1042 - Inferring the Initiation Site of Brain Necrosis from Proton Therapy via Inverse Mechanistic Modeling
Presenter(s)
N. Cogno1, K. D. Shah2, F. Ehret3, C. Beekman1, H. A. Shih4, H. Paganetti1, and I. Chamseddine1; 1Department of Radiation Oncology, Massachusetts General Hospital/Mass General Brigham and Harvard Medical School, Boston, MA, 2Department of Radiation Oncology and Winship Cancer Institute, Emory University, Atlanta, GA, 3Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt-Universität zu Berlin, Department of Radiation Oncology, Berlin, Germany, 4Department of Radiation Oncology, Mass General Brigham / Massachusetts General Hospital, Boston, MA
Purpose/Objective(s):
Brain radionecrosis (RN) is a side effect of proton radiotherapy (RT) that can cause significant neurological impairment and be life-threatening in severe cases. Understanding the drivers of heterogeneous biological responses to RT is essential for improving patient outcomes. However, current imaging techniques detect lesions after spatial expansion, whose progression is influenced by secondary processes such as inflammation, thereby obscuring the initial site of injury and its relationship with dose and linear energy transfer (LET).
The objectives of this work are: (1) to develop a mechanistic model that simulates lesion progression backward in time to identify sites of origin; and (2) to test whether predicted initiation sites exhibit higher LET than dose-matched healthy tissue.
Materials/Methods:
We developed a cellular automaton (CA) that accounts for vascular heterogeneity to model longitudinal changes in necrotic lesions observed on follow-up (FU) images. The simulation domain is the brain, discretized into automaton units (AUs), each corresponding to an MRI voxel and assigned a necrotic, healthy, or healed state. Lesions expand or shrink between FU scans, with each AU evolving independently according to its neighbors’ state and local capillary density. The model was first developed to reproduce forward RN progression and validated on lesions from 7 proton-treated patients.
For each patient, posterior distributions of the model parameters were inferred with Bayesian methods by running the model from the second FU (FU2) to the first (FU1). The model rules were then inverted to simulate lesion dynamics backward in time and Monte Carlo sampling was used to generate initiation probability maps.
Results:
Radial dispersion, defined as the distance of each voxel from the lesion center of mass in the joint dose-LET plane, was computed at FU1 and FU2. The median FU2/FU1 ratio of radial dispersion was 3.17 (range 1.17–4.99; n = 7), confirming lesion spread.
We ran 10,000 backward simulations using patient-specific parameters and the FU1 lesion as input. The model localized lesion initiation to a subset of FU1 voxels. Compared with dose-matched healthy voxels, predicted initiation regions showed higher mean LET in 6 of 7 patients.
For each predicted voxel, we computed the difference between its LET and that of the dose-matched control. We then calculated the probability-weighted mean difference, where positive values indicated higher LET in predicted voxels relative to controls. This analysis confirmed the trend, with positive mean differences in 6 of 7 patients and confidence intervals excluding zero.
Conclusion:
Our results show that incorporating vascular architecture when inferring necrosis initiation reveals a consistent association between higher LET and predicted lesion origin. This biologically grounded model provides a 3D framework for modulating dose-LET effects in radiation necrosis, with potential to improve risk prediction and guide treatment optimization.