SU-E-T-513: Clinical Implementation of a GPU Accelerated Pencil Beam Dose Calculation Algorithm

This work reports a clinical implementation of a GPU accelerated pencil beam dose calculation algorithm (GPU-PB).

Model parameters were determined using in-house scripts written in MATLAB. Dose distributions in a water phantom were calculated using a Pinnacle TPS for various open field sizes. Lateral profiles at 2-mm incremental depths were used to calculate PB kernel parameters. Weighted sum of squares were used for least-squares parameter fitting utilizing a Levenberg-marquardt algorithm. Weightings were adjusted based on goodness of fitting and in accordance with field size to suppress deviations due to horn effects. The scale factor was fitted iteratively. The calculated doses for two patient cases were analyzed with a 3D gamma method (3%/3mm).

Excellent agreement between Pinnacle calculation and GPU-PB calculation was achieved regarding PDD, profiles and output factor. For a head-neck 10-field step-and-shoot IMRT case, gamma passing rate was over 99% and maximum absolute dose value is 75.6 Gy vs. 73.9 Gy, differing by 2.3%. Similar results were obtained for a SBRT lung case. Gamma passing rate is a function of PB kernel cut-off distance and beamlet resolution. It appears that if the highest accuracy is desired, a resolution of 10 mm or better in the direction parallel to MLC travel and a cutoff distance of 10 cm or better should be used. The calculation time increases with both cut-off distance and beamlet resolution. For example, GPU calculation time is 1.36 seconds for 5 cm cut-off distance, and increases to 4.84 s for 15 cm cut-off distance.

A GPU accelerated PB dose calculation algorithm has been implemented using clinical measurement data. Excellent agreement with Pinnacle TPS has been achieved. Beamlet size and PB cut-off distance should be chosen according to desired dose calculation accuracy and speed.