Abstract ID: 113 Accurate extraction of tissues parameters for Monte Carlo simulations using multi-energy CT

Abstract

Purpose Robust tissue characterization is essential for accurate dose calculation [1] , [2] . In this work, we present a novel method called Bayesian eigentissue decomposition (BETD) [3] to extract Monte Carlo inputs from computed tomography (CT) data having an arbitrary number of energies. Method Principal component analysis is applied on a reference dataset of human tissues to define eigentissues which are used as an optimal base of materials representing tissue compositions. To ensure robustness against CT noise, the Bayesian estimator is constructed and resolves the maximum a posteriori fraction of eigentissues in each voxel. The performance of the method in deriving proton beam interaction properties is evaluated with dual-energy CT (DECT) data and compared to a state-of-the-art elemental composition parameterization. Comparison is made with several levels of noise and in the presence of statistical variations in tissue composition and density. The performance of the BETD to an arbitrary number of energies is also investigated by simulating CT data with two to five energy bins with equivalent noise levels. Results Using simulated noise-free CT numbers for 43 reference soft tissues, the BETD and parameterization methods give equivalent results for stopping powers estimation (0.11% and 0.13% respectively). However, when noise and tissue variation are present, the BETD reduces the RMS error on stopping powers from 2.79% for parameterization to 1.88% for the proposed approach. The BETD method also shows potential for using CT with more than 2 energies, where a number of four energy bins is shown to reduce proton beam range uncertainty by a factor of up to 1.5 compared to the parameterization method used with DECT. Conclusion This work proposes a general approach to determine elemental compositions and density for Monte Carlo inputs using CT data in a clinical context, where noise and tissues variations significantly degrade the performance of currently known methods.