Optimizing Dose Calculation Parameters for GPU-Based Treatment Planning
C Sutterley*, Q Gautier, Y Graves, N Li, X Jia, S Jiang, Center for Advanced Radiotherapy Technologies, University of California, San Diego, La Jolla, CASU-E-T-570 Sunday 3:00PM - 6:00PM Room: Exhibit Hall
Purpose: Graphics processing unit (GPU)-based treatment planning is clinically desirable given its extremely high efficiency. One technical barrier for clinical implementation is the relatively small memory of GPU cards. The goal of this study is to find the optimal parameters used in dose deposition matrix (DDM) computations via finite size pencil beam (FSPB) model for IMRT optimization that maintains plan quality while minimizing memory usage.
Methods: We used our in-house GPU-based treatment planning system, SCORE, to investigate the impact of parameters used in DDM calculations, i.e., beamlet size and dose cutoff value, on the resulting plan quality and memory usage. We varied beamlet width from 2mm to 6mm, while fixing beamlet height to 5mm, and dose cutoff value from 0.01% to 1% of the maximum beamlet dose. For each combination of parameters we used SCORE to create the DDM. A treatment plan is optimized using the DDM and the resulting final dose distribution is evaluated by comparing it with the ground truth distribution (obtained under 2mm beamlet width and 0.01% dose cutoff value) using gamma test (2mm/2%) passing rate.
Results: We tested various combinations of the parameters on head/neck, prostate, and lung cancer patients. It is found that plan quality is highly sensitive to the beamlet size, such that increasing beamlet width to 3mm will reduce the passing rate to below 90%. In contrast, when keeping beamlet width at 2mm, but increasing dose cutoff parameter to 0.1%, a passing rate of over 95% is still maintained. The memory usage is reduced by a factor of 5.5 on average.
Conclusion: The optimal parameter for DDM calculation using FSPB model is beamlet size of 2mm and dose cutoff value of 0.1%, under which high plan quality is maintained while GPU memory usage is greatly reduced.