A robust Monte Carlo treatment planning optimization algorithm for dose painting clinical implementation

[Purpose] In order to start accurate clinical trials based on treatments more aggressive than traditional margin approaches, a robust optimization algorithm has been developed for dose calculation full Monte Carlo-based. Specifically, the algorithm is presented here to manage uncertainties on dose painting from PET/CT image data. [Material and Methods] CARMEN platform was updated to allow heterogeneous dose prescription by means the recurrent both approaches. For the approach considered as true dose painting by number (tDPBN) where the restriction of dose to volumes makes no sense, it was necessary to develop a novel algorithm including an optimization method at the voxel level under Lineal Programming (LP) formulation. For the approach based on the discretization of functional information into several clusterings (DPBN), instead of a recurrent equidistant isolevel, we implemented several algorithms able to reflect the diffuse and multifocal nature of the uptake regions. For this study, the affinity propagation proposed by Foster in order to reduce random errors due to the PET images registration process. Full Monte Carlo simulations were performed for pre-optimization and final dose calculation for taking into account the interactions of particles by means an explicit transport along the beam modifiers in the linac head. Axesse/Synergy linacs of Elekta were modellized with the EGSnrc/BEAMnrc code. The dose calculation in patient was carried out with the BEAMDOSE code, a modified version of DOSXYZnr for calculate the specific beamlet dose contribution on each voxel. A grid calculation consisting on 256 × 256 voxels per slice was used from the interpolation of PET/CT images reconstructed by keeping a compromise with EARL (ResEARch4Life®) accreditation requirements. Linear Programming formulation at voxel level allowed stablishing a tractable robustness of the uncertainties related to the heterogeneous dose prescription, imposing lower and upper-bound constraints to each voxel in accordance to the clustering volume to which they belong. For tDPBN, an inverse planning schema was previously developed. For DPBN by clusterings approach, a specific direct aperture optimization (BIOMAP) based on the sequencing of biophysical maps has been modified to generate apertures with Boolean combinations of the clustering projections. [Results] DPBN by means inverse planning showed excellent QVHs, although they were achieved by means of solutions with high MUs (around 2000 MU/fraction with more than 350 segments). Secondary contribution from MLC meant a high dose spillage to the body, so cannot be directly assumed for clinical implementation. The robust solutions for DPBN by means BIOMAP allow the accurate clinical implementation (around 750 MU/fraction with less than 200 segments) with Q index (planned_dose_matrix/true_prescribed_dose_matrix) over 95%.Peer reviewe