of brain necrosis aims to redefine biological effective radiation dose to personalize treatment planning.
