Room: Stars at Night Ballroom 4
Purpose: To develop a fast, accurate photon dose calculation engine for IMRT and VMAT optimization using fast tensor calculations and parallel GPU computing.
Methods: A finite-sized pencil beam dose calculation algorithm was developed that is capable of massive parallelization through a tensor formulation. Using beam data, a scattering kernel and fluence scaling factor were fitted to account for both collimation scatter and low energy tissue scatter as well as both primary and off-axis fluence scaling. Heterogeneity within the medium is further accounted for by again scaling the fluence using the radiological depth. Considering the radiation field as a group of adjacent beamlets, the dose is given by the convolution of the scaled fluence and the scattering kernel. Each convolution is calculated on beam planes orthogonal to the primary beam axis to construct the final dose tensor.
Results: The model displays a very high accuracy of comparing the dose at depth along the central beam axis between the calculation and the measured beam data. An acceptable accuracy was achieved in comparing the dose cross profiles at depth. The calculation results in a shaper penumbral region that is magnified with increasing field size. Without GPU acceleration on a generic workstation, the calculation took around 120s for 2500 beamlets and 500s for 10000 beamlets at 1mm resolution along the depth axis. The vast majority of the calculation time was due to the ray-tracing algorithm.
Conclusion: The fluence scaled tensor kernel dose calculation algorithm has demonstrated to be a viable candidate to further accelerate and optimize for IMRT and VMAT optimization. The incorporation of massive parallelization through the use of a GPU will not only provide a speed up of a few orders of magnitude but will allow finer calculation resolutions to increase accuracy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by an NIH grant (#R01CA201212)