SU‐E‐T‐719: Multi‐Energetic, GPU‐Accelerated Superposition/Convolution

Purpose: To develop a novel superposition/convolution based algorithm which eliminates the poly‐energetic transport approximation. Methods: We leveraged the hardware functionality of the GPU texture unit to allow our modern dual‐source superposition/convolution based dose calculation engine to efficiently perform multiple transports simultaneously. We experimented with dividing the spectrum in half (dual‐energetic), quarters (quad‐energetic) or N energy bins (multi‐energetic). These divisions were applied after the TERMA was computed using the exact, full‐spectrum attenuation. We have benchmarked the dosimetric properties of poly‐, dual‐, quad‐ and multi‐energetic superposition against a series of Monte Carlo dose accuracy benchmarks based on the ICCR 2000 benchmark and have performed a manual commissioning for an Elekta Infinity operating at 6MV. Results: The performance cost of dual‐energetic superposition was 11%–50%. The performance cost of quad‐energetic superposition was 39%–151%. Performance varied depending on GPU architecture and cache effects. The slower performance of quad‐energetic superposition was due to a smaller CUDA block size and the use of a separate density texture: we normally pack TERMA and density into a single texture. TERMA performance costs were 1% and 10%, respectively. The traditional, poly‐energetic superposition overestimated dose, particularly within the first 10 cm and in bone/aluminum. Dual‐energetic superposition greatly reduced this overestimation. Quad‐energetic and multi‐energetic superposition produced nearly identical results. Good agreement was achieved in air, water, bone and aluminum; all methods had trouble matching the fall off in lung due to the small treatment field. We based our manual commissioning of our Elekta Infinity linear accelerator on a published spectrum. We modeled the extra‐focal source as being very soft, which necessitated a slight hardening of the primary source. Conclusions: We have completed a multi‐energetic, GPU‐accelerated superposition/convolution based algorithm, which improves accuracy over the traditional, poly‐energetic approach and allows the use of physically accurate spectrums.