Purpose: The introduction of graphical processing unit (GPU) acceleration for RayStation’s collapsed cone convolution (CCC) algorithm saw tremendous increases in the speed of dose calculations, but these speed increases and their effect on plan quality has not been extensively documented. The complexity of the optimization problem provides no guarantee of reaching acceptable plan results in reasonable timeframes. This study investigates the impact of hardware on radiation treatment plan quality, pertaining to time burdens and optimization setup.
Methods: Several complicated treatment plans were developed in RayStation v8A SP1 with optimization criteria that produce clinically acceptable plans without additional user input. These criteria have been optimized using different sets of initializations and optimization steps (1x240, 2x120…). These were evaluated for speed of computation and plan quality under GPU acceleration. CPU optimized plans were created under time constraints matching the accelerated plans, and results were compared to evaluate the impact of acceleration on plan quality.
Results: CCC calculations entail a five-fold increased runtime in the CPU-accelerated environment over GPU-accelerated. Total runtime increases quadratically for long, continuous optimization runs, with decreased differential returns on time invested. The 1x240 optimization resulted in a plan with the most volume exceeding Rx (58.7%), while the 4x60 optimization resulted in the least excess (6.92%).
Conclusion: The combination of quadratically increasing runtime for long optimizations and the superior metrics of multi-initialized plans suggests that this strategy is best, but multi-initialized plans give more of their runtime to CCC. Given the same amount of runtime, GPU-accelerated optimizations can perform more iterations than CPU optimization by completing CCC evaluations efficiently – producing higher quality treatment plans. This suggests that clinics using RayStation’s CCC algorithm with tight timeframes can produce better plans with improved hardware. Further study intends to investigate the effect of varying GPU hardware on plan quality and optimization times.