ESTRO 2021 Abstract Book

S1330

ESTRO 2021

approved plans with 3361 beams computed with the clinically implemented CPU-MC dose engine in the same TPS were used for the validation of the new algorithm. Most lesions were in the Central Nervous System (which included brain tumors and cranio-spinal irradiations, 205 patients), head and neck region (77 patients) and bone sites (65 patients). Gastro-intestinal, mediastinum and lung treatments were also included in the group. For each beam, the dose was recalculated using the new GPU-MC dose engine and compared to the original CPU-MC dose. Beam dose difference distributions were studied to ensure that the two doses were equal within the expected fluctuations of the MC statistical uncertainty, s , of each computation. Plan dose distributions were compared through a global gamma analysis (2%,2mm). Additionally, comparison was performed with respect to the dosimetric indices D 98 , D 50 and D 1 of all ROIs defined as targets and to D 1 of the external contours. A complete assessment of the computational time as a function of s and dose grid voxel size was done. Results The median value of the voxel-based mean beam dose differences between CPU- and GPU-MC was -0.01% and the median of the standard deviations was close to √2 both for simulations with an s of 0.5% and 1.0%. This shows that the two dose distributions can be considered equal. The gamma analysis showed a median passing rate of 100.0% for plans with s=0.5% and 99.9% for s=1.0% as shown in Figure 1. All the DVH indices of ROIs defined as targets (5460) showed an average difference lower than 0.04%: the largest difference was found for D 1 with a mean difference of -0.04% and a standard deviation of 0.31%. The same analysis was repeated for D 1 of all external contours, obtaining an average difference of -0.03% and a standard deviation of 0.06%. The majority of the plans were computed with 1.0% statistical uncertainty on a 2 mm dose calculation grid, for which the median computation time was 5.2 s. A complete assessment of the calculation time as a function of voxel size and s is shown in Figure 2.

Conclusion A validation of a clinical MC algorithm running on GPU was performed on a large pool of patients treated with pencil beam scanning proton therapy. We demonstrated that the differences with the previous CPU-based MC were only due to the intrinsic statistical fluctuations of the MC method, which translated to insignificant differences on plan dose level. This validation further highlights the significant impact the implementation on GPU will have on calculation times and, consequently, on future clinical workflows.