Room: Track 3
Graphics-processing-units (GPU) have important potential in radiotherapy treatment planning due to superior computing power vs. CPUs. However GPU suffers from limited memory and sophisticated programming. We have developed a GPU-accelerated treatment planning system (TPS) and analyzed its performance by generating an IMPT plan with 500 fields.
A two-field clinical plan for nasopharyngeal carcinoma was re-planned with 500 fields, a feat technically impossible with a conventional TPS. We placed 500 fields evenly in the angular space using Fibonacci sequence. In order to minimize memory usage, we employed sparse matrix compression and dynamic voxel spacing techniques. Full advantage of vector/matrix operation provided by GPU was used to speed up the computation. The influence matrices were calculated using a modified ray-casting analytical dose engine and the dose was optimized using a worst-case robust optimization method. Both were highly parallelized using CUDA running on a Linux workstation with 2 Xeon-E5-2680 CPUs and 4 Tesla-K80 GPUs. The GPU-accelerated TPS has been implemented as an EclipseTM plugin, which provides users with real-time interaction during the optimization.
It took 331 seconds to generate a plan containing 723,134 spots. Of 331 seconds, 61 seconds were used to compute the influence matrices for 13 uncertainty scenarios and 100 seconds were used for robust optimization with 16 dose-volume constraints for 13 organs-at-risk (OARs). Due to the larger number of fields, the resulting plan has dramatically better OAR protection (Unit: Gy[RBE], new vs. clinical) [Brain D1%: 15.3 vs. 52.4; Brainstem D1%: 6.5 vs. 52.4; Right eye Dmean:12.5 vs. 13; Right optic nerve D1%: 24.6 vs. 34.4] compared to the clinical plan.
Our GPU-accelerated TPS achieves high efficiency in speed and memory usage. It enables us to pursue advanced studies like beam angle optimization and proton arc treatment, providing an opportunity for significant improvement in IMPT plan quality.
Funding Support, Disclosures, and Conflict of Interest: Supported by the National Cancer Institute Career Developmental Award K25CA168984, by Arizona Biomedical Research Commission Investigator Award, by The Lawrence W. and Marilyn W. Matteson Fund for Cancer Research, and by The Kemper Marley Foundation.