Author: Jaryd Ricardo Christie, Anthony J. Doemer, William T. Hrinivich, Junghoon Lee, Calin Reamy, Kundan S Thind 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University, Johns Hopkins University, Henry Ford Health 🌍
Purpose: To develop and validate an open-source Python package for fast, widely compatible, and user-friendly photon dose calculation algorithm through comparisons with clinical treatment planning systems (TPS) for linear accelerators (linacs) from different vendors and institutions.
Methods: A GPU-accelerated dual-source collapsed cone convolution (CCC) superposition algorithm was developed and implemented in CUDA and compiled as a Python package. 6 MV beam modeling was performed for an Elekta VersaHD linac from Institution A and a Varian TrueBeam linac from Institution B. Computed dose was compared with RayStation for the Elekta linac and Eclipse for the Varian linac. Percent depth dose (PDDs), lateral beam profiles, and output factors (OF) for 2 cm to 40 cm square fields were compared in a digital water phantom, and localized prostate cancer VMAT plans were compared in terms of 3D Gamma pass rates at 3%/2 mm criteria using the patient CT for the Elekta linac and solid water phantom for the Varian linac.
Results: Following compilation, the algorithm could be included in new Python projects through a single import statement. For square fields, mean±standard deviation errors in PDDs, profiles, and OFs were 1.2±0.5%, 1.2±0.7%, and 0.1±0.3% for the Elekta linac and 0.9±0.3%, 1.1±0.3%, and 0.1±0.1% for the Varian linac. Elekta single-arc VMAT computation times for the algorithm and RayStation were 1.5 s and 5.8 s, respectively, and Gamma pass rate was 96.6%. Varian two-arc VMAT computation times for the algorithm and Eclipse were 3.5 s and 20.2 s, respectively, and Gamma pass rate was 99.5%.
Conclusion: The Python GPU CCC algorithm successfully modeled linacs from multiple vendors and institutions with fast computation time and is easily integrated into new Python projects. This tool holds promise to accelerate the development of advanced and generalizable auto-planning approaches, building on a growing ecosystem of advanced open-source algorithms.