Author: Girish Bal, Jan Kralj, Ayan Mitra, PhD, Ling Shao, Matjaz Subic, Yevgen Voronenko 👨🔬
Affiliation: RefleXion Medical, Cosylab 🌍
Purpose: This work enhances the efficiency of radiation therapy treatment planning by optimizing the beamlet dose matrix and full patient dose computations using GPU acceleration. The Collapsed-Cone Convolution/Superposition (CCCS) algorithm is implemented to reduce computation time.
Methods: The CCCS algorithm was fully implemented using Compute Unified Device Architecture (CUDA) for an NVIDIA RTX 6000 Ada GPU and integrated within a commercial Treatment Planning System (TPS). Key optimizations include dynamically identifying regions with non-zero TERMA (Total Energy Released per Unit Mass) values to focus convolutions on those regions, avoiding unnecessary calculations and improving efficiency. To preserve dose calculation fidelity, the GPU-accelerated algorithm integrates fluence profile modeling and scatter effects, leveraging the parallel processing power of the GPU to account for secondary radiation interactions efficiently. The GPU's high computational throughput enables the simultaneous processing of multiple beamlets, ensuring that scatter effects and fluence variations are calculated without compromising on detail or accuracy. CUDA streams enable asynchronous kernel execution and simultaneous data transfers, effectively overlapping computations and minimizing idle GPU times. To further optimize GPU performance, beamlets are processed in batches, reducing kernel launch overhead and enabling efficient handling of multiple angle combinations. Patient-specific CT data is adapted by converting Hounsfield Unit (HU) values into tissue densities and precomputing linear attenuation coefficients. These are stored in lookup tables in GPU texture memory for rapid TERMA calculations, reducing computational delays. Together, these optimizations ensure precise, high-speed dose calculations, enabling treatment planning workflows to deliver accurate results significantly faster than traditional CPU-based methods.
Results: The GPU-accelerated method achieved up to a 23-fold speedup in beamlet and full dose computations compared to 8-core CPU methods, with comparable dose distribution, measured using the Conformality Index (CI), Heterogeneity Index (HI), and target dose coverage.
Conclusion: This approach enables faster treatment plan computations, enhancing clinical workflows.