MRI Based Bayesian Personalization of a Tumor Growth Model

Area of Science:

• Mathematical biology
• Computational neuroscience
• Medical imaging analysis

Background:

• Mathematical modeling of brain tumor growth often uses reaction-diffusion models.
• Estimating model parameters is challenging due to identifiability issues, segmentation uncertainty, and model approximations.

Purpose of the Study:

• To analyze patient-specific parameter uncertainty in brain tumor growth models.
• To develop a more informative personalization approach using Bayesian inference.

Main Methods:

• Utilized a highly parallelized Lattice Boltzmann Method (LBM) for the reaction-diffusion equation.
• Employed Gaussian Process Hamiltonian Monte Carlo (GPHMC) for posterior probability estimation.
• Compared the Bayesian personalization with spherical asymptotic analysis and derivative-free optimization.

Main Results:

• The Bayesian personalization approach provided more informative results than traditional methods.
• It successfully sampled parameter posterior probabilities from patient-specific magnetic resonance images.
• Highlighted the presence of multiple modes and parameter correlations, which optimization methods missed.

Conclusions:

• Bayesian personalization offers a superior method for analyzing uncertainty in patient-specific tumor growth models.
• This approach yields richer insights into glioblastoma dynamics compared to existing techniques.
• The method demonstrates potential for improved clinical relevance in brain tumor research.